TerraMosaic Daily Digest: Feb 17, 2026
Daily Summary
This digest synthesizes 158 selected papers and tracks a tightening loop between observation, mechanism, and decision. Event-scale reconstructions (e.g., the 2023 Xi’an debris flow and the 2025 Dharali mudslide–flash-flood cascade) are paired with runout/hazard models to map impact footprints, while monitoring studies push beyond deformation maps by resolving subsurface structure and stability contrasts using surface-wave imaging, InSAR time series, and deep segmentation.
Across the portfolio, the strongest contributions are validation- and uncertainty-aware: impact-based flash-flood warnings are benchmarked against multi-source impact records to quantify false-alarm reductions, probabilistic fragility and threshold methods explicitly propagate rainfall and hydraulic uncertainty, and new open data infrastructures—such as a cloud-optimized Italian radar precipitation archive—lower the barrier to reproducible, near-real-time hazard analytics.
Key Trends
- Forensic event studies are becoming modelling testbeds: Recent disasters are reconstructed from trigger to runout using field evidence plus physically based simulations, yielding transferable constraints on mobility, deposition, and hazard extent.
- Monitoring is shifting from “where it moves” to “why it fails”: Geophysics and remote sensing are used to infer slip surfaces, stratigraphy, and seismic response (not just velocities), enabling structure-aware stability diagnosis.
- Hybrid physics–data workflows dominate mobility and failure prediction: Coupled solvers (e.g., SPH–DEM/DEM) and ML–numerical integrations are used to simulate multi-phase landslides and to link susceptibility with downstream hazard.
- Early-warning research is quantifying decision value under uncertainty: Studies evaluate warning systems with impact data and probabilistic thresholds, emphasizing spatial prioritization and measurable reductions in false alarms.
- Foundation-model tooling and open archives are entering operations: Knowledge-enhanced vision–language frameworks and cloud-native environmental datacubes support faster situational awareness and reproducible multi-hazard pipelines.
Selected Papers
This digest features 158 selected papers from 921 new papers analyzed (out of 2591 raw papers scanned; 921 papers after deduplication). Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Preliminary analysis of the deadly debris flow on August 11th, 2023, at Xi’an of China: characteristics, mechanism, and hazard assessment
Core Problem: A comprehensive understanding of the development, characteristics, causative factors, and hazard extent of the deadly Haogou debris flow event is needed to improve public perception of risks and early warning abilities in similar settings.
Key Innovation: Provided a preliminary analysis of the Haogou debris flow using field surveys, remote sensing, a physically based model (FSLAM), and FLO-2D simulations, identifying key causative factors (steep source, weathered rocks, heavy rain) and quantifying runout distance (1.2 km), velocity (6 m/s), and peak accumulation depth (3.7 m), highlighting the importance of improved monitoring and early warning.
2. Advanced characterization of landslide subsurface structure using mixed-source surface wave imaging: A case study of the Majiagou landslide in the three Gorges Reservoir area of China
Core Problem: A comprehensive understanding of the geological structure and site seismic response mechanisms of landslides is crucial for evaluating landslide hazards, guiding seismic design, and improving engineering protection measures, especially for deep-seated landslides, which often lack detailed 2-D spatial information.
Key Innovation: Used mixed-source surface wave imaging to construct high-resolution 2-D shear wave velocity profiles of the Majiagou landslide, integrating them with borehole data to precisely delineate a four-layer geological structure and identify 2-D sliding surfaces, and calculated natural frequencies and seismic amplification factors to reveal variations in seismic stability across different areas.
3. Cause and effect analysis of Dharali disaster in Uttarkashi District of Uttarakhand, India, on August 5, 2025
Core Problem: The Dharali disaster, a devastating mudslide-flash flood sequence, highlights the compounded hazard created by unstable geomorphology, climate-driven extreme rainfall, and poorly regulated infrastructure in the Himalayas, requiring a detailed cause and effect analysis for effective mitigation.
Key Innovation: Provided a cause and effect analysis of the Dharali disaster, identifying the interaction of fragile Himalayan geology, steep paraglacial terrain, evolving climatic drivers (intense localized rainfall, ITCZ shift), and anthropogenic interventions (road widening, blasting) as key factors, underscoring the need for geoscience-based land-use planning and strict control of human activities for hazard mitigation.
4. Multi-model landslide integrated hazard detection using SBAS-InSAR and U-Net segmentation
Core Problem: Challenges in detecting difficult deformation signals and achieving low spatial positioning accuracy when identifying potential landslide hazards in mountainous regions.
Key Innovation: Proposes a multi-tiered, multi-model approach combining SBAS-InSAR time-series data, spatial autocorrelation, susceptibility evaluation (Random Forest, Information Value), kernel density estimation, U-Net semantic segmentation of Sentinel-2 imagery, and a geomechanical evolutionary model to accurately identify potential landslide areas, validated with GNSS data and historical landslide distributions.
5. Integrating machine learning and numerical methods for enhanced landslide susceptibility and hazard mapping in the Bhotekoshi watershed, central Nepal
Core Problem: Traditional landslide susceptibility maps are limited in assessing risk to downslope settlements and infrastructure, leading to an underestimation of exposure.
Key Innovation: Integrates machine learning-based landslide susceptibility (Random Forest) with numerical runout modeling to provide a comprehensive landslide hazard assessment, significantly increasing the estimated exposure of population, buildings, and roads within very high hazard classes.
6. Quantifying the added value of impact-based warnings for flash flood monitoring using innovative multi-source impact data
Core Problem: The operational benefits of impact-based warnings (IBW) compared to traditional hazard-based warnings (HBW) for flash floods are not sufficiently validated, particularly regarding their ability to improve communication, prioritize areas, and reduce false alarms.
Key Innovation: Developed a validation framework using diverse multi-source impact data (legislative decrees, insurance claims, fire/rescue logs) over 13 years to compare IBW and HBW for flash floods. Demonstrated a clear added value of IBW, particularly at finer spatial scales, reducing false alarms by a factor of 2 to 3.
7. DRAD: A new model for dynamic real-time Avalanche detection from videos with residual depth-separable convolution and feature pyramid networks
Core Problem: Direct avalanche detection from videos remains unsolved, despite the application of deep learning for image classification and segmentation, hindering real-time monitoring of this destructive geohazard.
Key Innovation: Develops DRAD, a novel deep learning approach integrating depth-separable convolutions, feature pyramid networks, and residual networks, for efficient and effective real-time dynamic avalanche detection from videos, achieving 92.1% classification accuracy and robustness even with low-quality social media videos.
8. Simulation of soil-rock mixture fluidized landslide dynamics using the 3D-SPH-DEM method enhanced by reorientation and spherical cap cutting algorithms
Core Problem: Accurately characterizing irregular rock blocks and complex soil-rock coupling effects in Soil-rock mixture fluidized landslide (SRMFL) simulations, where traditional numerical methods face limitations in contact analysis.
Key Innovation: Proposes an improved 3D-SPH-DEM coupling method, incorporating a reorientation algorithm for rock feature surfaces and a spherical cap cutting algorithm for soil-rock contact volume. Validates the method by reproducing a real landslide event and investigates effects of rock content and shapes on SRMFL dynamics.
9. IT-DPC-SRI: A Cloud-Optimized Archive of Italian Radar Precipitation (2010-2025)
Core Problem: Historical fragmentation and inaccessibility of Italian weather radar precipitation data, hindering long-term analysis and operational use for disaster management.
Key Innovation: Creating IT-DPC-SRI, the first publicly available, long-term (2010-2025), cloud-optimized archive of Italian weather radar Surface Rainfall Intensity (SRI) observations, harmonized into a coherent Zarr datacube and made accessible for research and operational use by the Italian Civil Protection Department.
10. An interdisciplinary approach to the pre- and syn-eruptive magma dynamics during the Tajogaite monogenetic eruption (La Palma, 2021)
Core Problem: Understanding magma dynamics in magmatic systems where developed and monogenetic volcanoes coexist, and improving eruption forecasting in monogenetic systems.
Key Innovation: Integrated petrological, geochemical, and geophysical data to reconstruct multi-stage pre- and syn-eruptive processes of the 2021 Tajogaite eruption, revealing late-stage intrusions, magma ascent times, and the development of a crystal mush zone, highlighting the critical importance of integrated monitoring for improved eruption forecasting.
11. Bridging urban and catchment scales in urban pluvial flooding: a multi-resolution simulation tool
Core Problem: Traditional urban pluvial flood assessment methods often fail to predict detailed surface runoff in complex urban areas, either lacking precision (general hydrological estimations) or neglecting runoff contributions from the entire catchment (detailed hydrodynamic analyses focused only on urban areas).
Key Innovation: Introduces a new methodology and system that integrates GIS analysis and satellite sensing-derived datasets to generate high-resolution flood maps. This multi-resolution single simulation approach bridges the gap between hydrological basin tools and urban hydrodynamic systems, achieving high accuracy (80%) in simulating historical flood events by considering whole catchment runoff.
12. Soil infiltration and slope stability of shrub-covered loess slopes on the northeastern Qinghai–Tibet Plateau: experimental and numerical simulation
Core Problem: The impact mechanism of vegetation on slope soil water infiltration and stability in loess areas is unclear, hindering regional ecological restoration and shallow geological disaster prevention.
Key Innovation: Investigated the permeability characteristics and influencing factors of root-containing soils on loess slopes and, through numerical simulation, evaluated the hydrological response and stability of vegetation-covered slopes under rainfall, demonstrating how vegetation enhances slope stability.
13. Floristic structure and root dynamics in slope stabilization of landslide-prone areas in the Phewa watershed, Nepal
Core Problem: The Phewa watershed faces constant landslide threats, necessitating the identification of appropriate plant species that can enhance soil stability in landslide-prone areas.
Key Innovation: Assessed floral diversity and root traits of various plant species in landslide areas, identifying specific species like Saccharum spontaneum as optimal for enhancing soil stability and lowering landslide risk based on uprooting force and root characteristics.
14. Retrogressive thaw slump expansion and vegetation phenological response in the Qilian Mountains, northeastern Tibetan Plateau
Core Problem: Permafrost degradation driven by climate warming is accelerating landscape changes, with retrogressive thaw slumps (RTS) emerging as a critical disturbance, but their impacts on vegetation phenology at the watershed scale are poorly understood.
Key Innovation: Investigated the spatial heterogeneity of vegetation responses to RTS expansion in a Tibetan Plateau basin, revealing a nearly fivefold increase in RTS area and localized ecological resilience (increased vegetation vigor) in RTS-affected areas, possibly due to enhanced groundwater recharge.
15. Power-Law Evolution Behavior of Acoustic Emission Preceding Rockburst: Insights from Laboratory-Scale Investigation
Core Problem: Understanding the evolution behavior of acoustic emission (AE) preceding rockbursts is crucial for elucidating their underlying mechanisms and identifying reliable precursor characteristics.
Key Innovation: Revealed that AE count evolution follows a critical acceleration model with a power-law singularity (index p in [0.9, 1]) preceding rockbursts, insensitive to confining pressure. Discovered a significant scaling synchronization between macroscopic energy dissipation rate and AE radiation energy rate, establishing two universal laws: the critical acceleration law and the energy synchronization scaling law for rockbursts.
16. Retraction Note: On the Probabilistic Seismic Hazard Assessment in Kazakhstan
Core Problem: A prior study on probabilistic seismic hazard assessment (PSHA) in Kazakhstan has been withdrawn; readers need a clear record of the correction.
