TerraMosaic Daily Digest: Jan 30, 2026
Daily Summary
This digest synthesizes 261 selected papers and focuses on landslide process mechanics and slope evolution, flood generation, routing, and hydroclimatic forcing, high-resolution remote-sensing monitoring workflows. Top-ranked studies examine earthquake-triggered slope response and liquefaction, landslide susceptibility mapping, and satellite and LiDAR-based deformation monitoring.
Across the full set, evidence converges on mechanism-constrained analysis with operational relevance, especially for seismic source-to-ground response pathways and coastal and submarine hydro-geomechanics. The strongest contributions pair interpretable process evidence with monitoring or forecasting workflows that support warning design and risk prioritization.
Key Trends
- Landslide studies increasingly resolve process chains: Contributions connect triggering conditions, slope deformation, and mobility outcomes, improving the basis for warning thresholds and scenario testing.
- Flood analyses are becoming event-specific and process-based: Papers emphasize precipitation structure, antecedent wetness, and catchment controls rather than static hazard descriptors.
- Monitoring workflows rely on integrated remote-sensing products: Multi-source satellite and airborne observations are used for deformation retrieval, change detection, and rapid post-event mapping.
- Seismic hazard research links source behavior to ground response: Recurring topics connect rupture or loading conditions with geotechnical performance and consequence assessment.
- Coastal and submarine hazards are treated as coupled systems: Wave, mass-transport, and shoreline processes are analyzed together with engineering implications.
Selected Papers
This digest features 261 selected papers from 2956 RSS items scanned (1422 new papers after deduplication; 1422 analyzed). Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Interpretation and classification of landslides triggered by the 2024 Noto Peninsula Earthquake (M7.6) for landslide inventory
Core Problem: Existing landslide inventories for the 2024 Noto Peninsula Earthquake did not comprehensively classify landslide features (scar, transported, accumulated areas), limiting the understanding of earthquake-induced landslide characteristics.
Key Innovation: A reconnaissance study providing a more comprehensive interpretation and classification of earthquake-induced landslides in a concentrated area of the Noto Peninsula, categorizing them into six types and subtypes (e.g., deep-seated vs. shallow-disrupted scars) and analyzing their distribution with slope angle using a 1-m DEM, thereby enhancing the understanding of their geological and geomorphological characteristics.
2. A systematic framework for the integration of feature selection and artificial intelligence in landslide susceptibility assessment
Core Problem: The performance of AI-based landslide susceptibility mapping, particularly using ANNs, is highly dependent on the selection of conditioning factors, and a systematic evaluation of feature selection impact is lacking.
Key Innovation: A systematic framework integrating multiple feature selection algorithms (Random Forest, Relief-F, Mutual Information, Information Gain Ratio) with Artificial Neural Networks (ANN) to rigorously evaluate the impact of conditioning factors on landslide susceptibility maps, demonstrating that dimensionality reduction improves both accuracy and model interpretability, and providing reproducible guidance for land use decision-making.
3. Stability charts for unsaturated uniform slopes
Core Problem: Conventional slope stability charts do not adequately account for the strength imparted by suction stress in partially saturated slopes, making design considerations challenging in such scenarios.
Key Innovation: A set of straightforward and quick stability charts for uniform unsaturated slopes, developed using the kinematic approach of limit analysis and upper bound theorem, which account for changes in suction due to hydrological processes (infiltration, evaporation, water table fluctuations), eliminating the need for iterations and demonstrating efficacy in assessing slope stability under unsaturated conditions.
4. Sensitivity of urban seismic damage predictions to input data detail: an application to Sanremo, Italy
Core Problem: The quantitative impact of varying levels of detail in hazard, vulnerability, and exposure data on urban seismic damage predictions remains poorly constrained, leading to inaccurate evaluations when transitioning from large-scale to urban-scale applications.
Key Innovation: A multi-level comparative framework applied to Sanremo, Italy, quantifying the sensitivity of urban seismic damage predictions to input data detail by comparing estimates from large-scale datasets against refined local information. This demonstrates that local data, especially site-specific amplification effects and building characteristics, lead to significant discrepancies and improved accuracy in damage intensity and spatial distribution, providing a robust framework for disaster risk management.
5. Cross-domain coseismic landslide segmentation: local boundary enhancement & global pixel contrastive learning
Core Problem: Accurately segmenting coseismic landslides, especially across different data domains.
Key Innovation: Proposing a method combining local boundary enhancement and global pixel contrastive learning for cross-domain coseismic landslide segmentation.
6. Bidirectional Cross-Perception for Open-Vocabulary Semantic Segmentation in Remote Sensing Imagery
Core Problem: High-resolution remote sensing imagery, with its densely distributed objects and complex boundaries, places high demands on geometric localization and semantic prediction for open-vocabulary semantic segmentation (OVSS), which existing training-free methods struggle to meet.
Key Innovation: Proposes SDCI, a spatial-regularization-aware dual-branch collaborative inference framework for training-free OVSS, which introduces a cross-model attention fusion (CAF) module, a bidirectional cross-graph diffusion refinement (BCDR) module, and a convex-optimization-based superpixel collaborative prediction (CSCP) mechanism to enhance geometric localization and semantic prediction.
7. PINN-based short-term forecasting of fault slip evolution during the 2010 slow slip event in the Bungo Channel, Japan
Core Problem: Accurately monitoring and forecasting fault slip evolution for understanding earthquake cycles and assessing seismic hazards is challenging, especially for slow transient slips, and existing frictionally homogeneous models can lead to unstable forecasts.
Key Innovation: A physics-based data assimilation framework using Physics-Informed Neural Networks (PINNs) to calculate fault slip evolutions and optimize spatially heterogeneous frictional properties, successfully forecasting slow transient slip by integrating geodetic observations and fault mechanics, leading to stable slip evolution.
8. Few-Shot Domain Adaptation with Temporal References and Static Priors for Glacier Calving Front Delineation
Core Problem: State-of-the-art models for glacier calving front delineation perform poorly when applied to novel, out-of-distribution study sites, making their accuracy insufficient for scientific analysis and global monitoring.
Key Innovation: A few-shot domain adaptation strategy that incorporates spatial static prior knowledge and temporal reference images, significantly reducing delineation error for glacier calving fronts at novel study sites, enabling global-scale monitoring without architectural modifications.
9. Physics Informed Reconstruction of Four-Dimensional Atmospheric Wind Fields Using Multi-UAS Swarm Observations in a Synthetic Turbulent Environment
Core Problem: Accurate reconstruction of four-dimensional atmospheric wind fields is essential for applications like weather forecasting and hazard prediction, but conventional instruments leave spatio-temporal gaps within the lower atmospheric boundary layer, and individual UAS platforms provide only limited local measurements.
Key Innovation: Presents a framework for reconstructing 4D atmospheric wind fields using coordinated UAS swarm measurements, estimating local wind components with a Bi-LSTM and assimilating them into a Physics-Informed Neural Network (PINN) to reconstruct a continuous wind field, achieving accurate and scalable reconstruction for hazard prediction.
10. Simplified seepage rate estimation of zoned embankment dams
Core Problem: Under- or over-estimating seepage quantity in zoned embankment dams during design, often due to simplified 2D modeling, leads to discrepancies between predicted and measured values, impacting dam safety management and early warning systems.
Key Innovation: Proposed a modified method for estimating seepage rates in zoned embankment dams, improving upon previous studies and deriving a new closed-form solution. The approach enables realistic prediction of partitioned seepage rates within the core layer and accounts for seepage contributions from the foundation and abutments, verified with extensive measured data.
11. AI-Augmented Synthetic Aperture Radar (SAR) for ground deformation: Introducing the Adaptive Regional AI System (ARAIS)
Core Problem: Current AI-augmented SAR applications for large-scale ground deformation monitoring suffer from limited generalizability across diverse geological settings, insufficient incorporation of various deformation forces, and subjective definition of deformation regions, reducing reliability and practical usability.
Key Innovation: Introduction of the Adaptive Regional Artificial Intelligence System (ARAIS), a conceptual framework designed to dynamically select and apply the most suitable AI algorithms for ground deformation monitoring based on specific local geological features and external triggers, thereby providing flexible, context-aware analysis and supporting disaster risk management.
12. Suffusion-induced stratification exacerbation in reclaimed coral sand foundations
Core Problem: Dredged artificial island foundations are susceptible to suffusion-induced stratification and subsequent localized foundation subsidence and collapse, but the characteristics and triggering mechanisms of suffusion in these reclaimed coral sands are not fully understood.
Key Innovation: An integrated experimental-numerical approach systematically investigating suffusion characteristics in reclaimed coral sand foundations, revealing pronounced suffusion susceptibility in sandy gravel layers and demonstrating that pit dewatering and tidal action can trigger intensive suffusion, exacerbating stratification and leading to foundation subsidence and collapse, while also providing insights for mitigation strategies.
13. A coupled 3D DDA-MPM framework for soil-structure interaction modeling and its application in geotechnical hazards modeling
Core Problem: Need for an advanced numerical framework capable of accurately modeling complex soil-structure interaction, particularly for applications in geotechnical hazards.
Key Innovation: Development of a coupled 3D DDA-MPM (Discontinuous Deformation Analysis - Material Point Method) framework for soil-structure interaction modeling, with applications in geotechnical hazards.
14. Unsupervised Deep Learning for Environmental Risk Monitoring: Landslide Detection from Multi-Resolution Remote Sensing Imagery
Core Problem: Detecting landslides from multi-resolution remote sensing imagery, particularly using unsupervised methods for environmental risk monitoring.
Key Innovation: An unsupervised deep learning framework for landslide detection from multi-resolution remote sensing imagery, applicable to environmental risk monitoring.
15. Alpine plant trait combinations shape soil erosion dynamics and patterns
Core Problem: Investigating how combinations of alpine plant traits influence soil erosion dynamics and patterns.
Key Innovation: Revealing that alpine plant trait combinations significantly shape soil erosion dynamics and patterns.
16. Induced Earthquakes in the Southern Delaware Basin, Texas, Are Bound by a Geomechanically Controlled Maximum Magnitude
Core Problem: Understanding why widespread induced seismicity in the southern Delaware Basin exhibits a maximum magnitude truncation.
Key Innovation: Documenting that induced seismicity in the southern Delaware Basin is bound by a geomechanically controlled maximum magnitude, likely due to structurally constrained shallow faults with limited down-dip widths, which limits rupture dimensions.
17. The Role of Normal Stress and Shear Stress Heterogeneity in the Inferred Depth‐Independence of Stress Drop
Core Problem: The contradiction between the inferred depth-independence of earthquake stress drops and linear scaling predictions for frictional stick-slip instabilities, which assume increasing fault normal stress with depth.
Key Innovation: Examination of simulated earthquake sequences in continuum rate-and-state fault models showing a weaker dependence of stress drop on normal stress than simple friction, especially when fault dimensions are much larger than nucleation scales, potentially explaining the lack of inferred depth-dependence.
18. FloodUnet: A Rapid Spatio‐Temporal Prediction Model for Flood Evolution Based on an Enhanced U‐Net
Core Problem: Gaps in existing research for predicting dynamic flood evolution, including predicting maps from initial time steps, weak transferability for unseen breach scenarios, and potential enhancement of neural network frameworks.
Key Innovation: Proposing FloodUnet, an enhanced U-Net deep learning model for rapid and accurate spatio-temporal prediction of flood evolution, achieving high precision (RMSE 0.2m, NSE 0.9) and being three orders of magnitude faster than hydrodynamic models, with improved feature representation via residual modules and channel attention.
19. Multiscale modeling for coastal cities: addressing climate change impacts on flood events at urban-scale
Core Problem: Understanding how future climate scenarios may affect flooding in coastal cities, requiring high-resolution simulations of extreme storm surges, waves, and river discharges.
Key Innovation: Integrated modeling framework simulating extreme storm surges, waves, and river discharges at high urban resolution (up to 2m) for three European coastal cities under future climate scenarios, revealing flood trends driven by local geomorphic features, sea-level rise, and storm intensity changes, providing insights for adaptation.
20. Learning to Advect: A Neural Semi-Lagrangian Architecture for Weather Forecasting
Core Problem: Recent machine-learning weather forecasting models use monolithic architectures that implicitly represent distinct physical mechanisms, making long-range advection expensive and challenging to treat effectively.
Key Innovation: Presents PARADIS, a physics-inspired global weather prediction model that imposes inductive biases through functional decomposition into advection, diffusion, and reaction blocks, implementing advection via a Neural Semi-Lagrangian operator, achieving state-of-the-art forecast skill at a fraction of the training cost.
21. RSGround-R1: Rethinking Remote Sensing Visual Grounding through Spatial Reasoning
Core Problem: Remote Sensing Visual Grounding (RSVG) faces challenges in localizing target objects in large-scale aerial imagery due to the vast spatial scale and high semantic ambiguity, requiring enhanced spatial reasoning from MLLMs.
Key Innovation: RSGround-R1, a reasoning-guided, position-aware post-training framework, enhances spatial understanding for RSVG by using Chain-of-Thought Supervised Fine-Tuning, Reinforcement Fine-Tuning with a positional reward, and a spatial consistency guided optimization scheme, achieving superior performance and generalization.
22. SENDAI: A Hierarchical Sparse-measurement, EfficieNt Data AssImilation Framework
Core Problem: Spatiotemporal field reconstruction from observation-sparse deployment conditions faces challenges in bridging the gap with data-rich training regimes, especially when target domains exhibit distributional shifts and multi-scale dynamics.
Key Innovation: SENDAI, a hierarchical Sparse-measurement, EfficieNt Data AssImilation Framework, reconstructs full spatial states from hyper sparse sensor observations by combining simulation-derived priors with learned discrepancy corrections, achieving significant performance improvements in satellite remote sensing reconstruction and preserving diagnostically relevant structures.
23. LLM4Fluid: Large Language Models as Generalizable Neural Solvers for Fluid Dynamics
Core Problem: Existing deep learning approaches for fluid dynamics modeling suffer from limited generalization to unseen flow conditions and require retraining for new scenarios.
Key Innovation: LLM4Fluid, a spatio-temporal prediction framework, leverages LLMs as generalizable neural solvers for fluid dynamics by compressing flow fields into a latent space with physics-informed disentanglement, using a pretrained LLM for temporal prediction, and employing a modality alignment strategy, achieving state-of-the-art accuracy and zero-shot generalization.
24. Urban Neural Surface Reconstruction from Constrained Sparse Aerial Imagery with 3D SAR Fusion
Core Problem: Existing Neural Surface Reconstruction (NSR) methods for urban 3D reconstruction suffer from geometric ambiguity and instability, particularly under constrained sparse-view aerial imagery conditions, limiting their applicability in large-scale urban remote sensing.
Key Innovation: Presents the first urban NSR framework that fuses 3D SAR point clouds with aerial imagery, integrating radar-derived spatial constraints into an SDF-based NSR backbone to guide ray selection and adaptive sampling, significantly enhancing reconstruction accuracy, completeness, and robustness.
25. MetricAnything: Scaling Metric Depth Pretraining with Noisy Heterogeneous Sources
Core Problem: Scaling metric depth estimation pretraining is challenging due to heterogeneous sensor noise, camera-dependent biases, and metric ambiguity in noisy cross-source 3D data, hindering the development of robust vision foundation models for depth.
Key Innovation: Introduces Metric Anything, a simple and scalable pretraining framework that learns metric depth from noisy, diverse 3D sources using a Sparse Metric Prompt as a universal interface, demonstrating a clear scaling trend and achieving SOTA results on various depth-related tasks and boosting MLLM spatial intelligence.
26. MORPH: PDE Foundation Models with Arbitrary Data Modality
Core Problem: Existing PDE models struggle to seamlessly handle heterogeneous spatiotemporal datasets of varying data modality (1D-3D), different resolutions, and multiple fields with mixed scalar and vector components, limiting their flexibility and scalability.
Key Innovation: Introduces MORPH, a modality-agnostic, autoregressive foundation model for PDEs built on a convolutional vision transformer backbone. It combines component-wise convolution, inter-field cross-attention, and axial attentions to handle diverse data modalities and resolutions, outperforming models trained from scratch and strong baselines on various downstream prediction tasks.
27. Thermodynamically Consistent Viscoplastic Constitutive Model with Subloading Surface for Crushable Geomaterials
Core Problem: Crushable geomaterials, despite often being regarded as rate-insensitive, can exhibit rate-dependent behavior under high confining pressures, necessitating a thermodynamically consistent constitutive model.
Key Innovation: Develops a thermodynamically consistent viscoplastic constitutive model with a subloading surface specifically for crushable geomaterials, addressing their observed rate-dependent behavior when subjected to high confining pressures.
