TerraMosaic Daily Digest: Feb 5, 2026
Daily Summary
This digest synthesizes 217 selected papers and focuses on landslide process mechanics and slope evolution, freeze-thaw and cryosphere-driven instability, flood generation, routing, and hydroclimatic forcing. Top-ranked studies examine risk, fragility, and resilience assessment, cryosphere-driven slope and infrastructure instability, and infrastructure performance under multi-hazard stress.
Across the full set, evidence converges on mechanism-constrained analysis with operational relevance, especially for seismic source-to-ground response pathways and infrastructure-focused hazard performance. 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.
- Cryosphere and freeze-thaw effects remain first-order controls: Studies quantify thaw-related weakening and cold-region instability relevant to hazard evolution and design.
- Flood analyses are becoming event-specific and process-based: Papers emphasize precipitation structure, antecedent wetness, and catchment controls rather than static hazard descriptors.
- Seismic hazard research links source behavior to ground response: Recurring topics connect rupture or loading conditions with geotechnical performance and consequence assessment.
- Infrastructure-facing outputs are increasingly decision-ready: Asset performance is evaluated with uncertainty-aware frameworks to support mitigation and maintenance prioritization.
Selected Papers
This digest features 217 selected papers from 1166 RSS items analyzed across multiple journals. Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Modelling the effectiveness of GLOF DRM measures – a case study from the Ala-Archa valley, Kyrgyz Republic
Core Problem: The effectiveness of Disaster Risk Management (DRM) measures for Glacial Lake Outburst Floods (GLOFs) is insufficiently understood, hindering effective and target-oriented decision-making.
Key Innovation: Assessed the effectiveness of three GLOF DRM measures (lake lowering, a deflection dam, and a retention basin) using numerical modeling in the Ala-Archa catchment, developed a hazard reduction score, and proposed a conceptual framework and methodology for GLOF and debris flow risk management.
2. The World’s Largest Saddle Dam at Risk: Multisensor Geohazard Analysis and Downstream Impacts
Core Problem: Assessing the geohazard risks and potential catastrophic downstream impacts of the Grand Ethiopian Renaissance Dam (GERD) Saddle Dam, a critical auxiliary dam.
Key Innovation: Conducted a comprehensive multisensor geohazard analysis integrating GRACE, Sentinel-1/2, WorldView-3, hydrological modeling, PSI, geospatial analysis, and statistical techniques; identified critical structural vulnerabilities (groundwater infiltration, differential settlement, seepage), anomalous seismicity linked to volcanic activity and reservoir impoundment, and simulated catastrophic downstream flood risks to Sudan and Egypt.
3. Probabilistic analysis of thaw settlement and serviceability in Arctic embankments: A case study on the ITH
Core Problem: Permafrost degradation due to climate change presents a significant geohazard to Arctic transportation infrastructure, introducing high uncertainty in soil behavior and thermal response, making thaw settlement assessment challenging.
Key Innovation: Introduces a Python interface integrated with TEMP/W to conduct Monte Carlo simulations for probabilistic thaw settlement assessment and serviceability evaluation, providing a framework for risk-informed maintenance and reliability-based design.
4. Influence of soluble–insoluble rock interfaces on water and mud inrush mechanisms in water-rich tunnels: Insights from CFD–DEM coupled simulations
Core Problem: Water and mud inrush disasters are highly likely during tunnel excavation across soluble–insoluble rock interfaces, but their underlying mechanisms are poorly understood, hindering effective mitigation.
Key Innovation: Investigated the '12·8' water and mud inrush disaster using a cross-scale, dynamically coupled CFD–DEM framework, revealing a bidirectional feedback mechanism between fluid flow and particle migration, and proposing an effective active intervention combining inflow reduction with curtain grouting.
5. Testing Volcano Deformation Models Against 3D Seismic Reflection Imagery of Ancient Intrusions
Core Problem: The difficulty in comparing geodetic source parameters retrieved from analytical models of volcano deformation to actual, known intrusion geometries, which limits the validation of these critical forecasting tools.
Key Innovation: Used 3D seismic reflection data of an ancient laccolith to quantify buried intrusion geometries and overburden deformation, allowing for a direct comparison and validation of analytical volcano deformation models, revealing underestimation of emplacement depth.
6. Migration of Multimodal Deep Crustal Earthquake Swarm Beneath the Abu Volcano Group, Japan
Core Problem: Understanding the mechanisms and structural controls governing deep crustal earthquake swarms beneath active volcanic centers, which provide insights into deep magmatic processes and precursory volcanic seismicity.
Key Innovation: Analyzed a deep crustal earthquake swarm beneath the Abu Volcano Group, distinguishing deep low-frequency earthquakes from volcano-tectonic events, precisely relocating events to reveal a complex multi-cluster fault network, and demonstrating upward migration of DLFEs suggesting fluid/magma ascent triggering brittle failure.
7. Effect of reactive magnesia on erosion control of a lead contaminated lean clay: insights from runoff, infiltration and disintegration characteristics
Core Problem: Inadequate understanding of erosion-induced environmental risks of contaminated soil and effective remediation strategies, specifically for lead-contaminated silty clay.
Key Innovation: Investigating reactive magnesia (MgO) as a remediation material, demonstrating its effectiveness in decreasing chemically dissolved lead in runoff by 71%, reducing soil loss by 92.4%, lowering lead concentration in eroded particles, and significantly delaying soil disintegration.
8. Submarine power cables on soft structured clay seabed: gravel berm to reduce cable settlements and enhance protection
Core Problem: Mitigating damage to submarine power cables on soft seabed from risks like fishing gear, anchors, and underwater landslides, specifically by investigating the long-term settlement and stability of protective gravel berms.
Key Innovation: Development of a novel iterative method within PLAXIS finite element framework to adjust the modified compression index based on soil destructuration, providing more realistic settlement calculations for structured clays. The study also provides insights into berm geometry optimization for reducing cable settlement and enhancing protection against underwater landslides.
9. Continental mantle earthquakes of the world
Core Problem: Explaining and detecting continental mantle earthquakes, which are hard to explain by plate motions and difficult to detect.
Key Innovation: Developed a method using regional seismic waves and relative amplitudes to detect over 400 continental mantle earthquakes globally since 1990, revealing they occur under a wider range of tectonic and thermal conditions than previously thought.
10. Sampling effects on machine-learning performance and tectonic controls on landslide susceptibility: insights from the Adra River basin (SE Spain)
Core Problem: Developing an improved methodological framework for landslide susceptibility assessment, considering sampling effects on machine-learning performance and tectonic controls, in areas significantly affected by landslides.
Key Innovation: Development of an improved methodological framework for landslide susceptibility assessment, incorporating insights from sampling effects on machine-learning performance and tectonic controls in the Adra River basin.
11. Village geohazard resilience assessment based on contribution weight superposition and Shannon entropy method- a case study in Dechang County, China
Core Problem: The need for an effective framework to assess geohazard resilience at the village level to support evidence-based mitigation and planning.
Key Innovation: Introduction of a framework for evaluating geohazard resilience at the village level, utilizing a contribution weight superposition and Shannon entropy method, demonstrated through a case study.
12. Seasonal vertical surface thaw displacement in 2018 on Samoylov Island (Lena Delta, northeastern Siberia) measured by satellite SAR interferometry with X-, C- and L-band sensors
Core Problem: Quantifying seasonal vertical surface deformation (thaw subsidence and frost heave) in low-land permafrost regions using DInSAR is challenging due to influences from atmospheric, soil moisture, vegetation, and snow cover conditions, and the need to compare results across different SAR frequencies.
Key Innovation: Quantified seasonal vertical surface thaw displacement on Samoylov Island using multi-frequency DInSAR (X-, C-, and L-band), demonstrating good agreement between frequencies and relating deformation patterns to soil type and moisture conditions, while highlighting challenges in accurately capturing sub-pixel variability.
13. Runoff–sediment coupling in low vegetation coverage scenario: Insights from plot scale experiment in the loess environment
Core Problem: The specific role of vegetation cover pattern in runoff generation and sediment production, and their coupling, under low vegetation coverage scenarios in the loess environment remains unclear.
Key Innovation: Conducted plot-scale experiments in a loess environment to investigate runoff-sediment coupling under varying low vegetation coverage, revealing that increasing vegetation coverage within the low range (<10%) paradoxically promoted rill formation, enhanced concentrated flow, and intensified runoff and soil erosion, suggesting a critical vegetation coverage threshold for effective soil and water loss reduction.
14. Mechanical properties of Inuvik-Tuktoyaktuk highway built on permafrost over the years of operation from 2019 to 2024
Core Problem: Road infrastructure built over permafrost is increasingly vulnerable to climate change, with seasonal thawing causing significant degradation in mechanical performance and compromising structural integrity.
Key Innovation: Monitored the mechanical response of a granular road embankment on permafrost over multiple thawing seasons (2022, 2024), revealing strong seasonal trends in elastic modulus and demonstrating the impact of permafrost degradation on embankment stiffness and stress transmission.
15. Swelling-shrinkage of an unsaturated clay upon freezing: Experimental investigation and data-driven modelling
Core Problem: Current studies on clay deformation upon freezing predominantly focus on swelling, overlooking systematic analysis of freezing-induced shrinkage, and conventional data-driven models perform poorly with sparse data.
Key Innovation: Conducted freeze-thaw cycle tests on clayey soil to establish fundamental relationships for freezing-induced deformation (swelling/shrinkage) and proposed a hybrid data-driven model, integrating experimental evidence with physical mechanisms, for accurate prediction with sparse data.
16. A novel deep learning coupled model for extracting flood control scheduling rules for reservoir groups
Core Problem: Traditional rule curves for single reservoirs are inadequate for coordinating multiple, hydrologically linked reservoirs during extreme flood events, leading to suboptimal flood control scheduling.
Key Innovation: Developed BiTCN-STA-SHAP, a novel deep learning coupled model integrating a bidirectional temporal convolutional network, spatio-temporal attention, and SHAP for interpretability, which effectively extracts flood-control scheduling rules for reservoir groups, achieving superior performance (group-mean NSE = 0.97) and robustness under extreme stress tests, and revealing physically consistent patterns of influence.
17. Projecting future exposure to compound precipitation and wind extremes using Copula methods with Bayesian model averaging
Core Problem: Assessing intensity risk and projecting future exposure to Compound Precipitation and Wind Extremes (CPWEs) is challenging due to complex social-meteorological factor interactions and the inherent uncertainty of General Circulation Model (GCM) outputs.
Key Innovation: Established a novel CPWE risk assessment framework integrating Bayesian Model Averaging for GCM ensemble, Copula-based CPWE selection, and intensity assessment, which projects increased intensity and spatial heterogeneity of future CPWEs, and identifies persistent high exposure risk in specific regions, peaking around mid-century, providing insights for early warning and climate resilience.
18. Impacts of spatial-temporal rainfall structures and antecedent wetness on flood variability at the catchment scale
Core Problem: The precise influence of spatial-temporal rainfall structures and antecedent wetness on flood variability and flood quantile uncertainty at the catchment scale remains unclear, limiting accurate flood risk estimation.
Key Innovation: Developed a new approach using a multivariate distribution to generate synthetic event indicators and a data-driven runoff model, revealing that antecedent wetness has the greatest impact on flood peak variability (CV=0.23), followed by temporal (CV=0.16) and spatial (CV=0.05) rainfall structures, and emphasizing the need to incorporate these factors into flood risk estimation.
19. An improved semi-resolved CFD-DEM method for particle systems with wide mesh/particle size ratios
Core Problem: Traditional CFD-DEM methods are limited in simulating solid-liquid two-phase flows with particles of a wide size range, which are common in geotechnical engineering, hindering the understanding of mesoscopic mechanisms.
Key Innovation: Developed an improved semi-resolved CFD-DEM method with a dynamic coupling strategy that accurately simulates systems with wide mesh/particle size ratios (L/d), capturing both macroscopic phenomena and local flow field characteristics, and applied it to upward seepage in sandy soils to reveal how non-uniform flow drives preferential migration of fine particles, inducing piping in gap-graded soils.
20. Seismic microzonation of a region with complex surficial geology based on nonlinear site amplification modelling
Core Problem: Quantifying the unpredictable impacts of seismic hazards in regions with complex and heterogeneous surficial geology, such as Saguenay, by accurately modeling nonlinear site amplification for seismic microzonation.
Key Innovation: Application of a novel methodology for regional seismic microzonation based on 1D nonlinear ground response analyses, utilizing the nonlinear site period (TNL) as a more effective parameter, and developing optimal correlation equations between site amplification and various proxies to create robust seismic microzonation maps.
21. Experimental investigation on dynamic characteristics of gravel soil filler for railway subgrade affected by varying initial moisture content
Core Problem: The dynamic characteristics of gravel soil filler for railway subgrade, particularly under rainfall-earthquake coupling effects, are not fully understood, impacting railway engineering safety and stability.
Key Innovation: Conducted dynamic properties, dynamic strength, and static triaxial shear tests on gravel soil filler under varying unsaturated states and long-term vibration loading, revealing its dynamic response characteristics and providing basic design parameters for seismic response analysis of railway subgrade.
22. Frost Heave Characteristics of Qinghai–Xizang Silty Clay: Experimental and Numerical Modeling
Core Problem: Station platforms of the Qinghai–Xizang Railway experience persistent frost heave due to severe cold, temperature fluctuations, and groundwater, requiring investigation of governing mechanisms and effective mitigation strategies.
Key Innovation: Developed a one-dimensional thermal–hydraulic coupled freezing system, conducted staged freezing tests on silty clay, assessed three mitigation measures (geomembranes, gravel isolation layer), and successfully reproduced observed variations with numerical models, providing practical guidance for frost heave mitigation.
23. Mechanical deterioration of flawed rocks under freeze-thaw and cyclic loading
Core Problem: Understanding the mechanical deterioration and fracture mechanisms of flawed rocks subjected to combined freeze-thaw cycles and cyclic loading, particularly relevant for geotechnical engineering in high-cold mountainous regions.
Key Innovation: Experimental and numerical analysis revealing how freeze-thaw cycles and flaw angles affect frost heave pressure, mechanical properties, and fracture modes (shear-dominated to tensile-dominated) in flawed sandstone, leading to higher coupled damage and lower fatigue life. A numerical model considering cryogenic THM coupling is established.
24. The Magnetic Signature of Stress in Rocks
Core Problem: Lower stresses typical of earthquakes have been considered magnetically undetectable, limiting the ability to reconstruct paleostress fields in the elastic crust for seismic hazard assessment.
Key Innovation: Demonstrated that magnetic responses to sub-GPa stresses can be precisely calibrated, enabling three-dimensional paleostress reconstructions in rocks (even stresses of a few MPa can reset magnetic signals without heat or deformation), offering a non-destructive pathway for improving seismic hazard assessment.
25. Extreme Weather Nowcasting via Local Precipitation Pattern Prediction
Core Problem: Precipitation nowcasting is challenging due to pronounced spatial locality, intricate fine-scale rainfall structures, and variability in forecasting horizons; existing models are either computationally expensive or biased toward normal rainfall, and benchmark datasets are often skewed.