Key Innovation: This item is a retraction notice and does not report new methods or results; its value is in clarifying the publication record.
17. Seismic performance assessment of single and twin tunnels subjected to earthquake and rail loads
Core Problem: The seismic performance of underground single and twin tunnels, especially under combined earthquake and high-speed rail loads, needs comprehensive assessment to inform robust design strategies.
Key Innovation: Conducted a novel finite element analysis to demonstrate that twin tunnel systems are significantly more susceptible to damage from earthquakes than single tunnels, experiencing substantially higher stresses and displacements, thereby emphasizing the critical need for robust seismic design strategies for such configurations.
18. Stay-or-Relocate Model (STORM): An agent-based population displacement simulator applied to a multi-phase volcanic eruption scenario
Core Problem: Emergency management practitioners require adaptable decision support tools to account for fine-scale variations in hazard exposure, vulnerability, and demographics during and after natural hazard events, specifically for population displacement.
Key Innovation: Developed STORM, an agent-based model simulating resident decision-making (stay, relocate, return) during multi-phase volcanic eruptions, incorporating household characteristics and changing circumstances (damage, outages, access loss, school disruption, liveability), and quantifying accommodation selection and support requirements.
19. Fractal and wavelet diagnostics of flood resilience in transport corridors: Evidence from the 20–21 September 2025 Eastern Black Sea Floods, Türkiye
Core Problem: Evaluating the resilience of transport corridors exposed to extreme hydro-meteorological events and identifying critical segments prone to flood damage in mountainous regions.
Key Innovation: Introduces a diagnostic framework combining fractal (Higuchi Fractal Dimension in a sliding-window scheme) and wavelet analyses to detect spatial 'fragility windows' and capture multiscale irregularities along river corridors, demonstrating that morphological irregularities can serve as practical indicators of flood-induced fragility for improving flood resilience.
20. Instability mechanisms of deep tunnels induced by stiff discontinuities: insights from laboratory comparative tests and numerical simulation analysis
Core Problem: The types and scales of geological disasters induced by high in-situ stress in deep engineering excavations are strongly regulated by stiff discontinuities within rock masses, but the precise instability mechanisms and failure evolution processes are not fully understood.
Key Innovation: Conducted direct shear tests on various structural plane specimens and employed Particle Flow Code (PFC) to analyze the failure evolution of surrounding rock during tunnel excavation. Identified shear sliding along dip direction of structural planes as the dominant contributor to deformation and slip-shear composite failure, offering critical implications for understanding instability mechanisms and proposing a systematic strategy for rockburst prevention.
21. Probabilistic fragility analysis of unsaturated soil slope under rainfall infiltration considering stress-dependent water retention behaviour
Core Problem: Existing fragility analyses for rainfall-induced slope failure often neglect the impact of stress states on soil hydraulic properties, leading to underestimation of failure probability and risk.
Key Innovation: Conducts probabilistic fragility analysis of unsaturated soil slopes under rainfall, explicitly incorporating a stress-dependent water retention model and spatially variable soil parameters. Demonstrates that ignoring stress effects significantly underestimates failure probability and risk, especially for extreme rainfall.
22. Seismic response study of photovoltaic slopes considering soil plastic modulus evolution under far-field long-period ground motions
Core Problem: Existing designs for photovoltaic (PV) slopes on expressways often neglect the impact of far-field long-period (FFLP) ground motions on soil plastic modulus degradation and cumulative damage, leading to potential seismic safety issues.
Key Innovation: This study systematically investigates the dynamic response of PV slopes under various ground motions using a calibrated P2PSand constitutive model and a pile-soil coupled numerical method. It demonstrates that FFLP motions, particularly harmonic ones, cause significant cumulative damage and excessive permanent displacement of both the slope and PV structure, even at lower peak accelerations, highlighting a critical oversight in current design practices.
23. VLCE: A Knowledge-Enhanced Framework for Image Description in Disaster Assessment
Core Problem: Contemporary Vision Language Models (VLMs) produce details inadequately aligned with disaster assessment objectives due to a deficiency in domain knowledge and the absence of a refined descriptive process.
Key Innovation: Presents VLCE, a dedicated multimodal framework that integrates external semantic knowledge from ConceptNet and WordNet to improve image captioning, generating disaster-specific descriptions from satellite and UAV imagery for actionable intelligence in real-time disaster assessment.
24. Anatomy of incremental behaviour of granular materials induced by loading reversal: a stress probing analysis
Core Problem: The incremental behavior of granular materials under complex stress paths involving loading reversals is not fully understood, which limits the accuracy of constitutive formulations in geotechnical engineering.
Key Innovation: A series of stress probing tests, combined with the discrete-element method, was conducted to capture the evolution of strain response envelopes during loading reversal. This revealed a dramatic increase in plastic modulus with nearly unchanged plastic flow direction, guiding the development of a new constitutive model within the bounding surface framework.
25. Damage mechanisms in loess media under blast action: insights from experiments and numerical simulations
Core Problem: Understanding the macroscopic and microscopic physical and mechanical responses and damage mechanisms of loess media under explosive loading is crucial for engineering practices and geological disaster prevention in loess areas.
Key Innovation: Analyzes loess damage under explosive loading through field explosion tests and discrete element numerical simulations. The study confirms the particle expansion method's effectiveness, reveals the coupled influence of charge shape and particle gradation on crack propagation and damage range, and validates stress wave attenuation, providing a new analytical framework for blasting design and damage prediction in loess areas.
26. Study on the Correlation Mechanism Between Mode I/II Fracture Toughness Evolution and Freeze-Thaw Damage in Red Sandstone Under Freeze-Thaw Cycles
Core Problem: The differential degradation mechanisms of mode I and mode II fracture toughness in rock masses affected by freeze-thaw cycles, which contribute to strength reduction and instability in cold regions, are not fully clarified.
Key Innovation: Quantified the significant decrease in both KIC and KIIC (80.32% and 65.25% respectively after 60 F-T cycles) due to micropore expansion and coalescence. Established a prediction model for fracture toughness using a freeze-thaw damage variable based on P-wave velocity, highlighting its better suitability for predicting shear failure modes.
27. Correction: Empirical Reformulation of Sellmeijer’s Criterion for Backward Erosion Piping
Core Problem: A previously published empirical reformulation of Sellmeijer’s criterion for backward erosion piping required correction.
Key Innovation: This item is a correction notice; it does not introduce new science beyond amending the original publication.
28. Seismic performance of a foundation shoe connection for precast columns combining starter bolts and grouted corrugated ducts
Core Problem: Precast column-to-foundation connections in commercial and industrial buildings need to effectively resist both static and seismic forces, ensuring adequate plasticity and energy dissipation.
Key Innovation: Introduced and tested a novel precast column-to-foundation connection system combining steel bolted shoes and grouted intermediate rebars, demonstrating its superior resistance, energy dissipation capacity, and ductility under reverse cyclic loading, and proposed analytical and numerical models for its behavior.
29. Fire-induced changes in modal behavior and seismic implications for historical masonry mosques
Core Problem: The impact of fire on the dynamic behavior of historical masonry structures and its subsequent implications for their seismic vulnerability are not well understood, potentially leading to insufficient post-fire restoration practices.
Key Innovation: Employed a hybrid analysis framework to demonstrate that fire-induced stiffness degradation significantly alters the dynamic characteristics of historical masonry mosques, leading to reduced natural frequencies and amplified localized seismic vulnerability, highlighting the need for performance-based conservation strategies.
30. Effects of snowmelt runoff process on soil erosion in seasonal frozen districts
Core Problem: Understanding the mechanism by which snowmelt runoff processes contribute to soil erosion in seasonal frozen districts, particularly the role of runoff erosion forces.
Key Innovation: Investigating the laws of snowmelt infiltration and runoff variation under different climatic backgrounds, proposing an innovative method for calculating snowmelt runoff shear stress incorporating surface roughness, and verifying the enhancement of friction and drag forces on the surface due to increased runoff depth and velocity.
31. Bidirectional seismic kinematic coupling responses of large-diameter tunnels in sloping sites under Rayleigh waves: Analytical solutions using Green’s functions
Core Problem: Existing studies often overlook the additional bending moment and significant coupling effects generated by soil-tunnel interaction, particularly for large-diameter tunnels in sloping sites under Rayleigh-wave excitation, leading to an incomplete understanding of their bidirectional seismic response and kinematic coupling. Conventional solution techniques also struggle with the spatially and temporally varying nonhomogeneous loads.
Key Innovation: Developed a Green’s function framework and derived four novel Green’s functions to comprehensively evaluate bidirectional seismic responses and kinematic coupling effects in large-diameter tunnels in sloping sites. The solutions incorporate fixed-end constraints, rotational damping, and prestressing, providing a robust analytical framework for seismic design and advancing theoretical understanding of tunnel seismic resilience.
32. Hydro-mechanical coupling in fractured rocks: A numerical study using the implicit joint-continuum model
Core Problem: Accurately modeling coupled hydro-mechanical behavior in fractured rocks is challenging due to the discontinuous, stress-sensitive, and anisotropic nature of fractures, which localize flow and deformation.
Key Innovation: Employed the implicit joint-continuum model (IJCM) to efficiently investigate coupled hydro-mechanical behavior in fractured rocks, demonstrating its ability to reproduce permeability evolution under stress and highlighting the importance of tensor-based permeability formulations for fractured media analyses.
33. Simulation of flood processes under extreme precipitation scenarios based on an improved HEC-HMS model in a typical karst mountainous basin, southwest China
Core Problem: Effectively translating the characteristics of extreme precipitation into dynamic guidance for flood control practices in water conservancy projects, particularly in karst basins, remains a critical gap in existing studies.
Key Innovation: Proposed a new modeling framework combining extreme precipitation scenarios (considering spatial distribution, temporal allocation, and intensity) with an improved HEC-HMS model tailored for karst mountainous basins, successfully simulating flood processes, mitigating runoff overestimation, and providing dynamic guidance for flood control decision-making.
34. Field monitoring study on a highway geogrid-reinforced embankment in Xinjiang, China
Core Problem: Understanding the in-situ performance and stability of high geogrid-reinforced embankments under varying environmental conditions (temperature, humidity) and construction practices.
Key Innovation: Long-term field monitoring of a 19.3m high geogrid-reinforced embankment, revealing the influence of subzero temperatures on soil/interface strength, soil arching, and the impact of nonstandard compaction on geogrid strains, providing data for service performance evaluation.
35. Multi-Arrival Infrasound from Meteoroids: Fragmentation Signatures versus Propagation Effects in a Fine-Scale Layered Atmosphere
Core Problem: Ambiguity in infrasonic signatures of meteoroid fragmentation: distinguishing between complex breakup events and distorting effects of a layered atmosphere is critical for accurate energy estimates and source reconstruction.
Key Innovation: Uses pseudo-differential parabolic equation (PPE) simulations and a unique regional dataset to quantify atmospheric modification of acoustic waveforms, establishing diagnostic criteria to separate source physics from propagation artifacts, thereby improving infrasound's reliability for monitoring natural bolides and other events.
36. Green-NAS: A Global-Scale Multi-Objective Neural Architecture Search for Robust and Efficient Edge-Native Weather Forecasting
Core Problem: Developing robust and efficient global-scale weather forecasting models for low-resource, edge-native environments is challenging, requiring a balance between accuracy, computational energy costs, and carbon footprints.