28. Satellite-Based Fraction of Available Water Reveals Soil Moisture Deficits Preceding Major Wildfires
Core Problem: Critical soil moisture (SM) thresholds as predisposing factors for large wildfires are poorly characterized across different satellite data sources due to variations in volumetric SM measurements.
Key Innovation: Calculation of the Fraction of Available Water (FAW) from multiple satellite products (SMOS, AMSR2, SMAP, CCI) to establish robust thresholds (e.g., FAW < 0.50 for plant stress, FAW < 0.20 for extreme drought) for identifying soil moisture deficits that predispose areas to wildfire danger, demonstrated by analyzing conditions preceding major wildfires in Chile.
29. An automatic identification of rock mass discontinuity from 3D point clouds using multi-point clustering algorithm
Core Problem: Accurate and efficient identification of rock mass discontinuities from 3D point clouds is crucial for rock mass stability analysis, but existing methods may lack balance between accuracy and efficiency.
Key Innovation: A new method for automatic identification of rock discontinuities from 3D point clouds using a multi-point clustering algorithm (MPC), combined with KNN, PCA, RANSAC, and an improved SOM neural network, to accurately and efficiently recognize individual discontinuities and group discontinuity sets, essential for rock mass stability analysis.
30. Framework for predicting multi-sector business interruption losses solely from earthquake scenarios
Core Problem: Predicting multi-sector business interruption losses resulting from earthquake scenarios.
Key Innovation: Developing a framework to predict business interruption losses across multiple sectors based solely on earthquake scenarios.
31. Microbial functional shifts amplify the temperature sensitivity of soil nitrogen across the erosion-deposition continuum
Core Problem: Understanding how microbial functional shifts influence the temperature sensitivity of soil nitrogen in the context of erosion and deposition.
Key Innovation: Revealing that microbial functional shifts amplify the temperature sensitivity of soil nitrogen across the erosion-deposition continuum.
32. Spatiotemporal analysis of rainfall erosivity in Oklahoma
Core Problem: Analyzing the spatiotemporal patterns and variability of rainfall erosivity in Oklahoma.
Key Innovation: A spatiotemporal analysis of rainfall erosivity across Oklahoma.
33. Experimental study on macro-meso damage of jointed rock mass under sequential cyclic loading–unloading followed by freeze–thaw cycles
Core Problem: Understanding the macro-meso damage evolution in jointed rock masses subjected to combined cyclic loading-unloading and freeze-thaw cycles.
Key Innovation: Experimental investigation of the sequential effects of cyclic loading-unloading and freeze-thaw cycles on the macro-meso damage of jointed rock masses.
34. Submarine Fiber‐Optic Sensing Revels Monterey Paleocanyon Evolution With Multi‐Scale Ambient Noise Imaging
Core Problem: Understanding the evolutionary history of Monterey Canyon due to limited observational coverage and imaging resolution.
Key Innovation: Developing a novel multi-scale ambient noise imaging framework using submarine fiber-optic sensing to reveal nested paleocanyon geometries and their multi-phase evolutionary process, providing insights into landscape evolution at active continental margins.
35. P‐Wave Reverberations in the Water Column of the Chilean Subduction Trench
Core Problem: P-waves generated by subduction earthquakes can get trapped and reverberate in submarine trenches, contaminating seismic signals, and the mechanisms are not fully understood.
Key Innovation: Analysis of P-wave reverberations for 43 Chilean subduction earthquakes, showing they are generated when P-waves are trapped in the water column in front of the epicenter, regardless of rupture proximity to the trench.
36. Elevated Asthenospheric Temperature Driving the Partial Melting of Enriched Lithospheric Mantle Feeding the DM‐EM1 Mixing‐Type Magmatism of the Jingpohu Volcano in Eastern NE China: Evidence From Teleseismic P‐Wave Attenuation
Core Problem: The unclear origin of binary-mixing of DM and EM1 sources for Cenozoic intraplate volcanism in eastern Northeast China.
Key Innovation: First map of teleseismic P-wave attenuation across Northeast China, revealing strong attenuation and abnormally low asthenospheric Qp beneath Jingpohu volcano, indicating a high-temperature asthenosphere with partial melting, which likely induces lithospheric melting and feeds the volcanism.
37. When Do Riverine Systems “Feel the Burn”? Simulating How Burn Extent and Severity Modulate Hydrologic Controls on Biogeochemical Export
Core Problem: Understanding how burn extent and severity modulate post-fire hydrologic changes and their impact on streamflow, nitrate, and dissolved organic carbon export in riverine systems.
Key Innovation: Using the SWAT model to simulate 1830 wildfire scenarios, showing that area burned thresholds differ with burn severity and analyte, and that post-fire transport of dissolved organic carbon is sensitive to both, while nitrate is primarily sensitive to area burned, highlighting the role of flow pathways and soil properties.
38. A probabilistic approach to wildfire spread prediction using a denoising diffusion surrogate model
Core Problem: Wildfire spread prediction is inherently uncertain, and traditional models often provide only fixed outcomes, failing to capture the range of possible scenarios.
Key Innovation: Introduces the first denoising diffusion model for probabilistic wildfire spread prediction, a generative AI model that simulates a range of possible fire scenarios and captures inherent uncertainty, producing physically meaningful forecast ensembles.
39. SMKC: Sketch Based Kernel Correlation Images for Variable Cardinality Time Series Anomaly Detection
Core Problem: Conventional anomaly detection in multivariate time series fails when monitoring systems experience sensor churn, leading to variable cardinality and unseen values in data windows.
Key Innovation: Proposes SMKC, a framework that decouples dynamic input structure from the anomaly detector by using permutation-invariant feature hashing to sketch raw inputs into a fixed-size state sequence and constructing a hybrid kernel image to capture global temporal structure, enabling robust anomaly detection with variable cardinality.
40. HydroSense: A Dual-Microcontroller IoT Framework for Real-Time Multi-Parameter Water Quality Monitoring with Edge Processing and Cloud Analytics
Core Problem: The global water crisis demands affordable, accurate, and real-time water quality monitoring solutions, which traditional manual or expensive commercial systems fail to provide, especially in resource-constrained environments.
Key Innovation: HydroSense, a dual-microcontroller IoT framework, integrates six critical water quality parameters (including water level) into a unified, real-time monitoring system with edge processing and cloud analytics, achieving high accuracy and significant cost reduction.
41. Breaking the Regional Barrier: Inductive Semantic Topology Learning for Worldwide Air Quality Forecasting
Core Problem: Global air quality forecasting grapples with extreme spatial heterogeneity and the poor generalization of existing transductive models to unseen regions.
Key Innovation: Proposes OmniAir, a semantic topology learning framework tailored for global station-level prediction. It encodes invariant physical environmental attributes into generalizable station identities and dynamically constructs adaptive sparse topologies, effectively capturing long-range non-Euclidean correlations and physical diffusion patterns. Achieves state-of-the-art performance, high efficiency, and scalability, bridging monitoring gaps in data-sparse regions.
42. PLANING: A Loosely Coupled Triangle-Gaussian Framework for Streaming 3D Reconstruction
Core Problem: Streaming 3D reconstruction from monocular image sequences remains challenging, as existing methods typically favor either high-quality rendering or accurate geometry, but rarely both, and suffer from structural redundancy.
Key Innovation: Introduces PLANING, an efficient on-the-fly reconstruction framework built on a hybrid representation that loosely couples explicit geometric primitives with neural Gaussians, enabling decoupled geometry and appearance modeling, stable streaming reconstruction, and significantly faster processing with high quality.
43. An efficient, accurate, and interpretable machine learning method for computing probability of failure
Core Problem: Computing the probability of failure for complex systems, determined by a threshold condition on a computer model, requires an efficient, accurate, and interpretable machine learning method that minimizes expensive model evaluations while preserving decision boundary geometry.
Key Innovation: Introduction of the Penalized Profile Support Vector Machine based on the Gabriel edited set, an interpretable ML method that employs an adaptive sampling strategy to strategically allocate points near the failure boundary and builds a locally linear surrogate boundary consistent with its geometry, proving convergence and outperforming state-of-the-art methods.
44. Soft Masked Transformer for Point Cloud Processing with Skip Attention-Based Upsampling
Core Problem: Point cloud processing methods often overlook task-level context during the encoding stage, and effective communication between encoding and decoding layers for high-level tasks like segmentation is challenging, leading to increased network parameters and training time.
Key Innovation: Proposes SMTransformer, which integrates task-level information into a vector-based transformer using a soft mask generated from task-level queries and keys. It introduces a skip-attention-based up-sampling block for dynamic feature fusion and a shared position encoding strategy to reduce parameters and training time, achieving state-of-the-art semantic segmentation.
45. Adaptive Swarm Mesh Refinement using Deep Reinforcement Learning with Local Rewards
Core Problem: Classical Adaptive Mesh Refinement (AMR) in Finite Element Method (FEM) relies on heuristics or expensive error estimators, leading to suboptimal performance for complex simulations, and existing machine learning-based AMR methods do not scale well.
Key Innovation: Formulates AMR as a system of collaborating, homogeneous agents that iteratively split, enabling a spatial reward formulation focused on reducing maximum mesh element error. The proposed ASMR++ offers efficient, stable optimization and generates highly adaptive meshes that outperform heuristic and learned baselines.
46. From Limited Labels to Open Domains:An Efficient Learning Method for Drone-view Geo-Localization
Core Problem: Traditional supervised drone-view geo-localization (DVGL) methods heavily depend on paired training data and struggle with cross-view correlations from unpaired data. Adapting to new domains requires costly retraining, and unsupervised methods suffer from feature confusion.
Key Innovation: Proposes CDIKTNet, a cross-domain invariant knowledge transfer network with limited supervision. It consists of a cross-domain invariance sub-network (CDIS) to learn structural and spatial invariance from small paired data, and a cross-domain transfer sub-network (CDTS) using dual-path contrastive learning. This alleviates feature confusion and achieves state-of-the-art performance.
47. Towards Anomaly-Aware Pre-Training and Fine-Tuning for Graph Anomaly Detection
Core Problem: Graph anomaly detection (GAD) remains challenging due to label scarcity (high annotation cost) and homophily disparity at node and class levels.
Key Innovation: Introduces Anomaly-Aware Pre-Training and Fine-Tuning (APF), a framework that incorporates node-specific subgraphs (selected via Rayleigh Quotient) and learnable spectral polynomial filters during pre-training to enhance anomaly awareness. Fine-tuning uses a gated fusion mechanism and anomaly-aware regularization, achieving superior performance on 10 benchmark datasets.
48. Dealing with Uncertainty in Contextual Anomaly Detection
Core Problem: Contextual anomaly detection (CAD) needs to identify anomalies in a target variable conditioned on contextual variables, but existing methods often fail to explicitly model both aleatoric and epistemic uncertainties, limiting reliability and interpretability.
Key Innovation: Proposes Normalcy Score (NS), a novel CAD framework built on heteroscedastic Gaussian process regression, which explicitly models both uncertainties and provides confidence intervals for anomaly assessment. NS outperforms SOTA CAD methods in accuracy and interpretability, enabling uncertainty-driven decision-making.
49. Enhancing Semantic Segmentation with Continual Self-Supervised Pre-training
Core Problem: Adapting vision foundation models to low-data target domains for downstream dense prediction tasks like semantic segmentation through continual self-supervised pre-training remains underexplored.
Key Innovation: Proposes GLARE (Global Local and Regional Enforcement), a novel continual self-supervised pre-training task that enhances downstream semantic segmentation performance by introducing patch-level augmentations for local consistency and a regional consistency constraint leveraging spatial semantics. It efficiently updates lightweight adapter modules, consistently improving performance with minimal overhead.
50. MOMEMTO: Patch-based Memory Gate Model in Time Series Foundation Model
Core Problem: Reconstruction-based deep models for time series anomaly detection tend to over-generalize, accurately reconstructing unseen anomalies, and existing memory architectures for normal patterns suffer from high training costs and poor integration with time series foundation models (TFMs).
Key Innovation: Proposes MOMEMTO, an improved TFM for anomaly detection with a patch-based memory module to mitigate over-generalization. It captures representative normal patterns from multiple domains, enables joint fine-tuning, and achieves higher AUC and VUS scores compared to baselines, particularly in few-shot learning scenarios.
51. Expanding the Chaos: Neural Operator for Stochastic (Partial) Differential Equations
Core Problem: Learning solution operators for Stochastic Differential Equations (SDEs) and Stochastic Partial Differential Equations (SPDEs) with deep learning models for fast solvers and new perspectives.
Key Innovation: Designing neural operator (NO) architectures for SDEs and SPDEs based on Wiener-chaos expansions (WCE), which project driving noise paths onto orthonormal Wick-Hermite features and use NOs to parameterize chaos coefficients, enabling full trajectory reconstruction from noise in a single forward pass, and showing competitive accuracy across various tasks including flood forecasting.
52. Novel particle reconstruction and tracking algorithms to reveal 3D micromechanical behaviors of coral sands
Core Problem: The micromechanical behaviors of coral sands are poorly understood due to the inherent complexity of their highly irregular particle shapes, which pose significant difficulties for accurate three-dimensional (3D) particle reconstruction and tracking using X-ray tomography (µCT).
Key Innovation: Proposed a novel framework integrating large vision models and discrete label optimization for efficient and accurate 3D particle reconstruction with optimal transport for robust particle tracking. This framework effectively resolves tracking both particle breakage and internal voids, revealing heterogeneous local shear deformation and fabric anisotropy in coral sands.
53. Strength, Stiffness, and Hysteretic Damping of Steel Pile–Sand Interfaces under Cyclic Axial Loading
Core Problem: The most critical parameters of the soil–foundation interface (strength, stiffness, and hysteretic damping) in dynamic analysis involving offshore wind foundations are not fully understood, especially under cyclic axial loading.
Key Innovation: Investigates the behavior of the sand–pile interface under cyclic axial loading to determine its strength, stiffness, and hysteretic damping, which are critical parameters for dynamic analysis of offshore wind foundations.
54. Impact of geological conceptualization in predicting pore pressure reduction from urban excavations
Core Problem: The impact of geological conceptualization on the accuracy of predicting pore pressure reduction from urban excavations is not well understood, affecting ground stability assessments.
Key Innovation: Analysis of the impact of geological conceptualization on predicting pore pressure reduction from urban excavations, aiming to improve accuracy in geotechnical assessments.
55. Scale-dependent connectivity behavior in multi-clustered fracture systems
Core Problem: Understanding the scale-dependent connectivity behavior in multi-clustered fracture systems is crucial for various geological and engineering applications, but remains complex.
Key Innovation: Investigation of scale-dependent connectivity behavior in multi-clustered fracture systems to enhance understanding of fluid flow and mechanical properties.
56. Devising wireless flood-sensing networks for critical infrastructure facilities
Core Problem: Developing effective and reliable wireless sensing networks for monitoring flood conditions in critical infrastructure facilities.
Key Innovation: Proposing and devising wireless flood-sensing networks specifically tailored for critical infrastructure facilities.
57. Mapping land degradation in the Massili River Basin, Burkina Faso: a spatio-temporal analysis of contributing factors
Core Problem: Mapping land degradation and analyzing its spatio-temporal contributing factors in the Massili River Basin.
Key Innovation: A spatio-temporal analysis and mapping of land degradation and its contributing factors in the Massili River Basin, Burkina Faso.
58. Seasonal flooding amplifies the positive asymmetric response of ecosystem carbon exchange along the precipitation gradient in saline wetlands
Core Problem: Understanding how seasonal flooding impacts ecosystem carbon exchange in saline wetlands along a precipitation gradient.
Key Innovation: Demonstrating that seasonal flooding amplifies the positive asymmetric response of ecosystem carbon exchange along the precipitation gradient in saline wetlands.
59. Deterioration characteristics of red beds lithological interface under water-rock interaction and its influence on tunnel deformation
Core Problem: Understanding the deterioration characteristics of red beds lithological interfaces under water-rock interaction and their impact on tunnel deformation and stability.
Key Innovation: Investigated the specific mechanisms of rock deterioration at red beds interfaces due to water-rock interaction and quantified its influence on tunnel stability.
60. Three-dimensional identification and attribution of flash and slow agricultural droughts in the North China Plain
Core Problem: Accurately identifying and attributing different types of agricultural droughts (flash and slow) in a complex region.
Key Innovation: Developing a three-dimensional approach for the identification and attribution of flash and slow agricultural droughts.
61. The impact of rainfall characteristics on combined sewer overflows in wet weather
Core Problem: Quantifying how different rainfall characteristics influence the occurrence and magnitude of combined sewer overflows (CSOs).
Key Innovation: Analysis of the specific impact of rainfall characteristics on combined sewer overflows during wet weather.