Key Innovation: exPreCast, an efficient deterministic framework for generating finely detailed radar forecasts, integrating local spatiotemporal attention, a texture-preserving cubic dual upsampling decoder, and a temporal extractor, along with a newly constructed balanced radar dataset, achieving state-of-the-art performance across normal and extreme rainfall regimes.
26. On the Adversarial Robustness of Hydrological Models
Core Problem: Water managers require trustworthy and provably safe hydrological models for risk-averse operational settings, but simple prediction-observation comparisons are insufficient to understand model behavior and robustness under perturbations.
Key Innovation: Introduces and explores adversarial robustness analysis in hydrological modeling, evaluating how small, targeted perturbations to meteorological forcings induce substantial changes in simulated discharge. Finds that LSTMs generally demonstrate greater robustness than HBV models and that model responses are often approximately linear to perturbation size.
27. SOMA-1M: A Large-Scale SAR-Optical Multi-resolution Alignment Dataset for Multi-Task Remote Sensing
Core Problem: Existing SAR-optical benchmark datasets for remote sensing suffer from limitations such as single spatial resolution, insufficient data scale, and low alignment accuracy, hindering the training and generalization of multi-scale foundation models.
Key Innovation: Introduces SOMA-1M, a large-scale (1.3M pairs), pixel-level precisely aligned SAR-optical multi-resolution dataset (0.5m to 10m) covering 12 land cover categories. It includes a rigorous coarse-to-fine image matching framework and establishes comprehensive evaluation benchmarks for four hierarchical vision tasks, significantly enhancing performance for multimodal remote sensing.
28. Granite sliding on granite: friction, wear rates, surface topography, and the scale-dependence of rate-state effects
Core Problem: Understanding the fundamental tribological processes (friction, wear, rate-state effects) at granite-granite contacts, which serve as a model for tectonic faulting, and challenging common assumptions about their dominance by particulate wear.
Key Innovation: Demonstrates that high friction in granite-granite contacts is dominated by cold welding within plastically deformed asperity junctions, not particulate wear; shows negligible thermal and rate-dependent effects in macroscopic samples, rationalized by scale-dependence of pre-slip; and reproduces experimental features with molecular dynamics simulations.
29. Free-field response of a seawater-frozen seabed system under obliquely incident SV waves
Core Problem: Understanding and predicting the free-field seismic response of a coupled seawater-frozen seabed-bedrock system subjected to obliquely incident SV waves, particularly how seismic shaking threatens offshore infrastructure in shallow cold-region shelves.
Key Innovation: Derivation of an analytical free-field solution for the coupled system using LCAM for the frozen seabed, capturing P-SV mode conversion. The solution provides closed-form frequency-domain displacements and allows for parametric analyses, revealing strong influences of seawater, incident angle, temperature, porosity, and inter-phase constraint on transfer functions and amplification.
30. Experimental investigation on volumetric strain accumulation of saturated marine sands subjected to wave-induced complex cyclic loading
Core Problem: The need to understand and model the volumetric strain accumulation of saturated marine sands under complex wave-induced cyclic loading, which significantly affects the long-term settlement and stability of offshore foundations.
Key Innovation: Experimental investigation using a Hollow Cyclic Apparatus to characterize the dilatant behavior and volumetric strain accumulation of saturated marine sands under complex cyclic stress paths, leading to a stress-dependent normalized volumetric strain incremental model for estimating long-term settlement in offshore foundation soils.
31. Moisture migration, ice lenses and frost heave characteristics of soils under one-dimensional freezing action: A critical literature review
Core Problem: Soil frost heave seriously threatens engineering structures in cold regions, and there are remaining issues in understanding its micro-mechanisms, influencing factors, and modeling, especially for coarse-grained soils and soils with admixtures.
Key Innovation: Provided a comprehensive review synthesizing the state-of-the-art in moisture migration, cryostructure, and soil frost heave, including micro-mechanisms, impacting factors, and simulation models, while identifying future research directions to improve prevention and control of frost heave distresses.
32. Variations in saturated hydraulic conductivity and microstructural characteristics of loess, paleosol, and their contact zone under seepage conditions
Core Problem: The permeability characteristics of paleosol and "interfacial loess" (contact zone between loess and paleosol), particularly their saturated hydraulic conductivity (Ksat) and microstructural response to hydraulic gradient and infiltration time under seepage conditions, are less documented compared to single-layer loess.
Key Innovation: Investigated Ksat and microstructural changes in loess, paleosol, and interfacial loess under varying hydraulic gradients and infiltration times, revealing that interfacial loess Ksat is intermediate, Ksat values fluctuate based on soil type and density, and microstructural changes are more pronounced in interfacial loess, providing theoretical support for anti-seepage design and fill quality control in loess regions.
33. A dynamic multi-objective inversion framework for seepage parameters based on monitoring data: case study of an earth-rockfill dam
Core Problem: Traditional methods for inverting seepage parameters in earth-rockfill dams rely on static analysis, neglecting time-dependent effects during dam operation, which limits accurate seepage safety analysis.
Key Innovation: Developed a dynamic multi-objective inversion framework that uses surrogate models (OED-SVM, OED-XGBoost) and real-time updated multi-objective functions based on continuous hydraulic head monitoring data, achieving high predictive accuracy and providing optimal seepage parameters for dam safety analysis (total error between simulated and observed hydraulic head within 5%).
34. A unified contact model incorporating surface abrasion for DEM simulation of granular soil behavior across small to large strains
Core Problem: Current Discrete Element Method (DEM) contact models for granular soils often lead to inconsistencies when simulating mechanical behavior across different strain levels (small to large) due to the omission of surface degradation mechanisms like abrasion.
Key Innovation: Proposed a unified contact model that incorporates surface abrasion, allowing the normal force-displacement law to naturally transition from Hertzian at small strains to a linear-bounded relation at large strains, effectively capturing increased compressibility due to degradation and showing good agreement with experimental data for macroscopic mechanical behavior of granular soils.
35. Study on the influence of wide range pre-shear strain and pre-shear direction on the cyclic dynamic characteristics of sand using discrete element method
Core Problem: Understanding how wide-ranging pre-shear strain and pre-shear direction influence the cyclic dynamic characteristics, microstructure evolution, and liquefaction resistance of sand, which is crucial for predicting soil behavior under seismic loading.
Key Innovation: Utilization of discrete element method (PFC3D) to systematically analyze the effects of pre-shear strain and direction on sand's dynamic behavior, revealing the transition from flow sliding to cyclic activity, the non-monotonic influence on liquefaction resistance, the detrimental effect of opposing cyclic loading, and highlighting the dominant role of fabric anisotropy over void ratio changes.
36. Mechanical responses and reinforcement assessment of non-typical in-situ enlargement for closely spaced tunnel groups in weak surrounding rock
Core Problem: Closely spaced tunnel groups in weak surrounding rock, particularly with non-typical in-situ enlargement schemes, face amplified deformation and stability challenges due to complex excavation paths and construction sequences.
Key Innovation: Investigated a 'fill–then-excavate' scheme for tunnel enlargement, characterized mechanical behavior of mudstone/sandstone, integrated numerical simulations with in-situ monitoring, and assessed alternative backfilling and reinforcement (small-pipe grouting) strategies to optimize construction and reduce ground and structural deformations.
37. Impact of Spatial Scale on Optical Earth Observation‐Derived Seasonal Surface Water Extents
Core Problem: Limitations of spatial scale on Landsat-derived surface water extent products, particularly in detecting smaller water bodies and seasonal changes, which are instrumental for better understanding flood dynamics.
Key Innovation: Mapped seasonal surface water extents utilizing high-resolution (4.77m) PlanetScope Basemap imagery and machine learning, demonstrating that higher resolution imagery detects significantly more small water bodies (widths <50–70m) than moderate resolution Landsat products, improving understanding of flood dynamics.
38. Imposing Boundary Conditions on Neural Operators via Learned Function Extensions
Core Problem: Neural operators, while powerful for solving PDEs, struggle to handle general, highly variable, and complex non-homogeneous boundary conditions, especially when the solution operator exhibits strong sensitivity to boundary forcings.
Key Innovation: Proposes a general framework that conditions neural operators on complex non-homogeneous boundary conditions by mapping boundary data to latent pseudo-extensions over the entire spatial domain, enabling existing operator learning architectures to learn rich dependencies on complex BCs and input functions, achieving state-of-the-art accuracy on diverse PDE problems.
39. Benchmarking Artificial Intelligence Models for Daily Coastal Hypoxia Forecasting
Core Problem: Seasonal models for coastal hypoxia provide coarse forecasts, lacking the fine-scale variability needed for daily, responsive ecosystem management.
Key Innovation: A reproducible framework benchmarking four deep learning architectures (BiLSTM, Medformer, ST-Transformer, TCN) for daily hypoxia classification, demonstrating high accuracy and discriminative ability, with ST-Transformer achieving the highest performance.
40. Perception-Based Beliefs for POMDPs with Visual Observations
Core Problem: Real-world Partially Observable Markov Decision Processes (POMDPs) with high-dimensional visual observations (e.g., camera images) are intractable for traditional belief- and filtering-based solvers.
Key Innovation: Introduces the Perception-based Beliefs for POMDPs (PBP) framework, which integrates an image classifier to map visual observations to probability distributions over states, allowing POMDP solvers to handle high-dimensional visual data by incorporating these distributions directly into belief updates, and demonstrating improved performance and robustness compared to deep RL methods.
41. A Hybrid Autoencoder for Robust Heightmap Generation from Fused Lidar and Depth Data for Humanoid Robot Locomotion
Core Problem: Achieving reliable terrain perception for humanoid robots in unstructured environments, which is challenging due to the need for robust heightmap generation from potentially noisy and incomplete single-sensor data.
Key Innovation: A learning-based framework for robust heightmap generation using a hybrid Encoder-Decoder Structure (EDS) that fuses multimodal data from depth cameras, LiDAR, and IMUs, improving reconstruction accuracy and reducing mapping drift with temporal context.
42. SolarGPT-QA: A Domain-Adaptive Large Language Model for Educational Question Answering in Space Weather and Heliophysics
Core Problem: General-purpose LLMs lack domain-specific knowledge and pedagogical capability for explaining complex space science concepts, hindering effective education and forecasting related to solar activity and its impacts.
Key Innovation: SolarGPT-QA, a domain-adapted large language model built on LLaMA-3, trained with scientific literature and refined QA data, which outperforms general-purpose models in educational explanations for space weather and heliophysics, balancing scientific accuracy and educational effectiveness.
43. Regional frequency analysis of extreme ocean waves along the Indian coastline in a changing climate
Core Problem: Understanding the regional frequency of extreme ocean waves and assessing the impact of climate change on wave climate along the Indian coastline is crucial for coastal hazard management.
Key Innovation: Conducted regional frequency analysis of significant wave height data from ERA5 for 20 Indian coastal sites across different seasons and two distinct 20-year intervals (1941–1960 and 2001–2020) to evaluate climate change effects. Identified homogeneous regions and quantified substantial differences in extreme wave height statistics across the two datasets for the monsoon season.
44. Wind-driven wave overtopping on permeable revetments: spatial distribution and wave-induced load characteristics
Core Problem: Quantifying the spatial distribution of overtopping volume and dynamic wave-induced loads on the hinterland of permeable coastal revetments under wind-driven, extreme wave conditions (typhoons).
Key Innovation: Development of a coupled wind–wave–structure numerical model (integrating VOF, Darcy-Forchheimer, and waves2Foam) to systematically study wind-driven overtopping on dolos-armored seawalls, revealing Gaussian-type spatial distribution of overtopping and loads, and quantifying the influence of wind speed, water depth, and armor-layer thickness/placement on impact zones and load reduction.
45. High resolution 3D model-based topographic assessment of scour around a hollow artificial reef by integration monocular vision and image processing techniques
Core Problem: Evaluating scour around complex structures like hollow artificial reefs (HARs) is challenging due to their porous nature, intricate geometry, and the combined influence of hydrodynamic and geotechnical factors, making accurate 3D topographic assessment difficult.
Key Innovation: A novel workflow integrating laboratory experiments, a monocular vision system, and image processing techniques to generate high-density and high-accuracy 3D scour surface models around HARs, enabling detailed assessment of scour characteristics and proposing empirical equations for predicting scour region extent.
46. Assessment and Mapping of Soil Organic Carbon Stocks in Boreal Montane Forest Permafrost Region: A Case Study of Northeastern China
Core Problem: Evaluating carbon storage potential in boreal montane forest permafrost regions is challenging due to a lack of understanding of control mechanisms for soil organic carbon (SOC) stock changes and timely, spatially explicit estimates, especially given climate warming threats to permafrost stability.
Key Innovation: Investigation and mapping of SOC stocks in a boreal montane forest permafrost region, using partial least squares structural equation modeling to quantify the effects of topography, climate, vegetation, C input, permafrost conditions, and soil environment, and generating a spatial distribution map of SOC stocks.
47. Evaluating Long-Term Effectiveness of Managed Aquifer Recharge for Groundwater Recovery and Nitrate Mitigation in an Overexploited Aquifer System
Core Problem: Over-extraction of groundwater leads to continuous water level decline and significant groundwater depressions, but the long-term regional effectiveness of Managed Aquifer Recharge (MAR) for groundwater recovery and nitrate mitigation, especially considering geological heterogeneity, remains poorly quantified.
Key Innovation: Development and application of a coupled flow and multi-component reactive transport model to evaluate the long-term, regional impacts of MAR on groundwater recovery and nitrate evolution in the Xiong'an depression area, quantifying the dominance of physical dilution and the role of geological heterogeneity in creating localized reduction hotspots.
48. Validation of analog sensor measurements in hydrometeorological participatory monitoring in various tropical countries
Core Problem: Lack of open data and feasibility of traditional hydrometeorological data collection in remote tropical mountain regions, necessitating alternative participatory monitoring approaches.
Key Innovation: Validation of a participatory monitoring approach using low-cost analog sensors for hydrometeorological data (air temperature, relative humidity, rainfall, water level) in mountainous regions, showing ambivalent but promising results for certain parameters.
49. A distributed monitoring method for soil water content based on actively heated optical frequency domain reflectometry and physics-informed neural networks
Core Problem: Accurate retrieval of soil water content from complex thermal response data using actively heated optical frequency domain reflectometry is challenging due to limitations of conventional theoretical models and lack of physical consistency in purely data-driven methods.
Key Innovation: Developed a physics-informed neural network framework that integrates macroscopic physical principles into a correlation-based loss function, achieving high accuracy (R2 = 0.92, MAE = 0.0125 cm3·cm⁻3) and robustness in distributed soil water content retrieval, and enabling determination of thermal conductivity–water content relationship without prior calibration.
50. Design and performance evaluation of tie-down cables for mitigating seismic uplift in cable-stayed bridges
Core Problem: Earthquake-induced bearing uplift is a common seismic damage mode in cable-stayed bridges, and designing tie-down cables to mitigate this requires reconciling large horizontal deck displacements with vertical restraint.
Key Innovation: Developed a practical seismic design framework for tie-down cables, validated through comprehensive analyses, demonstrating effective uplift control and highlighting the need for capacity verification due to increased bearing forces.