Key Innovation: Introduces Green-NAS, a multi-objective Neural Architecture Search framework that explicitly minimizes computational energy costs and carbon footprints while optimizing for model accuracy and efficiency. It achieves high accuracy with significantly fewer parameters than other global models and demonstrates improved accuracy through transfer learning for data-limited cities.
37. Impact of the permeable breakwater on spatiotemporal siltation in a semi-enclosed harbor: case study of Binhai Harbor, China
Core Problem: Binhai Harbor faces pronounced siltation challenges with a distinctive and unconventional spatial distribution, making it difficult to identify sediment sources and mechanisms for effective mitigation.
Key Innovation: Systematically analyzed multi-period hydrodynamic, sediment, and bathymetric data, and developed a hydro-morphodynamic model incorporating breakwater permeability, revealing the mechanisms of central basin siltation and providing theoretical support for optimizing breakwater designs to suppress siltation.
38. Experimental Investigation on Multistage Fatigue Deformation and Failure Behavior of Transversely Isotropic Rock Under Triaxial Stress Disturbance and Unloading Conditions
Core Problem: Existing studies inadequately reveal the mechanical behaviors of soft and hard composite rock bodies under the complex coupling effect of disturbance and unloading during deep-buried tunneling, which significantly affects surrounding rock stability.
Key Innovation: Designed a triaxial multi-stage fatigue–unloading–confining pressure test (MFT–MSUCP) to investigate composite rock behavior. Found that MFT–MSUCP significantly decreased strength and axial strain compared to MFT, with strength being highly sensitive to unloading. Revealed distinct fracture patterns and quantified fracture volume changes using 3D CT reconstruction.
39. The colonization, succession and ecological processes of saxicolous moss crust in the Qinling Mountains
Core Problem: The need for effective methods to restore engineered and rocky slopes and enhance their stability against erosive forces.
Key Innovation: Characterizing the synergistic microbe-plant interactions in saxicolous moss crust formation, revealing its mechanical and chemical contributions to rock substrate reshaping and ecosystem stability, and providing a theoretical framework for bio-restoration strategies on rocky slopes.
40. Transient aerodynamic loads and spatiotemporal evolution of the temperature field in cold-region high-speed railway tunnels
Core Problem: High-speed railway tunnels in cold regions are susceptible to severe frost damage, with their internal temperature field evolution significantly affected by train-induced aerodynamic effects.
Key Innovation: Employs a CFD dynamic mesh approach to elucidate coupled evolution mechanisms between transient aerodynamic loads and the tunnel temperature field, clarifying the regulatory effect of pressure waves on airflow, and proposing a correction coefficient for anti-freeze insulation length based on train frequency and external air temperature.
41. Theoretical analysis for ground stress and tunnel deformation during surcharge-induced consolidation considering arching effect
Core Problem: Existing tunnel load theories neglect the long-term effects of soil deformation characteristics and arching effects during reclamation surcharge-induced ground consolidation, leading to errors in calculating soil stress around embedded shield tunnels and underestimating deformation risks.
Key Innovation: Presents analytical solutions for ground stress and tunnel deformation, explicitly considering soil creep consolidation and the evolution of the soil arching effect under reclamation surcharge conditions. The method was validated with field data and provides insights into critical parameters influencing soil and structural response, highlighting the significant impact of the arching effect.
42. Key geometric surface of the normal rock joint deformation:Aperture into an equivalent rough composite surface based on BEM
Core Problem: Existing theoretical studies on rock joint closure predominantly analyze contact between two rough surfaces, which is insufficient for analyzing stress-induced contact behavior when both surfaces deform, and fails to conveniently utilize relevant parameters for stress variations.
Key Innovation: Employed an equivalent theory to transform the interaction of two rough surfaces into a rough composite surface and a rigid plane, combined with the Boundary Element Method (BEM), enabling accurate elastoplastic simulation of joint stress variations, convenient expression of roughness parameters, and precise calculation of displacement, thus providing significant theoretical value for investigating joint deformation mechanisms relevant to rock mass stability.
43. Hypoplastic modeling of sand behavior accounting for memory surface and evolving fabric anisotropy under cyclic loading conditions
Core Problem: Existing hypoplastic models for granular soils exhibit limitations in accurately reproducing soil responses during unloading and reloading phases under both drained and undrained cyclic loading conditions, which is critical for geotechnical engineering design.
Key Innovation: Proposed a hypoplastic model for granular soils that integrates the memory surface (MS) concept with the anisotropic critical state theory (ACST) and intergranular strain (IS), incorporating a memory surface defined by dissipation energy and an evolving deviatoric fabric tensor, thereby significantly enhancing simulation capabilities for sand behavior under various cyclic loading conditions relevant to seismic-induced ground deformation and liquefaction.
44. One-dimensional nonlinear site response in deep soil deposits: A modified Hybrid–Hyperbolic model and prediction-bias analysis
Core Problem: Accurately predicting seismic response in deep soil deposits is challenging due to conventional constitutive models' inability to capture nonlinear behavior across the full strain range, and an incomplete understanding of bias mechanisms in deep deposit simulations.
Key Innovation: This research develops a modified hybrid-hyperbolic (HH) backbone model coupled with a modified damping model to accurately represent hysteretic soil behavior across the full strain range. It identifies that model boundary and soil-column thickness significantly influence spectral amplification, and that fundamental-frequency overestimation caused by rigid boundary assumptions can be mitigated by extending the soil column into bedrock or increasing its thickness.
45. Learning Data-Efficient and Generalizable Neural Operators via Fundamental Physics Knowledge
Core Problem: Existing neural operators (NOs) for modeling physical systems primarily learn from target PDEs, overlooking fundamental physical principles, which limits data efficiency, predictive accuracy, and out-of-distribution (OOD) generalization.
Key Innovation: Proposes a multiphysics training framework that jointly learns from both original PDEs and their simplified basic forms, enhancing data efficiency, reducing predictive errors, and improving OOD generalization for neural operators across various 1D/2D/3D PDE problems.
46. Joint Enhancement and Classification using Coupled Diffusion Models of Signals and Logits
Core Problem: Robust classification in noisy environments is challenging because standard approaches treat signal enhancement and classification as separate, sequential stages, failing to leverage semantic information from the classifier's output during denoising.
Key Innovation: A general, domain-agnostic framework that integrates two interacting diffusion models (one for the input signal and one for the classifier's output logits) to enable mutual guidance, where signal enhancement refines class estimation and evolving class logits guide signal reconstruction towards discriminative regions, improving classification accuracy under diverse noise conditions.
47. Symbolic recovery of PDEs from measurement data
Core Problem: Accurately identifying symbolic PDE models representing underlying physical laws from indirect and noisy measurement data, which is crucial for interpretability and understanding in scientific domains.
Key Innovation: Using neural network architectures based on rational functions for symbolic representation of physical laws, with an identifiability result showing unique reconstruction of the simplest physical law in the limit of noiseless, complete measurements, promoting interpretability and sparsity.
48. The Stationarity Bias: Stratified Stress-Testing for Time-Series Imputation in Regulated Dynamical Systems
Core Problem: Time-series imputation benchmarks suffer from a 'Stationarity Bias' where uniform random masking and shape-agnostic metrics overemphasize performance in easy, low-entropy stationary regimes, masking poor performance during critical transients.
Key Innovation: Formalizing the Stationarity Bias and proposing a 'Stratified Stress-Test' that partitions evaluation into Stationary and Transient regimes, demonstrating that deep learning models are essential for preserving morphological fidelity during transients, and deriving empirical missingness distributions for robust training.
49. Stabilizing Test-Time Adaptation of High-Dimensional Simulation Surrogates via D-Optimal Statistics
Core Problem: Distribution shifts between training and deployment often cause severe performance degradation in machine learning surrogates used to accelerate costly simulations, and existing Test-Time Adaptation (TTA) methods are unstable for high-dimensional, unstructured regression problems.
Key Innovation: A Test-Time Adaptation (TTA) framework based on storing maximally informative (D-optimal) statistics, which jointly enables stable adaptation and principled parameter selection at test time, yielding significant out-of-distribution improvements for high-dimensional simulation regression and generative design optimization.
50. SOON: Symmetric Orthogonal Operator Network for Global Subseasonal-to-Seasonal Climate Forecasting
Core Problem: Accurate global Subseasonal-to-Seasonal (S2S) climate forecasting remains challenging due to chaotic atmospheric dynamics and suboptimal modeling of anisotropic dynamics by existing models.
Key Innovation: Proposing the Symmetric Orthogonal Operator Network (SOON) which uses an Anisotropic Embedding strategy and SOON Blocks with symmetric decomposition of Zonal and Meridional Operators to improve forecasting accuracy and computational efficiency for global S2S climate forecasting, critical for disaster preparedness.
51. Score-based change point detection via tracking the best of infinitely many experts
Core Problem: Developing robust nonparametric online change point detection algorithms that can handle complex data distributions and adapt to changing environments remains a challenge.
Key Innovation: Proposed a nonparametric online change point detection algorithm based on sequential score function estimation and a tailored 'tracking the best expert' approach (a version of the fixed share forecaster for infinite experts and quadratic loss), demonstrating promising results and providing rigorous high-probability bounds.
52. SSL4EO-S12 v1.1: A Multimodal, Multiseasonal Dataset for Pretraining, Updated
Core Problem: The need for a robust, multimodal, and multitemporal Earth Observation dataset for pretraining large-scale foundation models, with previous versions having geospatial alignment inaccuracies and inefficient data structures.
Key Innovation: Presents SSL4EO-S12 v1.1, an updated dataset fixing previous issues, offering analysis-ready data loading, maintaining extensive spatial coverage, and adding new modalities (elevation, land-cover, vegetation) to support multimodal pre-training for Earth Observation.
53. APCoTTA: Continual Test-Time Adaptation for Semantic Segmentation of Airborne LiDAR Point Clouds
Core Problem: Fixed models for Airborne Laser Scanning (ALS) point cloud semantic segmentation suffer performance degradation due to continuous domain shifts, and existing Continuous Test-Time Adaptation (CTTA) methods are underexplored for this data type, facing challenges like catastrophic forgetting and error accumulation.
Key Innovation: Proposal of APCoTTA, a novel CTTA framework tailored for ALS point cloud semantic segmentation, which uses gradient-driven layer selection, entropy-based consistency loss, and random parameter interpolation to achieve superior performance on new benchmarks by mitigating forgetting and error accumulation.
54. FedX: Explanation-Guided Pruning for Communication-Efficient Federated Learning in Remote Sensing
Core Problem: Communication overhead in Federated Learning (FL) for remote sensing (RS) image classification tasks due to the frequent exchange of large model updates between clients and a central server.
Key Innovation: Proposes FedX, an explanation-guided pruning strategy that uses backpropagation-based explanation methods to estimate model component importance and prunes the least relevant ones, significantly reducing communication overhead and enhancing generalization in FL for RS tasks.
55. Morephy-Net: An Evolutionary Multi-objective Optimization for Replica-Exchange-based Physics-informed Neural Operator Learning Networks
Core Problem: Existing physics-informed neural networks and operator-learning models face challenges in balancing data/physics losses, maintaining robustness under noisy/sparse observations, and providing reliable uncertainty quantification when solving parametric partial differential equations (PDEs).