62. A comparative analysis of different physically based modeling approaches to predict the fate of sewer leaks and their effects on urban aquifer recharge and contamination
Core Problem: Accurately predicting the movement and impact of sewer leaks on urban groundwater systems and potential contamination.
Key Innovation: A comparative analysis of physically based modeling approaches for predicting the fate of sewer leaks and their effects on urban aquifers.
63. A modified nonlinear seepage-stress coupling model of heterogeneous pore-fracture dual medium for engineering geomaterials
Core Problem: Developing an improved model to accurately simulate coupled seepage and stress behavior in complex heterogeneous geomaterials.
Key Innovation: A modified nonlinear seepage-stress coupling model for heterogeneous pore-fracture dual medium in engineering geomaterials.
64. The Influence of Elevated Temperature on Time‐Dependent Compaction Creep in Bleursville Sandstone
Core Problem: Systematically exploring the influence of realistic crustal temperatures on sandstone compaction creep over longer time scales.
Key Innovation: Experimental investigation of time-dependent compaction creep in Bleursville sandstone at elevated temperatures (up to 150°C), revealing reduced yield stress and significantly higher creep rates, and identifying a shift in deformation mechanisms from subcritical cracking to pressure solution at higher temperatures.
65. Fracture‐Assisted Pressure Solution Creep of Granite: An Example From the Mont Blanc Massif, Western Alps
Core Problem: Understanding the microstructure and ductile deformation mechanisms (fracture-assisted pressure solution creep) in deformed granites like those in the Mont Blanc massif.
Key Innovation: Quantitative analysis of finite strain and development of a viscous creep model to explain fracture-assisted pressure solution creep in Mont Blanc granite, linking microstructure to deformation mechanisms and suggesting its role in ductile deformation and shear zone geometry.
66. Variable Flood Discharge Constrains Autochthonous Organic Carbon Preservation in Deltas: Insights From Physical Experiments
Core Problem: Unclear how predicted shifts in precipitation frequency and magnitude will influence delta and wetland development and organic carbon preservation.
Key Innovation: Physical delta experiments showing that variable flood discharge significantly reduces delta-top area, increases slopes, enhances channel mobility, and limits shoreline movement, leading to a substantial difference in preserved organic material.
67. Resolving Convection Doubles Sahel's Contribution to Global Dust Emission During the Monsoon Season
Core Problem: Current climate models struggle to capture dust emissions, partly due to their inability to resolve convection, leading to underestimation of contributions from key regions like the Sahel.
Key Innovation: Analysis showing that resolving convection in a global storm-resolving model doubles the Sahel's contribution to global dust emissions during the monsoon season by increasing friction velocity and decreasing soil moisture, due to improved representation of mesoscale convective systems.
68. A Diagnostic Framework and Data Inventory to Analyze Human Intervention on Streamflow
Core Problem: Data constraints often force simplifying assumptions in hydrological models regarding human impacts on streamflow, introducing biases and obscuring true human influences.
Key Innovation: Compilation of Data_IH2, a data inventory of human interventions (reservoir operations, inter-basin transfers, water supplies) for the Contiguous United States, and development of a modeling framework leveraging this data and the Budyko hypothesis to diagnose which management activities most strongly modify streamflow regimes.
69. NZ-BeachTopo30: A national-scale and full-coverage 30 m beach topography dataset for New Zealand reconstructed by fusing ICESat-2 and Sentinel-2
Core Problem: A critical gap in global coastal data, specifically the lack of a seamless, nationwide map of New Zealand's beach elevations.
Key Innovation: Production of the first seamless, nationwide map of New Zealand's beach elevations at 30-meter resolution by fusing ICESat-2 and Sentinel-2 data, rigorously validated, expanding usable coastal data, and providing an essential foundation for planning against sea-level rise and coastal erosion.
70. Urban agglomerations environmental heterogeneity and heatwave risks: spatiotemporal insights from remote sensing and public sentiment analysis
Core Problem: Understanding the spatiotemporal heatwave risks in urban agglomerations, considering environmental heterogeneity.
Key Innovation: Gaining insights into heatwave risks using remote sensing and public sentiment analysis.
71. Predict-Project-Renoise: Sampling Diffusion Models under Hard Constraints
Core Problem: Neural emulators based on diffusion models cannot guarantee physical accuracy or constraint satisfaction, which is critical for scientific applications.
Key Innovation: Introduces a constrained sampling framework that enforces hard constraints at generation time by defining a constrained forward process and proposes Predict-Project-Renoise (PPR), an iterative algorithm that alternates between denoising predictions, projecting onto the feasible set, and renoising, reducing constraint violations by over an order of magnitude.
72. Model-Free Neural State Estimation in Nonlinear Dynamical Systems: A Comparative Study of Neural Architectures and Classical Filters
Core Problem: Neural network models are increasingly used for state estimation in control and decision-making but their behavior as principled filters in nonlinear dynamical systems, especially when trained purely from data without explicit system dynamics, is unclear.
Key Innovation: A systematic empirical comparison of Transformer-based models, state-space neural networks, and recurrent architectures against particle filters and nonlinear Kalman filters in multiple nonlinear scenarios. It shows that neural models (especially SSMs) achieve state estimation performance approaching strong nonlinear Kalman filters and outperform weaker classical baselines with higher inference throughput, despite lacking access to system models.
73. Gaussian Belief Propagation Network for Depth Completion
Core Problem: Effectively handling the sparse and irregular nature of input depth data in deep networks for depth completion remains a significant challenge, limiting performance, especially under high sparsity.
Key Innovation: Gaussian Belief Propagation Network (GBPN), a novel hybrid framework integrating deep learning with probabilistic graphical models for end-to-end depth completion. It dynamically constructs a scene-specific Markov Random Field (MRF) via a Graphical Model Construction Network (GMCN) that learns adaptive non-local edges, and enhances GBP with a serial & parallel message passing scheme, achieving SOTA performance and robustness.
74. Physics-Guided Tiny-Mamba Transformer for Reliability-Aware Early Fault Warning
Core Problem: Reliability-centered prognostics for rotating machinery requires early warning signals that are accurate under nonstationary conditions, domain shifts, and class imbalance, while maintaining a small and predictable false-alarm rate.
Key Innovation: Physics-Guided Tiny-Mamba Transformer (PG-TMT), a compact tri-branch encoder for online condition monitoring. It combines convolutional, Mamba state-space, and lightweight Transformer branches, derives an analytic temporal-to-spectral mapping for physical plausibility, and uses extreme-value theory (EVT) with a dual-threshold hysteresis for reliability-aware decision rules, achieving higher precision-recall AUC and shorter mean time-to-detect.
75. SR$^{2}$-Net: A General Plug-and-Play Model for Spectral Refinement in Hyperspectral Image Super-Resolution
Core Problem: Existing HSI-SR methods often neglect spectral consistency across bands, leading to spurious oscillations and physically implausible artifacts, and addressing this through network architecture design sacrifices generality and flexibility.
Key Innovation: SR$^{2}$-Net (Spectral Rectification Super-Resolution Network), a lightweight plug-and-play rectifier that can be attached to various HSI-SR models without architectural modification. It uses Hierarchical Spectral-Spatial Synergy Attention (H-S$^{3}$A) and Manifold Consistency Rectification (MCR) to reinforce cross-band interactions and constrain spectra to a physically plausible manifold, along with a degradation-consistency loss, achieving consistent improvements in spectral fidelity.
76. DA-SPS: A Dual-stage Network based on Singular Spectrum Analysis, Patching-strategy and Spearman-correlation for Multivariate Time-series Prediction
Core Problem: Existing multivariate time-series forecasting methods do not effectively consider the impact of extraneous variables on target variable prediction and fail to fully extract complex sequence information based on various time patterns.
Key Innovation: DA-SPS, a dual-stage network that uses Singular Spectrum Analysis (SSA), LSTM, and P-Conv-LSTM for target variable processing, and Spearman correlation with an L-Attention module for extraneous variable processing, combining results for improved multivariate time-series prediction.
77. Sim-MSTNet: sim2real based Multi-task SpatioTemporal Network Traffic Forecasting
Core Problem: Network traffic forecasting suffers from poor generalization with limited data, and multi-task learning methods struggle with task imbalance and negative transfer when modeling various service types.
Key Innovation: Sim-MSTNet, a sim2real based multi-task spatiotemporal network traffic forecasting model that uses a simulator for synthetic data generation, domain randomization for reducing the sim-to-real gap, and attention-based mechanisms with dynamic loss weighting for multi-task learning.
78. From Implicit Ambiguity to Explicit Solidity: Diagnosing Interior Geometric Degradation in Neural Radiance Fields for Dense 3D Scene Understanding
Core Problem: Neural Radiance Fields (NeRFs) suffer from 'Interior Geometric Degradation' (IGD) in dense, self-occluding scenes, reconstructing hollow or fragmented structures instead of solid interiors, leading to systematic instance undercounting and unreliable quantitative 3D analysis.
Key Innovation: An explicit geometric pipeline based on Sparse Voxel Rasterization (SVRaster), initialized from SfM feature geometry, which projects 2D instance masks onto an explicit voxel grid and enforces geometric separation via recursive splitting to preserve physical solidity and improve instance recovery in dense 3D scenes.
79. From Consistency to Complementarity: Aligned and Disentangled Multi-modal Learning for Time Series Understanding and Reasoning
Core Problem: Effective cross-modal integration in multi-modal LLMs for time series understanding is challenging due to fine-grained temporal misalignment and severe entanglement between shared and modality-specific semantics, hindering localized interpretation and complementary reasoning.
Key Innovation: MADI, a multi-modal LLM enhanced with fine-grained alignment and disentangled interaction, featuring Patch-level Alignment for cross-modal correspondence, Discrete Disentangled Interaction for separating and synergizing semantics, and Critical-token Highlighting for emphasizing query-relevant signals, improving time series understanding and reasoning.
80. Accurate Network Traffic Matrix Prediction via LEAD: an LLM-Enhanced Adapter-Based Conditional Diffusion Model
Core Problem: Accurate Network Traffic Matrix (TM) forecasting is challenging due to stochastic, non-linear, and bursty network dynamics, with existing discriminative models suffering from over-smoothing and limited uncertainty awareness, leading to poor fidelity under extreme bursts.
Key Innovation: LEAD, a Large Language Model (LLM)-Enhanced Adapter-based conditional Diffusion model that transforms traffic matrices into RGB images ('Traffic-to-Image' paradigm), uses a 'Frozen LLM with Trainable Adapter' for temporal semantics, and a Dual-Conditioning Strategy to guide the diffusion model for accurate, dynamic network traffic matrix generation.
81. Multi-Modal Time Series Prediction via Mixture of Modulated Experts
Core Problem: Accurate multi-modal time series forecasting is challenging due to complex, evolving dynamics, and existing methods relying on token-level fusion are ill-suited for scarce high-quality time-text pairs or substantial time series variation, complicating cross-modal alignment.
Key Innovation: Expert Modulation, a new paradigm for multi-modal time series prediction that conditions both routing and expert computation on textual signals, enabling direct and efficient cross-modal control over expert behavior, demonstrating substantial improvements in forecasting.
82. Seg-MoE: Multi-Resolution Segment-wise Mixture-of-Experts for Time Series Forecasting Transformers
Core Problem: Transformer-based models for time-series forecasting struggle to scale efficiently while capturing long-term temporal dynamics, as existing Mixture-of-Experts (MoE) approaches use token-wise routing that fails to exploit temporal data's locality and continuity.
Key Innovation: Seg-MoE, a sparse Mixture-of-Experts design for time-series forecasting Transformers, routes and processes contiguous time-step segments, allowing experts to model intra-segment interactions, achieving state-of-the-art accuracy across multivariate long-term forecasting benchmarks.
83. The Ensemble Inverse Problem: Applications and Methods
Core Problem: A new multivariate statistical problem, the Ensemble Inverse Problem (EIP), arises in fields like high energy physics, full waveform inversion (FWI), and inverse imaging, requiring inversion for an ensemble distributed according to a pushforward of a prior under a forward process.
Key Innovation: Proposes non-iterative inference-time methods using a new class of conditional generative models (ensemble inverse generative models) that leverage ensemble information in the observation set to construct posterior samplers, avoiding explicit iterative use of the forward model at inference time.
84. BLO-Inst: Bi-Level Optimization Based Alignment of YOLO and SAM for Robust Instance Segmentation
Core Problem: Integrating object detectors as prompt generators for SAM suffers from objective mismatch (detectors optimized for localization don't provide optimal SAM prompts) and alignment overfitting, hindering fully automated and generalizable instance segmentation.
Key Innovation: Introduces BLO-Inst, a unified bi-level optimization framework that aligns detection and segmentation objectives by formulating it as a nested optimization problem, effectively transforming the detector into a segmentation-aware prompt generator for improved mask quality.
85. Scale-Equivariant Imaging: Self-Supervised Learning for Image Super-Resolution and Deblurring
Core Problem: Existing self-supervised methods for imaging inverse problems like super-resolution and deblurring fail to obtain competitive performances because invariance to roto-translations is insufficient to learn from measurements containing only low-frequency information.
Key Innovation: Proposes scale-equivariant imaging, a new self-supervised approach that leverages the approximate scale-invariance of many image distributions to recover high-frequency information lost in the measurement process. It outperforms other self-supervised methods and achieves performance on par with fully supervised learning.
86. An explainable vision transformer with transfer learning based efficient drought stress identification
Core Problem: Early detection of drought stress is critical for reducing crop loss, and while Vision Transformers (ViTs) offer potential for capturing subtle indicators from aerial imagery, their decision-making process needs to be interpretable.
Key Innovation: Proposes an explainable deep learning pipeline leveraging ViTs for drought stress detection in potato crops using aerial imagery. It combines ViT with SVM or uses an end-to-end ViT, and visualizes attention maps to explain the model's decision-making process, achieving high accuracy and interpretability.
87. EROAM: Event-based Camera Rotational Odometry and Mapping in Real-time
Core Problem: Existing event-based camera odometry and mapping approaches struggle with real-time, accurate camera rotation estimation, often relying on less robust methods like event generation models or contrast maximization.
Key Innovation: Presents EROAM, a novel event-based rotational odometry and mapping system that achieves real-time, accurate camera rotation estimation. It employs a spherical event representation and introduces Event Spherical Iterative Closest Point (ES-ICP), featuring efficient map management and parallel point-to-line optimization.
88. Visual Localization via Semantic Structures in Autonomous Photovoltaic Power Plant Inspection
Core Problem: Automating UAV inspection of PV power plants requires precise navigation to capture optimal images, which is challenging due to the need for accurate localization relative to the plant structures.
Key Innovation: Presents a novel localization pipeline that directly integrates PV module detection with UAV navigation. It uses detections to identify power plant structures, associates them with a power plant model, and infers UAV position, demonstrating robustness and real-time applicability for inspection.
89. Neural Force Field: Few-shot Learning of Generalized Physical Reasoning
Core Problem: Current AI models struggle with few-shot learning and generalization in Out-of-distribution (OOD) physical reasoning settings because they cannot abstract core physical principles from observations.
Key Innovation: Presents Neural Force Field (NFF), a framework extending Neural Ordinary Differential Equation (NODE) to learn complex object interactions through continuous explicit force field representations. NFF, trained with few examples, achieves strong generalization to unseen scenarios in physical reasoning tasks, enabling efficient planning and rapid adaptation.
90. Beyond Retraining: Training-Free Unknown Class Filtering for Source-Free Open Set Domain Adaptation of Vision-Language Models
Core Problem: Vision-language models (VLMs) struggle to reject samples from emerging unknown classes in open-set domain adaptation, especially when only unlabeled data is available. Existing retraining methods distort VLM geometry, and score-based detectors suffer from threshold sensitivity.
Key Innovation: Proposes VLM-OpenXpert, a training-free, plug-and-play inference pipeline. It includes SUFF (SVD on high-confidence unknowns to extract an 'unknown subspace' for feature suppression) and BGAT (Box-Cox transform and bimodal Gaussian mixture for adaptive thresholding). It matches or outperforms retraining-heavy state-of-the-art methods.
91. Predictability-Aware Compression and Decompression Framework for Multichannel Time Series Data with Latent Seasonality
Core Problem: Real-world multichannel time series prediction faces growing demands for efficiency (runtime, communication cost) while maintaining accuracy, especially in edge and cloud environments.
Key Innovation: Proposes a predictability-aware compression-decompression framework using a circular seasonal key matrix with orthogonality to capture latent seasonality and mitigate reconstruction errors. This framework is time-efficient, accuracy-preserving, and compatible with diverse predictors, achieving superior overall performance on various datasets.