51. DIC-Based Investigation into Failure Mechanisms and Geogrid Optimization of Eccentrically Loaded Footings above Void
Core Problem: Ground loss from tunnelling can create localized subsurface voids, threatening the serviceability and safety of shallow foundations, especially under eccentric loading.
Key Innovation: Performed DIC-assisted laboratory model tests to investigate failure mechanisms of eccentrically loaded footings above voids, providing comprehensive full-field displacement maps and optimizing geogrid reinforcement parameters (embedment depth and length ratios) to significantly improve bearing capacity.
52. Measurement of ice content in silty clay based on elastic wave velocity
Core Problem: A continuous and non-destructive method is needed to accurately measure ice content in silty clay, which is crucial for understanding frozen soil behavior and related engineering applications.
Key Innovation: Establishment of a strong relationship between elastic wave velocity and ice content in silty clay, leading to the development and validation of predictive models for ice content using wave velocity, initial water content, and temperature, offering a feasible and efficient non-destructive measurement approach.
53. Enhancement of mechanical properties in reactive magnesia-based carbonated-solidified soil using carbonic anhydrase
Core Problem: Conventional reactive magnesia-based carbonation for ground improvement suffers from slow CO2 absorption kinetics and hydrolysis, and existing enhancement strategies are often costly or limited in performance.
Key Innovation: Incorporation of carbonic anhydrase (CA) with reactive MgO to significantly accelerate CO2 hydration and enhance the unconfined compressive strength of carbonated-solidified sandy and silty soils (63.6%-424% increase), offering a low-carbon, efficient, and eco-friendly approach for soil stabilization and carbon sequestration.
54. A General-Purpose Diversified 2D Seismic Image Dataset from NAMSS
Core Problem: There is a need for a large, diverse, and geographically distributed 2D seismic image dataset to support modern machine learning research in geophysics, allowing for robust evaluation of generalization to unseen geological and acquisition conditions.
Key Innovation: The creation and release of the Unicamp-NAMSS dataset, comprising 2588 cleaned and standardized 2D seismic sections from 122 survey areas, designed for machine learning model pretraining, transfer learning, and benchmarking in seismic interpretation.
55. Physics as the Inductive Bias for Causal Discovery
Core Problem: Causal discovery in complex real-world dynamical systems often struggles with identifiability, stability, and robustness, especially when systems exhibit feedback, cyclic interactions, and non-stationary data, and many methods are formulated under restrictive assumptions.
Key Innovation: An integrative causal discovery framework for dynamical systems that leverages partial physical knowledge (e.g., ODEs) as an inductive bias by modeling system evolution as an SDE, where known physics informs the drift term and unknown causal couplings are in the diffusion term, leading to improved causal graph recovery and more stable estimates.
56. Multi-Aspect Mining and Anomaly Detection for Heterogeneous Tensor Streams
Core Problem: Existing tensor decomposition and anomaly detection methods cannot effectively handle heterogeneous tensor streams (mixed categorical and continuous attributes) without distorting data properties, nor can they track temporal dynamics for accurate group anomaly detection in event streams.
Key Innovation: Proposes HeteroComp, a method that continuously summarizes heterogeneous tensor streams into latent 'components' and their temporal dynamics using Gaussian process priors, enabling accurate and efficient group anomaly detection without relying on data stream length.
57. Learning, Solving and Optimizing PDEs with TensorGalerkin: an efficient high-performance Galerkin assembly algorithm
Core Problem: The need for an efficient, high-performance, and unified algorithmic framework for numerical solution, constrained optimization, and physics-informed learning of Partial Differential Equations (PDEs) with a variational structure, especially on unstructured meshes.
Key Innovation: TensorGalerkin, a novel, highly-optimized, and GPU-compliant framework for linear system assembly in Galerkin discretizations of PDEs, which provides significant computational efficiency and accuracy gains for solving, optimizing, and learning PDEs.
58. Wid3R: Wide Field-of-View 3D Reconstruction via Camera Model Conditioning
Core Problem: Prior visual geometry reconstruction methods typically assume rectified or pinhole camera inputs, limiting their applicability to real-world scenarios involving wide field-of-view cameras (fisheye, panoramic) and often requiring careful calibration and undistortion.
Key Innovation: Presents Wid3R, a generalizable multi-view 3D estimation neural network that supports wide field-of-view camera models. It leverages a ray representation with spherical harmonics and a novel camera model token for distortion-aware 3D reconstruction, enabling direct 3D reconstruction from 360 imagery.
59. MTPano: Multi-Task Panoramic Scene Understanding via Label-Free Integration of Dense Prediction Priors
Core Problem: Comprehensive panoramic scene understanding is challenging due to the scarcity of high-resolution, multi-task annotations and the difficulty of adapting perspective foundation models to panoramic domains due to geometric distortions and coordinate system discrepancies.
Key Innovation: Introduces MTPano, a robust multi-task panoramic foundation model with a label-free training pipeline. It leverages perspective dense priors for pseudo-label generation and employs a Panoramic Dual BridgeNet with geometry-aware modulation layers to disentangle and integrate features for various dense prediction tasks, achieving state-of-the-art performance.
60. Smoothness Errors in Dynamics Models and How to Avoid Them
Core Problem: Graph Neural Networks (GNNs) for dynamics modeling, particularly for partial differential equations (PDEs) over surfaces, suffer from oversmoothing, while existing solutions like unitary graph convolutions can be overconstraining for physical systems where smoothness naturally increases.
Key Innovation: Systematically studies smoothing effects in GNNs and proposes relaxed unitary convolutions. These balance smoothness preservation with the natural smoothing required for physical systems, outperforming baselines on PDEs (heat, wave equations) and weather forecasting.
61. NeVStereo: A NeRF-Driven NVS-Stereo Architecture for High-Fidelity 3D Tasks
Core Problem: It remains non-trivial to obtain a single framework that can jointly deliver accurate camera poses, reliable depth, high-quality novel view synthesis (NVS), and accurate 3D surfaces from casually captured multi-view RGB-only inputs, as existing methods often have limitations in one or more of these aspects.
Key Innovation: Presented NeVStereo, a NeRF-driven NVS-stereo architecture that jointly delivers camera poses, multi-view depth, novel view synthesis, and surface reconstruction. This design combines NeRF-based NVS, confidence-guided depth estimation, NeRF-coupled bundle adjustment, and iterative refinement, achieving consistently strong zero-shot performance with improved accuracy across various 3D tasks.
62. LD-SLRO: Latent Diffusion Structured Light for 3-D Reconstruction of Highly Reflective Objects
Core Problem: 3-D reconstruction of highly reflective objects using fringe projection profilometry is a significant challenge, as specular reflection and indirect illumination often lead to severe distortion or loss of projected fringe patterns.
Key Innovation: Proposed LD-SLRO, a latent diffusion-based structured light method for reflective objects. It encodes phase-shifted fringe images to extract latent representations, which then condition a latent diffusion model to probabilistically suppress reflection-induced artifacts and recover lost fringe information, significantly improving both fringe quality and 3-D reconstruction accuracy.
63. A Comparative Study of 3D Person Detection: Sensor Modalities and Robustness in Diverse Indoor and Outdoor Environments
Core Problem: The need for a systematic evaluation of 3D person detection performance and robustness across different sensor modalities (camera-only, LiDAR-only, camera-LiDAR fusion) in diverse indoor and outdoor environments, beyond typical autonomous driving contexts.
Key Innovation: A comprehensive comparative study evaluating three representative 3D detection models (BEVDepth, PointPillars, DAL) across sensor modalities, analyzing their behavior under varying occlusion, distance, sensor corruptions, and misalignments, demonstrating the superior performance of fusion-based approaches while highlighting their vulnerabilities.
64. Empowering Time Series Analysis with Large-Scale Multimodal Pretraining
Core Problem: Existing time series foundation models primarily rely on unimodal pretraining, lacking complementary modalities to enhance understanding. Key challenges include the absence of a unified multimodal pretraining paradigm, large-scale multimodal corpora, and effective integration of heterogeneous modalities for generalization.
Key Innovation: Proposes a multimodal pretraining paradigm leveraging endogenous (derived images/text) and exogenous (real-world news) modalities for time series analysis. It curates MM-TS, the first large-scale multimodal time series dataset, and introduces HORAI, a frequency-enhanced multimodal foundation model, achieving state-of-the-art zero-shot performance in forecasting and anomaly detection.
65. Stable but Wrong: When More Data Degrades Scientific Conclusions
Core Problem: The implicit belief that more data always leads to more reliable scientific conclusions can fail when the reliability of observations degrades in an intrinsically unobservable manner.
Key Innovation: Identifies a structural regime where standard inference procedures converge smoothly but systematically to incorrect conclusions despite conventional diagnostic checks, demonstrating that additional data can amplify error when observational integrity is compromised unobservably, thus revealing an intrinsic limit of data-driven science.
66. Depth as Prior Knowledge for Object Detection
Core Problem: Detecting small and distant objects remains challenging for object detectors due to scale variation, low resolution, and background clutter, and existing depth-informed approaches require complex architectural modifications.
Key Innovation: DepthPrior, a framework that uses depth as prior knowledge (via Depth-Based Loss Weighting, Depth-Based Loss Stratification, and Depth-Aware Confidence Thresholding) without modifying detector architectures, achieving significant improvements for small objects in various benchmarks.
67. Visualizing the loss landscapes of physics-informed neural networks
Core Problem: A lack of understanding and visualization of the loss landscapes for physics-informed neural networks (PINNs), which are crucial for scientific machine learning but differ from traditional data-driven ML in their loss function definition.
Key Innovation: A comprehensive review and empirical investigation of loss landscapes for PINNs, demonstrating that they share properties with data-driven classification problems and that different physics loss formulations often yield similar, smooth, and well-conditioned landscapes.
68. EoCD: Encoder only Remote Sensing Change Detection
Core Problem: Existing remote sensing change detection methods are computationally expensive and complex due to reliance on Siamese encoders and sophisticated decoders, or early fusion methods that still require complex decoders and show inferior performance.
Key Innovation: EoCD, an encoder-only change detection method that performs early fusion of temporal data and replaces the decoder with a parameter-free multiscale feature fusion module, significantly reducing model complexity while maintaining optimal balance between performance and prediction speed.
69. Splat and Distill: Augmenting Teachers with Feed-Forward 3D Reconstruction For 3D-Aware Distillation
Core Problem: Vision Foundation Models (VFMs) lack robust 3D awareness, limiting their applicability in tasks requiring geometrically grounded knowledge, and prior 3D distillation methods are slow or suffer from feature-averaging artifacts.
Key Innovation: Splat and Distill instills 3D awareness into 2D VFMs by augmenting a teacher with a fast, feed-forward 3D Gaussian reconstruction pipeline, which lifts 2D features into 3D and then 'splats' them onto novel viewpoints to supervise a student, significantly outperforming prior works in 3D awareness and enhancing 2D feature semantic richness.
70. Thinking with Geometry: Active Geometry Integration for Spatial Reasoning
Core Problem: Existing Multimodal Large Language Model (MLLM) strategies for integrating geometric priors for spatial reasoning are passive, leading to semantic-geometry misalignment and redundant signals.
Key Innovation: GeoThinker, a framework that enables MLLMs to actively and selectively retrieve geometric evidence conditioned on internal reasoning demands via Spatial-Grounded Fusion and Importance Gating, significantly improving spatial intelligence and generalization.
71. Predicting Camera Pose from Perspective Descriptions for Spatial Reasoning
Core Problem: Multi-image spatial reasoning remains challenging for Multimodal Large Language Models (MLLMs), particularly perspective taking, which requires building a coherent 3D understanding across viewpoints and reasoning from language-specified viewpoints.
Key Innovation: CAMCUE, a pose-aware multi-image framework that uses camera pose as an explicit geometric anchor for cross-view fusion and novel-view reasoning, enabling direct grounding of natural-language viewpoint descriptions to target camera poses and synthesizing imagined target views, significantly improving spatial reasoning and reducing inference time.
72. Decision-Focused Sequential Experimental Design: A Directional Uncertainty-Guided Approach
Core Problem: Traditional sequential experimental design in the predict-then-optimize paradigm focuses on maximizing prediction accuracy, leading to inefficiency because performance is ultimately evaluated by decision loss, creating a mismatch between prediction error and decision-making utility.
Key Innovation: Proposes a decision-focused sequential experimental design approach that uses a computationally tractable directional-based metric to quantify predictive uncertainty, avoiding the need for an optimization oracle. This design criterion achieves strong consistency, convergence guarantees, and earlier stopping times compared to decision-blind designs, optimizing data collection for improved downstream decisions.
73. Reduced-Order Surrogates for Forced Flexible Mesh Coastal-Ocean Models
Core Problem: Traditional physics-based coastal-ocean models are computationally expensive, hindering workflows like ensemble forecasting and long climate simulations, and existing POD-based surrogates have limitations.
Key Innovation: Introduces a flexible Koopman autoencoder formulation that incorporates meteorological forcings and boundary conditions, systematically comparing it against POD-based surrogates. The Koopman autoencoder with temporal unrolling yields superior accuracy (relative RMSE 0.01-0.13, R^2 0.65-0.996) and achieves inference speed-ups of 300-1400x, making long-term, high-resolution coastal-ocean modeling feasible.
74. In-context Time Series Predictor
Core Problem: Existing Transformer-based or LLM-based time series forecasting methods may not fully utilize in-context learning capabilities, can be parameter-inefficient, and suffer from overfitting.
Key Innovation: A novel 'in-context time series predictor' that reformulates time series forecasting tasks as input tokens for Transformer-based LLMs, aligning with in-context mechanisms, achieving parameter efficiency, and superior performance across various data settings (full-data, few-shot, zero-shot).
75. A Policy Gradient-Based Sequence-to-Sequence Method for Time Series Prediction
Core Problem: Sequence-to-sequence models for multi-step-ahead time series prediction suffer from exposure bias (mismatch between training and inference conditions) and compounding errors when using previously generated predictions as inputs.
Key Innovation: A new training paradigm using a policy gradient-based method to learn an adaptive input selection strategy for sequence-to-sequence prediction models, where a trainable policy network dynamically chooses beneficial inputs to maximize long-term prediction performance, enhancing accuracy and stability.
76. WAVE: Weighted Autoregressive Varying Gate for Time Series Forecasting
Core Problem: Existing autoregressive Transformer models for time series forecasting often struggle to effectively capture both long-range and local temporal patterns, limiting their predictive performance.
Key Innovation: Introduction of WAVE (Weighted Autoregressive Varying gatE), an attention mechanism with both Autoregressive (AR) and Moving-average (MA) components, which enhances and decouples the ability of various attention mechanisms to capture diverse temporal patterns, achieving state-of-the-art results in time series forecasting.
77. Linear Transformers as VAR Models: Aligning Autoregressive Attention Mechanisms with Autoregressive Forecasting
Core Problem: Existing multi-layer Transformers for autoregressive time series forecasting (TSF) often misalign with autoregressive objectives, obscuring underlying VAR structures and hindering interpretability and generalization ability.