Key Innovation: Morephy-Net integrates evolutionary multi-objective optimization (to avoid ad hoc loss weighting), replica-exchange stochastic gradient Langevin dynamics (to enhance global exploration and stabilize training), and Bayesian uncertainty quantification to solve parametric PDEs robustly in noisy data regimes.
56. GenDA: Generative Data Assimilation on Complex Urban Areas via Classifier-Free Diffusion Guidance
Core Problem: Reconstructing high-resolution urban wind flow from sparse sensor data is challenging, yet essential for assessing air quality, heat dispersion, and pedestrian comfort. Existing methods struggle with obstacle-aware reconstruction and generalization across unseen geometries.
Key Innovation: Proposes GenDA, a generative data assimilation framework using a multiscale graph-based diffusion architecture. It interprets classifier-free guidance as a learned posterior reconstruction mechanism, enabling obstacle-aware reconstruction and generalization across unseen geometries, wind directions, and mesh resolutions, significantly improving reconstruction accuracy.
57. Causally constrained reduced-order neural models of complex turbulent dynamical systems
Core Problem: Reduced-order neural emulators of complex turbulent systems, such as those in climate dynamics, often develop spurious, noncausal dependencies, limiting their ability to accurately respond to external forcings when trained on unforced data.
Key Innovation: A flexible framework based on response theory and score matching is introduced to suppress spurious noncausal dependencies in neural emulators. This enhances their ability to respond to both weak and strong external forcings, demonstrated using the stochastic Charney-DeVore model relevant for low-frequency atmospheric variability.
58. The compressibility of kaolin clay: a micromechanical perspective based on DEM
Core Problem: Understanding the complex micromechanical mechanisms governing the compressibility of kaolin clays, especially the role of physicochemical forces, is crucial for accurate macroscopic soil behavior prediction.
Key Innovation: Formulated a contact model for DEM simulations incorporating long-range van der Waals attraction and electrical double layer repulsion to reproduce interplatelet physicochemical forces, revealing their significant role at low stresses and the evolution of clay fabric during compression.
59. Robust and efficient inland vessel draft detection based on multi-task learning and weighted directed graphs
Core Problem: Existing inland vessel draft detection methods lack cross-scenario generalization and robustness, especially when hull surfaces are stained or draft scale characters are missing, leading to detection errors and missed detections.
Key Innovation: Proposed VDDNet, a high-precision detection method based on multi-task learning (YOLOv11, CBAM for segmentation) and weighted directed graphs for reconstructing missing draft scale information, achieving robust, high-accuracy (MAVD 2.006 pixels, MADDE 0.04 m), and real-time (80 FPS) vessel draft detection under complex conditions.
60. Petrogenesis of Huichizi granite and hosted amphibolite lens: implications for early Paleozoic evolution of the North Qinling orogenic belt
Core Problem: The genesis of Early Paleozoic granite and its relationship with multi-stage metamorphism in the North Qinling Tectonic Belt (NQTB) are heavily debated, limiting understanding of the Qinling Orogenic Belt's evolution.
Key Innovation: Conducted systematic petrology, geochemistry, and zircon chronology studies on Huichizi granite and a hosted lenticle, revealing three metamorphic ages, adakitic signatures of the granite derived from partial melting of Neoproterozoic mafic rocks, and a multi-stage tectonic evolution involving subduction, magma emplacement, and later metamorphism/anatexis.
61. Improving understanding of drought events in the Dongting Lake Basin: insights from GRACE and optical satellites
Core Problem: Traditional monitoring approaches face limitations in capturing water storage anomalies and effectively monitoring spatiotemporal drought evolution characteristics in the Dongting Lake Basin, a critical region for water security.
Key Innovation: Integrated three GRACE TWSA solutions via the Bayesian Three-Cornered Hat (BTCH) fusion method to achieve higher accuracy in monitoring terrestrial water storage anomalies, revealing a critical depletion post-2018 and effectively tracking drought evolution, corroborated by optical satellite data showing dramatic water surface area reduction, providing technical support for water resources management.
62. Automatic detection of in-channel wood using UAV-based deep learning: a scalable approach for river monitoring
Core Problem: Traditional methods for detecting and mapping in-channel wood are labor-intensive, costly, and spatially limited, hindering understanding of river processes and assessment of associated risks.
Key Innovation: Developed a UAV-based deep learning framework, systematically benchmarking state-of-the-art semantic segmentation architectures, demonstrating that Attention U-Net with RGB imagery provides highly accurate and transferable automatic detection of in-channel wood, advancing scalable river monitoring.
63. Erosion rates in rock salt exposures with diverse karren monitored by erosion pins, close-range photogrammetry and terrestrial laser scanner
Core Problem: Published data on solutional erosion in rock salt exposures are scarce, hindering the understanding of emergent salt diapir evolution and its practical implications (e.g., long-term geostorage).
Key Innovation: Monitored erosion rates in rock salt exposures using erosion pins, terrestrial laser scanner (TLS), and ground-based photogrammetry, establishing a robust relationship between rainfall and slope-normal erosion. Developed empirical relationships to model geomorphic evolution and found photogrammetry more effective than TLS for small-scale erosion assessment.
64. Dam-induced changes in multi-fraction sediment recovery in the middle-lower Yangtze River
Core Problem: The impact of dam-induced sediment starvation on the sediment grain size distribution (GSD) and subsequent morphological changes in downstream river systems has received less attention.
Key Innovation: Investigated the effects of the Three Gorges Dam (TGD) on multi-fraction sediment transport and bed recovery in the Yangtze River, identifying three distinct dam-induced sediment regimes (static armored gravel bed, active bed armoring, strong erosion) and highlighting the increasing role of tributaries and lakes as sediment sources.
65. Implicit methods for reliability analysis of phased-mission systems subject to cascading deterministic common cause failures
Core Problem: Reliability analysis of Phased-Mission Systems (PMSs) is complicated by Cascading Deterministic Common Cause Failures (CDCCFs), where initial failures can trigger a domino effect, and existing methods struggle to handle complex system structures (e.g., loops) and arbitrary time-to-failure distributions.
Key Innovation: Two implicit approaches utilizing multi-valued decision diagrams for reliability analysis of PMSs subject to CDCCFs, specifically designed for systems with no-loop and Hamiltonian loop structures, which can handle arbitrary time-to-failure distributions and both external/internal common causes.
66. Microstructures in oxidic Oxisol under intensive cultivation
Core Problem: Understanding the long-term impacts of intensive conventional cultivation and irrigation on the microstructural stability and hierarchy of Oxisol soils.
Key Innovation: Integrating micromorphology, SR-μCT, and micro-chemical analysis to demonstrate that intensive cultivation systematically transitions stable coalesced blocks to individualized, degraded microaggregates, promoting kaolinite dispersion, and altering aggregate size, roundness, and porosity over 43 years.
67. Linear time-lag effects and nonlinear interactions of global drought-flood abrupt alternation in responses to multiple factors
Core Problem: Research on the driving mechanisms of global drought-flood abrupt alternation (DFAA) events is limited, often focusing only on linear effects of climate change and atmospheric circulation, neglecting other factors like surface energy fluxes and nonlinear interactions.
Key Innovation: Incorporated multiple factors (including surface energy fluxes) and used a revised DFAA index, Pearson correlation, multiple linear regression, and interpretable machine learning to systematically analyze linear time-lag effects and nonlinear interactions on global DFAA, significantly increasing explanatory power and revealing key nonlinear threshold regulations for improved prediction and disaster prevention.
68. Size Transferability of Graph Transformers with Convolutional Positional Encodings
Core Problem: Understanding and ensuring the transferability and scalability of Graph Transformers (GTs) from small to larger graphs, particularly when using GNN-based positional encodings.
Key Innovation: Establishing theoretical connections between GTs with GNN positional encodings and Manifold Neural Networks, proving that GTs inherit transferability guarantees and exhibit scalable behavior, demonstrated with an application to shortest path distance estimation over terrains.
69. Benchmarking IoT Time-Series AD with Event-Level Augmentations
Core Problem: Existing anomaly detection (AD) for safety-critical IoT time series is often evaluated using point-level results on curated datasets, which limits its practical value for model selection under realistic perturbations.
Key Innovation: Introduction of a unified evaluation protocol with event-level augmentations (sensor dropout, drift, noise, shifts) and sensor-level probing for root-cause analysis, applied to benchmark 14 models on diverse datasets to improve practical model selection.
70. RPT-SR: Regional Prior attention Transformer for infrared image Super-Resolution
Core Problem: General-purpose super-resolution models, especially Vision Transformers, are inefficient in infrared imaging scenarios with fixed viewpoints (e.g., surveillance), as they fail to exploit strong, persistent spatial priors inherent in such scenes.
Key Innovation: RPT-SR, a Regional Prior attention Transformer, which explicitly encodes scene layout information into the attention mechanism using a dual-token framework (learnable regional prior tokens and local tokens) to dynamically modulate local reconstruction, achieving state-of-the-art performance in infrared image super-resolution.
71. CEPAE: Conditional Entropy-Penalized Autoencoders for Time Series Counterfactuals
Core Problem: Accurately performing counterfactual inference on time series data, especially when impacted by specific events, is crucial for decision-making, but existing autoencoder-based methods are not specifically tailored for time series or designed to encourage disentangled representations.
Key Innovation: Introduces CEPAE (Conditional Entropy-Penalized Autoencoder), a novel autoencoder-based approach for time series counterfactual inference that employs an entropy penalization loss over the latent space to encourage disentangled data representations, outperforming other approaches on various datasets.
72. Neural Network-Based Parameter Estimation of a Labour Market Agent-Based Model
Core Problem: Parameter estimation in large-scale Agent-Based Models (ABMs) is computationally challenging due to the vast parameter space, limiting their use as decision-support tools.
Key Innovation: Evaluates and demonstrates the effectiveness of a neural network-based simulation-based inference (SBI) framework for parameter estimation in ABMs, showing it can recover original parameters and improve efficiency compared to traditional Bayesian methods, applied to a labour market ABM.
73. Beyond ReLU: Bifurcation, Oversmoothing, and Topological Priors
Core Problem: Deep Graph Neural Networks (GNNs) suffer from oversmoothing, where node features converge to a homogeneous, non-informative state, limiting their depth and representational power.
Key Innovation: Re-framing oversmoothing as a bifurcation problem, theoretically discovering that replacing monotone activations induces a bifurcation that destabilizes the homogeneous state and creates stable, non-homogeneous patterns that resist oversmoothing, along with a bifurcation-aware initialization.
74. Criteria-first, semantics-later: reproducible structure discovery in image-based sciences
Core Problem: The dominant 'semantics-first' paradigm in image-based science (recovering structure by predicting labels) fails under conditions requiring open-ended discovery, cross-comparability, and long-term monitoring due to label drift and lack of reproducibility.
Key Innovation: Proposal of a 'criteria-first, semantics-later' framework for reproducible structure discovery in image-based sciences, separating criterion-defined, semantics-free structure extraction from downstream semantic mapping, providing a domain-general scaffold for stable partitions and structural fields, and enabling plural interpretations and explicit crosswalks.
75. Solving Parameter-Robust Avoid Problems with Unknown Feasibility using Reinforcement Learning
Core Problem: Applying reinforcement learning to reachability problems is mismatched because RL optimizes expected returns, potentially performing poorly on low-probability but still unsafe states, and robust optimization requires known feasibility, which is often unknown a priori.