92. SDSC:A Structure-Aware Metric for Semantic Signal Representation Learning
Core Problem: Most Self-Supervised Learning (SSL) methods for signals use distance-based objectives (e.g., MSE) that are sensitive to amplitude, invariant to waveform polarity, and unbounded, hindering semantic alignment and interpretability.
Key Innovation: Proposes Signal Dice Similarity Coefficient (SDSC), a structure-aware metric quantifying structural agreement based on signed amplitudes. SDSC, used as a loss function, achieves comparable or improved performance over MSE in forecasting and classification, particularly in low-resource scenarios, by enhancing structural fidelity in signal representations.
93. BIR-Adapter: A parameter-efficient diffusion adapter for blind image restoration
Core Problem: Diffusion-based image restoration methods often rely on auxiliary feature extractors or extensive fine-tuning, leading to high computational cost and parameter overhead, despite large-scale pre-trained diffusion models retaining informative representations under degradations.
Key Innovation: Introduces BIR-Adapter, a parameter-efficient, plug-and-play attention mechanism that substantially reduces trained parameters (up to 36x fewer) for blind image restoration. It also proposes a sampling guidance mechanism to mitigate hallucinations, achieving competitive or superior performance compared to SOTA methods and enabling seamless integration into existing models.
94. Physics-Informed Neural Networks for Real-Time Gas Crossover Prediction in PEM Electrolyzers: First Application with Multi-Membrane Validation
Core Problem: Current physics-based models for hydrogen crossover in PEM electrolyzers require extensive calibration and computational resources that preclude real-time implementation, while purely data-driven approaches fail to extrapolate beyond training conditions, which is critical for dynamic electrolyzer operation.
Key Innovation: Presents the first application of physics-informed neural networks (PINNs) for hydrogen crossover prediction, trained on augmented measurements and constrained by a constitutive physics model embedded in the loss function, achieving exceptional accuracy and sub-millisecond inference for real-time control and maintaining high R^2 when predicting crossover beyond training range.
95. SSCATeR: Sparse Scatter-Based Convolution Algorithm with Temporal Data Recycling for Real-Time 3D Object Detection in LiDAR Point Clouds
Core Problem: Real-time 3D object detection in LiDAR point clouds is computationally intensive, as existing sparse convolution techniques process entire frames even when only small regions change, leading to inefficiency.
Key Innovation: Proposes SSCATeR, a Sparse Scatter-Based Convolution Algorithm with Temporal Data Recycling, which leverages the continuous sweeping motion of LiDAR scanning to concentrate object detection efforts on specific regions that receive a change in point data, storing convolution results between passes to ignore unchanged regions, significantly reducing convolution operations while maintaining accuracy.
96. Reliable Deep Learning for Small-Scale Classifications: Experiments on Real-World Image Datasets from Bangladesh
Core Problem: Complex CNN architectures often overfit on small datasets, making them unreliable for real-world image recognition tasks with limited data.
Key Innovation: Demonstrating that a compact CNN architecture can achieve high classification accuracy, efficient convergence, and low computational overhead on small, real-world image datasets, effectively capturing discriminative features and generalizing robustly for small-class image classification tasks.
97. Real-Time Pulsatile Flow Prediction for Realistic, Diverse Intracranial Aneurysm Morphologies using a Graph Transformer and Steady-Flow Data Augmentation
Core Problem: The clinical translation of fluid mechanical markers for intracranial aneurysms is hindered by the time-consuming and specialized nature of Computational Fluid Dynamics (CFD), making real-time prediction difficult.
Key Innovation: A Graph Transformer model that incorporates temporal information, supervised by large CFD data, to accurately predict Wall Shear Stress (WSS) across the cardiac cycle from IA surface meshes in real-time, further enhanced by injecting low-cost steady-state CFD data as augmentation to improve performance with small pulsatile data samples.
98. Temporally-Similar Structure-Aware Spatiotemporal Fusion of Satellite Images
Core Problem: Satellite images are often degraded by noise, and existing noise-robust ST fusion methods fail to capture fine spatial structures, leading to oversmoothing and artifacts.
Key Innovation: Introduces TSSTF, incorporating Temporally-Guided Total Variation (TGTV) and Temporally-Guided Edge Constraint (TGEC) to promote spatial piecewise smoothness and consistency in edge locations, outperforming state-of-the-art methods under noisy conditions.
99. Threshold criteria for silt movement and suspension on sloped bed under waves
Core Problem: Defining the conditions (threshold criteria) under which silt moves and becomes suspended on sloped beds under wave action.
Key Innovation: Establishing threshold criteria for silt movement and suspension on sloped beds under waves.
100. Soft soil deformation and erosion in riser touchdown zone: insights from T-bar penetration tests
Core Problem: Understanding soft soil deformation and erosion processes in the riser touchdown zone.
Key Innovation: Gaining insights into soft soil deformation and erosion through T-bar penetration tests.
101. Granular cementation failure mapped at microscale: novel apparatus quantifies compression-shear-bending interactions in marine calcareous sands
Core Problem: Limited understanding and quantification of granular cementation failure mechanisms, particularly compression-shear-bending interactions, at the microscale in marine calcareous sands.
Key Innovation: Development of a novel apparatus to map granular cementation failure at the microscale and quantify compression-shear-bending interactions in marine calcareous sands.
102. Effects of Multiscale Fluid Flows on Wave Propagation in Partially Saturated Porous Rock
Core Problem: The effects of multiscale fluid-flow mechanisms on wave propagation in partially saturated porous rock are not fully understood, limiting the comprehensive modeling of loss mechanisms across microscopic, mesoscopic, and macroscopic scales.
Key Innovation: A unified model integrating the Rayleigh model of bubble oscillations and the squirt-flow model into Santos’ poroelasticity theory to investigate multiscale fluid-flow effects on wave propagation in partially saturated porous rock. This model reveals three distinct P1-wave attenuation peaks and a three-stage phase velocity dispersion, providing insights into wave propagation, reflection, and surface waves under various boundary conditions.
103. Safety evaluation of operational metro shield tunnels using improved game theory and dynamic variable weight theory
Core Problem: Conducting a comprehensive safety evaluation for operational metro shield tunnels, considering complex interactions and uncertainties.
Key Innovation: An improved safety evaluation method for metro shield tunnels combining game theory and dynamic variable weight theory.
104. TIDA-SR: A time-conditioned deformable attention network for DEM super-resolution in cloud-covered mountainous regions
Core Problem: Obtaining high-resolution Digital Elevation Models (DEMs) in cloud-covered mountainous regions, which is challenging for remote sensing and crucial for geohazard analysis.
Key Innovation: TIDA-SR, a time-conditioned deformable attention network for DEM super-resolution, specifically designed for cloud-covered mountainous regions.
105. Mapping herbivore-accessible biomass across a heterogeneous mountain landscape using multisensor high-resolution UAV data
Core Problem: Mapping herbivore-accessible biomass across complex and heterogeneous mountain landscapes.
Key Innovation: Mapping herbivore-accessible biomass across a heterogeneous mountain landscape using multisensor high-resolution UAV data.
106. Enhancing aboveground biomass density estimates in tropical forests using GEDI waveform data
Core Problem: Improving the accuracy of aboveground biomass density estimates in tropical forests.
Key Innovation: Enhancing aboveground biomass density estimates in tropical forests by utilizing GEDI waveform data.
107. Kilometer-scale simulation of atmospheric in-cloud ground icing over complex terrain in Norwegian Arctic
Core Problem: Simulating atmospheric in-cloud ground icing over complex terrain at a kilometer scale in the Norwegian Arctic.
Key Innovation: Developed and applied a kilometer-scale simulation model to predict ground icing in a complex Arctic environment, which can inform geohazard assessments.
108. A physically-based model for particle transport over urban road surface with consideration of raindrop erosion and runoff effect
Core Problem: Accurately modeling particle transport on urban road surfaces, accounting for complex hydrological processes.
Key Innovation: A physically-based model incorporating raindrop erosion and runoff effects for urban road surface particle transport.
109. Insights into LNAPL saturation distribution in capillary zone based on light transmission and mechanical analysis
Core Problem: Characterizing the complex saturation distribution of Light Non-Aqueous Phase Liquids (LNAPL) in the capillary zone.
Key Innovation: Utilizing light transmission and mechanical analysis to gain insights into LNAPL saturation distribution in the capillary zone.
110. Performance assessment of an electrokinetic barrier-aquifer system using a dual-domain solute transport model
Core Problem: Evaluating the effectiveness of electrokinetic barriers for remediating contaminated aquifers.
Key Innovation: Performance assessment of an electrokinetic barrier-aquifer system using a dual-domain solute transport model.
111. Physics-informed neural networks (PINNs) for dynamic pile-soil interaction problems
Core Problem: Efficiently and accurately solving complex dynamic pile-soil interaction problems in geotechnical engineering.
Key Innovation: Application of Physics-informed neural networks (PINNs) to model dynamic pile-soil interaction problems.
112. A novel infrared thermography-based method for measuring suction at evaporating soil surfaces
Core Problem: Developing a novel method for measuring suction at evaporating soil surfaces.
Key Innovation: A novel infrared thermography-based method for measuring suction at evaporating soil surfaces.
113. Confronting Historical Precipitation Trends in Models With Observations: Forced Signal and Atmospheric Internal Variability
Core Problem: Evaluating climate models' ability to simulate historical precipitation changes and distinguishing between forced signals and atmospheric internal variability.
Key Innovation: Using multiple atmospheric models and ensemble simulations to estimate forced signals and internal variability in precipitation trends since 1980, finding consistency in forced trends across models, significant regional influence of internal variability, and identifying regions where wetting/drying are likely forcing-driven.
114. Field‐Scale Soil Moisture Predictions in Real Time Using In Situ Sensor Measurements in an Inverse Modeling Framework: SWIM2
Core Problem: The need for accurate, real-time soil moisture predictions for irrigation optimization and water management, despite challenges like sensor bias and model input uncertainty.
Key Innovation: Introduction of SWIM2, a digital twin integrating continuous sensor data and soil samples with an FAO-based soil water balance model using Bayesian inverse modeling (DREAM(ZS)), achieving robust, accurate, and precise real-time soil moisture predictions for a 7-day horizon, even with minimal prior knowledge and sensor bias.
115. Comparing Snow Water Equivalent Estimations From Long Short‐Term Memory Networks and Physics‐Based Models in the Western United States
Core Problem: Accurate estimation of snow water equivalent (SWE) is critical but challenging due to the spatial variability and complexity of snow processes, and the application of machine learning (LSTM) to SWE remains underexplored.
Key Innovation: Development and training of an LSTM model to estimate SWE at point locations in the western U.S., demonstrating superior performance in magnitude and temporal accuracy, higher adaptability across snowpack regimes and erroneous forcing, and reliance on physically relevant characteristics, compared to physics-based models.
116. Detecting the occurrence of preferential flow in soils with stable water isotopes
Core Problem: Detecting and understanding preferential flow paths in soils, where water bypasses slower flow, is challenging but crucial for water resource management and environmental protection.
Key Innovation: Introduces a new, simple, and non-invasive method using stable water isotopes in soil samples to detect and identify the occurrence of preferential flow in soils over large areas.
117. Conditional Denoising Model as a Physical Surrogate Model
Core Problem: Surrogate modeling for complex physical systems typically faces a trade-off between data-fitting accuracy and physical consistency, with physics-consistent approaches often failing to guarantee strict adherence to governing equations.
Key Innovation: Introduces the Conditional Denoising Model (CDM), a generative model that learns the geometry of the physical manifold by training to restore clean states from noisy ones, achieving higher parameter and data efficiency and adhering to physical constraints more strictly than baselines, despite never seeing governing equations during training.
118. PHDME: Physics-Informed Diffusion Models without Explicit Governing Equations
Core Problem: Diffusion models are unreliable in sparse data regimes for forecasting dynamical systems, and existing physics-informed machine learning (PIML) methods typically require explicit governing equations, which are often only partially known.
Key Innovation: PHDME, a port-Hamiltonian diffusion framework that leverages structural priors without requiring full knowledge of governing equations. It trains a Gaussian process distributed Port-Hamiltonian system (GP-dPHS) on limited observations to capture energy-based dynamics, generates a physically consistent artificial dataset, and informs the diffusion model with a structured physics residual loss.
119. GeoRC: A Benchmark for Geolocation Reasoning Chains
Core Problem: Vision Language Models (VLMs) are good at geolocation prediction but often fail to explain their reasoning, hallucinating scene attributes, indicating a "reasoning horizon gap" in extracting fine-grained visual attributes from high-resolution images.
Key Innovation: GeoRC, the first benchmark for geolocation reasoning chains, created with expert GeoGuessr players, providing 800 ground truth reasoning chains for 500 Google Street View scenes. It evaluates VLM-generated reasoning chains, revealing that while large VLMs rival humans in prediction, they lag significantly in producing auditable reasoning chains, especially open-weight models.
120. WorldBench: Disambiguating Physics for Diagnostic Evaluation of World Models
Core Problem: Existing physics-based video benchmarks for generative "world models" suffer from entanglement, evaluating multiple physical laws simultaneously, which limits their diagnostic capability for assessing physical fidelity.
Key Innovation: WorldBench, a novel video-based benchmark for concept-specific, disentangled evaluation of world models' physical understanding. It includes benchmarks for intuitive physical understanding and low-level physical constants, revealing specific failure patterns in SOTA models and highlighting their lack of physical consistency for reliable real-world interactions.
121. PILD: Physics-Informed Learning via Diffusion
Core Problem: Diffusion models, being purely data-driven, have limited applicability in engineering and scientific problems where physical laws must be followed, especially when data is sparse.
Key Innovation: PILD (Physics-Informed Learning via Diffusion), a framework that unifies diffusion modeling and first-principles physical constraints. It introduces a virtual residual observation sampled from a Laplace distribution to supervise generation and incorporates a conditional embedding module to inject physical information into the denoising network, improving accuracy, stability, and generalization across various engineering and scientific tasks.
122. TimeSliver : Symbolic-Linear Decomposition for Explainable Time Series Classification
Core Problem: Explaining model decisions in time series classification by identifying the influence of every temporal segment is challenging, as existing post-hoc methods suffer from reference state sensitivity and struggle with sequential dependencies, while attention weights often lack faithfulness.
Key Innovation: TimeSliver, a novel explainability-driven deep learning framework that jointly uses raw time-series data and its symbolic abstraction to construct a representation where each element linearly encodes the contribution of each temporal segment. It outperforms other temporal attribution methods and achieves competitive predictive performance on various time-series datasets.
123. Epistemic Uncertainty Quantification for Pre-trained VLMs via Riemannian Flow Matching
Core Problem: Pre-trained Vision-Language Models (VLMs) are deterministic and lack intrinsic mechanisms to quantify epistemic uncertainty, which is crucial for understanding model ignorance and reliability.
Key Innovation: REPVLM, a method that quantifies epistemic uncertainty in VLMs by computing the probability density on the hyperspherical manifold of VLM embeddings using Riemannian Flow Matching, achieving near-perfect correlation with prediction error and enabling scalable out-of-distribution detection.
124. SmartMeterFM: Unifying Smart Meter Data Generative Tasks Using Flow Matching Models
Core Problem: Diverse generative tasks for smart meter data (synthetic generation, imputation, super-resolution) require dedicated models, leading to redundancy and inefficiency, while data availability is limited by privacy or corruption.
Key Innovation: SmartMeterFM, a novel approach using flow matching models, unifies diverse smart meter data generative tasks (imputation, super-resolution, synthetic generation) into a single conditional generation model, eliminating retraining and producing realistic, consistent high-dimensional time series data.
125. Goal-Driven Adaptive Sampling Strategies for Machine Learning Models Predicting Fields
Core Problem: Ensuring a desired level of accuracy for machine learning models predicting fields at minimal computational cost (e.g., as few black-box samples as possible) remains a challenge, as active learning strategies are often limited to scalar quantities or specific model types.
Key Innovation: Proposes a model-agnostic active learning strategy for machine learning models capable of predicting fields. It combines a Gaussian process model for a scalar reference value and simultaneously aims at reducing epistemic model error and the difference between scalar and field predictions, showcasing high accuracy at significantly smaller cost.
126. MoHETS: Long-term Time Series Forecasting with Mixture-of-Heterogeneous-Experts
Core Problem: Real-world multivariate time series exhibit intricate multi-scale structures (global trends, local periodicities, non-stationary regimes), making long-horizon forecasting challenging, and existing sparse Mixture-of-Experts (MoE) approaches typically rely on homogeneous MLP experts that poorly capture diverse temporal dynamics.