Key Innovation: Shows that linear attention layers can be interpreted as dynamic VAR structures and proposes Structural Aligned Mixture of VAR (SAMoVAR), a linear Transformer variant that aligns its architecture with autoregressive objectives using interpretable dynamic VAR weights, leading to improved performance, interpretability, and computational efficiency for multivariate TSF.
78. GenIAS: Generator for Instantiating Anomalies in time Series
Core Problem: Existing synthetic anomaly injection methods for time series anomaly detection (TSAD) rely on ad hoc, hand-crafted strategies that fail to capture diverse and complex anomalous patterns, particularly in multivariate settings.
Key Innovation: Proposes GenIAS (Generator for Instantiating Anomalies in Time Series), which generates realistic and diverse anomalies via a novel learnable perturbation in the latent space of a variational autoencoder, enabling injection across different temporal segments and scales, and improving distinguishability through a tunable prior.
79. Hierarchical Time Series Forecasting with Robust Reconciliation
Core Problem: In hierarchical time series forecasting, existing reconciliation methods rely on an estimated covariance matrix of forecast errors, which, due to estimation uncertainty from finite samples, can degrade forecast performance.
Key Innovation: A robust optimization framework for hierarchical reconciliation that explicitly accounts for uncertainty in the estimated covariance matrix by minimizing the worst-case average of weighted squared residuals over an uncertainty set, formulated as a semidefinite optimization problem, leading to improved forecast performance.
80. Differentiable Constraint-Based Causal Discovery
Core Problem: Existing causal discovery methods (constraint-based and score-based) have limitations: constraint-based methods struggle with small sample sizes, while score-based methods often forgo explicit conditional independence testing.
Key Innovation: A novel causal discovery method that employs gradient-based optimization of conditional independence constraints, enabled by differentiable d-separation scores derived through percolation theory and soft logic, demonstrating robust performance in low-sample regimes.
81. TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting
Core Problem: Foundation models for zero-shot time series forecasting face challenges in efficient long-horizon prediction and reproducibility, with existing synthetic-only approaches often underperforming on challenging benchmarks.
Key Innovation: TempoPFN, a univariate time series foundation model based on linear RNNs, exclusively pre-trained on diverse synthetic data, featuring a GatedDeltaProduct architecture with state-weaving for fully parallelizable training and robust temporal state-tracking, achieving top-tier zero-shot forecasting performance efficiently.
82. A Representer Theorem for Hawkes Processes via Penalized Least Squares Minimization
Core Problem: Estimating latent triggering kernels for linear multivariate Hawkes processes, crucial for understanding event interaction structures, typically involves solving costly infinite-dimensional optimization problems in kernel methods.
Key Innovation: A novel representer theorem for Hawkes processes under penalized least squares minimization, where optimal triggering kernel estimators are linear combinations of transformed kernels with analytically fixed unity dual coefficients, leading to a highly efficient estimator for large-scale event sequence data.
83. Streaming Operator Inference for Model Reduction of Large-Scale Dynamical Systems
Core Problem: Traditional Operator Inference (OpInf) for model reduction of large-scale dynamical systems operates as a batch method, requiring all data to be loaded into memory simultaneously, which is a barrier for large datasets and prevents online model updates with new data.
Key Innovation: Proposes Streaming OpInf, which learns reduced models from sequential data streams using incremental SVD for adaptive basis construction and recursive LS for streaming operator updates. This eliminates the need to store complete datasets, enables online model adaptation, reduces memory by over 99%, and achieves orders-of-magnitude faster predictions with comparable accuracy to batch OpInf.
84. Semi-analytical seepage solution for suction bucket installation in soils with depth-dependent anisotropic permeability
Core Problem: Lack of effective and accurate design tools to analyze soil seepage behavior during suction-assisted penetration of suction bucket foundations, especially considering depth-dependent anisotropic permeability and skirt wall thickness.
Key Innovation: Proposal of a novel semi-analytical solution for the seepage field around suction buckets, incorporating depth-dependent anisotropic permeability and skirt wall thickness, validated against finite element simulations, providing insights into seepage flow patterns and soil penetration resistance for optimizing foundation design.
85. Numerical investigation of flow patterns and sheltering effects around jacket foundations under varying attack angles
Core Problem: Understanding the correlation between flow characteristics and local scour around jacket foundations under varying flow attack angles, which is crucial for predicting and mitigating scour threats to offshore wind installations.
Key Innovation: Numerical modeling to examine flow-structure interactions and scour around jacket foundations under different attack angles, demonstrating how sheltering effects and contracted flow influence scour volumes and depths. A dimensionless sheltering coefficient (Csh) is proposed, establishing a strong linear relationship with total scour volumes, providing a tool for engineering practice.
86. Interpretation of soil deformation characteristics based on in-situ test: A comprehensive review
Core Problem: Traditional laboratory tests for soil deformation properties are limited by sample disturbance, scale effects, and inability to reproduce in-situ stress-strain history, leading to inaccurate characterization of soil stiffness.
Key Innovation: A comprehensive review synthesizing recent developments in mechanisms, operational principles, and interpretation of in-situ tests for soil deformation parameters, analyzing strain-dependent nonlinearity, dynamic loading effects, and illustrating a hyperbolic degradation model for stiffness prediction.
87. Detection and Feature Analysis of Martian Dust Devil Based on Multimodel Deep Learning
Core Problem: Existing studies on Martian dust devils are primarily regional or short-term, lacking a global and long-term understanding of their distribution and behavior, and automated identification and 3D scale estimation are challenging.
Key Innovation: Development of the Martian dust devils ultra approach, combining a detection-cascade algorithm, semantic segmentation, and a 3D algorithm for automated identification and 3D scale estimation of Martian dust devils, demonstrating improved accuracy and efficiency, and revealing their spatial distribution and morphology.
88. Causal Attention and Frequency Domain Gating Fusion Enhanced Dual-Branch Optical and SAR Images Change Detection
Core Problem: Change detection in heterogeneous optical and SAR images faces specific issues including structural inconsistency between modalities, speckle noise in SAR images, and insufficient utilization of complementary cross-modal features, limiting feature representation and fusion.
Key Innovation: Introduction of CAFGFNet, an end-to-end dual-branch network for optical and SAR images change detection, featuring modality-specific encoding, a Causal Attention mechanism for temporal dependencies, and a frequency-domain gating fusion module for adaptive fusion of complementary frequency components, achieving high-precision change detection.
89. OpenLandMap-soildb: global soil information at 30 m spatial resolution for 2000–2022+ based on spatiotemporal Machine Learning and harmonized legacy soil samples and observations
Core Problem: There is increasing interest in global dynamic soil information with changes in soil properties mapped over time and at high spatial resolution, but such comprehensive datasets are often lacking.
Key Innovation: Developed OpenLandMap-soildb, a global dynamic dataset of key soil properties (SOC, bulk density, pH, texture, soil types) at 30m spatial resolution for 2000-2022+, using spatiotemporal Machine Learning and a large compilation of harmonized legacy soil samples, providing a robust foundation for tracking land degradation.
90. Editorial: Geophysical electromagnetic exploration theory, technology and application
Core Problem: The ongoing challenges in geophysical electromagnetic exploration, such as the demand for greater exploration depth, higher data resolution, and more reliable interpretations.
Key Innovation: A comprehensive review and presentation of the latest advancements in geophysical electromagnetic exploration, covering emerging theoretical frameworks, innovative numerical simulation methods, and advanced inversion techniques, including applications like AI-driven inversion and urban seismic risk assessment.
91. Travel time resilience of national multimodal transport systems under extreme events: A passenger-oriented framework
Core Problem: Existing resilience assessments for transport systems largely focus on single-mode or regional scales, failing to address the large-scale, multi-modal, timetable-dependent nature of National Multimodal Transport Systems (NMTSs) under extreme events.
Key Innovation: A passenger-oriented framework to assess NMTS resilience under extreme events, constructing a dynamic functionality network and formulating a resilience model with efficient computational methods, demonstrated with case studies including a heavy rainstorm.
92. Integrating ensemble learning and rocky desertification indices improves accuracy and interpretability of soil thickness prediction in karst landscapes
Core Problem: Spatially predicting soil thickness in complex karst landscapes is challenging due to high heterogeneity, intricate impacts, and rocky desertification, limiting understanding of hydrological and ecological processes.
Key Innovation: Integrated interpretable machine learning (ML) with rocky desertification indices (RIs) to improve soil thickness prediction in karst regions, demonstrating that RIs significantly boost model explanatory power and consistency, and stacking ensembles reduce errors, providing a scalable framework for soil management.
93. Hydraulic convergence-confinement method
Core Problem: Groundwater control is a primary concern and key challenge during tunnel excavation, but hydraulic considerations in tunnel design have largely been limited to evaluating groundwater inflow rates and water pressure, lacking a comprehensive understanding of hydraulic equilibrium and drainage control mechanisms.
Key Innovation: Development of the hydraulic convergence-confinement concept (HCC) through theoretical analysis, numerical simulations, and small-scale model tests, providing an effective tool for explaining hydraulic behavior, interpreting hydraulic equilibrium, and offering a conceptual basis for hydraulic tunnel design and limited-drainage tunnel concepts.
94. Validation and applicability analysis of a novel soft-shell contact model for coated granular materials: discrete element modelling and experimental study
Core Problem: Accurately simulating the mechanical behavior of soft-coated granular materials (e.g., polymer-reinforced ballast, methane hydrate sediments) using Discrete Element Method (DEM) is challenging because traditional models treat them as homogenized particles, leading to inaccuracies.
Key Innovation: Developed and validated a novel Soft-Shell (SS) contact model embedded in EDEM for soft-coated spherical particles, demonstrating that it accurately matches experimental triaxial shear behavior (macroscopic strength and deformation) of silicone-coated steel balls, unlike modulus homogenization models, and suggested applicability ranges for key parameters.
95. Influence of sampling disturbance and laboratory reconsolidation procedures on soils with varying plasticity
Core Problem: Sampling disturbance significantly alters the mechanical properties of soils, particularly undrained shear behavior, and the effectiveness of laboratory reconsolidation procedures for restoring intact shear strength varies with soil plasticity and disturbance severity.
Key Innovation: Comprehensive numerical and experimental investigation demonstrating that sampling disturbance reduces initial stiffness and alters shear behavior, with low-PI soils being more susceptible. It also shows the SHANSEP method's superior efficacy for reconsolidation, proposing a critical consolidation pressure parameter (nc) to optimize intact shear strength recovery.
96. Impact of Fluid‐Induced Pore Geometry Alteration on Acoustic Velocity in Carbonate Rocks
Core Problem: Standard porosity-velocity models often fail to accurately represent the elastic properties and acoustic velocity in carbonate rocks because they do not adequately account for the impact of fluid-induced pore geometry alteration (precipitation and dissolution).
Key Innovation: Demonstrated through laboratory experiments that pore geometry (e.g., grain-contact cementation vs. intrapore crystal growth, and dissolution patterns) has a greater influence on elastic properties and acoustic velocity than porosity alone, challenging traditional models and highlighting the need for quantitative pore geometry parameters in rock physics models.
97. Strength α‐Quartz: New Results From High Pressure In Situ X‐Ray Diffraction Experiments
Core Problem: Accurately determining the flow strength of α-quartz, a key mineral controlling continental crust strength, under various temperature and pressure conditions, and validating existing flow laws.
Key Innovation: Presented new experimental data on α-quartz flow strength using in situ synchrotron x-ray diffraction during uniaxial deformation, providing a novel method for monitoring deformation and broadly confirming previous mechanical data, contributing to improved flow laws for the Earth's crust.
98. Darkened Snow Triggers Different Snowmelt Responses Over Contrasting Water Years in Great Salt Lake Headwater Basins
Core Problem: Quantifying the cumulative impact of snow darkening by light-absorbing particles on snowmelt over contrasting water years at a spatially distributed scale, as previous studies were mostly point-scale.
Key Innovation: Using a spatially distributed process-based snow model (iSnobal) to simulate snowmelt under different albedos, revealing that melt sensitivity to snow darkening shows consistent spatial patterns but its magnitude is controlled by seasonal meteorological variability, with pronounced shifts at subalpine elevations.
99. DCER: Dual-Stage Compression and Energy-Based Reconstruction
Core Problem: Multimodal fusion systems face robustness challenges due to noisy inputs degrading representation quality and missing modalities causing prediction failures.
Key Innovation: DCER, a dual-stage compression and energy-based reconstruction framework that addresses both noisy inputs (via within-modality frequency transforms and cross-modality bottleneck tokens) and missing modalities (via energy-based reconstruction), achieving state-of-the-art performance in multimodal fusion benchmarks.
100. LISA: Laplacian In-context Spectral Analysis
Core Problem: Adapting Laplacian-based time-series models at inference-time, especially under changing dynamics, to improve forecasting and autoregressive rollout performance using only an observed prefix.
Key Innovation: LISA (Laplacian In-context Spectral Analysis), a method that combines delay-coordinate embeddings and Laplacian spectral learning to produce diffusion-coordinate state representations, with lightweight latent-space residual adapters for inference-time adaptation of time-series models, improving performance under changing dynamics.
101. CyIN: Cyclic Informative Latent Space for Bridging Complete and Incomplete Multimodal Learning
Core Problem: Multimodal machine learning models, typically trained on perfectly paired data, suffer significant performance drops and lack robustness when deployed with dynamic and unpredictable missing modalities in real-world scenarios.
Key Innovation: Proposes CyIN (Cyclic INformative Learning framework) that bridges complete and incomplete multimodal learning by building an informative latent space using cyclic token- and label-level Information Bottleneck and employing cross-modal cyclic translation to reconstruct missing modalities, enabling robust performance in diverse incomplete scenarios.
102. Transolver-3: Scaling Up Transformer Solvers to Industrial-Scale Geometries
Core Problem: Scaling deep learning-based neural PDE solvers to industrial-scale geometries with over $10^8$ cells remains a fundamental challenge due to prohibitive memory complexity.
Key Innovation: Transolver-3, a highly scalable framework for high-fidelity physics simulations, introducing faster slice and deslice via matrix multiplication associative property and geometry slice tiling, combined with amortized training and physical state caching, enabling handling meshes with over 160 million cells.
103. Boosting SAM for Cross-Domain Few-Shot Segmentation via Conditional Point Sparsification
Core Problem: Dense points used by SAM-based methods perform poorly in Cross-Domain Few-Shot Segmentation (CD-FSS), particularly in medical or satellite domains, due to large domain shifts disrupting point-image interactions.
Key Innovation: Conditional Point Sparsification (CPS), a training-free approach that adaptively guides SAM interactions for cross-domain images by leveraging ground-truth masks from reference exemplars to sparsify dense matched points, leading to more accurate segmentation results.
104. Learning with Adaptive Prototype Manifolds for Out-of-Distribution Detection
Core Problem: Existing prototype-based Out-of-Distribution (OOD) detection methods suffer from the Static Homogeneity Assumption (fixed representational resources) and the Learning-Inference Disconnect (discarding rich prototype quality knowledge at inference), limiting their capacity and performance.
Key Innovation: Introduces APEX (Adaptive Prototype for eXtensive OOD Detection), which optimizes the learned feature manifold. It features an Adaptive Prototype Manifold (APM) using the Minimum Description Length principle to determine optimal prototype complexity per class, and Posterior-Aware OOD Scoring (PAOS) to quantify prototype quality, achieving state-of-the-art OOD detection.