Key Innovation: Feasibility-Guided Exploration (FGE), a method that simultaneously identifies a subset of feasible initial conditions under which a safe policy exists and learns a policy to solve the reachability problem over this set, demonstrating significantly improved coverage in robust avoid problems.
76. Fluids You Can Trust: Property-Preserving Operator Learning for Incompressible Flows
Core Problem: Traditional numerical solvers for incompressible flows are computationally expensive, and existing neural operators fail to exactly enforce physical properties like incompressibility.
Key Innovation: A novel property-preserving kernel-based operator learning method that analytically enforces physical properties (incompressibility, periodicity, turbulence) for incompressible flows, achieving higher accuracy and faster training than neural operators.
77. Uni-Flow: a unified autoregressive-diffusion model for complex multiscale flows
Core Problem: Existing physics-informed machine learning approaches struggle to simultaneously maintain long-term temporal evolution and resolve fine-scale structure in complex multiscale spatiotemporal flows (chaotic, turbulent, physiological regimes).
Key Innovation: Uni-Flow, a unified autoregressive-diffusion framework that explicitly separates temporal evolution (low-resolution latent dynamics) from spatial refinement (high-resolution physical field reconstruction), enabling faster-than-real-time, accurate modeling of complex multiscale flows across diverse benchmarks.
78. Neural-POD: A Plug-and-Play Neural Operator Framework for Infinite-Dimensional Functional Nonlinear Proper Orthogonal Decomposition
Core Problem: AI for Science is often limited by 'discretization,' where learned representations are restricted to specific grids, and classical Proper Orthogonal Decomposition (POD) is limited to linear subspace approximations.
Key Innovation: Neural-POD, a plug-and-play neural operator framework that constructs nonlinear, orthogonal basis functions in infinite-dimensional space, enabling resolution-invariant learning, generalization to unseen parameters, and capturing nonlinear structures in complex spatiotemporal systems like Burgers' and Navier-Stokes equations.
79. NeuroLifting: Neural Inference on Markov Random Fields at Scale
Core Problem: Inference in large-scale Markov Random Fields (MRFs) is computationally challenging, with traditional methods often failing to balance efficiency and solution quality, especially as problem scale increases.
Key Innovation: Introduced NeuroLifting, a novel technique that leverages Graph Neural Networks (GNNs) to reparameterize decision variables in MRFs, enabling efficient and parallelizable optimization via gradient descent, demonstrating superior solution quality on large-scale MRFs with linear computational complexity growth.
80. Multispectral airborne laser scanning for tree species classification: a benchmark of machine learning and deep learning algorithms
Core Problem: Challenges in leveraging deep learning techniques for automated point cloud processing and identifying rare tree species in class-imbalanced datasets using multispectral airborne laser scanning (ALS).
Key Innovation: A comprehensive benchmark of deep learning and traditional machine learning methods for tree species classification using high-density multispectral ALS data, demonstrating that point-based deep learning (specifically a point transformer model) outperforms other approaches.
81. Randomness and signal propagation in physics-informed neural networks (PINNs): A neural PDE perspective
Core Problem: The implications of statistically random weight matrices in trained Physics-informed neural networks (PINNs) for signal propagation and stability remain unsatisfactorily understood, hindering interpretability and reliable application.
Key Innovation: Analyzes the spectral and statistical properties of trained PINN weights, showing consistency with random matrix theory, and demonstrates through neural PDEs how numerical stability of associated discretizations governs signal propagation and overall network stability.
82. Flock: A Knowledge Graph Foundation Model via Learning on Random Walks
Core Problem: Existing Knowledge Graph Foundation Models (KGFMs) for zero-shot link prediction are limited by deterministic equivariance, preventing them from distinguishing structurally similar but semantically distinct relations, thus limiting expressive power.
Key Innovation: Proposes Flock, a KGFM leveraging probabilistic node-relation equivariance, which iteratively samples random walks, encodes them with a sequence model, and aggregates representations, achieving state-of-the-art performance on zero-shot link prediction and acting as a universal approximator for isomorphism-invariant link-level functions.
83. TabImpute: Universal Zero-Shot Imputation for Tabular Data
Core Problem: Missing data is a widespread problem in tabular settings, and existing imputation solutions lack universality, exhibiting large performance variance and requiring time-consuming hyperparameter tuning, especially in small datasets.
Key Innovation: Introduces TabImpute, a pre-trained transformer building on TabPFN, which provides accurate and fast zero-shot imputations for tabular data without fitting or hyperparameter tuning, enabled by a novel entry-wise featurization and a diverse synthetic training data generation pipeline.
84. Can Multimodal LLMs Perform Time Series Anomaly Detection?
Core Problem: The potential of multimodal LLMs (MLLMs) for time series anomaly detection (TSAD) is largely unexplored, and existing studies oversimplify the problem by not fully addressing multi-granular anomalies and irregular time series.
Key Innovation: Builds VisualTimeAnomaly, a benchmark to investigate zero-shot MLLM capabilities for TSAD across various anomaly types and irregular sampling, and proposes TSAD-Agents, a multi-agent framework that leverages MLLMs and traditional methods for automatic, adaptive TSAD.
85. A LoD of Gaussians: Unified Training and Rendering for Ultra-Large Scale Reconstruction with External Memory
Core Problem: Existing Gaussian Splatting methods struggle with scaling to ultra-large scenes due to chunking artifacts, training complexities across scales, and GPU memory limitations, hindering seamless multi-scale reconstruction and interactive visualization.
Key Innovation: Introduces "A LoD of Gaussians," a framework for unified training and rendering of ultra-large-scale Gaussian scenes on a single consumer-grade GPU without partitioning. It uses out-of-core storage, a Level-of-Detail (LoD) representation, a hybrid data structure, and a caching/view scheduling system for efficient, seamless multi-scale reconstruction and interactive visualization.
86. Towards Geometric and Textural Consistency 3D Scene Generation via Single Image-guided Model Generation and Layout Optimization
Core Problem: Generating high-quality 3D scenes from a single RGB image remains a significant challenge, as current approaches often struggle to ensure both object generation quality and scene coherence in multi-object scenarios.
Key Innovation: A novel three-stage framework for 3D scene generation that achieves explicit geometric representations and high-quality textural details from a single image. It includes image instance segmentation/inpainting, pseudo-stereo viewpoint construction for camera/depth estimation with model selection, and layout optimization via Chamfer distance minimization for precise alignment.
87. Agents of Discovery
Core Problem: Modern scientific research faces increasing data volumes and complexity, requiring sophisticated data analysis tools and workflows, which are often special-purpose. Automating routine analysis components is needed to counteract this complexity.
Key Innovation: Investigates using Large Language Models (LLMs) to create a team of agents that jointly solve data analysis-based research problems by creating code to operate standard tools and libraries. This system demonstrates the capacity to automate routine analysis components, achieving performance comparable to human state-of-the-art results in anomaly detection.
88. PolySHAP: Extending KernelSHAP with Interaction-Informed Polynomial Regression
Core Problem: KernelSHAP, a leading method for approximating Shapley values in XAI, assumes a linear game, limiting its accuracy for models with non-linear feature interactions, while exact computation is prohibitively expensive.
Key Innovation: PolySHAP extends KernelSHAP by approximating the game with higher-degree polynomials to capture non-linear feature interactions, yielding empirically better and consistent Shapley value estimates. It also provides the first strong theoretical justification for the practical performance of paired sampling.
89. Grappa: Gradient-Only Communication for Scalable Graph Neural Network Training
Core Problem: Cross-partition edges in distributed Graph Neural Network (GNN) training lead to high communication costs, overwhelming networks as graphs deepen and partition counts grow, thus limiting scalability.
Key Innovation: Grappa is a distributed GNN training framework that enforces gradient-only communication, periodically repartitions, and applies a lightweight coverage-corrected gradient aggregation. This achieves up to 13x faster training, better accuracy for deeper models, and trillion-edge scalability on commodity hardware.
90. Tabular Foundation Models Can Learn Association Rules
Core Problem: Classical Association Rule Mining (ARM) suffers from rule explosion and poor scalability, while recent neural approaches perform poorly in low-data regimes, hindering knowledge discovery in tabular data.
Key Innovation: A model-agnostic framework is introduced to extract association rules from any conditional probabilistic model over tabular data, leveraging Tabular Foundation Models (TFMs). TabProbe, an instantiation, uses TFMs to produce concise, high-quality association rules with strong predictive performance, robustly even in low-data settings.
91. In situ quantification of soil thermal properties by integrating active-distributed temperature sensing and Bayesian inference
Core Problem: Conventional methods for soil thermal property estimation suffer from low resolution and high uncertainty, limiting accurate soil thermal behavior simulations and subsurface investigations relevant to geotechnical and geothermal applications.
Key Innovation: Developed an in situ framework integrating active-distributed temperature sensing (DTS) with Bayesian inference and a composite-medium line-source (CMLS) model to provide distributed, depth-resolved, and uncertainty-quantified measurements of soil thermal properties.
92. Prediction of ice resistance in floating ice floes using hybrid temporal convolution-recurrent neural networks
Core Problem: Accurate and efficient assessment of ice resistance in floating ice floes, critical for polar navigation and marine structure design, is constrained by limited accuracy of empirical formulations and high computational cost of CFD-DEM simulations.
Key Innovation: Developed hybrid temporal surrogate models (TCN-LSTM and TCN-GRU) based on CFD-DEM simulations, achieving high prediction accuracy (R2=0.97) and significantly reducing computational time (approx. 49% reduction) for ice resistance assessment, enabling rapid design screening.
93. Nonparametric prediction of ship maneuvering motions based on the heterogeneous integration model
Core Problem: Single nonparametric models for predicting ship maneuvering motions often fail to simultaneously ensure prediction accuracy and cross-condition robustness under complex sea states in maritime engineering and intelligent navigation.
Key Innovation: Proposed a heterogeneous integration model combining multiple neural networks (FC-NBEAN, Bi-LSTM) and a genetic algorithm for adaptive sub-model weighting, with an ARMA-based residual module, achieving significantly improved prediction accuracy (31.2% RMSE, 40.1% MAE reduction) and cross-condition robustness for multi-DOF ship motions.
94. Using smoothed particle hydrodynamics with open boundaries to model floating offshore wind turbines in realistic extreme conditions
Core Problem: Accurately predicting the dynamics of floating offshore wind turbines (FOWTs) in realistic extreme conditions, such as combined wave-current and breaking waves, is challenging but crucial for effective design and operation.
Key Innovation: Developed and demonstrated a novel SPH-based framework with open boundaries to model FOWTs in realistic extreme conditions, including focused waves on sheared currents and spilling breaking waves, and showed its capability for design refinement to reduce extreme pitch angles and mooring tensions.
95. Quantifying evaporation of intercepted rainfall: a hybrid correction approach for eddy-covariance measurements
Core Problem: Eddy-covariance (EC) measurements systematically underestimate evaporation during and shortly after rainfall events, leading to unreliable vapor flux data and an imbalance in energy and water budgets, particularly for intercepted rainfall.
Key Innovation: Development of a hybrid correction approach that complements EC measurements with a Rutter-based 2D model's estimates of evaporation under interception conditions, providing accurate evaporation rates from intercepted liquid precipitation and improving the parametrization of surface fluxes in weather and climate models.