Key Innovation: Proposes MoHETS, an encoder-only Transformer that integrates sparse Mixture-of-Heterogeneous-Experts (MoHE) layers, combining a shared depthwise-convolution expert with routed Fourier-based experts. Incorporates exogenous information via cross-attention and uses a lightweight convolutional patch decoder. Achieves state-of-the-art performance on multivariate benchmarks, reducing average MSE by 12%.
127. Generalized Information Gathering Under Dynamics Uncertainty
Core Problem: Existing active information gathering methods for unknown dynamical systems derive bespoke costs for specific modeling choices, lacking a unifying framework that decouples these choices from the information-gathering cost.
Key Innovation: Proposes a unifying framework that explicitly exposes causal dependencies and derives a general information-gathering cost based on Massey's directed information, proving mutual information as a special case and providing theoretical justification for active learning approaches.
128. Negatives-Dominant Contrastive Learning for Generalization in Imbalanced Domains
Core Problem: Imbalanced Domain Generalization (IDG) is underexplored due to the technical complexity of handling entangled domain and label shifts under heterogeneous long-tailed distributions, leading to biased decision boundaries.
Key Innovation: Proposes Negative-Dominant Contrastive Learning (NDCL) for IDG, theoretically establishing a generalization bound and enhancing inter-class decision-boundary separation by emphasizing negatives, while encouraging intra-class compactness and enforcing posterior consistency across domains.
129. SMOG: Scalable Meta-Learning for Multi-Objective Bayesian Optimization
Core Problem: Methods for meta-learned priors in multi-objective Bayesian optimization remain largely unexplored, despite the availability of historical data from related optimization tasks and the need to accelerate optimization with limited measurement budgets.
Key Innovation: Proposes SMOG, a scalable and modular meta-learning model based on a multi-output Gaussian process that explicitly learns correlations between objectives, building a structured joint Gaussian process prior and supporting hierarchical, parallel training for efficient meta-learning.
130. Efficient Causal Structure Learning via Modular Subgraph Integration
Core Problem: Learning causal structures from observational data is computationally intensive in high-dimensional settings due to the super-exponential growth of the search space and increasing computational demands.
Key Innovation: VISTA (Voting-based Integration of Subgraph Topologies for Acyclicity), a modular framework that decomposes global causal structure learning into local subgraphs based on Markov Blankets, integrates them via a weighted voting mechanism, and ensures acyclicity, achieving notable improvements in accuracy and efficiency.
131. Parametric Hyperbolic Conservation Laws: A Unified Framework for Conservation, Entropy Stability, and Hyperbolicity
Core Problem: Learning hyperbolic systems directly from data is challenging, as existing approaches typically enforce only conservation or rely on prior knowledge, failing to guarantee fundamental physical properties like entropy stability and hyperbolicity.
Key Innovation: Introduction of SymCLaw, a parametric hyperbolic conservation law that parameterizes flux functions to guarantee real eigenvalues and complete eigenvectors, embeds entropy-stable design principles by jointly learning a convex entropy function, and provides a corresponding entropy-stable numerical flux scheme, generalizing to unseen initial conditions and maintaining stability.
132. Position: Certifiable State Integrity in Cyber-Physical Systems -- Why Modular Sovereignty Solves the Plasticity-Stability Paradox
Core Problem: Deploying monolithic foundation models in safety-critical Cyber-Physical Systems (CPS) faces challenges like catastrophic forgetting, residual spectral bias, and lack of formal verification, hindering certifiable state integrity across non-stationary lifecycle dynamics.
Key Innovation: Advocating for a 'Modular Sovereignty' paradigm (HYDRA) – a library of compact, frozen regime-specific specialists combined via uncertainty-aware blending – to ensure regime-conditional validity, rigorous disentanglement of uncertainties, and modular auditability, offering a certifiable path for robust state integrity in CPS.
133. Bulk-Calibrated Credal Ambiguity Sets: Fast, Tractable Decision Making under Out-of-Sample Contamination
Core Problem: Huber (linear-vacuous) contamination in distributionally robust optimization (DRO) can make the worst-case risk infinite and the DRO objective vacuous unless strong boundedness or support assumptions are imposed.
Key Innovation: Introduction of bulk-calibrated credal ambiguity sets that learn a high-mass bulk set from data while considering contamination inside the bulk and bounding the remaining tail contribution separately, leading to a closed-form, finite mean+sup robust objective and tractable linear or second-order cone programs for decision making under out-of-sample contamination.
134. 4D-CAAL: 4D Radar-Camera Calibration and Auto-Labeling for Autonomous Driving
Core Problem: Accurate extrinsic calibration of 4D radar and cameras, and the labor-intensive, unreliable manual labeling of sparse radar data, are significant challenges for developing robust radar-based perception algorithms in autonomous driving.
Key Innovation: Proposes 4D-CAAL, a unified framework featuring a novel dual-purpose calibration target (checkerboard for camera, corner reflector for radar) and a robust correspondence matching algorithm for accurate extrinsic calibration, coupled with an auto-labeling pipeline that transfers camera annotations to radar point clouds, significantly reducing manual effort.
135. VSE: Variational state estimation of complex model-free process
Core Problem: Estimating the state of complex dynamical processes from noisy nonlinear measurements is challenging when a suitable physics-based model characterizing the temporal evolution of the process state is unavailable (model-free).
Key Innovation: Designs a Variational State Estimation (VSE) method that provides a closed-form Gaussian posterior of an underlying complex model-free dynamical process from noisy nonlinear measurements, using a recurrent neural network (RNN) for computationally simple inference, and an additional RNN in the learning phase for mutual improvement, demonstrating competitive performance against particle filters and data-driven methods.
136. Efficient4D: Fast Dynamic 3D Object Generation from a Single-view Video
Core Problem: Generating dynamic 3D objects from a single-view video is challenging due to the lack of 4D labeled data, and extending image-to-3D pipelines is slow and expensive due to back-propagating information-limited supervision signals.
Key Innovation: Proposes Efficient4D, a framework that generates high-quality spacetime-consistent images under different camera views, then uses them as labeled data to directly reconstruct 4D content through a 4D Gaussian splatting model. It achieves a 10-fold speed increase over prior art while preserving quality.
137. CMOOD: Concept-based Multi-label OOD Detection
Core Problem: Existing Out-Of-Distribution (OOD) detection methods struggle with complex, multi-label settings, failing to capture intricate semantic relationships and generalize to unseen label combinations, especially in zero-shot scenarios.
Key Innovation: Introduces CMOOD, a novel zero-shot multi-label OOD detection framework. It leverages pre-trained vision-language models, enhancing them with a concept-based label expansion strategy and a new scoring function. This models complex label dependencies and precisely differentiates OOD samples without additional training.
138. CleanSurvival: Automated data preprocessing for time-to-event models using reinforcement learning
Core Problem: Data preprocessing is a critical yet frequently neglected aspect of machine learning, especially for specialized tasks like survival (time-to-event) analysis, which lacks tailored, automated solutions.
Key Innovation: Presents 'CleanSurvival', a reinforcement-learning-based solution for optimizing preprocessing pipelines specifically for survival analysis. It uses Q-learning to select optimal combinations of data imputation, outlier detection, and feature extraction techniques for various time-to-event models, achieving superior predictive performance and faster model finding.
139. Rethinking Multimodal Learning from the Perspective of Mitigating Classification Ability Disproportion
Core Problem: Multimodal learning (MML) is significantly constrained by modality imbalance, leading to suboptimal performance because existing approaches overlook the inherent disproportion in model classification ability between weak and strong modalities.
Key Innovation: Proposes a novel multimodal learning approach that dynamically balances the classification ability of weak and strong modalities using a sustained boosting algorithm and an adaptive classifier assignment strategy. This mitigates the imbalance issue and achieves superior performance on various datasets.
140. Progressively Deformable 2D Gaussian Splatting for Video Representation at Arbitrary Resolutions
Core Problem: Implicit neural representations (INRs) for video struggle with scalable decoding across rates and resolutions, often requiring retraining or multi-branch designs, and structured pruning fails to provide permutation-invariant progressive transmission.
Key Innovation: Proposes D2GV-AR, a deformable 2D Gaussian video representation that enables arbitrary-scale rendering and any-ratio progressive coding within a single model. It uses a canonical set of 2D Gaussian primitives with temporal evolution modeled by neural ODEs, scale-aware grouping, and D-optimal subset pruning.
141. Plain Transformers Can be Powerful Graph Learners
Core Problem: Most advanced Graph Transformers (GTs) have strayed far from plain Transformers, integrating message-passing or sophisticated attention mechanisms, which hinders the easy adoption of training advances developed for Transformers in other domains.
Key Innovation: Demonstrates that the plain Transformer architecture can be a powerful graph learner with three simple modifications: simplified L2 attention, adaptive root-mean-square normalization, and an MLP-based stem for graph positional encoding. This 'Powerful Plain Graph Transformer' (PPGT) shows noteworthy expressivity and strong empirical performance.
142. An Evidence-Based Post-Hoc Adjustment Framework for Anomaly Detection Under Data Contamination
Core Problem: Unsupervised anomaly detection (AD) methods typically assume clean training data, yet real-world datasets often contain undetected or mislabeled anomalies, leading to significant performance degradation, and existing solutions require access to training pipelines, data, or prior knowledge.
Key Innovation: Proposes EPHAD, a simple yet effective test-time adaptation framework that updates the outputs of AD models trained on contaminated datasets using evidence gathered at test time, integrating prior knowledge captured by the AD model with evidence derived from multimodal foundation models (like CLIP), classical AD methods, or domain-specific knowledge.
143. A Comparative Study on How Data Normalization Affects Zero-Shot Generalization in Time Series Foundation Models
Core Problem: Input normalization methods for Time-Series Foundation Models (TSFMs) are overlooked, despite time-series data exhibiting significant scale variation across domains and channels, coupled with non-stationarity, which can undermine TSFM performance regardless of architectural complexity.
Key Innovation: Through systematic evaluation across four architecturally diverse TSFMs, empirically establishes REVIN as the most efficient approach, reducing zero-shot MASE significantly and matching the best in-domain accuracy without any dataset-level preprocessing, while highlighting its effect utilization depends on architectural design choices and optimization objective.
144. Physics-Aware Heterogeneous GNN Architecture for Real-Time BESS Optimization in Unbalanced Distribution Systems
Core Problem: Existing deep learning approaches for BESS optimization in three-phase unbalanced distribution grids often lack explicit three-phase representation, making it difficult to accurately model phase-specific dynamics and enforce operational constraints, leading to infeasible dispatch solutions.
Key Innovation: Demonstrates that by embedding detailed three-phase grid information into heterogeneous graph nodes, diverse GNN architectures can jointly predict network state variables with high accuracy, and incorporates a physics-informed loss function with soft penalties for critical battery constraints (SoC and C-rate limits) during training, ensuring reliable and constraint-compliant dispatch.
145. Efficient Spike-driven Transformer for High-performance Drone-View Geo-Localization
Core Problem: Traditional ANN-based drone-view geo-localization (DVGL) has high power consumption, while spiking neural networks (SNNs) for DVGL suffer from critical information loss and difficulty learning long-range dependencies due to spike-driven sparsity.
Key Innovation: SpikeViMFormer, the first SNN framework for DVGL, featuring a lightweight spike-driven transformer backbone, a spike-driven selective attention (SSA) block for feature enhancement, a spike-driven hybrid state space (SHS) block for long-range dependencies, and a hierarchical re-ranking alignment learning (HRAL) strategy for optimization.
146. VAE with Hyperspherical Coordinates: Improving Anomaly Detection from Hypervolume-Compressed Latent Space
Core Problem: In high-dimensional latent spaces, standard Variational Autoencoders (VAEs) distribute latent vectors on hypersphere "equators," challenging the detection of out-of-distribution (abnormal) latent vectors and limiting their generative capacity.
Key Innovation: Formulating the latent variables of a VAE using hyperspherical coordinates, which allows compressing latent vectors towards a given direction on the hypersphere, thereby creating a more expressive approximate posterior and significantly improving both fully unsupervised and OOD anomaly detection abilities.
147. Estimating Spatiotemporally Continuous GEDI Aboveground Biomass Density During 2015–2020 From Multisource Data Using Machine Learning and Deep Learning
Core Problem: GEDI Aboveground Biomass Density (AGBD) products are spatially discontinuous, limiting their practical application for quantifying terrestrial carbon exchange.
Key Innovation: A machine learning and deep learning method to estimate spatiotemporally continuous GEDI AGBD and its uncertainty from multisource data (optical, SAR, LiDAR, topographic, climatic), demonstrating that CNN_spatial outperforms other methods and produces more accurate AGBD maps over time.
148. A Decade of GNSS Signal Disruptions in SMAP-R Full-Polarimetric Observations Worldwide
Core Problem: Global Navigation Satellite System–Reflectometry (GNSS-R) signals, crucial for Earth surface observations, are vulnerable to radio-frequency interference (RFI) and signal disruptions, particularly in conflict zones, which degrades data quality and hinders geophysical parameter retrieval.
Key Innovation: An examination of worldwide GNSS signal disruptions (RFI) using SMAP-R full-polarimetric observations over a decade, demonstrating extensive RFI detection in conflict zones, quantifying its impact on signal-to-noise ratio and polarimetric responses, and defining a flagging methodology for SMAP-R data.
149. PKBFM: A Prior Knowledge-Based Deep Learning Framework for Bamboo Forest Mapping
Core Problem: Precise bamboo forest mapping (BFP) is challenging due to difficulty in obtaining field samples, reflectance similarity with other forests, and challenges in synthesizing spatial and temporal information.
Key Innovation: PKBFM, a novel prior knowledge-based deep learning framework for high-precision bamboo forest mapping, which integrates preliminary label acquisition, noise label filtering, and refined recognition using an attention-based temporal-spatial network (TSNet), significantly reducing reliance on field samples.
150. Graph Feature Refinement and Adaptive Matching Network for Cross-Scene Hyperspectral Image Classification
Core Problem: Cross-scene hyperspectral image (HSI) classification faces significant challenges due to domain shifts from varying imaging conditions and environmental factors, and existing methods often overlook structural dependencies and have limited adaptability.
Key Innovation: A novel approach for cross-scene HSI classification that uses graph feature refinement and an adaptive matching network, incorporating a spectral refinement augmentation module, a structure-guided matching mechanism, and a multigraph contrastive learning strategy to achieve strong robustness and enhanced generalization.
151. Evaluation of TBM tunneling efficiency based on a new rock brittleness index considering energy conversion
Core Problem: Existing rock brittleness indices may not fully capture the energy conversion mechanisms relevant to Tunnel Boring Machine (TBM) tunneling efficiency, making it difficult to accurately evaluate rock excavatability and the effects of moisture.
Key Innovation: A new rock brittleness index (BI_N) that considers energy conversion mechanisms, effectively characterizing rock brittleness, its decrease with moisture content, and its correlation with TBM specific energy (SE) and Cerchar Abrasivity Index (CAI), providing a better evaluation of TBM tunneling efficiency.
152. Construction of artificial water conduit by pulse hydraulic fracturing for water pre-drainage in roof fissured sandstone aquifer
Core Problem: Need for effective water pre-drainage in roof fissured sandstone aquifers, likely in mining or underground construction, to manage water inflow.
Key Innovation: Proposing and developing a method for constructing artificial water conduits using pulse hydraulic fracturing for water pre-drainage.
153. Characteristics and Size Effects of 3-D Micro-fracture Structures of Tectonically Deformed Shale Based on X-ray Tomography and Digital Image Processing Technology: Implications for Estimating Permeability REV Size of Intensely Fractured Shale
Core Problem: Shale reservoirs with intense tectonic transformation exhibit significant heterogeneity and anisotropy in fracture structures and seepage properties, hindering shale gas development, and the permeability Representative Elementary Volume (REV) size of intensely fractured shale is difficult to estimate.
Key Innovation: A qualitative and quantitative investigation of 3-D micro-fracture structures of tectonically deformed shale (TDS) using X-ray tomography and digital image processing, analyzing size effects of fracture void percentage, spatial attitude angles, and permeability coefficients to estimate permeability REV sizes for various TDSs, providing insights for shale gas development.
154. Hybrid Strengthening Method and Mechanism of Sand Powder 3D Printed Rock Analog
Core Problem: Conventional fabrication techniques for rock analogs struggle to replicate complex defect structures and achieve sufficient compressive strength, limiting their applicability for physical modeling in geomechanical applications related to rock mass stability.
Key Innovation: A hybrid strengthening method for sand powder 3D printed rock analogs, combining particle gradation optimization and glass fiber reinforcement, which significantly enhances uniaxial compressive strength (up to 174.69% increase) and accurately replicates brittle-ductile transition and strength-stiffness correlation of natural rocks, providing a scalable pathway for fabricating high-fidelity rock analogs for geotechnical applications.