105. Multi-AD: Cross-Domain Unsupervised Anomaly Detection for Medical and Industrial Applications
Core Problem: Traditional deep learning models for anomaly detection often lack annotated data, especially in cross-domain applications like medical diagnosis and industrial defect detection, making robust unsupervised anomaly detection challenging.
Key Innovation: Proposed Multi-AD, a convolutional neural network (CNN) model for robust unsupervised anomaly detection across medical and industrial images. It employs squeeze-and-excitation (SE) blocks, knowledge distillation, and a discriminator network, integrating multi-scale features and a teacher-student architecture to achieve superior performance in both image-level and pixel-level anomaly detection.
106. VGGT-Motion: Motion-Aware Calibration-Free Monocular SLAM for Long-Range Consistency
Core Problem: Calibration-free monocular SLAM suffers from severe scale drift on long sequences due to motion-agnostic partitioning causing zero-motion drift, and conventional geometric alignment being computationally expensive.
Key Innovation: VGGT-Motion, a calibration-free SLAM system that achieves efficient and robust global consistency over kilometer-scale trajectories by proposing a motion-aware submap construction mechanism, an anchor-driven direct Sim(3) registration strategy, and a lightweight submap-level pose graph optimization.
107. Mapper-GIN: Lightweight Structural Graph Abstraction for Corrupted 3D Point Cloud Classification
Core Problem: Robust 3D point cloud classification, especially under noise and transformation corruptions, typically requires scaling up backbones or specialized data augmentation.
Key Innovation: Mapper-GIN, a lightweight pipeline that improves robustness through structural graph abstraction using the Mapper algorithm to partition point clouds into overlapping regions, constructing a region graph, and performing classification with a Graph Isomorphism Network, achieving competitive and stable accuracy with only 0.5M parameters.
108. LoGoSeg: Integrating Local and Global Features for Open-Vocabulary Semantic Segmentation
Core Problem: Existing open-vocabulary semantic segmentation (OVSS) methods, despite leveraging vision-language models, often suffer from imprecise spatial alignment, object hallucination, or missed detections due to reliance on image-level pretraining and lack of strong object priors.
Key Innovation: LoGoSeg, an efficient single-stage framework that integrates an object existence prior, a region-aware alignment module, and a dual-stream fusion mechanism to achieve precise region-level visual-textual correspondences and reduce hallucinations in OVSS without external mask proposals or additional backbones.
109. CAViT -- Channel-Aware Vision Transformer for Dynamic Feature Fusion
Core Problem: Vision Transformers (ViTs) rely on static multilayer perceptrons (MLPs) for channel-wise mixing, which limits their adaptability to input content and representational expressiveness across diverse computer vision tasks.
Key Innovation: CAViT, a dual-attention architecture that replaces the static MLP with a dynamic, attention-based mechanism for channel-wise feature interaction, allowing the model to dynamically recalibrate feature representations based on global image context and enhancing representational expressiveness.
110. End-to-End Compression for Tabular Foundation Models
Core Problem: In-context learning tabular foundation models, despite their performance, suffer from quadratic complexity due to their transformer architecture, leading to high overhead in training and inference time and limiting their capacity for large-scale datasets.
Key Innovation: Proposes TACO, an end-to-end tabular compression model that compresses the training dataset into a latent space. This significantly reduces inference time (up to 94x faster) and memory consumption (up to 97% less) compared to state-of-the-art tabular transformers, while maintaining performance and scaling better with increased dataset sizes.
111. Probabilistic Multi-Regional Solar Power Forecasting with Any-Quantile Recurrent Neural Networks
Core Problem: Increasing PV generation introduces significant uncertainty into power system operation, requiring probabilistic forecasting beyond deterministic point predictions, especially for multi-regional systems.
Key Innovation: Proposes an Any-Quantile Recurrent Neural Network (AQ-RNN) framework for multi-regional PV power forecasting that provides calibrated conditional quantiles at arbitrary probability levels, effectively exploiting spatial dependencies and outperforming baselines in accuracy, calibration, and prediction interval quality.
112. Large-scale Score-based Variational Posterior Inference for Bayesian Deep Neural Networks
Core Problem: The inadequacy of most existing score-based variational inference methods for large-scale Bayesian Deep Neural Networks (BNNs) due to computational and technical reasons.
Key Innovation: A novel scalable score-based variational inference method for BNNs that combines score matching loss and a proximal penalty term, avoids reparametrized sampling, and allows for noisy unbiased mini-batch scores, making it scalable to large-scale networks.
113. Contour Refinement using Discrete Diffusion in Low Data Regime
Core Problem: Robust boundary detection of irregular and translucent objects is challenging, especially in low data regimes and with limited computational resources, hindering applications in areas like environmental monitoring.
Key Innovation: A lightweight discrete diffusion contour refinement pipeline using a CNN with self-attention, conditioned on a segmentation mask, iteratively denoising a sparse contour representation. It includes novel adaptations for low-data efficacy and inference efficiency, outperforming baselines on medical imaging and a custom wildfire dataset.
114. Rule-Based Spatial Mixture-of-Experts U-Net for Explainable Edge Detection
Core Problem: Deep learning models for edge detection, despite high performance, lack transparency ("black box" nature), hindering their use in safety-critical applications where verification is mandatory.
Key Innovation: Proposes the Rule-Based Spatial Mixture-of-Experts U-Net (sMoE U-Net) with Spatially-Adaptive Mixture-of-Experts blocks and a Takagi-Sugeno-Kang (TSK) Fuzzy Head, achieving competitive performance while providing pixel-level explainability through "Rule Firing Maps" and "Strategy Maps."
115. VLN-Pilot: Large Vision-Language Model as an Autonomous Indoor Drone Operator
Core Problem: Traditional drone navigation relies on rule-based or geometric path-planning, lacking language-driven semantic understanding and context-awareness for complex, human-friendly control in GPS-denied indoor environments.
Key Innovation: Introduces VLN-Pilot, a framework where a large Vision-and-Language Model (VLLM) acts as an autonomous indoor drone pilot, interpreting natural language instructions and visual observations to plan and execute context-aware flight behaviors, enabling high success rates in complex navigation tasks.
116. UAV Trajectory Optimization via Improved Noisy Deep Q-Network
Core Problem: Standard deep reinforcement learning methods for Unmanned Aerial Vehicle (UAV) trajectory optimization often suffer from limited exploration ability and training instability in simulated environments.
Key Innovation: Proposes an Improved Noisy Deep Q-Network (Noisy DQN) that enhances exploration through residual NoisyLinear layers with adaptive noise scheduling and improves training stability via smooth loss and soft target network updates, achieving faster convergence and higher rewards for UAV trajectory optimization.
117. Self-Portrait of the Focusing Process in Speckle: III. Tailoring Complex Spatio-Temporal Focusing Laws To Overcome Reverberations in Reflection Imaging
Core Problem: Multiple scattering (reverberations) in complex media fundamentally limits the depth and resolution of reflection imaging.
Key Innovation: Extends the distortion matrix concept to the frequency domain and uses an iterative phase reversal process in space-time Fourier space to compensate for reverberations, thereby optimizing axial and transverse resolution in confocal imaging through complex media.
118. PhysicsAgentABM: Physics-Guided Generative Agent-Based Modeling
Core Problem: LLM-based multi-agent systems are expensive and poorly calibrated for timestep-aligned state-transition simulation, while classical ABMs struggle with rich individual-level signals and non-stationary behaviors.
Key Innovation: PhysicsAgentABM, which shifts inference to behaviorally coherent agent clusters, uses state-specialized symbolic agents for mechanistic priors, a multimodal neural transition model, and uncertainty-aware epistemic fusion for calibrated cluster-level transition distributions, enabling scalable and calibrated simulation with LLMs.
119. A Differential and Pointwise Control Approach to Reinforcement Learning
Core Problem: Reinforcement learning in continuous state-action spaces faces challenges like poor sample efficiency and lack of pathwise physical consistency, especially in scientific computing.
Key Innovation: Differential Reinforcement Learning (Differential RL), a novel framework that reformulates RL via a differential dual formulation, inducing a Hamiltonian structure that embeds physics priors and ensures consistent trajectories. It also introduces Differential Policy Optimization (dfPO), a pointwise, stage-wise algorithm for improved sample efficiency and dynamic alignment.
120. From Link Prediction to Forecasting: Addressing Challenges in Batch-based Temporal Graph Learning
Core Problem: Traditional batch-oriented evaluation for dynamic link prediction in temporal graphs leads to information loss/leakage and inconsistent tasks due to fixed-sized batches, skewing model performance.
Key Innovation: Reformulating dynamic link prediction as a link forecasting task that better accounts for temporal information, mitigating issues of batch-based evaluation and enabling fairer comparison of methods.
121. MVGS: Multi-view Regulated Gaussian Splatting for Novel View Synthesis
Core Problem: Traditional 3D Gaussian Splatting (3DGS) methods, relying on single-view supervision, often overfit specific training views, leading to suboptimal novel-view synthesis and imprecise 3D geometries.
Key Innovation: A multi-view regulated 3DGS optimization method (MVGS) that incorporates a multi-view training strategy, cross-intrinsic guidance, cross-ray densification, and multi-view augmented densification to improve overall accuracy, prevent overfitting, and enhance 3D reconstruction quality.
122. Efficient Scene Modeling via Structure-Aware and Region-Prioritized 3D Gaussians
Core Problem: Existing 3D Gaussian Splatting (3DGS) methods primarily rely on photometric supervision, resulting in irregular spatial distribution of Gaussians and inefficient optimization that overlooks underlying geometric context, impacting fidelity and efficiency.
Key Innovation: Introduction of Mini-Splatting2, an efficient scene modeling framework that couples structure-aware distribution (enforcing spatial regularity and sparsity) and region-prioritized optimization (improving training discrimination through geometric saliency), leading to more compact, faster-converging, and high-fidelity 3D Gaussian models.
123. Physics-Driven Local-Whole Elastic Deformation Modeling for Point Cloud Representation Learning
Core Problem: Existing data-driven point cloud representation learning methods primarily focus on spatial distribution, overlooking the crucial relationship between local features and the whole structure, which limits accuracy and interpretability, especially for objects undergoing deformation.
Key Innovation: A dual-task encoder-decoder framework that integrates a physics-driven elastic deformation mechanism with data-driven implicit fields to learn fine-grained features and model the structural relationship between local regions and the whole shape, guided by physics-based loss functions.
124. Guided Diffusion Sampling on Function Spaces with Applications to PDEs
Core Problem: Recovering complete solutions in PDE-based inverse problems from extremely sparse or noisy measurements, and accurately capturing posterior distributions in function spaces under minimal supervision and severe data scarcity.
Key Innovation: A general framework (FunDPS) for conditional sampling in PDE-based inverse problems using a function-space diffusion model and plug-and-play guidance, which trains a discretization-agnostic denoising model with neural operators and extends Tweedie's formula to infinite-dimensional Banach spaces.
125. A Contrastive Learning Foundation Model Based on Perfectly Aligned Sample Pairs for Remote Sensing Images
Core Problem: Existing Contrastive Learning (CL) methods for Self-Supervised Learning (SSL) in computer vision require specific adaptation for Remote Sensing (RS) images due to the significant domain gap, leading to potential semantic inconsistency and lower efficiency.
Key Innovation: "PerA," a novel self-supervised contrastive learning foundation model for RS images that generates semantically "Perfectly Aligned" sample pairs using spatially disjoint masks, leading to high-quality, memory-efficient, and adaptable RS features for various downstream tasks.
126. Plug-and-play linear attention with provable guarantees for training-free image restoration
Core Problem: The quadratic complexity of Multi-head self-attention (MHSA) in vision Transformers creates a bottleneck for real-time and resource-constrained image restoration deployment.
Key Innovation: PnP-Nystra, a training-free Nyström-based linear attention module, which acts as a plug-and-play replacement for MHSA in pretrained image restoration Transformers, offering significant speedups with minimal quality drop and provable kernel approximation error guarantees.
127. Multi-Agent Inverted Transformer for Flight Trajectory Prediction
Core Problem: Predicting multi-agent flight trajectories is challenging due to the need to model both individual aircraft behaviors over time and complex interactions between flights, while also generating explainable prediction outcomes.
Key Innovation: Proposal of MAIFormer, a Multi-Agent Inverted Transformer, featuring masked multivariate attention for individual aircraft spatio-temporal patterns and agent attention for social patterns among multiple agents, achieving superior performance and interpretability in real-world flight trajectory prediction.
128. RefAM: Attention Magnets for Zero-Shot Referral Segmentation
Core Problem: Achieving strong performance in referring segmentation without fine-tuning or complex architectural modifications, by effectively leveraging features from large-scale generative diffusion models.
Key Innovation: RefAM, a training-free grounding framework that exploits attention scores from diffusion transformers, using insights like filtering stop words as 'attention magnets' and handling 'global attention sinks' (GAS) to produce sharper and more accurate grounding maps for zero-shot referring image and video segmentation.
129. An Attention-based Feature Memory Design for Energy-Efficient Continual Learning
Core Problem: Mitigating catastrophic forgetting in continual learning for tabular data streams on resource-constrained edge devices, while also ensuring energy and memory efficiency.
Key Innovation: AttenMLP, an attention-based feature memory design that integrates feature replay, context retrieval, and sliding buffer updates within a minibatch training framework, achieving comparable accuracy to state-of-the-art models while substantially reducing energy consumption for continual learning on tabular data streams.
130. Softly Constrained Denoisers for Diffusion Models
Core Problem: Diffusion models struggle to produce samples that respect constraints, and existing regularization or guidance methods bias the generative model away from the true data distribution, especially when constraints are misspecified in scientific applications.
Key Innovation: Proposes integrating guidance-inspired adjustments directly into the denoiser, achieving a soft inductive bias towards constraint-compliant samples that improves compliance while maintaining flexibility to deviate from misspecified constraints.
131. See Less, See Right: Bi-directional Perceptual Shaping For Multimodal Reasoning
Core Problem: Large vision-language models (VLMs) often overlook fine-grained visual evidence, generalize poorly across domains, and incur high inference-time costs for multimodal reasoning.
Key Innovation: Proposes Bi-directional Perceptual Shaping (BiPS), which transforms question-conditioned masked views into bidirectional where-to-look signals to shape perception during training, improving VLM performance and out-of-domain generalization.
132. Active Perception Agent for Omnimodal Audio-Video Understanding
Core Problem: Omnimodal large language models face challenges in fine-grained cross-modal understanding and multimodal alignment, often relying on rigid workflows and passive response generation.
Key Innovation: Introduces OmniAgent, the first fully active perception agent that dynamically orchestrates specialized unimodal tools and employs a coarse-to-fine audio-guided perception paradigm to achieve more fine-grained omnimodal reasoning and state-of-the-art performance in audio-video understanding.
133. FUSE-Flow: Scalable Real-Time Multi-View Point Cloud Reconstruction Using Confidence
Core Problem: Fusing large-scale multi-view depth observations into high-quality point clouds under strict real-time constraints remains challenging due to computational complexity, memory usage, and limited scalability of existing methods.