96. Deepest-ever rock core extracted from under Antarctic ice sheet
Core Problem: Understanding the past retreat of the West Antarctic Ice Sheet to better predict its future behavior and potential contributions to sea-level rise.
Key Innovation: Extraction of the deepest-ever rock core from beneath the Antarctic ice sheet, providing crucial data for analyzing historical ice sheet dynamics.
97. Resilience analysis of four-state engineering systems under the framework of continuous-time Markov chain model
Core Problem: Most existing resilience measures for engineering systems are probability-based and can only be computed by simulation, lacking closed-form analytical expressions for resistance and recovery from disruptive events.
Key Innovation: Derives closed-form expressions for three resilience metrics (survival function of resistance time, probability of recovery before breakdown, survival function of recovery time) for four-state engineering systems using a continuous-time Markov chain model, offering analytical insights for improving system resilience.
98. System reliability analysis with active learning: A one-step look-ahead policy based on expected uncertainty reduction function
Core Problem: Kriging-based active learning methods for system reliability analysis face challenges at the system level due to interactions among multiple limit state functions, making efficient and accurate failure probability estimation difficult.
Key Innovation: Proposes a novel Kriging-based active learning method for system reliability analysis, developing an Expected Uncertainty Reduction (EUR) function to evaluate the potential contribution of candidate samples in improving failure probability estimation accuracy, achieving high accuracy with fewer performance function evaluations.
99. Uncertainty-aware remaining useful life prediction and PPO based optimal maintenance scheduling in industrial IoT
Core Problem: Predictive maintenance in industrial IoT faces challenges in accurately predicting Remaining Useful Life (RUL) due to inherent uncertainties, and subsequently, in designing optimal maintenance schedules that account for these uncertainties and avoid issues like Q-value overestimation or training instability.
Key Innovation: A framework that combines a non-parametric probabilistic prediction method (Quantile Regression integrated with GRU and CNN) for uncertainty-aware RUL prediction with a Proximal Policy Optimization (PPO) approach for designing optimal adaptive maintenance schedules, demonstrating improved RUL prediction and reduced maintenance schedule time.
100. Divergent responses of soil organic carbon stocks in different layers to global changes on the Tibetan Plateau
Core Problem: Understanding the spatial distribution of soil organic carbon (SOC) on the Tibetan Plateau and its layer-specific responses to future climate change, which is crucial for regional carbon cycling and ecosystem management.
Key Innovation: Using an RFE-RF approach with 372 soil profiles to model SOC density at different depths, identifying aridity index as the most influential driver, and projecting significant declines in SOC stocks under the SSP5–8.5 scenario, particularly in deeper layers, while showing surface SOC increase in arid-desert regions.
101. Multi-scale mechanical behavior and damage mechanisms of granite after seawater cyclic thermal shock
Core Problem: Assessing the long-term stability and integrity of Enhanced Geothermal Systems (EGS) in coastal environments requires a comprehensive understanding of granite's multi-scale mechanical behavior under thermal-hydraulic conditions.
Key Innovation: Systematically investigated the multi-scale mechanical and microstructural evolution of granite under seawater cyclic thermal shock, identifying mineral-specific damage mechanisms and developing a thermo-mechanical-chemical coupled statistical damage constitutive model for EGS stability assessment.
102. Opportunities for remote sensing to monitor freshwater water quality after extreme weather events: A systematic review
Core Problem: Inadequately explored multi-dimensional impacts of extreme weather events on inland water quality and challenges in using remote sensing to capture these effects.
Key Innovation: A systematic review identifying research gaps in remote sensing for water quality monitoring after extreme weather events, highlighting the need for integrated approaches to quantify specific changes, expand geographic coverage, and leverage advanced technologies.
103. A novel large language model–inspired transformer framework for reconstructing subsurface soil properties at untested locations from sparse cone penetrometer test data
Core Problem: Accurate characterization of subsurface conditions from sparse CPT data is challenging due to soil spatial variability and limited datasets.
Key Innovation: Proposes a novel geotechnical self-supervised learning framework (GSSLF) inspired by large language models, integrating a self-supervised masking strategy and a tailored transformer-based architecture (GeoFormer) with a modified attention mechanism to reconstruct subsurface soil properties from sparse CPT data, providing robust accuracy and uncertainty estimation.
104. AMOC Weakening Shapes Ocean Heat Storage Patterns Under Strong Idealized Warming
Core Problem: A comprehensive understanding of how a weakening Atlantic Meridional Overturning Circulation (AMOC) influences ocean heat storage and redistribution in a warming climate remains elusive.
Key Innovation: AMOC weakening significantly influences both vertical and inter-basin heat redistribution, leading to a 3% increase in global ocean heat uptake and altered storage patterns (more in Indo-Pacific and intermediate layers, less in deep Atlantic).
105. Linking Hadley Cell Instability to Slow Equatorial Motions in Reanalysis and CMIP6 Models
Core Problem: The processes driving the growth of slow-moving tropical phenomena like the Madden-Julian Oscillation (MJO) and convectively-coupled equatorial Rossby (ER) waves, and their link to Hadley Cell instability, are not fully understood.
Key Innovation: Wave activity (MJO and ER waves) is strongest when a Hadley Cell-linked instability metric peaks, with stronger correlations for ER waves, suggesting common driving mechanisms with tropical depression-type waves.
106. Tropical TCV as a Process Diagnostic: Connecting Probability to Processes in km–Scale Models Via Moisture Budget Statistics
Core Problem: Km-scale climate models with parameterized convection often fail to realistically produce the observed bimodal total column vapor (TCV) distribution and confine rainfall to a moist mode, leading to rainfall biases.
Key Innovation: The TCV distribution serves as a valuable process diagnostic, revealing that the lack of bimodality in parameterized models stems from an incorrect relationship between moisture flux convergence and TCV in non-raining environments.
107. Low Cloud Dispersion Effects by Anthropogenic Aerosols in Polluted Air
Core Problem: The dispersion effect of anthropogenic aerosols on cloud albedo, specifically how aerosols impact the relative dispersion (ε) of cloud droplet size distributions in polluted continental clouds, remains a major source of uncertainty in climate forcing.
Key Innovation: In-situ measurements and machine learning analysis in polluted East China show that aerosol number concentration and hygroscopicity dominate changes in cloud droplet dispersion under specific conditions, enhancing the aerosol indirect effect by >11%.
108. Refine Now, Query Fast: A Decoupled Refinement Paradigm for Implicit Neural Fields
Core Problem: Implicit Neural Representations (INRs) for large 3D scientific simulations face a critical fidelity-speed dilemma: deep MLPs incur high inference costs, while efficient embedding-based models lack sufficient expressiveness.
Key Innovation: Proposes the Decoupled Representation Refinement (DRR) architectural paradigm, which leverages a deep refiner network and non-parametric transformations in a one-time offline process to encode rich representations into a compact, efficient embedding structure, decoupling slow neural networks from the fast inference path. It also introduces DRR-Net and Variational Pairs (VP) for data augmentation.
109. tensorFM: Low-Rank Approximations of Cross-Order Feature Interactions
Core Problem: Efficiently capturing high-order interactions between attributes in tabular categorical data for prediction problems, which is crucial in various applications like click-through rate prediction.
Key Innovation: tensorFM, a new model that generalizes field-weighted factorization machines and efficiently captures high-order feature interactions via a low-rank tensor approximation, demonstrating competitive performance and low latency.
110. DAV-GSWT: Diffusion-Active-View Sampling for Data-Efficient Gaussian Splatting Wang Tiles
Core Problem: Existing 3D Gaussian Splatting methods for generating expansive landscapes are constrained by a reliance on densely sampled exemplar reconstructions, leading to high data volume requirements.
Key Innovation: DAV-GSWT is a data-efficient framework that leverages diffusion priors and active view sampling to synthesize high-fidelity Gaussian Splatting Wang Tiles from minimal input observations, integrating hierarchical uncertainty quantification and generative diffusion models to identify informative viewpoints and hallucinate missing structural details.
111. GMAIL: Generative Modality Alignment for generated Image Learning
Core Problem: Indiscriminate use of highly realistic generated images for training machine learning models can cause mode collapse due to modality discrepancies between real and synthetic domains.
Key Innovation: GMAIL is a novel framework that explicitly treats generated images as a separate modality, bridging the two distinct modalities in the same latent space through a multi-modal learning approach to effectively leverage generative models and boost performance across various vision-language tasks.
112. Fractional-Order Federated Learning
Core Problem: Federated learning (FL) suffers from significant drawbacks including slow convergence, high communication cost, and instability when dealing with non-independent-and-identically-distributed (non-IID) client data.
Key Innovation: FOFedAvg is a novel FedAvg variation that incorporates Fractional-Order Stochastic Gradient Descent (FOSGD) to capture long-range relationships and deeper historical information, improving communication efficiency, accelerating convergence, and mitigating instability caused by heterogeneous non-IID client data.
113. Bridging Day and Night: Target-Class Hallucination Suppression in Unpaired Image Translation
Core Problem: Day-to-night unpaired image translation is challenging due to large appearance shifts and lack of pixel-level supervision, often introducing semantic hallucinations of target-class objects (e.g., traffic signs, vehicles) that degrade downstream performance.
Key Innovation: A novel framework that detects and suppresses target-class hallucinations during unpaired image translation by designing a dual-head discriminator for semantic segmentation and introducing class-specific prototypes as semantic anchors, performing iterative refinement to preserve object semantics.
114. Doubly Stochastic Mean-Shift Clustering
Core Problem: Standard Mean-Shift algorithms are highly sensitive to the bandwidth hyperparameter, particularly in data-scarce regimes, leading to fragmentation and spurious modes (over-segmentation).
Key Innovation: Doubly Stochastic Mean-Shift (DSMS) introduces randomness in both trajectory updates and kernel bandwidth by drawing data samples and radius from a continuous uniform distribution, acting as an implicit regularization mechanism to improve stability and prevent over-segmentation in sparse clustering scenarios.
115. Efficient Generative Modeling beyond Memoryless Diffusion via Adjoint Schr\"odinger Bridge Matching
Core Problem: Diffusion models often yield highly curved trajectories and noisy score targets due to an uninformative, memoryless forward process that induces independent data-noise coupling, leading to inefficient sampling.
Key Innovation: Adjoint Schr"odinger Bridge Matching (ASBM) is a generative modeling framework that recovers optimal trajectories by learning the Schr"odinger Bridge forward dynamic through a data-to-energy sampling perspective and then learning the backward generative dynamic, producing significantly straighter and more efficient sampling paths.
116. Semantic-Guided 3D Gaussian Splatting for Transient Object Removal
Core Problem: Transient objects in casual multi-view captures cause ghosting artifacts in 3D Gaussian Splatting (3DGS) reconstructions, and existing solutions suffer from significant memory costs or vulnerability to parallax ambiguity.
Key Innovation: Proposes a semantic filtering framework for category-aware transient removal in 3DGS using vision-language models (CLIP), which resolves parallax ambiguity by identifying object categories independently of motion patterns, leading to improved reconstruction quality with minimal memory overhead.
117. An Industrial Dataset for Scene Acquisitions and Functional Schematics Alignment
Core Problem: Manual alignment of functional schematics with 2D/3D scene acquisitions for old industrial facilities is tedious, complex, and does not scale, hindering digital twin creation.
Key Innovation: Introduction of IRIS-v2, a comprehensive dataset for industrial scene acquisition and schematic alignment, and a method combining segmentation and graph matching to reduce alignment time for digital twin creation.