155. A Novel Coal Testing System for Low-Temperature Fluid Cyclic Fracturing and Wetting Enhancement and Its Application
Core Problem: Liquid CO₂ fracturing (LCF) for coal seam permeability enhancement faces challenges in high-pressure phase behavior control and lacks a systematic testing system, while the mechanisms of liquid CO₂ and water cyclic fracturing (LCWCF) are unclear.
Key Innovation: Development of a novel coal testing system for low-temperature fluid cyclic fracturing and wetting enhancement, capable of reproducing in situ coal seam conditions and enabling synchronous acquisition of pressure and acoustic emission data. This system demonstrates that LCWCF achieves higher fracturing efficiency and faster fracture initiation compared to hydraulic fracturing and LCF, providing an experimental basis for advanced permeability enhancement technologies.
156. Experimental Investigation of Macro- and Microscopic Degradation and Damage Mechanisms in Granite Subjected to Microwave Radiation under True Triaxial Compression
Core Problem: Efficient fragmentation of hard rock in "three-deep" engineering is challenging, and the degradation and damage mechanisms of granite subjected to microwave radiation under true triaxial compression are not fully understood across multiple scales.
Key Innovation: A multi-scale experimental investigation combining true triaxial compression tests with nanoindentation experiments to elucidate the macro- and microscopic degradation and damage mechanisms in granite subjected to microwave radiation, establishing a cross-scale degradation and fracture mechanism by analyzing heterogeneous mineral responses and energy dissipation, and revealing the dominant role of microwave power in mechanical degradation.
157. Fracture Network Evolution Mechanisms in Coal Through Cyclic Progressive CO<sub>2</sub> Shear Fracturing Under True Triaxial Stress
Core Problem: Enhancing coal seam permeability and promoting complex, shear-dominated fracture networks for coalbed methane (CBM) recovery is challenging, and the fracture network evolution mechanisms under different CO2 fracturing schemes are not fully understood.
Key Innovation: An investigation into fracture network evolution mechanisms in coal under three CO2 fracturing schemes (CCI, SCI, CPI) under true triaxial stress, demonstrating that the cyclic progressive injection (CPI) scheme generates numerous shear microcracks and a highly complex fracture network, leading to a proposed three-tier fracture network technology for CBM recovery.
158. Numerical Modeling of 3D Hydraulic Fractures Propagation in Bedding-Riched Shale Oil Reservoir
Core Problem: Natural bedding fractures in shale oil reservoirs restrict fracture height during hydraulic fracturing, and understanding the characteristics and mechanics of 3D hydraulic fracture propagation in laminated shale is crucial for optimizing shale oil recovery.
Key Innovation: A 3D hydraulic fracture propagation model using a hybrid finite-discrete element method (FDEM) that considers shale matrix mechanics, locally high-density natural bedding fractures, and in-situ stress, to quantitatively study fracture propagation mechanics and optimize injection parameters (pumping rate, fluid viscosity) for enhancing fracture complexity and connectivity in laminated shale reservoirs for improved oil recovery.
159. Mechanical properties of masonry structures in Portugal: proposal of new correlations with the Masonry Quality Index (MQI)
Core Problem: Lack of comprehensive understanding and standardized correlations for the mechanical properties of masonry structures, particularly in Portugal, hindering accurate assessment of structural quality.
Key Innovation: Proposal of new correlations between the mechanical properties of masonry structures and the Masonry Quality Index (MQI) in Portugal.
160. Macro-modelling approach for nonlinear analysis of coupled concrete shear walls with symmetrical and asymmetrical openings
Core Problem: Challenges in accurately performing nonlinear analysis of coupled concrete shear walls, especially those with symmetrical and asymmetrical openings, using existing modeling approaches.
Key Innovation: Development of a macro-modelling approach for nonlinear analysis of coupled concrete shear walls, including those with symmetrical and asymmetrical openings.
161. Development of predictive expressions for drift-based damage states for precast columns in socket foundations
Core Problem: Need for reliable predictive expressions to assess drift-based damage states for precast columns in socket foundations, crucial for structural performance evaluation.
Key Innovation: Development of predictive expressions for drift-based damage states specifically for precast columns in socket foundations.
162. Glacial valley erosion controls on subsurface hydrology: Speleogenesis of Stortuvhola cave, sub-Arctic Norway
Core Problem: Understanding the controls of glacial valley erosion on subsurface hydrology and its role in speleogenesis, particularly in sub-Arctic environments.
Key Innovation: Investigation of glacial valley erosion controls on subsurface hydrology and speleogenesis, using Stortuvhola cave in sub-Arctic Norway as a case study.
163. Railway track performance prediction considering track-drainage interdependencies
Core Problem: Predicting railway track performance by explicitly considering the interdependencies between track conditions and drainage systems.
Key Innovation: A model for railway track performance prediction that integrates track-drainage interdependencies.
164. Knowledge-data fusion for water supply pipe failure prediction: A hybrid physics-informed and data-driven method
Core Problem: Accurately predicting failures in water supply pipes by combining domain knowledge with data-driven approaches.
Key Innovation: A hybrid physics-informed and data-driven method for water supply pipe failure prediction, using knowledge-data fusion.
165. Advancing vegetation segmentation from ALS point clouds: From benchmarking to GreenSegNet-A
Core Problem: Accurately segmenting vegetation from airborne laser scanning (ALS) point clouds.
Key Innovation: GreenSegNet-A, an advanced method for vegetation segmentation from ALS point clouds, developed through benchmarking.
166. Estimation of grassland canopy cover at quadrat and plot scales using multi-scale UAV imagery
Core Problem: Accurately estimating grassland canopy cover at different spatial scales using UAV imagery.
Key Innovation: Estimation of grassland canopy cover at quadrat and plot scales using multi-scale UAV imagery.
167. Estimating tree diameter at breast height (DBH) from UAV data: A comparison of oblique–Vertical imagery fusion and allometric modeling
Core Problem: Accurately estimating tree diameter at breast height (DBH) from UAV data.
Key Innovation: A comparative analysis of oblique-vertical imagery fusion and allometric modeling for estimating tree DBH from UAV data.
168. Individual tree detection from aerial RGB images using transfer learning semantic segmentation and simulated illumination template matching in the Yellow River Delta
Core Problem: Detecting individual trees from aerial RGB images accurately.
Key Innovation: Individual tree detection from aerial RGB images using transfer learning semantic segmentation and simulated illumination template matching.
169. <em>rTwig</em>: An R package to correct overestimated small branches and twigs in quantitative structure models of trees
Core Problem: Correcting the overestimation of small branches and twigs in quantitative structure models of trees.
Key Innovation: The `rTwig` R package designed to correct overestimated small branches and twigs in tree quantitative structure models.
170. Principle analysis and device development of a novel technology for preventing lockset failure and cable bolt ejection
Core Problem: Developing a novel technology to prevent lockset failure and cable bolt ejection in underground engineering applications.
Key Innovation: Developed a new technology and device to enhance the stability and safety of cable bolts in underground structures.
171. Multiscale simulation and machine learning sensitivity analysis of pore-structure controls on solute transport in heterogeneous porous media
Core Problem: Understanding the influence of pore-structure on solute transport in heterogeneous porous media.
Key Innovation: Utilizing multiscale simulation and machine learning sensitivity analysis to investigate pore-structure controls on solute transport.
172. Impact responses of layered rocks in tunnel excavations: Hydro-mechanical analysis
Core Problem: Analyzing the hydro-mechanical impact responses of layered rocks during tunnel excavations.
Key Innovation: Hydro-mechanical analysis of impact responses in layered rocks during tunnel excavations.
173. Scaling of capillary pressure-saturation curve in porous media under various wetting conditions
Core Problem: Understanding the scaling of capillary pressure-saturation curves in porous media under various wetting conditions.
Key Innovation: Investigating the scaling of capillary pressure-saturation curves in porous media under various wetting conditions.
174. Hydrogravimetry Enables Quantification of Alpine Groundwater Dynamics
Core Problem: Quantification of groundwater processes in alpine hydrological systems remains highly uncertain, despite their critical role and increasing importance under climate change.
Key Innovation: Using terrestrial time-lapse gravimetry (TLG) to measure groundwater storage changes (GWSC) in an alpine headwater catchment, revealing a spatially variable, dynamic system with greater GWSC magnitudes in higher-elevation regions, demonstrating TLG's utility for alpine groundwater investigations.
175. Processes Governing the Ablation of Intercepted Snow
Core Problem: Current modeling approaches for interception and ablation of snow in forest canopies have uncertain transferability and omit key processes, limiting accuracy.
Key Innovation: In situ observations from a needleleaf forest revealing strong associations between canopy snow load, wind shear stress, and snowmelt with unloading, leading to a new canopy snow ablation model that significantly improves performance in simulating canopy snow load across various meteorological conditions.
176. Automated stratigraphic interpretation from drillhole lithological descriptions with uncertainty quantification: litho2strat 1.0
Core Problem: Historical drillhole data often lacks stratigraphic interpretations, limiting its use for subsurface geological modeling.
Key Innovation: Developed "litho2strat 1.0", an automated method that converts drillhole lithological descriptions into stratigraphic interpretations, quantifies uncertainty, and correlates multiple drillholes, successfully tested on South Australian data.
177. High resolution monthly precipitation isotope estimates across Australia from machine learning
Core Problem: High-resolution, long-term estimates of monthly precipitation stable isotope variability are often unavailable, limiting research in various fields.
Key Innovation: Produced high-resolution monthly precipitation stable isotope estimates across Australia from 1962–2023 using a random forest machine learning approach, demonstrating high predictive skill and making the outputs freely available.
178. Order-Aware Test-Time Adaptation: Leveraging Temporal Dynamics for Robust Streaming Inference
Core Problem: Existing Test-Time Adaptation (TTA) methods typically treat test-time data streams as independent samples, overlooking the valuable supervisory signal inherent in temporal dynamics, leading to suboptimal adaptation to distribution shifts.
Key Innovation: Introduces Order-Aware Test-Time Adaptation (OATTA), which formulates TTA as a gradient-free recursive Bayesian estimation task using a learned dynamic transition matrix as a temporal prior, consistently boosting established baselines by leveraging temporal dynamics for robust streaming inference.
179. The Powers of Precision: Structure-Informed Detection in Complex Systems -- From Customer Churn to Seizure Onset
Core Problem: Early detection of emergent phenomena in complex systems by identifying and utilizing hidden causal interactions, especially when the data-generating process is unknown and partially observed.
Key Innovation: A machine learning method that learns an optimal feature representation from a one-parameter family of estimators (powers of empirical covariance/precision matrix) to tune into underlying structure, combined with a supervised learning module. It achieves structural consistency and competitive results in seizure detection and churn prediction.
180. AC2L-GAD: Active Counterfactual Contrastive Learning for Graph Anomaly Detection
Core Problem: Graph anomaly detection faces challenges from label scarcity and extreme class imbalance, and existing graph contrastive learning methods suffer from random augmentations breaking semantic consistency and uninformative negative sampling.
Key Innovation: AC2L-GAD, an Active Counterfactual Contrastive Learning framework that uses information-theoretic active selection and counterfactual generation to identify complex nodes, generate anomaly-preserving positive augmentations, and normal negative counterparts for hard contrasts, while reducing computational overhead.
181. Graph-Free Root Cause Analysis
Core Problem: Existing graph-free Root Cause Analysis (RCA) methods incorrectly assume the root cause has the highest anomaly score, failing when faults propagate and downstream anomalies become larger.
Key Innovation: PRISM, a simple and efficient framework for graph-free RCA that formulates a class of component-based systems with theoretical guarantees, achieving high accuracy and speed by not relying on the highest anomaly score assumption.
182. Synthetic Pattern Generation and Detection of Financial Activities using Graph Autoencoders
Core Problem: Detecting illicit financial activities (e.g., money laundering) in transaction networks is challenging due to the scarcity of labeled real-world data and privacy constraints.
Key Innovation: Investigating Graph Autoencoders (GAEs) trained on synthetically generated data for illicit financial activity patterns, demonstrating that graph-based representation learning on synthetic data can effectively learn and distinguish these topological patterns, overcoming data limitations.
183. Evaluating Prediction Uncertainty Estimates from BatchEnsemble
Core Problem: Deep learning models often struggle with accurate and computationally feasible uncertainty estimation, frequently underestimating it.
Key Innovation: Investigating BatchEnsemble as a scalable method for uncertainty estimation, introducing GRUBE (a BatchEnsemble GRU cell) for sequential modeling, demonstrating that it matches deep ensembles' performance with fewer parameters and reduced computational cost.
184. CORDS: Continuous Representations of Discrete Structures
Core Problem: Learning problems requiring prediction of variable-sized sets of objects face challenges with existing methods that rely on padded representations or explicit set size inference.
Key Innovation: CORDS (Continuous Representations of Discrete Structures) is a novel strategy that casts variable-sized set prediction as a continuous inference problem, using an invertible mapping to transform spatial objects into continuous density and feature fields, enabling robust handling of unknown set sizes.
185. Unifying Heterogeneous Degradations: Uncertainty-Aware Diffusion Bridge Model for All-in-One Image Restoration
Core Problem: All-in-One Image Restoration (AiOIR) struggles to reconcile conflicting optimization objectives across heterogeneous degradations, leading to suboptimal adaptation due to coarse control mechanisms.
Key Innovation: Uncertainty-Aware Diffusion Bridge Model (UDBM) reformulates AiOIR as a stochastic transport problem steered by pixel-wise uncertainty, using a relaxed diffusion bridge and a dual modulation strategy (noise and path schedules) to achieve state-of-the-art performance across diverse restoration tasks.
186. TabClustPFN: A Prior-Fitted Network for Tabular Data Clustering
Core Problem: Clustering tabular data is challenging due to heterogeneous feature types, diverse data-generating mechanisms, and the lack of transferable inductive biases, making it difficult to extend prior-fitted networks (PFNs) to this unsupervised task.
Key Innovation: TabClustPFN, a prior-fitted network for tabular data clustering, performs amortized Bayesian inference over cluster assignments and cardinality, pretrained on synthetic data, enabling it to cluster unseen datasets in a single forward pass without retraining, outperforming baselines.
187. When Gradient Optimization Is Not Enough: $\dagger$ Dispersive and Anchoring Geometric Regularizer for Multimodal Learning
Core Problem: Multimodal learning models often suffer from geometric pathologies like intra-modal representation collapse and cross-modal inconsistency, degrading robustness and fusion despite strong optimization.
Key Innovation: A lightweight geometry-aware regularization framework (regName) that enforces intra-modal dispersive regularization (promoting diversity) and inter-modal anchoring regularization (bounding cross-modal drift), consistently improving multimodal and unimodal performance by explicitly regulating representation geometry.
188. Synthetic-to-Real Domain Bridging for Single-View 3D Reconstruction of Ships for Maritime Monitoring
Core Problem: Efficient 3D reconstruction of ships for maritime monitoring is challenging because most state-of-the-art methods require multi-view supervision, annotated 3D ground truth, or are computationally intensive, making them impractical for real-time deployment.
Key Innovation: Presents an efficient pipeline for single-view 3D reconstruction of real ships by training entirely on synthetic data and requiring only a single view at inference. Uses the Splatter Image network, fine-tuned on synthetic and custom datasets, and integrates segmentation, postprocessing, and georeferenced placement on an interactive web map.
189. Blind Ultrasound Image Enhancement via Self-Supervised Physics-Guided Degradation Modeling
Core Problem: Ultrasound (US) image interpretation is hampered by multiplicative speckle, acquisition blur, and scanner/operator-dependent artifacts, but supervised enhancement methods are impractical due to the lack of clean targets or known degradations in practice.
Key Innovation: Presents a blind, self-supervised enhancement framework that jointly deconvolves and denoises B-mode images using a Swin Convolutional U-Net trained with a physics-guided degradation model. Synthesizes inputs by convolving with a Gaussian PSF surrogate and injecting noise. Achieves highest PSNR/SSIM across noise levels and consistently boosts Dice for segmentation.
190. Zero-Shot Video Restoration and Enhancement with Assistance of Video Diffusion Models
Core Problem: Although diffusion-based zero-shot image restoration and enhancement methods have achieved great success, applying them to video restoration or enhancement leads to severe temporal flickering.
Key Innovation: Proposes the first framework that utilizes rapidly-developed video diffusion models to assist image-based methods in maintaining more temporal consistency for zero-shot video restoration and enhancement. Introduces homologous latents fusion, heterogenous latents fusion, and a COT-based fusion ratio strategy, along with temporal-strengthening post-processing. The method is training-free and applicable to any diffusion-based image method.