Key Innovation: Proposes FUSE-Flow, a frame-wise, stateless, and linearly scalable point cloud streaming reconstruction framework that generates and fuses point cloud fragments using measurement confidence and 3D distance consistency, and introduces an adaptive spatial hashing-based weighted aggregation method for large-scale multi-camera efficiency, achieving high-throughput, low-latency, and improved geometric fidelity.
134. ARM: Refining Multivariate Forecasting with Adaptive Temporal-Contextual Learning
Core Problem: Long-term time series forecasting (LTSF) for multivariate inputs is challenged by the inefficiency of existing Transformer models to correctly capture complex temporal-contextual and series-wise relationships, often underperforming univariate counterparts.
Key Innovation: ARM, an enhanced multivariate temporal-contextual adaptive learning architecture for LTSF, which employs Adaptive Univariate Effect Learning (AUEL), Random Dropping (RD), and Multi-kernel Local Smoothing (MKLS) to better handle individual series patterns and inter-series dependencies, achieving superior performance on benchmarks.
135. CATS: Enhancing Multivariate Time Series Forecasting by Constructing Auxiliary Time Series as Exogenous Variables
Core Problem: Multivariate Time Series Forecasting (MTSF) models often underperform univariate ones due to their deficiency in effectively representing and incorporating inter-series relationships.
Key Innovation: CATS, a method that constructs Auxiliary Time Series (ATS) from Original Time Series (OTS) to function as a 2D temporal-contextual attention mechanism, effectively representing and incorporating inter-series relationships for forecasting, achieving state-of-the-art results with reduced complexity and parameters.
136. YOLO-based Bearing Fault Diagnosis With Continuous Wavelet Transform
Core Problem: Traditional bearing fault diagnosis methods may lack locality awareness and spatially interpretable decision-making, and global classification approaches might not fully capture transient fault signatures.
Key Innovation: A locality-aware framework using Continuous Wavelet Transform (CWT) to convert 1D vibration signals into 2D time-frequency spectrograms, then applying YOLO (v9, v10, v11) for object detection to localize and classify fault types, providing region-aware visualization and strong cross-dataset generalization.
137. Mechanical behavior of marine coral sand - coral clay mixtures reinforced with bionic honeycomb polymer grid: Experimental and artificial intelligence methods
Core Problem: Accurately predicting the strength of marine coral sand-coral clay mixtures, especially when reinforced with bionic honeycomb polymer grids, to ensure the stability of reclaimed island foundations and offshore infrastructure.
Key Innovation: Development and validation of a CNN-LSTM model for accurate strength prediction of reinforced marine composite foundation soil, identifying key influencing factors using SHAP, deriving an empirical formula for rapid engineering use, and creating a user-friendly GUI for practitioners.
138. An efficient SPH-FEM model for three-dimensional fluid-structure interactions via direct particle-surface mesh coupling
Core Problem: Precise simulation of three-dimensional fluid-structure interaction (FSI) phenomena is computationally challenging, especially regarding efficient data exchange and handling complex geometries in SPH-FEM coupling.
Key Innovation: A bidirectional SPH-FEM coupling strategy based on a single-layer particle boundary technique that constructs a direct mapping between surface meshes and SPH particles, eliminating additional boundary treatments and significantly improving data exchange efficiency for 3D FSI problems with complex geometries.
139. LightKD-SAR: Lightweight Architecture With Knowledge Distillation for High-Performance SAR Object Detection
Core Problem: Conventional SAR object detection methods require high computational and memory costs, limiting deployment in resource-constrained environments, while existing lightweight detectors struggle with SAR imagery challenges like sparse objects, speckle noise, and multiscale variations.
Key Innovation: Proposal of LightKD-SAR, a lightweight SAR object detection framework combining an efficient network architecture (customized inverted residual modules, optimized feature extraction/fusion) with enhanced instance selection based knowledge distillation, achieving a superior tradeoff between detection accuracy and computational efficiency.
140. Joint Shifted Attention and Cross-Guided Feature Fusion for Remote Sensing Scene Classification
Core Problem: High-resolution remote sensing images present significant challenges for scene classification due to high inter-class similarity and intra-class diversity.
Key Innovation: Proposed SACGNet, a novel network with a shift self-attention module and a cross-guided feature compensation module, along with a structure-aware learning rate scheduling strategy (SGLR), achieving superior classification performance on remote sensing scene classification datasets.
141. Mamba-Enhanced Background Suppression Diffusion Model for Hyperspectral Anomaly Detection
Core Problem: Hyperspectral anomaly detection faces a significant challenge in separating scarce, small, and subtle anomalous targets from complex backgrounds, with generative models for background suppression remaining underexplored.
Key Innovation: Proposed the Mamba-enhanced background suppression diffusion model (MBSDM), which integrates Mamba into a diffusion-based framework to capture complex background features and guide their suppression, thereby highlighting anomalous targets, and introduced a new loss function to mitigate training bias.
142. A 1 km soil organic carbon density dataset with depth of 20 and 100 cm from 1985 to 2020 in China
Core Problem: Need for a high-resolution, spatiotemporal dataset of Soil Organic Carbon Density (SOCD) in China to trace its dynamics and understand its role in the global carbon cycle and ecosystem health.
Key Innovation: Produced a high-resolution (1 km) SOCD dataset for China (1985-2020) at two depths (0-20 and 0-100 cm) by integrating diverse data sources with a Random Forest approach and a climate zoning strategy, demonstrating strong agreement with independent validations.
143. LEX v1.6.0: a new large-eddy simulation model in JAX with GPU acceleration and automatic differentiation
Core Problem: Current numerical weather models operating in the "gray zone" (kilometer-scale resolution) struggle to accurately represent subgrid-scale (SGS) turbulence, which significantly influences simulated weather systems.
Key Innovation: Developed LEX v1.6.0, a new large-eddy simulation (LES) model in JAX with GPU acceleration and automatic differentiation, enabling the training of deep learning-based SGS turbulence models (e.g., using an autoencoder) to accurately simulate atmospheric phenomena at gray-zone resolutions comparable to benchmark LES.
144. Machine learning slashes the testing needed to work out battery lifetimes
Core Problem: The time-consuming bottleneck in battery development related to testing and determining battery lifetimes.
Key Innovation: Development of a machine-learning method that reasons and adapts, significantly reducing the testing needed to determine battery lifetimes.
145. Evolving glacier landsystems during Patagonian Ice Sheet recession influenced by a changing topographic setting
Core Problem: Investigating the geomorphic signature and evolving glacier landsystems of the Río Cisnes palaeo-outlet glacier in Patagonia during its recession over a reverse-bed slope.
Key Innovation: Interpreted landforms within a glacial landsystems framework using process-form regimes from modern analogues, developed a conceptual model demonstrating temporary and permanent changes in glacial landsystems (active temperate outlet lobe, temperate glaciolacustrine-calving outlet glacier, temperate cirque and valley glacier landsystems), and highlighted the influence of complex regional topography and ice-dammed lake development on landsystem evolution.
146. Variability and probabilistic modeling of external blast loads on cylindrical shells
Core Problem: Deterministic load models in current blast-resistant design codes fail to quantify the significant variability of external blast loads on cylindrical shells, hindering reliability-based blast design.
Key Innovation: Systematically investigates and quantifies the variability of external blast loads, decoupling repeatability and charge orientation errors. Develops spatial distribution models for the coefficient of variation and proposes a probabilistic load model guided by the 95% fractile value, directly integrable with structural reliability design codes for probabilistic blast-resistant design.
147. L2M-Reg: Building-level uncertainty-aware registration of outdoor LiDAR point clouds and semantic 3D city models
Core Problem: Achieving accurate building-level registration between outdoor LiDAR point clouds and semantic 3D city models is challenging, primarily due to the generalization uncertainty inherent in Level of Detail 2 (LoD2) city models.
Key Innovation: L2M-Reg, a plane-based fine registration method, is proposed that explicitly accounts for model uncertainty. It involves establishing reliable plane correspondence, building a pseudo-plane-constrained Gauss–Helmert model, and adaptively estimating vertical translation, leading to more accurate and computationally efficient LiDAR-to-Model registration at the building level.
148. Integrated evaluation of snow density reanalysis products in the Northern Hemisphere
Core Problem: Systematic evaluations of snow density in widely used reanalysis datasets are lacking, hindering their applicability in climate modeling, hydrological studies, and accurate representation of snow cover physical characteristics.
Key Innovation: Conducted a comprehensive evaluation of five reanalysis datasets using 4,319 snow stations, identifying ERA5-Land and GLDAS-Noah as best performers, and revealing that reanalysis products fail to reproduce observed interannual trends in snow density and exhibit region- and dataset-specific impacts of snow density biases on snow depth biases.
149. A grain-based FDEM with finite-width microcracks for modeling granite deformation and strength
Core Problem: Existing numerical approaches struggle to simultaneously capture finite-width microcrack closure, mineralogical heterogeneity, and realistic compressive-to-tensile strength ratios in crystalline rocks, limiting accurate modeling of their macroscopic mechanical behavior.
Key Innovation: Development of a 3D grain-based FDEM model that explicitly represents polycrystalline mineral structure and embeds finite-width microcracks, enabling accurate reproduction of nonlinear compaction, strength characteristics, temperature-induced strength degradation, and stress-controlled hydraulic fracture propagation in granite.
150. Influence of hydro-mechanical parameter coupling on the dynamic behavior of unsaturated railway subgrades
Core Problem: Conventional analyses of unsaturated railway subgrades often use inconsistent hydro-mechanical parameters, leading to unreliable predictions of dynamic response under train loads.
Key Innovation: Performed joint laboratory measurements of SWCC and shear modulus for railway subgrade materials, integrated these matched parameters into a 2.5D FEM, and demonstrated that accurate coupling is essential for reliable prediction of dynamic behavior, revealing a nonlinear shear modulus-saturation relationship and its impact on vibration.
151. Quantifying Regime Transition of Mineral Precipitation in a Microfluidic Fracture
Core Problem: Understanding the patterns and transitions of mineral precipitation dynamics in fractured media, which influences subsurface fluid flow and processes like karst evolution.
Key Innovation: Developed an experimental system to observe CaCO3 precipitation in microfluidic fractures, revealing two distinct patterns (bands vs. clusters) governed by fluid shear forces and particle repulsive forces, and demonstrating their impact on permeability.
152. Interdecadal Increase in Spring‐To‐Summer Persistence of Central American Precipitation Anomalies
Core Problem: Most studies evaluate Central American spring and summer precipitation in isolation, overlooking interdecadal shifts in cross-seasonal persistence, which impacts agriculture and ecosystems.
Key Innovation: Demonstrated an interdecadal increase in spring-to-summer persistence of Central American precipitation anomalies since the 2000s, attributing it to a regime change in ocean-atmosphere coupling (warm tropical Atlantic and cold tropical Central-Eastern Pacific SST anomalies).
153. The Current State of Undergraduate Hydrology Courses in North America: A Path Forward
Core Problem: The need to enhance workforce development, research, and training in hydrology to advance the future of hydrologic science.
Key Innovation: An assessment of the current state of undergraduate hydrology courses in North America and proposing a path forward to improve education and training in the field (based on title and implied content).
154. Quantile-Physics Hybrid Framework for Safe-Speed Recommendation under Diverse Weather Conditions Leveraging Connected Vehicle and Road Weather Information Systems Data
Core Problem: Recommending real-time safe driving speed intervals on freeways under diverse and inclement weather conditions to reduce crash risk, considering the impact on driver visibility and tire-road surface friction.
Key Innovation: A hybrid predictive framework that combines Quantile Regression Forests (QRF) to estimate vehicle speed distributions with a physics-based upper speed limit (derived from real-time road grip and visibility) to recommend safe speed intervals, leveraging high-resolution Connected Vehicle and Road Weather Information System data.
155. Cross-talk based multi-task learning for fault classification of physically coupled machine system
Core Problem: Fault classification in machine systems is challenging because fault conditions are physically coupled with other variables, and existing methods often rely solely on direct fault labels, ignoring additional embedded information.
Key Innovation: A multi-task learning (MTL) framework with a cross-talk structure that jointly learns fault conditions and related physical variables, allowing controlled information exchange and outperforming single-task and shared-trunk models on drone and motor fault datasets.
156. LOBSTgER-enhance: an underwater image enhancement pipeline
Core Problem: Underwater photography suffers from significant inherent challenges including reduced contrast, spatial blur, and wavelength-dependent color distortions, requiring heavy post-processing.
Key Innovation: LOBSTgER-enhance, an image-to-image pipeline that learns to reverse underwater degradations by introducing a synthetic corruption pipeline and using diffusion-based generation, achieving high perceptual consistency and generalization.
157. SpectraKAN: Conditioning Spectral Operators
Core Problem: Existing spectral neural operators rely on static Fourier kernels, limiting their ability to capture multi-scale, regime-dependent, and anisotropic dynamics governed by the global state of the system.
Key Innovation: SpectraKAN, a neural operator that conditions the spectral operator on the input itself, modulating a multi-scale Fourier trunk via single-query cross-attention, enabling adaptive behavior and achieving state-of-the-art performance on PDE benchmarks.
158. Cross-Domain Few-Shot Segmentation via Multi-view Progressive Adaptation
Core Problem: Existing Cross-Domain Few-Shot Segmentation (CD-FSS) methods exhibit constrained performance due to limited target samples and substantial domain gaps, hindering effective adaptation from source to target domains.
Key Innovation: Multi-view Progressive Adaptation (MPA), which progressively adapts few-shot capability to target domains from both data (Hybrid Progressive Augmentation) and strategy (Dual-chain Multi-view Prediction) perspectives, achieving robust and accurate adaptation by enforcing prediction consistency across diverse and complex views.
159. Balanced Anomaly-guided Ego-graph Diffusion Model for Inductive Graph Anomaly Detection
Core Problem: Graph anomaly detection (GAD) methods struggle with dynamic, evolving networks due to transductive learning paradigms and suffer from biased models caused by extreme class imbalance of anomalous nodes.
Key Innovation: Proposes a data-centric framework for inductive GAD that integrates a discrete ego-graph diffusion model to capture local anomaly topology and a curriculum anomaly augmentation mechanism to dynamically adjust synthetic data generation, improving detection and generalization.
160. ReGLA: Efficient Receptive-Field Modeling with Gated Linear Attention Network
Core Problem: Balancing accuracy and latency in lightweight models, particularly Transformer-based architectures, for high-resolution image processing tasks.
Key Innovation: Introduction of ReGLA, a series of lightweight hybrid networks integrating efficient convolutions and ReLU-based gated linear attention, featuring an Efficient Large Receptive Field (ELRF) module, ReLU Gated Modulated Attention (RGMA) module, and a multi-teacher distillation strategy to achieve state-of-the-art performance with high efficiency on various vision benchmarks.
161. HealthMamba: An Uncertainty-aware Spatiotemporal Graph State Space Model for Effective and Reliable Healthcare Facility Visit Prediction
Core Problem: Existing healthcare facility visit prediction methods often formulate the task as a simple time-series forecasting problem, neglecting intrinsic spatial dependencies between different facility types and failing to provide reliable predictions under abnormal situations.