118. Continuous-Time Piecewise-Linear Recurrent Neural Networks
Core Problem: Existing piecewise-linear recurrent neural networks (PLRNNs) for dynamical systems reconstruction are discrete-time, which is inconsistent with continuous-time physical processes and struggles with irregularly sampled data.
Key Innovation: Development of continuous-time PLRNNs (cPLRNNs) with a novel training and simulation algorithm that bypasses numerical integration, allowing semi-analytical determination of topological objects and outperforming discrete-time PLRNNs and Neural ODEs on dynamical systems reconstruction benchmarks.
119. Relative Geometry of Neural Forecasters: Linking Accuracy and Alignment in Learned Latent Geometry
Core Problem: How neural networks internally represent the underlying latent geometry of complex dynamical systems they forecast remains poorly understood, making it difficult to compare and improve their internal mechanisms.
Key Innovation: Introduction of anchor-based, geometry-agnostic relative embeddings to study representational alignment in neural forecasters, revealing reproducible family-level structure and linking alignment to forecasting accuracy across various canonical dynamical systems.
120. Controlled oscillation modeling using port-Hamiltonian neural networks
Core Problem: Data-driven methods for learning dynamical systems often fail to learn underlying conservation laws, limiting their generalization, and existing port-Hamiltonian neural networks often do not consider power-preserving discretizations.
Key Innovation: Proposal to use a second-order discrete gradient method embedded in the learning of dynamical systems with port-Hamiltonian neural networks, demonstrating improved performance over Runge-Kutta methods for controlled oscillation modeling across various systems.
121. Meteorological data and Sky Images meets Neural Models for Photovoltaic Power Forecasting
Core Problem: Accurately forecasting photovoltaic power is challenging due to the inherent variability of solar energy production, especially for ramp events and under cloudy conditions, requiring improved robustness and extended forecasting capabilities.
Key Innovation: A multimodal hybrid approach combining sky images, photovoltaic energy history, and meteorological data with deep neural models for both nowcasting and forecasting, significantly improving prediction accuracy, particularly on cloudy days, and enhancing robustness.
122. Reconstructing Carbon Monoxide Reanalysis with Machine Learning
Core Problem: Maintaining the quality of atmospheric composition reanalysis products (e.g., Carbon Monoxide) when satellite observational data becomes unavailable or is discontinued.
Key Innovation: Investigating machine learning methods to predict monthly-mean total column of Carbon Monoxide re-analysis from control model simulations, aiming to compensate for data losses from satellite instruments.
123. When Remembering and Planning are Worth it: Navigating under Change
Core Problem: Spatial navigation for AI agents in non-stationary, uncertain environments with limited sensing is challenging, requiring robust memory, learning, and planning strategies.
Key Innovation: Explores a range of strategies and finds that an architecture incorporating multiple strategies (for exploration, search, and planning) and utilizing non-stationary probability learning to update episodic memories for building imperfect maps and planning on the fly, significantly improves efficiency in challenging tasks.
124. Sparse Additive Model Pruning for Order-Based Causal Structure Learning
Core Problem: In order-based causal structure learning, the pruning step (removing spurious edges) is a computational bottleneck and can harm estimation quality due to repeated model fitting and multiple testing.
Key Innovation: Introduces a new pruning method based on sparse additive models, enabling direct pruning without hypothesis testing, and proposes an efficient algorithm combining randomized tree embedding with group-wise sparse regression, achieving significantly faster and comparably accurate results.
125. The First Instrumentally Documented Fall of an Iron Meteorite: atmospheric trajectory and ground impact
Core Problem: Iron meteorite falls are rare, and until recently, no iron meteorite had a reliably determined pre-atmospheric orbit or instrumentally documented fall, limiting the understanding of their atmospheric entry dynamics and impact characteristics.
Key Innovation: The first instrumentally recorded and recovered fall of an iron meteorite, allowing for the reconstruction of its luminous trajectory using optical, infrasound, and seismic data, and the identification of distinct aerodynamic properties of iron meteoroids to improve entry models and recovery predictions.
126. Beyond Labels: Information-Efficient Human-in-the-Loop Learning using Ranking and Selection Queries
Core Problem: Traditional human-in-the-loop machine learning often limits human experts to simple labeling, which is information-inefficient and fails to capture the nuances of human judgment.
Key Innovation: Develops a human-in-the-loop framework that leverages rich query types (item ranking and exemplar selection) with probabilistic human response models and active learning algorithms to significantly reduce sample complexity and learning time compared to traditional label-only active learning.
127. Ensemble-size-dependence of deep-learning post-processing methods that minimize an (un)fair score: motivating examples and a proof-of-concept solution
Core Problem: Existing distribution-aware deep-learning post-processing methods for ensemble forecasts can be sensitive to ensemble size, leading to unfairness and unreliability (over-dispersion), especially when minimizing scores like aCRPS.
Key Innovation: Introduced 'trajectory transformers' (an adaptation of PoET) that applies self-attention over lead time while preserving conditional independence, achieving ensemble-size independence and improving forecast reliability for weekly mean T2m forecasts.
128. Individualized Federated Learning for Traffic Prediction with Error Driven Aggregation
Core Problem: Current Federated Learning for Traffic Prediction (FLTP) frameworks lack real-time model updating and rely on a single global model, failing to adapt to continuously incoming data and neglecting the non-IID characteristics of traffic data from different locations.
Key Innovation: Proposed NeighborFL, an individualized real-time federated learning scheme that uses a haversine distance-based and error-driven heuristic to group personalized local models, creating location-aware and tailored prediction models for each client, demonstrating improved real-time prediction accuracy.
129. Digital Twin Generation from Visual Data: A Survey
Core Problem: The need for a comprehensive overview of state-of-the-art methodologies for generating digital twins from visual data, including their advantages, limitations, and challenges.
Key Innovation: A comprehensive survey analyzing recent advances in generating digital twins from visual data, covering various approaches like 3D Gaussian Splatting and generative inpainting, and identifying key challenges and future research directions.
130. MMS-VPR: Multimodal Street-Level Visual Place Recognition Dataset and Benchmark
Core Problem: Existing visual place recognition (VPR) datasets lack multimodal diversity, underrepresent dense pedestrian street scenes, and have limited temporal coverage, especially in non-Western urban contexts.
Key Innovation: Introduction of MMS-VPR, a large-scale multimodal dataset for street-level place recognition in pedestrian-only environments, and MMS-VPRlib, a unified benchmarking platform for evaluating multimodal VPR methods.
131. Calibrated and uncertain? Evaluating uncertainty estimates in binary classification models
Core Problem: Difficulty in rigorously evaluating and understanding the qualitative properties of uncertainty estimates produced by increasingly complex data models like deep learning techniques for classification, which is crucial for scientific validity.
Key Innovation: Uses a unifying framework of approximate Bayesian inference and empirical tests on synthetic datasets to investigate and compare qualitative properties (calibration, out-of-distribution uncertainty) of six probabilistic machine learning algorithms, highlighting that deep learning methods struggle with consistent out-of-distribution uncertainty.
132. Out of Distribution Detection for Efficient Continual Learning in Quality Prediction for Arc Welding
Core Problem: Current machine learning models for quality prediction in dynamic manufacturing environments struggle with inherent distribution shifts, leading to critical limitations in maintaining robust performance.
Key Innovation: Extends the VQ-VAE Transformer architecture by leveraging its autoregressive loss as a reliable out-of-distribution (OOD) detection mechanism, integrating it with continual learning to optimize model adaptation and minimize costly labeling requirements.
133. GenFacts-Generative Counterfactual Explanations for Multi-Variate Time Series
Core Problem: Existing methods for generating counterfactual explanations for multivariate time series often produce invalid, implausible, or unintuitive results, limiting their effectiveness in enhancing model transparency.
Key Innovation: Introduces GenFacts, a generative framework based on a class-discriminative variational autoencoder, integrating contrastive and classification-consistency objectives, prototype-based initialization, and realism-constrained optimization to produce more plausible and interpretable counterfactuals for multivariate time series.
134. Generating Findings for Jaw Cysts in Dental Panoramic Radiographs Using a GPT-Based VLM: A Preliminary Study on Building a Two-Stage Self-Correction Loop with Structured Output (SLSO) Framework
Core Problem: Vision-language models (VLMs) face challenges in generating reliable and accurate radiological findings for medical image interpretation in clinical practice, often producing inconsistent or hallucinated outputs.
Key Innovation: Proposes a Self-correction Loop with Structured Output (SLSO) framework for GPT-based VLMs, integrating image analysis, structured data generation, consistency checking, and iterative regeneration to enhance the accuracy and reliability of AI-generated findings for jaw cysts in dental panoramic radiographs.
135. LeafNet: A Large-Scale Dataset and Comprehensive Benchmark for Foundational Vision-Language Understanding of Plant Diseases
Core Problem: The application of Vision-Language Models (VLMs) to domain-specific agricultural tasks like plant pathology is limited by the lack of large-scale, comprehensive multimodal image-text datasets and benchmarks.
Key Innovation: Introduces LeafNet, a comprehensive multimodal dataset (186,000 images, 97 disease classes, 13,950 Q&A pairs) and LeafBench, a VQA benchmark for plant disease understanding. Benchmarking reveals significant disparities in VLM capabilities, showing VLMs outperform vision-only models and highlighting gaps in fine-grained pathogen/species identification.
136. Multimodal Integrated Knowledge Transfer to Large Language Models through Preference Optimization with Biomedical Applications
Core Problem: Scarcity of high-quality multimodal biomedical data limits effective fine-tuning of LLMs for specialized tasks, making it hard to transfer multimodal domain-specific insights to unimodal LLMs.
Key Innovation: Introduces MINT (Multimodal Integrated kNowledge Transfer), a framework that aligns unimodal LLMs with domain-specific decision patterns from multimodal data through preference optimization (e.g., ORPO), enabling LLMs to perform predictive tasks using unimodal inputs while retaining multimodal knowledge, demonstrated effectively in rare genetic disease prediction and tissue classification.
137. ZeroScene: A Zero-Shot Framework for 3D Scene Generation from a Single Image and Controllable Texture Editing
Core Problem: Single image 3D scene reconstruction methods struggle with both individual asset quality and overall scene coherence, while texture editing lacks local continuity and multi-view consistency.
Key Innovation: Proposes ZeroScene, a zero-shot framework leveraging large vision models for single image-to-3D scene reconstruction and controllable texture editing, ensuring geometric/appearance accuracy, scene layout reconstruction, and detailed, consistent textures.
138. APG-DDQN: Expert-guided deep reinforcement learning for AUSV local path planning with modal switching in unstructured environments
Core Problem: Efficient and intelligent local path planning for amphibious unmanned surface vehicles (AUSVs) in unknown, unstructured environments is a prerequisite for task completion.
Key Innovation: An APG-DDQN framework that integrates expert knowledge with deep reinforcement learning to improve AUSV path planning efficiency and quality, featuring an environment perception module, a modal-aware artificial potential field for pre-training, and a carefully designed state/action/reward space.
139. Travel-time-weighted adaptive simplification of sound speed profiles for high-accuracy underwater acoustic ray tracing
Core Problem: High-resolution sound speed profiles (SSPs) are essential for accurate underwater acoustic applications but impose significant computational, storage, and transmission burdens, necessitating simplification while preserving acoustic propagation characteristics.