191. Bridging Graph Structure and Knowledge-Guided Editing for Interpretable Temporal Knowledge Graph Reasoning
Core Problem: Existing LLM-based Temporal Knowledge Graph Reasoning (TKGR) methods prioritize contextual over structural relations, leading to difficulties in extracting relevant subgraphs from dynamic graphs and producing hallucination-prone inferences with temporal inconsistencies.
Key Innovation: Introduces IGETR, a hybrid framework combining GNNs for structured temporal modeling and LLMs for contextual understanding, using a three-stage pipeline (temporal GNN for candidate paths, LLM-guided editing, path integration) to achieve SOTA performance and interpretability.
192. Investigating Batch Inference in a Sequential Monte Carlo Framework for Neural Networks
Core Problem: Particle-based Bayesian inference methods like Sequential Monte Carlo (SMC) for neural networks are computationally expensive due to their typical reliance on full-batch data for likelihood and gradient evaluations.
Key Innovation: Explores data annealing methods to gradually introduce mini-batches into SMC samplers, achieving up to 6x faster training with minimal accuracy loss on benchmark image classification problems.
193. Cross-Fusion Distance: A Novel Metric for Measuring Fusion and Separability Between Data Groups in Representation Space
Core Problem: Quantifying degrees of fusion and separability between data groups in representation space, especially under domain shift, is challenging because existing distributional distance metrics conflate fusion-altering and fusion-preserving factors, leading to uninformative measures.
Key Innovation: Introduces Cross-Fusion Distance (CFD), a principled measure that isolates fusion-altering geometry while remaining robust to fusion-preserving variations, with linear computational complexity, and aligns more closely with downstream generalization degradation than alternatives.
194. Making Foundation Models Probabilistic via Singular Value Ensembles
Core Problem: Foundation models often yield overconfident, uncalibrated predictions, and the standard approach of training explicit ensembles for epistemic uncertainty quantification is computationally prohibitive for large models.
Key Innovation: Proposes Singular Value Ensemble (SVE), a parameter-efficient implicit ensemble method that freezes singular vectors of weight matrices and trains only per-member singular values, achieving uncertainty quantification comparable to explicit deep ensembles with less than 1% parameter increase.
195. Prior-Informed Flow Matching for Graph Reconstruction
Core Problem: Reconstructing graphs from partial observations remains a key challenge, as classical embedding methods often lack global consistency and modern generative models struggle to incorporate structural priors.
Key Innovation: Introduces Prior-Informed Flow Matching (PIFM), a conditional flow model that integrates embedding-based priors with continuous-time flow matching, using a permutation equivariant distortion-perception theory to refine initial estimates and learn a global coupling, consistently enhancing classical embeddings.
196. Late Breaking Results: Conversion of Neural Networks into Logic Flows for Edge Computing
Core Problem: Inefficient execution of neural networks on resource-constrained edge CPUs, which are better suited for control flow logic than large-scale multiply-accumulate (MAC) operations.
Key Innovation: A novel method to convert neural networks into equivalent decision trees, then compress decision paths into logic flows (if/else structures) with reduced MAC operations, achieving up to 14.9% latency reduction on a simulated RISC-V CPU without accuracy degradation.
197. Latent-IMH: Efficient Bayesian Inference for Inverse Problems with Approximate Operators
Core Problem: Sampling from posterior distributions in Bayesian linear inverse problems is computationally expensive when the parameters-to-observables operator (A) is costly to evaluate.
Key Innovation: Latent-IMH, a sampling method based on Metropolis-Hastings independence (IMH) that generates intermediate latent variables using a cost-effective approximation (Ã) and then refines them with the exact A, shifting computational cost to an offline phase and achieving orders of magnitude faster performance than state-of-the-art methods.
198. Towards regularized learning from functional data with covariate shift
Core Problem: Reliable learning from functional data in vector-valued regression is significantly challenged by covariate shift, where the input distributions of training and test data differ.
Key Innovation: Developing a general regularization framework for unsupervised domain adaptation in vector-valued regression under covariate shift, utilizing vector-valued reproducing kernel Hilbert spaces (vRKHS), establishing optimal convergence rates, and proposing an aggregation-based approach for selecting tuning parameters.
199. InspecSafe-V1: A Multimodal Benchmark for Safety Assessment in Industrial Inspection Scenarios
Core Problem: Reliable perception and safety assessment for AI systems in complex and dynamic industrial sites is a key bottleneck due to limitations in public datasets (simulated, single-modality, lack of fine-grained object-level annotations).
Key Innovation: Release of InspecSafe-V1, the first multimodal benchmark dataset for industrial inspection safety assessment, collected from routine operations of real inspection robots in real-world environments, covering five scenarios with pixel-level segmentation, semantic scene descriptions, safety labels, and seven synchronized sensing modalities.
200. ZipMoE: Efficient On-Device MoE Serving via Lossless Compression and Cache-Affinity Scheduling
Core Problem: The prohibitive memory footprint of Mixture-of-Experts (MoE) architectures severely impedes their practical deployment on resource-constrained edge devices, especially when model behavior must be preserved without lossy quantization.
Key Innovation: ZipMoE, an efficient and semantically lossless on-device MoE serving system that exploits the synergy between hardware properties and statistical redundancy via a caching-scheduling co-design, achieving up to 72.77% inference latency reduction and up to 6.76x higher throughput.
201. ViTMAlis: Towards Latency-Critical Mobile Video Analytics with Vision Transformers
Core Problem: Deploying Vision Transformers (ViTs) in latency-critical mobile video analytics (MVA) is challenging due to substantial inference delays, particularly for dense prediction tasks requiring high-resolution inputs, which exacerbates ViTs' quadratic computational complexity.
Key Innovation: Proposes ViTMAlis, a ViT-native device-to-edge offloading framework that uses a dynamic mixed-resolution inference strategy to adapt to network conditions and video content, jointly reducing transmission and inference delays while improving user-perceived rendering accuracy.
202. A Decomposable Forward Process in Diffusion Models for Time-Series Forecasting
Core Problem: Standard diffusion models for time-series forecasting struggle to preserve structured temporal patterns like seasonality effectively, as their noise injection process can degrade long-term patterns.
Key Innovation: Introduces a model-agnostic forward diffusion process that decomposes signals into spectral components (e.g., Fourier or Wavelet transform), staging noise injection according to component energy to maintain high signal-to-noise ratios for dominant frequencies, thereby improving the recoverability of long-term patterns and forecast quality with negligible computational overhead.
203. Parametrized Power-Iteration Clustering for Directed Graphs
Core Problem: Vertex-level clustering for directed graphs (digraphs) is challenging because edge directionality breaks key assumptions of popular spectral methods, which also incur the overhead of eigen-decomposition.
Key Innovation: Proposes Parametrized Power-Iteration Clustering (ParPIC), a random-walk-based method for weakly connected digraphs. It uses parametrized reversible random walk operators, automatic diffusion time tuning, and efficient embedding truncation to achieve competitive accuracy and improved scalability.
204. Fair Graph Machine Learning under Adversarial Missingness Processes
Core Problem: Existing work on fair Graph Neural Networks (GNNs) often assumes sensitive attributes are fully observed or missing completely at random, failing to account for adversarial missingness processes that can inadvertently disguise a fair model through imputation.
Key Innovation: Proposes Better Fair than Sorry (BFtS), a fair missing data imputation model for sensitive attributes. It uses a 3-player adversarial scheme where imputations approximate the worst-case scenario for fairness, achieving a better fairness x accuracy trade-off.
205. One Model, Any Conjunctive Query: Graph Neural Networks for Answering Queries over Incomplete Knowledge Graphs
Core Problem: Answering conjunctive queries over incomplete knowledge graphs is challenging, particularly predicting answers not explicitly present and generalizing to large queries of arbitrary structure.
Key Innovation: Proposes AnyCQ, a Graph Neural Network (GNN) model trained using a reinforcement learning objective to classify and retrieve answers to any conjunctive query on any knowledge graph. It generalizes to large queries and effectively transfers to novel knowledge graphs.
206. WL Tests Are Far from All We Need: Revisiting WL-Test Hardness and GNN Expressive Power from a Distributed Computation Perspective
Core Problem: The Weisfeiler-Lehman (WL) test perspective on GNN expressive power has gaps: it's unclear if WL tests are sufficiently primitive for understanding GNN expressivity, and WL-induced equivalence does not align well with characterizing GNN-computable function classes.
Key Innovation: Strengthens hardness results for the vanilla WL test, showing it's not primitive enough for constant-depth GNNs. Proposes an alternative framework based on an extended CONGEST model to study GNN expressivity, identifying implicit shortcuts and establishing further results for WL tests with augmented graphs.
207. Do graph neural network states contain graph properties?
Core Problem: While Graph Neural Networks (GNNs) achieve state-of-the-art performance, their non-Euclidean nature makes it difficult to interpret their internal representations, and existing explainability methods rarely investigate model-based approaches or probe embeddings for structural graph properties.
Key Innovation: Presents a model-agnostic explainability pipeline for GNNs employing diagnostic classifiers. It proposes considering graph-theoretic properties as features to study the emergence of representations in GNNs, aiming to probe and interpret learned representations across various architectures and datasets.
208. ASAP: Exploiting the Satisficing Generalization Edge in Neural Combinatorial Optimization
Core Problem: Deep Reinforcement Learning (DRL) solvers for Combinatorial Optimization (CO) problems often exhibit brittleness and poor generalization when facing distribution shifts.
Key Innovation: Proposes ASAP (Adaptive Selection After Proposal), a generic framework that decomposes decision-making into a robust proposal policy and an adaptable selection policy. It exploits the 'Satisficing Generalization Edge' and uses a two-phase training framework with MAML for fast online adaptation, improving generalization and adaptation on out-of-distribution instances.
209. One-Shot Federated Learning with Classifier-Free Diffusion Models
Core Problem: One-shot federated learning (OSFL) often relies on classifier-guided diffusion models, which require training auxiliary classifier models at each client, introducing additional computation overhead and communication costs.
Key Innovation: Introduces OSCAR (One-Shot Federated Learning with Classifier-Free Diffusion Models), a novel OSFL approach that eliminates auxiliary models. It uses foundation models to devise category-specific data representations at clients, integrated into a classifier-free diffusion model pipeline for server-side data generation, outperforming state-of-the-art while reducing communication load.
210. Redefining Neural Operators in $d+1$ Dimensions for Embedding Evolution
Core Problem: Existing Neural Operators (NOs) struggle with efficient embedding evolution, often relying on brute-force lengthening, leading to increased computation.
Key Innovation: Introduces an auxiliary dimension to explicitly model embedding evolution in operator form, redefining NOs in d+1 dimensions. Develops Schr"odingerised Kernel Neural Operator (SKNO) using Fourier-based operators for d+1 dimensional evolution, outperforming baselines and showing resolution invariance.
211. FastDINOv2: Frequency Based Curriculum Learning Improves Robustness and Training Speed
Core Problem: Reproducing large-scale vision foundation models like DINOv2 is computationally demanding, and there's a need to accelerate convergence and improve robustness to common corruptions during pre-training.
Key Innovation: Proposes a novel pre-training strategy for DINOv2 involving a frequency filtering curriculum (low-frequency first) and Gaussian noise patching augmentation, which reduces pre-training time and FLOPs by 1.6x and 2.25x respectively, while matching robustness and maintaining competitive linear probing performance.
212. CycleDiff: Cycle Diffusion Models for Unpaired Image-to-image Translation
Core Problem: Existing diffusion-based image-to-image translation methods struggle with aligning the diffusion and translation processes, leading to separate training or shallow integration, which can cause local minima and limit effectiveness.
Key Innovation: Proposes CycleDiff, a novel joint learning framework that aligns diffusion and translation processes by extracting image components with diffusion models and employing a time-dependent translation network. This enables global optimization, achieving improved fidelity and structural consistency across diverse cross-modality translation tasks.
213. Residual Reservoir Memory Networks
Core Problem: Conventional Reservoir Computing (RC) models often struggle with enhanced long-term propagation of input and memory capacity in time-series tasks.
Key Innovation: Introduces Residual Reservoir Memory Networks (ResRMNs), combining a linear memory reservoir with a non-linear reservoir based on residual orthogonal connections along the temporal dimension. This enhances long-term propagation and memory capacity, outperforming conventional RC models on time-series and pixel-level 1-D classification tasks.
214. Deep Residual Echo State Networks: exploring residual orthogonal connections in untrained Recurrent Neural Networks
Core Problem: Traditional Echo State Networks (ESNs) often struggle with long-term information processing and memory capacity, limiting their effectiveness for complex time series tasks.
Key Innovation: Introduces Deep Residual Echo State Networks (DeepResESNs), a novel class of deep untrained RNNs leveraging a hierarchy of untrained residual recurrent layers with different orthogonal configurations. This significantly boosts memory capacity and long-term temporal modeling, outperforming traditional shallow and deep RC models on various time series tasks.
215. Rotary Position Encodings for Graphs
Core Problem: Efficiently injecting structural information into attention mechanisms for graph-structured data, particularly for long-range dependencies, is a challenge for graph neural networks.
Key Innovation: Introduces Wave-Induced Rotary Encodings (WIRE), which apply rotary position encodings (RoPE) to graph-structured data by rotating tokens based on the graph Laplacian spectrum. WIRE efficiently injects structural information, boosting performance in graph learning tasks, recovering regular RoPE on grids, and depending asymptotically on graph effective resistance.
216. Entropy Guided Dynamic Patch Segmentation for Time Series Transformers
Core Problem: Existing patch-based transformers for time series modeling rely on temporally-agnostic patch construction, which fractures temporal coherence by splitting natural transitions across boundaries, disrupting short-term dependencies and weakening representation learning.
Key Innovation: Proposes EntroPE, a novel Entropy-Guided Dynamic Patch Encoder, as a temporally informed framework that dynamically detects transition points via conditional entropy and places patch boundaries, preserving temporal structure while retaining computational benefits, and uses an Adaptive Patch Encoder to capture intra-patch dependencies.
217. Efficient Test-Time Adaptation through Latent Subspace Coefficients Search
Core Problem: Most test-time adaptation (TTA) methods are unfriendly to edge deployment due to high latency and memory overhead, as they rely on backpropagation, activation buffering, or test-time mini-batches.
Key Innovation: Proposes ELaTTA, a gradient-free framework for single-instance TTA under strict on-device constraints, which freezes model weights and adapts each test sample by optimizing a low-dimensional coefficient vector in a source-induced principal latent subspace, pre-computed offline, and uses CMA-ES for optimization.
218. Align & Invert: Solving Inverse Problems with Diffusion and Flow-based Models via Representation Alignment
Core Problem: Enforcing alignment between internal representations of diffusion or flow-based generative models and pretrained self-supervised encoders has been shown to improve convergence and sample quality, but its application to inverse problems where ground-truth signals are unavailable needs further exploration.
Key Innovation: Proposes applying representation alignment (REPA) between diffusion or flow-based models and a DINOv2 visual encoder to guide the reconstruction process at inference time for inverse problems, empirically showing it substantially enhances reconstruction quality and perceptual realism, and providing theoretical results on REPA regularization.
219. City Navigation in the Wild: Exploring Emergent Navigation from Web-Scale Knowledge in MLLMs
Core Problem: Current MLLM evaluation benchmarks for embodied agents are predominantly language-centric or simulated, lacking the nuanced, knowledge-intensive reasoning required for practical, real-world navigation in complex environments.
Key Innovation: Introduction of Sparsely Grounded Visual Navigation and the CityNav benchmark for real-world MLLM evaluation, and Verbalization of Path (VoP) which grounds agent reasoning by probing city-scale cognitive maps from MLLMs, substantially enhancing navigation success.
220. Kernel Alignment-based Multi-view Unsupervised Feature Selection with Sample-level Adaptive Graph Learning
Core Problem: Existing multi-view unsupervised feature selection (MUFS) methods often overlook complex nonlinear dependencies among features and fail to account for differences in local neighborhood clarity among samples when fusing similarity graphs, limiting effectiveness.
Key Innovation: KAFUSE (Kernel Alignment-based multi-view unsupervised FeatUre selection with Sample-level adaptive graph lEarning), which employs kernel alignment with an orthogonal constraint to reduce both linear and nonlinear feature redundancy, and learns a cross-view consistent similarity graph by applying sample-level fusion to adapt view weights for each sample.
221. Practical Insights into Semi-Supervised Object Detection Approaches
Core Problem: A lack of comprehensive understanding and comparison of state-of-the-art semi-supervised object detection (SSOD) approaches, particularly regarding their performance variations with limited labeled data and their trade-offs in practical low-data regimes.
Key Innovation: A comprehensive comparison of three state-of-the-art SSOD approaches (MixPL, Semi-DETR, Consistent-Teacher) on multiple datasets, including a custom specialized dataset, providing practical insights into their performance, trade-offs, and suitability for low-data regimes.