Key Innovation: HealthMamba, an uncertainty-aware spatiotemporal framework comprising a Unified Spatiotemporal Context Encoder, a novel Graph State Space Model (GraphMamba) for hierarchical spatiotemporal modeling, and a comprehensive uncertainty quantification module, achieving around 6.0% improvement in prediction accuracy and 3.5% in uncertainty quantification over state-of-the-art baselines.
162. Fast-SAM3D: 3Dfy Anything in Images but Faster
Core Problem: The prohibitive inference latency of SAM3D, hindering its deployment for scalable, open-world 3D reconstruction, due to generic acceleration strategies failing to account for the pipeline's inherent multi-level heterogeneity (kinematic distinctiveness, texture sparsity, spectral variance).
Key Innovation: Fast-SAM3D, a training-free framework that dynamically aligns computation with instantaneous generation complexity through three heterogeneity-aware mechanisms: Modality-Aware Step Caching, Joint Spatiotemporal Token Carving, and Spectral-Aware Token Aggregation, delivering up to 2.67x end-to-end speedup with negligible fidelity loss.
163. Accelerated Sequential Flow Matching: A Bayesian Filtering Perspective
Core Problem: Sequential prediction from streaming observations using diffusion and flow-matching models incurs substantial inference latency due to repeated sampling from a non-informative initial distribution.
Key Innovation: Introduces Sequential Flow Matching, a Bayesian filtering framework that treats streaming inference as learning a probability flow. By initializing generation from the previous posterior, it provides a principled warm start that significantly accelerates sampling while maintaining competitive performance.
164. Bayesian Neighborhood Adaptation for Graph Neural Networks
Core Problem: Determining the optimal neighborhood scope (number of hops) for information aggregation in Graph Neural Networks (GNNs) is critical but existing two-stage approaches are time-consuming, biased, and do not adaptively handle both homophilic and heterophilic graphs.
Key Innovation: Proposes a Bayesian framework that models GNN message-passing as a stochastic process, treating the number of hops as a beta process. This allows for adaptive inference of the most plausible neighborhood scope simultaneously with GNN parameter optimization, improving expressivity and achieving competitive performance.
165. A Decomposition-based State Space Model for Multivariate Time-Series Forecasting
Core Problem: Multivariate time series (MTS) forecasting is challenging because real-world sequences intertwine slow trends, multi-rate seasonalities, and irregular residuals, and existing methods often rely on rigid, hand-crafted decompositions or generic end-to-end architectures that entangle components.
Key Innovation: Proposes DecompSSM, an end-to-end decomposition framework using three parallel deep state space model branches to capture trend, seasonal, and residual components, featuring adaptive temporal scales via an input-dependent predictor, a refinement module for shared cross-variable context, and an auxiliary loss that enforces reconstruction and orthogonality.
166. TSBOW: Traffic Surveillance Benchmark for Occluded Vehicles Under Various Weather Conditions
Core Problem: Existing traffic surveillance datasets for vehicle detection lack comprehensive coverage of extreme weather conditions, which are intensifying due to global warming, degrading CCTV signal quality, disrupting traffic flow, and increasing accident rates.
Key Innovation: Introduced TSBOW, a comprehensive dataset comprising over 32 hours of real-world traffic data with 48,000+ manually annotated and 3.2 million semi-labeled frames, specifically designed to enhance occluded vehicle detection across diverse and extreme annual weather scenarios. This resource establishes a benchmark for advancing Intelligent Transportation Systems.
167. Visual Implicit Geometry Transformer for Autonomous Driving
Core Problem: Autonomous driving systems require scalable, architecturally simple, and generalizable geometric models to estimate continuous 3D occupancy fields from diverse surround-view camera configurations, ideally without costly manual annotations.
Key Innovation: Introduces ViGT, a calibration-free Visual Implicit Geometry Transformer that estimates continuous 3D occupancy fields in a birds-eye-view (BEV) from multiple camera views. It employs a self-supervised training procedure leveraging synchronized image-LiDAR pairs, achieving state-of-the-art performance on pointmap estimation across five large-scale autonomous driving datasets.
168. Geometric Observability Index: An Operator-Theoretic Framework for Per-Feature Sensitivity, Weak Observability, and Dynamic Effects in SE(3) Pose Estimation
Core Problem: Classical sensitivity tools for camera pose estimation (e.g., SLAM, structure-from-motion) do not explain how individual image features influence the pose estimate or why dynamic/inconsistent observations disproportionately distort these systems.
Key Innovation: The Geometric Observability Index (GOI), a unified operator-theoretic framework extending influence function theory to matrix Lie groups, which quantifies the contribution of a single measurement and reveals the correspondence between weak observability and amplified sensitivity in SE(3) pose estimation.
169. Multi-instance robust fitting for non-classical geometric models
Core Problem: Most existing robust fitting methods are designed for classical geometric models and primarily focus on reconstructing a single instance, making them inadequate for robustly handling and reconstructing multiple instances of non-classical geometric models from noisy data.
Key Innovation: A novel multi-instance robust fitting method for non-classical geometric models, formulated as an optimization problem with a model-to-data error-based estimator (capable of handling outliers without a predefined threshold) and a meta-heuristic optimizer to seek the global optimum.
170. Almost Asymptotically Optimal Active Clustering Through Pairwise Observations
Core Problem: Efficiently clustering items into an unknown number of groups using noisy and actively collected pairwise observations, while providing theoretical guarantees on the number of queries needed.
Key Innovation: Establishes a fundamental lower bound on the expected number of queries for active clustering and designs an asymptotically optimal algorithm that leverages a Generalized Likelihood Ratio (GLR) statistic, demonstrating near-optimal performance in practice.
171. Poster: Camera Tampering Detection for Outdoor IoT Systems
Core Problem: Outdoor IoT camera systems are susceptible to tampering (vandalism or environmental conditions), undermining their monitoring effectiveness, especially when capturing still images without continuous frames.
Key Innovation: Proposing two approaches (rule-based and deep-learning-based) for detecting tampered images in outdoor IoT systems, evaluating their performance, and providing publicly available datasets to support method development.
172. Self-Supervised Learning with a Multi-Task Latent Space Objective
Core Problem: Instability in predictor-based self-supervised learning architectures when using multi-crop strategies, particularly with small local crops, hindering performance gains.
Key Innovation: Assigning a separate predictor to each view type to stabilize multi-crop training, and extending this to a multi-task formulation of asymmetric Siamese SSL that combines global, local, and masked views, leading to improved visual representation learning.
173. Tuning Out-of-Distribution (OOD) Detectors Without Given OOD Data
Core Problem: Existing out-of-distribution (OOD) detectors are often tuned using a separate, ad-hoc OOD dataset that may not be available, representative, or can introduce significant variance in performance.
Key Innovation: Introduces and formalizes the problem of tuning OOD detectors without a given OOD dataset, proposing a new generic approach that requires no extra data beyond that used to train the neural network, consistently improving performance across higher-parameter OOD detector families.
174. Dimensionality Reduction on Riemannian Manifolds in Data Analysis
Core Problem: Traditional Euclidean dimensionality reduction methods may not faithfully represent data that inherently lies on curved Riemannian manifolds, leading to suboptimal embeddings and analysis.
Key Innovation: Investigates and extends Riemannian geometry-based dimensionality reduction methods, including Principal Geodesis Analysis (PGA) and discriminant analysis, which exploit geodesic distances and tangent space representations to achieve more faithful low-dimensional embeddings and improved classification performance for manifold-valued data.
175. Breaking Symmetry Bottlenecks in GNN Readouts
Core Problem: Graph Neural Networks (GNNs) are fundamentally limited in distinguishing non-isomorphic graphs due to a bottleneck at the readout stage, where linear permutation-invariant readouts erase non-trivial symmetry-aware components of node embeddings.
Key Innovation: Introduces projector-based invariant readouts that decompose node representations into symmetry-aware channels and summarize them with nonlinear invariant statistics, provably overcoming the expressivity barrier of traditional readouts and significantly improving GNN performance across multiple benchmarks.
176. MambaVF: State Space Model for Efficient Video Fusion
Core Problem: Existing video fusion methods are computationally expensive and have limited scalability due to heavy reliance on optical flow estimation and feature warping.
Key Innovation: MambaVF, an efficient video fusion framework based on state space models (SSMs), reformulates fusion as a sequential state update process, capturing long-range temporal dependencies with linear complexity and significantly reducing computation and memory costs, achieving state-of-the-art performance across various video fusion tasks.
177. V-Retrver: Evidence-Driven Agentic Reasoning for Universal Multimodal Retrieval
Core Problem: Existing Multimodal Large Language Model (MLLM) approaches for universal multimodal retrieval are language-driven, rely on static visual encodings, and lack the ability to actively verify fine-grained visual evidence, leading to speculative reasoning.
Key Innovation: V-Retrver, an evidence-driven agentic reasoning framework that enables MLLMs to selectively acquire visual evidence via external visual tools, performing interleaved hypothesis generation and targeted visual verification, leading to improved retrieval accuracy and reasoning reliability.
178. SwimBird: Eliciting Switchable Reasoning Mode in Hybrid Autoregressive MLLMs
Core Problem: Most Multimodal Large Language Models (MLLMs) perform reasoning primarily with textual Chain-of-Thought, limiting effectiveness on vision-intensive tasks and lacking adaptive reasoning modes for different user queries.
Key Innovation: SwimBird, a reasoning-switchable MLLM that dynamically selects among text-only, vision-only, and interleaved vision-text reasoning modes, enabled by a hybrid autoregressive formulation and a curated supervised fine-tuning dataset, leading to improved performance on vision-dense tasks while preserving textual logic.
179. Large-Ensemble Simulations Reveal Links Between Atmospheric Blocking Frequency and Sea Surface Temperature Variability
Core Problem: Isolating the influence of sea surface temperature (SST) from chaotic internal atmospheric variability on atmospheric blocking events, which drive persistent weather extremes, remains a challenge.
Key Innovation: Uses century-long, large-ensemble simulations with deep-learning general circulation models to filter internal atmospheric noise, revealing robust teleconnections between Greenland blocking frequency and North Atlantic SST/El Niño patterns, and SST-forced trends in blocking frequency.
180. Interpreting Manifolds and Graph Neural Embeddings from Internet of Things Traffic Flows
Core Problem: The difficulty in interpreting complex and evolving IoT network topologies using traditional tools and the opaque nature of Graph Neural Network (GNN) internal representations, which hinders human understanding for security-critical operations.
Key Innovation: An interpretable pipeline that maps high-dimensional GNN embeddings of IoT traffic flows onto a latent manifold, generating directly visualizable low-dimensional representations for monitoring evolving network states, detecting intrusions, and highlighting phenomena like concept drift.
181. Wedge Sampling: Efficient Tensor Completion with Nearly-Linear Sample Complexity
Core Problem: Efficiently recovering low-rank tensors from a subset of entries, where uniform sampling often requires high sample complexity for effective initialization in polynomial-time algorithms.
Key Innovation: Wedge Sampling, a non-adaptive scheme that promotes structured length-two patterns, strengthening the spectral signal and enabling nearly-linear sample complexity for both weak and exact tensor recovery, improving over uniform sampling.
182. Image inpainting for corrupted images by using the semi-super resolution GAN
Core Problem: Deep learning models face a significant challenge in restoring corrupted images, especially when the extent of corruption is large, impacting image enhancement techniques.
Key Innovation: Introduction of a Generative Adversarial Network (GAN) and a distinct variant called Semi-Super Resolution GAN (SSRGAN) for learning and replicating missing pixels in corrupted images, demonstrating robustness across varying levels of pixel corruption.
183. Energy Guided smoothness to improve Robustness in Graph Classification
Core Problem: Graph Neural Networks (GNNs) often exhibit poor robustness to label noise in graph classification tasks, particularly on low-order graphs, with low label coverage, or when over-parameterized.
Key Innovation: Establishment of empirical and theoretical links between GNN robustness and the reduction of total Dirichlet Energy, and introduction of two training strategies (inductive bias in weight matrices via negative eigenvalue removal and a loss penalty for learned smoothness) to enhance GNN robustness without performance degradation in noise-free settings.
184. RAD: Region-Aware Diffusion Models for Image Inpainting
Core Problem: Existing diffusion models for image inpainting are often slow, require nested loops, or additional components for conditioning, limiting their efficiency and performance.
Key Innovation: Proposes Region-Aware Diffusion Models (RAD) with a pixel-specific noise schedule, enabling asynchronous local generation while considering global context, leading to significantly faster inference (up to 100 times) and state-of-the-art results in image inpainting.
185. DPMambaIR: All-in-One Image Restoration via Degradation-Aware Prompt State Space Model
Core Problem: Existing All-in-One image restoration approaches lack fine-grained modeling of degradation information and struggle to balance multi-task conflicts when addressing diverse image degradation problems with a single model.
Key Innovation: Proposes DPMambaIR, an All-in-One image restoration framework that introduces a fine-grained degradation extractor and a Degradation-Aware Prompt State Space Model (DP-SSM). DP-SSM leverages fine-grained degradation features as dynamic prompts, enhancing adaptability, complemented by a High-Frequency Enhancement Block (HEB), achieving state-of-the-art performance on mixed degradation datasets.
186. Feature Engineering is Not Dead: Reviving Classical Machine Learning with Entropy, HOG, and LBP Feature Fusion for Image Classification
Core Problem: The need for interpretable and computationally efficient image classification methods, as an alternative to complex deep learning models, particularly for tasks where classical machine learning can still provide competitive performance.
Key Innovation: A novel feature fusion approach for image classification combining multiscale, multi-orientation Permutation Entropy (PE) with Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP), creating a compact, interpretable, and effective hand-crafted feature set for SVM classifiers that achieves competitive performance on benchmark datasets.
187. Symplectic convolutional neural networks
Core Problem: Standard convolutional neural networks (CNNs) do not inherently preserve symplectic properties, which are crucial for accurately modeling physical systems governed by Hamiltonian dynamics.
Key Innovation: A new symplectic CNN architecture that leverages symplectic neural networks and proper symplectic decomposition to parameterize convolution layers and introduce a symplectic pooling layer, ensuring the network remains symplectic and improving performance on physical equations.
188. Progressive multi-fidelity learning with neural networks for physical system predictions
Core Problem: Acquiring accurate and sufficient high-fidelity data for physical system predictions is expensive and challenging, especially when data types are diverse and not concurrently available, leading to difficulties in building accurate surrogate models.
Key Innovation: A progressive multi-fidelity surrogate model that sequentially incorporates diverse data types using tailored encoders and a dual connection system (concatenations and additive connections) to exploit correlations and ensure additive corrections without performance degradation.
189. GMAC: Global Multi-View Constraint for Automatic Multi-Camera Extrinsic Calibration
Core Problem: Existing methods for automatic multi-camera extrinsic calibration lack robustness and applicability in complex dynamic environments or online scenarios, making them difficult to deploy.
Key Innovation: Proposes GMAC, a framework that models extrinsics as global variables constrained by latent multi-view geometric structure, prunes and reconfigures existing networks to support extrinsic prediction, and jointly optimizes cross-view reprojection and multi-view cycle consistency, achieving accurate and stable calibration without explicit 3D reconstruction or manual intervention.