Key Innovation: Introduction of the Travel-Time-Weighted Adaptive Simplification (TTWAS) algorithm, a novel physics-informed method for SSP layering derived from the acoustic ray travel-time equation, which outperforms conventional geometric methods in accuracy and automatically determines the optimal number of layers.
140. Analysis of tensile mechanical properties of LNG cryogenic flexible pipe
Core Problem: LNG cryogenic flexible pipes, critical for floating LNG systems, are prone to failure under significant tensile loads caused by extreme marine environmental conditions (wind, wave, current) due to their complex multi-layer structure, making tensile mechanical property analysis challenging.
Key Innovation: Established a fine finite element model and constructed an experimental platform to accurately analyze the tensile mechanical properties of LNG cryogenic flexible pipes, validating the model and investigating the influence of structural parameters and friction coefficients, and revealing distinct nonlinear characteristics and increased tensile stiffness in cryogenic conditions.
141. Offshore wind power prediction based on improved Hankel-DMD-LSTM hybrid model: Strengthening mode selection and residual compensation mechanism
Core Problem: Accurate forecasting of offshore wind power is challenging due to the volatility, intermittency, non-stationary, and multi-scale fluctuations of wind energy signals, which hinders grid operation and planning.
Key Innovation: Development of an improved Hankel-DMD-LSTM hybrid model for offshore wind power prediction, which strengthens mode selection using a discrete-time coefficient weighting (DTW) criterion for physically interpretable modes, and integrates an LSTM network for data-driven residual compensation, achieving high accuracy (R2 = 0.937) and low forecasting error.
142. ShipSeer: Pushing accuracy-performance boundaries in ship motion prediction with spectral and multi-component analysis
Core Problem: Accurate long-term ship motion prediction is challenging due to nonlinear ship dynamics, long-term dependencies, multi-step error accumulation, and limited onboard computational resources, impacting maritime safety and operational efficiency.
Key Innovation: Introduction of ShipSeer, a resource-efficient MLP-based multi-input multi-output forecasting framework that combines spectral and multi-component analysis to jointly predict eight ship motion states over long horizons, outperforming state-of-the-art models in accuracy and inference speed on real-world datasets.
143. LEX v1.6.0: a new large-eddy simulation model in JAX with GPU acceleration and automatic differentiation
Core Problem: Current large-eddy simulation (LES) models struggle in the 'gray zone' resolution where subgrid-scale turbulence significantly influences simulated weather systems, requiring novel frameworks for developing new SGS approaches.
Key Innovation: Development of LEX v1.6.0, a new LES model in JAX with GPU acceleration and automatic differentiation, which, when combined with a trained deep learning-based subgrid-scale (SGS) turbulence model, accurately simulates atmospheric phenomena like thermal bubble expansion at gray-zone resolutions, comparable to benchmark LES.
144. Using markets to adapt to climate change
Core Problem: Developing effective and scalable strategies for societal adaptation to the impacts of climate change.
Key Innovation: Proposing the utilization of market-based mechanisms as a tool to facilitate adaptation to climate change.
145. Prediction of the compressive strength of tailings backfill using an EO-LightGBM model: performance comparison and feature importance analysis
Core Problem: Accurate prediction of the compressive strength of tailings-based concrete is essential for mix design, engineering applications, and ensuring structural safety and long-term durability, but traditional methods may lack sufficient accuracy and generalization capability.
Key Innovation: Developed an EO-LightGBM model that accurately predicts the 7-day compressive strength of tailings concrete, achieving superior prediction accuracy (R2 of 0.94) and generalization capability compared to traditional machine learning methods, and identified key influential factors (weight concentration, tailings ratio, cement-sand ratio).
146. Utilizing gridded soil sampling and electromagnetic induction to map salinity in coastal farm fields of North Carolina, USA
Core Problem: Saltwater intrusion is intensifying in coastal agricultural fields, leading to variable patterns of soil salinization that are challenging to monitor across large areas due to heterogeneity, necessitating validation of non-destructive mapping tools.
Key Innovation: Evaluated electromagnetic induction (EMI) for mapping salinity in high-carbon agricultural fields, demonstrating its robustness for characterizing salinity patterns. The study found statistically significant correlations between EMI-derived EC values and both field kit and laboratory EC data, offering a scalable approach for monitoring saltwater intrusion.
147. Measuring and predicting supply chain resilience under disaster and pandemic disruptions: Evidence from Japan’s automotive sector
Core Problem: Quantifying and predicting supply chain resilience (SCR) of Japanese automotive manufacturers under disruptions from natural disasters and pandemics, integrating disaster risk and operational disruption perspectives.
Key Innovation: Developed a hybrid analytical framework combining a logistic regression scorecard and an LSTM model to assess SCR using publicly available data. Identified long employee tenure, stable production, and strong quality control as critical enablers of supply chain continuity and disaster resilience.
148. A system-level semantic reliability framework for diagnosing governance failures in high-risk socio-technical systems
Core Problem: Traditional reliability approaches for high-risk socio-technical systems (e.g., mining, construction) rely on accident records or audits, providing limited insight into governance and safety culture dimensions.
Key Innovation: Develops a semantic diagnostic framework that interprets corporate social responsibility (CSR) reports as indicators of governance performance, using topic modeling and network/multivariate analyses to track governance maturity, coupling, and structural change patterns across industries.
149. A hybrid deep learning and large language models framework for ship collision accident analysis
Core Problem: Effectively analyzing ship collision accidents by leveraging both structured accident data and unstructured accident reports, which traditional methods struggle to integrate comprehensively.
Key Innovation: A novel hybrid framework (DLSCA) combining a Hierarchical graph attention network with Transformer encoding (HGAT-Transformer) for structured data analysis and a fine-tuned LLM with retrieval-augmented generation for accident report analysis, demonstrating consistent results with real accident reports.
150. A novel data-driven health status assessment model based on multiple criteria appraisal recommendation with three-parameter interval grey number
Core Problem: Accurately assessing health status from large-scale historical data, particularly in extracting knowledge (IF-THEN rules) and representing data uncertainties within a multiple criteria appraisal recommendation (MCAR) framework.
Key Innovation: An improved MCAR framework that embeds interval rule-bases for knowledge extraction and defines three-parameter interval grey numbers (TPIGN) with a new distance to capture data uncertainties, using interval evidential reasoning (IER) for accurate appraisals, validated on lithium-ion batteries and turbofan engines.
151. Hierarchical Physics-Embedded Fusion Framework for Multi-Sensor Prognostics with Application to Diamond Wire Breakage and Extended Validation
Core Problem: Accurately predicting equipment remaining useful life (RUL) in complex industrial environments requires effective multi-sensor data fusion and explicit incorporation of domain-specific physical knowledge, which existing methods often overlook.
Key Innovation: A Hierarchical Physics-Embedded Fusion (HPEF) framework that explicitly embeds physical information into each modeling stage (feature extraction, data fusion, decision-making) to integrate multi-sensor data for RUL prediction, demonstrated on diamond wire breakage, enhancing physical consistency and decision reliability.
152. Spatial-temporal graph hybrid neural network for remaining useful life prediction of aero-engines
Core Problem: Effectively fusing multi-sensor data to improve RUL prediction accuracy for aero-engines is challenging due to complex operating conditions, signal variability, and potential loss of degradation information in sequential network structures.
Key Innovation: A spatial-temporal graph hybrid neural network combining Graph Convolutional Network (GCN) and Graph Transformer to effectively fuse multi-sensor information and extract spatio-temporal features for RUL prediction of aero-engines, enhancing the model's ability to capture spatio-temporal correlations.
153. A Physics-Informed Recurrent Neural Network for Long Sequence Remaining Useful Life Prediction of Rolling Bearing with Implicit Function
Core Problem: Accurate long-sequence RUL prediction for rolling bearings is challenging due to dynamic perturbations, error accumulation in multistep prediction, and lack of interpretability in data-driven approaches.
Key Innovation: A Physics-Informed Recurrent Neural Network (PIRNN) integrating a variational recurrent neural network (VRNN) and an implicit function solver (IFS) to capture long-range temporal dependencies, mitigate error propagation, and constrain the solution space using physics-governing PDEs for improved RUL prediction of rolling bearings.
154. High-resolution evolution of sediment yields in the Valencia and Menorca basins during the Neogene: Relation with climate, tectonics and Ebro Basin opening
Core Problem: A lack of high-resolution quantitative sedimentary history for the Valencia and Menorca basins and a clear understanding of the local versus global triggers (climate, tectonics, Ebro Basin opening) impacting sediment yield over the Neogene.
Key Innovation: Provides a high-resolution quantitative sedimentary history for the last 23 Ma, identifying successive trends in sediment flux and demonstrating how sediment yields record multi-timescale geomorphologic processes related to climate and tectonics, supporting a three-phase interpretation for the Ebro system.
155. Runoff-sediment-driven hydro-geomorphological impacts on the saltmarsh dynamics of Yellow River Delta
Core Problem: The hydro-geomorphological thresholds that drive saltmarsh expansion and contraction, particularly the impact thresholds of hydrological connectivity, remain poorly quantified, hindering effective management of their ecosystem services.
Key Innovation: Development of an innovative framework integrating remote sensing, numerical modeling, SHAP values, and ecological optima to quantify the effects of runoff-sediment-driven hydro-geomorphological processes on saltmarsh vegetation dynamics, identifying dominant controls (water salinity, hydrological connectivity, elevation) and distinct habitat preferences for different species, providing actionable guidance for water-sediment regulation and wetland restoration.
156. Different land uses in the Brazilian semi-arid: impacts and drivers on organic matter fractions and soil carbon
Core Problem: Assessing the impact of different land uses (pasture, agriculture) on soil organic matter fractions and carbon stocks in the Brazilian semi-arid Caatinga biome, where soils are susceptible to degradation.
Key Innovation: Quantifying the reduction in particulate and mineral-associated organic carbon stocks due to conventional agriculture, observing initial SOC increase followed by decline in pastures, and identifying soil texture and precipitation as key drivers influencing carbon dynamics in the region.
157. Optimized prediction of peak floor acceleration, peak inter-story drift ratio, and residual inter-story drift ratio in steel moment-resisting frames using innovative techniques
Core Problem: The construction industry requires precise and efficient prediction of key seismic response indicators for buildings, such as Peak Floor Acceleration (PFA), Peak Inter-story Drift Ratio (PIDR), and Residual Inter-story Drift Ratio (RIDR).
Key Innovation: This study utilizes various machine learning algorithms (e.g., DTR, ENR, HGBR, RFR) combined with optimization algorithms (e.g., MFO, PO, BO) to create highly accurate hybrid models for predicting PFA, PIDR, and RIDR in steel moment-resisting frames. It also incorporates feature selection and outlier detection to optimize the calculation process and reduce runtime, achieving high R2 values (e.g., 0.9921 for PFA) for improved prediction performance.
158. Effect of state evolution on the mechanical characteristics of coral sand and associated constitutive model
Core Problem: The classical critical state framework struggles to describe the mechanical response of coral sand due to its pronounced crushability and irregular morphology, leading to a density-dependent critical state line.
Key Innovation: Extensive investigation into the effects of density and gradation on coral sand's critical state evolution, leading to the development of a simplified constitutive model (SANISAND04-St) that incorporates an established state-evolution law, effectively describing its complex mechanical properties.