222. DrivIng: A Large-Scale Multimodal Driving Dataset with Full Digital Twin Integration
Core Problem: Existing autonomous driving datasets often lack a high-fidelity digital twin, limiting systematic testing, edge-case simulation, sensor modification, and sim-to-real evaluations for robust perception algorithm development.
Key Innovation: DrivIng, a large-scale multimodal driving dataset with a complete geo-referenced digital twin of a ~18 km route, providing continuous multi-sensor recordings and 3D bounding box annotations, enabling 1-to-1 transfer of real traffic into simulation for realistic and flexible scenario testing.
223. A Multi-Object Tracking framework with feature backtracking under severe occlusion
Core Problem: Robustly tracking multiple objects even under severe occlusion.
Key Innovation: A multi-object tracking framework that uses feature backtracking to handle severe occlusion.
224. Comparative Study of Urban Thermal Landscape Identification Methods Based on Landsat Imagery
Core Problem: Inconsistent results and hindered cross-study comparability among different urban thermal landscape (UTL) identification methods based on land surface temperature (LST) images.
Key Innovation: A comparative study of four UTL identification methods using Landsat 9 LST imagery, evaluating their performance and the influence of spatial extent, recommending MSD and Jenks methods for their effectiveness in identifying UHI core areas and illustrating thermal gradients, respectively.
225. Physical Attributes Embedded Prototypical Network for Incremental SAR Automatic Target Recognition
Core Problem: Automatic target recognition (ATR) systems in SAR applications face catastrophic forgetting when new target classes emerge, as adapting to new data alters the feature space and degrades performance on old data.
Key Innovation: PAEPN (Physical Attributes Embedded Prototypical Network), a novel class-incremental SAR ATR method that embeds physical attributes (from electromagnetic scattering and geometric priors) into the deep model to achieve stable representations, mitigate catastrophic forgetting, and enhance stability and interpretability.
226. Refining False Detections: An Efficient Network Focused on Error-Prone Regions in Infrared Small Target Detection
Core Problem: Existing deep learning methods for infrared small target detection (IRSTD) struggle with accurately identifying boundary pixels of patch-like and point-like targets, leading to misclassifications and overlooked detections, despite good localization.
Key Innovation: RefineFD, an efficient two-stage IRSTD network that detects targets from coarse segmentation to fine refinement by directly focusing on uncertain, error-prone regions, achieving better accuracy and lower computational cost than current methods.
227. Physics-Guided Polarimetric Feature Selection Based on Dual-Branch Networks for PolSAR Agriculture Field Classification
Core Problem: Existing deep learning methods for PolSAR land cover classification suffer from information loss and redundancy in polarimetric features, making it difficult to select discriminative subsets.
Key Innovation: A novel Physics-guided dual-branch network with a local group-adaptive feature selection mechanism and a multifamily diversity constraint to select physically consistent and complementary polarimetric features for improved classification performance and interpretability.
228. Cross-View Visible-Thermal Object-Level Change Detection for Small Vehicles
Core Problem: Existing object-level change detection methods in remote sensing struggle with cross-view misalignment and multimodal feature discrepancy in real-world, unaligned datasets.
Key Innovation: A cross-view multimodal object-level change detection (CMOCD) framework with a coarse-to-fine feature alignment (CFFA) module and a visible-thermal feature fusion (VTFF) module, along with a new CVOCD dataset, to improve change detection performance for small vehicles.
229. Impacts of FES2022 and AOD1B RL07 Background Models on KBR- and LRI-Based GRACE-FO Monthly Gravity Field Estimations
Core Problem: Temporal gravity field solutions from GRACE-FO are limited by aliasing effects from imperfect background models, and the combined impact of updated models (FES2022, AOD1B RL07) on KBR- and LRI-based estimations is not well quantified.
Key Innovation: Quantitative assessment of the influence of FES2022 and AOD1B RL07 background models on GRACE-FO gravity field estimations, demonstrating that these models reduce noise and enhance temporal consistency, particularly for LRI-based solutions, thereby improving monthly gravity field solutions.
230. Details-Enhanced Multiscale Network for Building Change Detection in Remote Sensing Images
Core Problem: Existing multiscale change detection methods in remote sensing images often produce blurry change boundaries because they do not directly enhance or enrich high-frequency features.
Key Innovation: A multiscale enhanced change detection method based on detail supplementation (MEDS-CD) that uses a multiscale edge-guided adaptive filter to extract and compensate high-frequency details, combined with a multiscale convolution modulation module and fuzzy weighting strategy, to achieve better segmentation performance and sharper change boundaries.
231. Projection-Evidence Collaborative Optimization for Cross-Modal Few-Shot SAR Target Detection
Core Problem: Few-shot SAR target detection in cross-modal scenarios (optical to SAR) struggles with high-dimensional noisy features, nonlinear modality differences, and inadequate modeling of epistemic uncertainty due to sample scarcity, leading to overconfident detection errors.
Key Innovation: A projection-evidence collaborative optimization (PECO) method that includes a projection distribution alignment module to reduce cross-modal discrepancies and a dynamic uncertainty calibration module to model class probabilities with a Dirichlet evidence distribution, jointly optimizing uncertainties to mitigate overconfidence errors in scarce-sample settings.
232. BRSMamba: Boundary-Aware Mamba for Forest and Shrub Segmentation From Diverse Satellite Imagery
Core Problem: Vegetation segmentation in remote sensing imagery is challenging due to scale variance, spectral ambiguity, and complex boundaries, with existing CNNs limited in long-range dependency modeling and Transformers being computationally expensive. Vision state-space models (SSMs) lack sufficient boundary information modeling and efficient multidirectional scanning.
Key Innovation: BRSMamba, a boundary-aware network integrating two novel modules with noncausal state-space duality (NC-SSD) to enhance boundary preservation and global context modeling for vegetation segmentation. It uses a boundary-subject fusion perception module and a boundary-body resolution module to inject boundary awareness into the NC-SSD state-transition matrix.
233. A Novel Data-Driven Approach to Leaf Area Index Modeling Using High-Fidelity Simulation-Based Full-Waveform LiDAR Data
Core Problem: Conventional LAI estimation methods (field-based, passive optical remote sensing) suffer from scalability, labor intensity, and spectral saturation limitations, especially in dense canopies.
Key Innovation: A framework for LAI estimation using high-fidelity full-waveform LiDAR simulations, evaluating both an interpretable empirical model and a data-driven convolutional neural network (CNN), demonstrating the CNN's superior accuracy and robust performance across scales, and validating both approaches with real LVIS data.
234. Numerical Simulation of Wellbore Closure Due to Shale Creep-Part I: A Poro-visco-elastoplastic Approach
Core Problem: Plug and abandonment (P&A) of oil and gas wells is challenging due to high costs and environmental risks from underground fluid leakage, necessitating improved practices, including understanding wellbore closure due to shale creep.
Key Innovation: Development of two numerical models (plane strain and axisymmetric) using a poro-visco-elastoplastic approach in ABAQUS to simulate wellbore closure due to shale creep, demonstrating that shale formations can fully close annular space and form an impermeable barrier within two years under specific conditions, thereby improving P&A practices.
235. Uncertainty-aware sensorless anomaly detection using a reliable indicator from position-guided multi-step deep decomposition network
Core Problem: Detecting anomalies in systems without direct sensor measurements, while accounting for uncertainty.
Key Innovation: A sensorless anomaly detection method using a reliable indicator derived from a position-guided multi-step deep decomposition network, with uncertainty awareness.
236. Analyzing power network vulnerability considering spatial heterogeneous demand under extreme heat
Core Problem: Assessing the vulnerability of power networks to extreme heat conditions, considering variations in spatial demand.
Key Innovation: A method for analyzing power network vulnerability under extreme heat, incorporating spatially heterogeneous demand.
237. Source-free domain adaptation for cross-domain remaining useful life prediction: A distributed federated learning perspective
Core Problem: Predicting the remaining useful life of components across different operational domains without access to source domain data, while maintaining data privacy.
Key Innovation: A source-free domain adaptation approach for cross-domain remaining useful life prediction, utilizing a distributed federated learning framework.
238. Integrating real and virtual graphs: a dual joint network method for anomaly detection in discrete manufacturing systems
Core Problem: Improving anomaly detection in discrete manufacturing systems by leveraging both real-world data and virtual system representations.
Key Innovation: A dual joint network method that integrates real and virtual graphs for enhanced anomaly detection in discrete manufacturing systems.
239. A Bayesian hierarchical spatio-temporal generalized extreme value modeling for safety analysis from traffic conflicts
Core Problem: Analyzing safety from traffic conflicts using a model that accounts for spatio-temporal variations and extreme events.
Key Innovation: A Bayesian hierarchical spatio-temporal generalized extreme value modeling approach for safety analysis of traffic conflicts.
240. Multi-scale signal transformer with signal processing-based attention interpretation for fault diagnosis of rotating machinery under variable speed conditions
Core Problem: Diagnosing faults in rotating machinery, especially under variable speed conditions, with improved interpretability.
Key Innovation: A multi-scale signal transformer with signal processing-based attention interpretation for fault diagnosis of rotating machinery.
241. Comparison of joint networks in limestones interbedded in shales
Core Problem: Understanding and comparing the characteristics of joint networks in different lithologies (limestones and shales).
Key Innovation: A comparative study of joint networks in limestones interbedded in shales, providing foundational knowledge for rock mass stability.
242. Synthetic learning for primitive-based building model reconstruction from point clouds
Core Problem: Reconstructing building models from point clouds efficiently and accurately.
Key Innovation: A synthetic learning approach for primitive-based building model reconstruction from point clouds.
243. EnBoT-SORT: Hierarchical fusion-association tracking with pseudo-sample generation for dense thermal infrared UAVs
Core Problem: Tracking objects in dense thermal infrared UAV imagery effectively.
Key Innovation: EnBoT-SORT, a hierarchical fusion-association tracking method with pseudo-sample generation for dense thermal infrared UAVs.
244. The rayleigh effects on the atmospheric correction of the ultraviolet imager on HY-1C and HY-1D satellites
Core Problem: Understanding and correcting Rayleigh effects during atmospheric correction for ultraviolet imagers on HY-1C and HY-1D satellites.
Key Innovation: Analysis of Rayleigh effects on the atmospheric correction of ultraviolet imagers on HY-1C and HY-1D satellites.
245. Improved hybrid algorithm for land surface temperature retrieval from Chinese GF-5B satellite
Core Problem: Improving the accuracy of land surface temperature (LST) retrieval from Chinese GF-5B satellite data.
Key Innovation: An improved hybrid algorithm for land surface temperature retrieval from the Chinese GF-5B satellite.
246. A spectral-preserving resampling for spatial upscaling of hyperspectral imagery
Core Problem: Spatially upscaling hyperspectral imagery while preserving its spectral information.
Key Innovation: A spectral-preserving resampling method for spatial upscaling of hyperspectral imagery.
247. Balancing drinking water security and conservation: A spatial multi-objective optimization framework for regional groundwater management under global change
Core Problem: Developing sustainable groundwater management strategies that balance water security and conservation under global change.
Key Innovation: A spatial multi-objective optimization framework for regional groundwater management to balance water security and conservation.
248. Necessity-strategy-response framework in lake sediment pollution control considering long-term effects
Core Problem: Developing a comprehensive and effective framework for controlling lake sediment pollution, accounting for long-term impacts.
Key Innovation: A necessity-strategy-response framework for lake sediment pollution control that considers long-term effects.
249. CFD-DEM investigation on particle clogging in porous media at different scales: From single pore to digital rock
Core Problem: Understanding the mechanisms of particle clogging in porous media across a range of scales.
Key Innovation: CFD-DEM investigation of particle clogging in porous media, spanning from single pore to digital rock scales.
250. Energy piles thermo-mechanical response to cyclic loading under varying temperature gradients
Core Problem: Understanding the thermo-mechanical behavior of energy piles under cyclic loading and varying temperature gradients.
Key Innovation: Investigating the response of energy piles to combined thermal and cyclic mechanical loading.
251. Yield surface coefficients for different foundation embedment and soil drained shear strengths using genetic and neural network analyses
Core Problem: Determining yield surface coefficients for foundations under varying embedment and soil shear strengths.
Key Innovation: Using genetic and neural network analyses to derive yield surface coefficients for foundations considering embedment and soil drained shear strengths.
252. Predictability of Storms in an Idealized Climate Revealed by Machine Learning
Core Problem: Insufficient understanding of factors governing the predictability of midlatitude storms.
Key Innovation: Using a Convolutional Neural Network (CNN) to predict and quantify uncertainty in storm intensity growth and trajectory, revealing that strong baroclinicity and enhanced jet meanders degrade predictability.
253. Tropical Cyclone Outer‐to‐Inner Brightness Temperature Ratio: A New Size‐Adaptive Parameter Reflecting Storm Intensity and Its Change
Core Problem: Challenging to compare convective structures across tropical cyclones due to size variations, and to link these structures to intensity and evolution.
Key Innovation: Introduction of the Outer-to-Inner Brightness-temperature Ratio (TBR), a size-adaptive metric strongly correlated with TC intensity and intensification, providing a tool to link convection features to TC evolution.
254. A Novel Soil Chronometer: Uranium Comminution Ages Measure Soil Production Rates in a Deep Granitic Weathering Profile
Core Problem: Quantifying soil production rates (SPRs) remains difficult due to limited tools, as conventional methods like 10Be dating rely on environmental conditions.
Key Innovation: Presentation of a new framework for measuring SPRs based on uranium comminution ages, validated at a deep granitic weathering profile, which does not rely on environmental conditions and provides a consistent SPR estimate.
255. On the Construction of Moho Reflected Shear Wave Phases From Ambient Noise
Core Problem: The physical origin and generation mechanisms of Moho reflected shear wave phases (SmS) identified in ambient noise cross-correlation functions remain poorly understood.
Key Innovation: Investigation of SmS amplitudes correlating with coastal ocean wave height, suggesting an oceanic origin, and showing that crustal heterogeneities facilitate SmS signal construction by altering incident angles of wavefields, elucidating key mechanisms for improved extraction and application.
256. Quantifying the Influence of Climate on Storm Activity Using Machine Learning
Core Problem: The relative contributions of seasonal climatology versus synoptic conditions in controlling averaged and individual midlatitude storm activity remain poorly quantified.
Key Innovation: Using ERA-5 reanalysis data and convolutional neural networks to assess the relative importance of climatic versus synoptic conditions, showing climatic conditions dominate mean storm activity, while synoptic conditions dominate individual storm characteristics, and long-term climate trends contribute more to heat anomalies than intensity variability.
257. A Robust and Efficient Continuous‐Differentiable Seepage Face Boundary Condition for Dynamic Groundwater Modeling
Core Problem: The standard mathematical representation of seepage boundaries in groundwater models often requires unnecessarily small timestep sizes for convergence, leading to low computational efficiency.
Key Innovation: Presentation of a continuous-differentiable seepage face (CDSF) equation that replaces the conventional mixed boundary condition with a head-dependent Robin boundary condition, improving numerical stability and computational performance in saturated flow models, as demonstrated through verification models.
258. Spatiotemporal mapping of invasive yellow sweetclover blooms using Sentinel-2 and high-resolution drone imagery
Core Problem: Identifying and mapping the spread of invasive plants like yellow sweetclover is crucial for land management and ecosystem protection.
Key Innovation: Developed spatiotemporal maps of invasive yellow sweetclover blooms using Sentinel-2 and high-resolution drone imagery, combined with field data and machine learning, to aid land managers in detection and control.
259. Insights into evapotranspiration partitioning based on hydrological observations using the generalized proportionality hypothesis
Core Problem: Determining the individual components of evapotranspiration (from plants, soil, water bodies) is difficult, hindering understanding of plant-water interactions and improvement of water/climate models.
Key Innovation: Developed a new method using hydrological observations (streamflow, rainfall) and the generalized proportionality hypothesis to partition evapotranspiration, providing insights into plant water use across different vegetation types and climate conditions.
260. Kilometer-scale convection-allowing model emulation using generative diffusion modeling
Core Problem: Emulating kilometer-scale convection-allowing atmospheric models efficiently.
Key Innovation: Using generative diffusion modeling for efficient emulation of complex atmospheric models.
261. Atmospheric CO2 concentration prediction based on bidirectional long short-term memory
Core Problem: Accurately predicting atmospheric CO2 concentration, which is challenging due to its nonlinear and time-dependent behavior.
Key Innovation: Developing a time series prediction framework based on Bidirectional Long Short-Term Memory (BILSTM) that demonstrates superior accuracy, stability, and robustness for CO2 forecasting.