190. SharpTimeGS: Sharp and Stable Dynamic Gaussian Splatting via Lifespan Modulation
Core Problem: Existing Gaussian-based methods for novel view synthesis of dynamic scenes struggle to maintain a balance between long-term static and short-term dynamic regions in both representation and optimization, leading to instability and compromised fidelity.
Key Innovation: SharpTimeGS, a lifespan-aware 4D Gaussian framework, introduces a learnable lifespan parameter that modulates temporal visibility and motion, effectively decoupling motion magnitude from temporal duration. This improves long-term stability, dynamic fidelity, and optimizes static/dynamic regions more effectively, achieving state-of-the-art real-time 4D reconstruction.
191. Root Cause Analysis of Outliers with Missing Structural Knowledge
Core Problem: Identifying the root cause of anomalies (changes in causal mechanisms) is challenging, especially when only few or a single sample from the post-intervention distribution is available, and structural causal knowledge (the causal graph) is missing.
Key Innovation: Simple, efficient methods for Root Cause Analysis (RCA) in polytrees with a single root cause, providing guarantees for a traversal algorithm when the causal graph is known, and causally justifying the heuristic of identifying root causes as variables with the highest marginal anomaly scores when the graph is unknown.
192. RANGER: A Monocular Zero-Shot Semantic Navigation Framework through Contextual Adaptation
Core Problem: Existing zero-shot object goal navigation methods for robots rely heavily on precise depth/pose information (limiting real-world applicability) and lack in-context learning (ICL) capability for rapid adaptation to new environments.
Key Innovation: Introduces RANGER, a zero-shot, open-vocabulary semantic navigation framework using only a monocular camera. It leverages 3D foundation models to eliminate depth/pose dependency and exhibits strong ICL capability by adapting from short videos, achieving competitive navigation performance and superior adaptability on HM3D and real-world environments.
193. Statsformer: Validated Ensemble Learning with LLM-Derived Semantic Priors
Core Problem: Existing approaches for integrating Large Language Model (LLM) knowledge into supervised statistical learning are limited by unvalidated heuristics, sensitivity to hallucination, and embedding semantic information within a single fixed learner.
Key Innovation: Introduces Statsformer, a guardrailed ensemble architecture that adaptively calibrates LLM-derived feature priors via cross-validation, mitigating hallucination and improving performance across diverse prediction tasks with an oracle-style guarantee.
194. The influence of initial structure on the evolution of permeability functions of a lean clay for a wide range of void ratios
Core Problem: The evolution of the permeability–void ratio function for clayey soils with different initial microstructures remains unclear, impacting geotechnical applications.
Key Innovation: Developed a low-compliance permeameter and conducted extensive tests on lean clay with distinct initial microstructures, revealing how initial structure significantly influences the permeability–void ratio curve and providing guidance for compaction conditions in applications like dams and landfill liners.
195. Response of Terrestrial Water Storage to Climate: Global Spatial Patterns and Driving Mechanism
Core Problem: Traditional statistical methods for understanding terrestrial water storage (TWS) response to climate often overlook time-lagged and nonlinear relationships, limiting comprehensive global spatial pattern analysis.
Key Innovation: Developed a climate zones-based explainable AI framework (LSTM with SHapley Additive exPlanations) to investigate the relative contribution of climate variables to TWS globally, revealing significant lagged responses and spatial heterogeneity in dominant drivers.
196. Image-based classification of stream stage to support ephemeral stream monitoring
Core Problem: Observing stage and calculating discharge in intermittent rivers and ephemeral streams (IRES) is technologically and methodologically challenging, despite their increasing prevalence due to climate change.
Key Innovation: Development of a low-cost, transferable image-based classification method using Logistic Regression to classify relative stream stage (no water, low, high) from field camera imagery, providing categorical flow states for ephemeral stream monitoring and quality control of continuous stage data.
197. Daily briefing: Tumours use neurons as hotline to the brain
Core Problem: The abstract covers multiple topics, including the need for new methods to study cities.
Key Innovation: Remote sensors are providing researchers with new ways to study cities.
198. Synthesizing scientific literature with retrieval-augmented language models
Core Problem: The challenge for researchers to efficiently and accurately synthesize the growing body of scientific literature.
Key Innovation: OpenScholar, a specialized open-source retrieval-augmented language model that answers scientific queries by identifying relevant passages and synthesizing citation-backed responses, demonstrating superior correctness and citation accuracy compared to other LLMs.
199. Satellite data can help transform food systems: Satellite data can be used to increase yields and improve sustainability in smallholder agricultural systems
Core Problem: The need to increase yields and improve sustainability in smallholder agricultural systems.
Key Innovation: Utilizing satellite data as a tool to enhance agricultural productivity and environmental sustainability in smallholder farming.
200. Geomorphological characterisation, pattern, and distribution of ice-margin positions of the former Scandinavian Ice Sheet
Core Problem: Systematically reconstructing former ice-margin positions of the Scandinavian Ice Sheet and understanding the glaciological processes and climatic responses that shaped them.
Key Innovation: Scrutinized high-resolution digital terrain models to map ~51,000 pieces of ice marginal evidence, categorizing them by dominant landform type, and investigated spatial patterns to suggest controls on landform formation (sediment cover, climate, marine/lake environment), providing seamless, internally-consistent maps and GIS data for ice sheet modeling.
201. Research on risk prevention and control of coal mine gas explosion using bayesian network and system dynamics: An optimization model for safety investment decision-making
Core Problem: Mitigating catastrophic coal mine gas explosion (CMGE) accidents and efficiently allocating safety-related resources requires an optimized framework for safety investment decision-making that considers complex causal factors.
Key Innovation: An optimization framework integrating System Dynamics (SD) with Bayesian Network (BN) to identify key causal factors of CMGE accidents and formulate an optimal safety investment portfolio, demonstrating the significant influence of safety management on coal mine safety.
202. Cause analysis and evolution characteristics of different types of railway accidents based on system dynamics theory
Core Problem: Understanding the key causes and evolution characteristics of various railway accidents (conflict, derailment, collision, fire, explosion) to improve safety and mitigate risks, as these accidents occur frequently and pose serious threats.
Key Innovation: Application of System Dynamics to analyze 547 railway accidents, identifying high-frequency/hazard chains, establishing causal loop/stock-flow diagrams, defining accident evolution periods (latent, acceleration, critical, outbreak), and introducing a 'false safety period' to quantify the impact of delayed control measures on risk accumulation.
203. Analysis of urban hydrogen-blended natural gas pipeline leak failure and accident evolution based on the combination of causal inference and probabilistic machine learning
Core Problem: Assessing the critical safety threats and risks associated with leakage failure and accident evolution in urban hydrogen-blended natural gas pipelines, a pivotal technology for energy transition that poses significant safety concerns.
Key Innovation: A novel framework integrating causal inference (Bow-Tie analysis) with probabilistic machine learning (Bayesian networks) for full quantitative risk assessment of pipeline leakage failure and accident evolution, enabling systematic failure factor identification and dynamic accident progression simulation.
204. Automated characterization of microfracture systems in organic-rich shales and their influence on porosity using convolutional neural networks on FIB-SEM images: A review
Core Problem: Characterizing microfracture systems and nanopore networks in organic-rich shales is challenging due to their inherent heterogeneity, complex morphology, and nano-scale features, hindering understanding of fluid transport and hydrocarbon storage.
Key Innovation: Evaluated state-of-the-art deep learning architectures for automated high-precision segmentation of microfractures and pore systems in FIB-SEM images of shales, finding KiU-Net superior (94% accuracy) for volumetric reconstruction and analysis of pore connectivity.
205. Remote sensing of tropical forest recovery: A review and decision-support framework for multi-sensor integration
Core Problem: Accurately monitoring tropical forest recovery is challenging because common proxies often overestimate success, and single-sensor approaches have limitations in capturing structural, compositional, and functional dimensions.
Key Innovation: Synthesized advances in multi-sensor remote sensing (LiDAR, optical, SAR, passive microwave, thermal) for comprehensive tropical forest recovery monitoring and presented a decision-support framework for integrating these data streams to overcome single-sensor blind spots and guide ecologically robust assessments.
206. Improving AMSR2 vegetation optical depth retrievals via land parameter retrieval model parameter optimisation
Core Problem: Traditional methods for estimating Vegetation Optical Depth (VOD) from microwave observations (AMSR2) using the Land Parameter Retrieval Model (LPRM) have limitations in accuracy and consistency, especially regarding parameter interactions and effective temperature inputs.
Key Innovation: The study proposes optimizing key LPRM parameters (surface roughness, effective temperature, single scattering albedo) through a global sensitivity analysis and evaluates two effective-temperature input scenarios. This optimization reduces model residuals, strengthens VOD-LAI agreement, and improves tracking of VWC variability, particularly in forests.
207. Adaptive image zoom-in with bounding box transformation for UAV object detection
Core Problem: Detecting small and sparse objects in UAV-captured images is challenging, as their size and sparsity hinder the optimization of effective object detectors.
Key Innovation: An adaptive zoom-in framework is proposed for UAV object detection. It features a lightweight offset prediction scheme with a novel box-based zooming objective to learn non-uniform zooming, coupled with a corner-aligned bounding box transformation method. This approach adaptively zooms into objects, significantly improving detection accuracy with minimal latency.
208. Impacts of using solar spectra adjusted for solar cycle variability in the radiometric correction of retrieved multi-band parameters from ETM+/Landsat-7 data
Core Problem: Radiometric correction of satellite data (ETM+/Landsat-7) might be affected by solar cycle variability, potentially impacting the accuracy of retrieved multi-band parameters.
Key Innovation: Investigates the impacts of using solar spectra adjusted for solar cycle variability in the radiometric correction process, aiming to improve the accuracy of retrieved multi-band parameters from ETM+/Landsat-7 data.
209. Adaptive Neighborhood Aggregation Algorithm for PolInSAR Forest Height Estimation
Core Problem: Single-baseline PolInSAR configurations for forest height estimation using the RVoG model suffer from parameter solution rank deficiency, leading to inaccurate results, especially in complex heterogeneous forest areas.
Key Innovation: Proposes an Adaptive Neighborhood Aggregation Algorithm (ANAA) that dynamically constructs adaptive windows based on scattering mechanism similarity (Wishart distance) and integrates multi-pixel observations to simultaneously solve for shared forest heights, significantly improving estimation accuracy and robustness.
210. Assessing the combination of passive and active microwave satellite observations (1.4 to 36 GHz) to estimate above ground biomass (AGB) globally
Core Problem: Accurately estimating Above Ground Biomass (AGB) globally using satellite observations, and evaluating the potential of combining different microwave observations and auxiliary data for this purpose.
Key Innovation: Developed a Neural Network inversion method to estimate global AGB by combining passive and active microwave satellite observations (1.4 to 36 GHz) with auxiliary data (NDVI, LST, SM), achieving high accuracy (R2 of 0.88, RMSE of 30 Mg/ha) and demonstrating potential for long-term AGB dynamics estimation.
211. The spatiotemporal evolution and driving factors of global surface temperature from 1940 to 2022
Core Problem: Understanding the spatiotemporal variations in global surface temperature (GST) and quantitatively identifying its complex, multi-factorial driving mechanisms from 1940 to 2022.
Key Innovation: Utilizes a unified quantitative-spatial framework (correlation analysis + Geographical Detector model) to assess driving mechanisms, identifying atmospheric water vapor as the dominant driver, amplified by synergistic interactions with NDVI and TP, and strong modulation by El Niño/La Niña. Projects continued GST increase using LSTM.
212. Direct and indirect impact assessments of climate changes and human activities on runoff changes in 31 source catchments of Yellow River Basin, China
Core Problem: Existing attribution assessments of runoff changes primarily focus on direct impacts of climate change and human activities, neglecting their indirect effects on underlying surface conditions and subsequent runoff changes.
Key Innovation: Developed an improved Budyko framework combining elasticity coefficient and principal component regression to comprehensively assess both direct and indirect impacts, revealing that human activities (77.4%) are the primary drivers of underlying surface parameter changes, and quantifying the contributions of direct climate (54.7%), indirect climate (12.2%), and human activities (33.1%) to runoff changes in the Yellow River Basin.
213. Recharge elevation, residence time and renewability of groundwater in the Upper Awash valley, Ethiopia: Applying environmental tracers in a highly populated volcanic basin
Core Problem: Groundwater management in the Upper Awash Basin is impaired by a lack of detailed hydrogeological characterization of complex volcanic aquifers, leading to falling water levels due to abstraction.
Key Innovation: Applies environmental tracers (stable isotopes, trace gases, 14C, hydrochemistry) to characterize groundwater recharge elevations, residence times (hundreds of years), and renewability, concluding that recharge is mostly local and providing a methodology applicable to other high-relief basins lacking detailed geological models.
214. Experimental and seismic performance study of composite window-type viscoelastic dampers
Core Problem: Limitations of conventional viscoelastic dampers in terms of installation flexibility, architectural integration, and maintenance costs, necessitating a novel design to enhance energy dissipation and structural adaptability under seismic loading.
Key Innovation: Introduction of a novel Composite Window-type Viscoelastic Damper (CWVED) with a window-integrated design, validated through analytical derivation and full-scale experimental testing, and demonstrated through nonlinear time-history analysis to significantly reduce inter-story drift and base shear under seismic loading.
215. A multiple tuned mass damper strategy for vibration control of monopile supported offshore wind turbines under seismic excitation
Core Problem: Offshore wind turbines (OWTs) are vulnerable to seismic risks, as earthquakes can excite high-frequency structural modes that are typically dormant under normal operational loads, requiring an effective vibration control strategy.
Key Innovation: Proposal and systematic evaluation of a 4-MTMD (multiple tuned mass damper) strategy targeting the first four bending modes of monopile-supported OWTs, demonstrating superior overall performance in controlling horizontal accelerations and providing comprehensive displacement reduction under various seismic excitations, particularly high-frequency ones.
216. Numerical investigation of train-induced ground vibrations in ballastless embankment considering wheel–rail geometric nonlinearity
Core Problem: Existing numerical models for train-induced ground vibrations often use simplified wheel-rail interaction, leading to unreliable predictions and underestimation of track displacement, especially at high speeds.
Key Innovation: Developed a novel 3D VTEG FE model incorporating rail irregularities and fully coupled wheel-rail geometric interaction, validated against field data, which accurately predicts ground vibrations and track displacement, and quantifies the benefits of foundation reinforcement in mitigating resonance-like behavior.
217. Effects of specimen preparation methods on polymer–montmorillonite interactions and hydraulic conductivity of polymer-modified bentonite–sand mixtures
Core Problem: The interaction mechanisms between polymers and montmorillonite, and their impact on the hydraulic conductivity of polymer-modified bentonite-sand mixtures, are not fully understood, especially concerning different specimen preparation methods.
Key Innovation: Comprehensive investigation revealing that wet blending leads to polymer intercalation and higher swelling potential, while dry blending results in phase-separated interaction, with both methods yielding distinct hydraulic conductivity characteristics, providing conceptual models for their performance.