TerraMosaic Daily Digest: Mar 4, 2026
Daily Summary
Research published on March 4, 2026 reveals a decisive shift from static hazard description to process-resolved geohazard science. Across landslides, debris avalanches, rockfall, and hydro-geomorphic events, the strongest studies couple physically interpretable models with high-resolution observations to resolve initiation, mobility, and downstream impact pathways.
Three contributions define the day. First, failure dynamics are parameterized with stronger material realism, from permeability-controlled debris-avalanche mobility and freeze-thaw mudstone degradation to landslide-induced surge-wave response. Second, monitoring pipelines are becoming simulation-ready: TLS gap reconstruction and 3D discontinuity extraction now feed directly into quantitative hazard models. Third, risk analysis is expanding from hazard intensity to consequence structure, incorporating exposure bias correction, operational warning tools, and community-grounded mitigation design.
Key Trends
Method development is converging on physics-aware, data-fused, and decision-oriented geohazard intelligence.
- Mobility modelling is becoming mechanism-explicit: new studies isolate first-order controls such as permeability, compressibility, weak-layer geometry, and substrate condition, improving the physical credibility of runout and deformation forecasts.
- Measurement-to-model integration is accelerating: advances in TLS, UAV photogrammetry, and 3D structural extraction increasingly produce model-ready boundary and discontinuity data, reducing the gap between field sensing and predictive simulation.
- Hazard chains are treated as coupled systems: landslide deformation, surge-wave generation, and infrastructure response are modeled jointly, enabling consequence assessment that better matches real cascading behavior.
- Early warning frameworks are moving beyond fixed thresholds: multi-source time-series fusion, transfer learning, and uncertainty-aware prediction are replacing single-indicator heuristics for stage recognition and short-term warning.
- Risk science is broadening to implementation conditions: recent work combines physical hazard metrics with exposure structure, governance constraints, and local perception, improving the operational relevance of mitigation strategies.
Selected Papers
This digest features 191 selected papers from 1061 papers analyzed across multiple journals. Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. The 15 September 2022 floods in northern Marche (Central Italy): disaster analysis, case studies and mitigation strategies for hydro-geomorphological hazard
Core Problem: Understanding the causes and impacts of the exceptional 2022 flood and landslide event in northern Marche, Italy, and identifying effective intervention measures to reduce future hydraulic risk, especially given poor maintenance and previous events.
Key Innovation: Presented a systematic and detailed disaster analysis of the 2022 floods and landslides in northern Marche, providing case studies and identifying innovative mitigation measures for reducing hydraulic risk and improving land-use planning and public awareness.
2. Comprehensive assessment of Himalayan glacial lakes concerning their distribution, dynamics, and hazard potential
Core Problem: The need for a comprehensive assessment of the distribution, growth, and Glacial Lake Outburst Flood (GLOF) hazard of glacial lakes across major Himalayan river basins.
Key Innovation: Examines the distribution, growth, and GLOF hazard of glacial lakes across major Himalayan river basins, assessing basin-wise GLOF susceptibility using glacial lake abundance and spatial characteristics.
3. A symmetry-guided mirror reconstruction method for filling large gaps in TLS surveys of mining subsidence
Core Problem: Terrestrial Laser Scanning (TLS) surveys of mining subsidence basins often yield incomplete point cloud data due to complex topography, steep slopes, and occlusion, hindering accurate monitoring of this anthropogenic geological hazard.
Key Innovation: Proposes a symmetry-guided mirrored raster reconstruction (MRR) method, enhanced with a systematic deviation angle, to effectively reconstruct large gaps in TLS-derived DEMs of mining subsidence basins, providing a reliable and scalable solution for routine subsidence monitoring.
4. Three-dimensional (3D) laser scanning–based identification of rock mass discontinuities for rockfall modeling using 3D discontinuous deformation analysis
Core Problem: Accurately identifying rock mass discontinuities and integrating this data into realistic 3D rockfall modeling to enhance reliability of simulations for hazard assessment.
Key Innovation: Presents an integrated framework combining 3D laser scanning for efficient and accurate identification of rock mass discontinuities with 3D Discontinuous Deformation Analysis (DDA) for quantitative rockfall modeling, demonstrated on a steep rock slope.
5. Modeling the Impact of Initiation Parameters on the Dynamics of the 2012 Te Maari Laterally Confined Debris Avalanche
Core Problem: A key challenge in modeling volcanic debris avalanches is choosing input parameters that produce reliable and realistic results, particularly regarding the complex, evolving rheological behavior driven by dynamic changes in pore pressure, internal resistance, and material properties.
Key Innovation: Used the D-Claw model to examine the effect of initial material properties (hydraulic permeability, compressibility, dilatancy) on the 2012 Te Maari debris avalanche. Found that flow mobility is primarily controlled by hydraulic permeability and its interactions, and that different mechanical processes can compensate to produce comparable runout, highlighting the need for future models to incorporate coupled, time-evolving internal material properties for improved hazard assessments.
6. The EAWS matrix, a decision support tool to determine the regional avalanche danger level (Part B): operational testing and use
Core Problem: Ensuring transparency and consistency in regional avalanche danger level assessment across European avalanche warning services, and identifying inconsistencies in the operational application of the EAWS Matrix.
Key Innovation: Analyzed the operational use of the revised EAWS Matrix by 26 European avalanche warning services, identifying consistent danger level assignments, transition zones, and areas needing harmonization (e.g., wet-snow stability assessment), providing empirically based guidance for refinement.
7. Mitigating Mazuku hazards: implementation and effectiveness of local dry-gas degassing measures in the Goma area (Virunga Volcanic Province)
Core Problem: Mitigation of carbon dioxide diffuse degassing hazards (Mazuku) is underexplored, and the effectiveness of local mitigation measures in active volcanic regions like Goma needs assessment, especially considering community perceptions and socio-economic realities.
Key Innovation: Used a mixed-methods approach to assess household perceptions of the implementation and effectiveness of local Mazuku risk mitigation measures in Goma, identifying three categories of measures and highlighting the importance of co-creating context-relevant strategies with local communities.
8. Sea level much higher than assumed in most coastal hazard assessments
Core Problem: Most coastal hazard assessments inadequately handle sea-level and land elevation data, leading to misjudgments of sea level relative to coastal elevation and underestimation of impacts.
Key Innovation: Shows through meta-analyses that measured coastal sea level is significantly higher than assumed in most hazard assessments (based on geoid models), particularly in the Global South, implying substantially greater land and population exposure to a given sea-level rise.
9. Assessment of cyclone risk in madagascar: multi-layer benefit-of-doubt approach
Core Problem: Madagascar's high exposure and multidimensional vulnerability to cyclones necessitate a robust cyclone risk reduction policy, requiring a regional-scale assessment of cyclone risk and its origins.
Key Innovation: Proposed the Hurricane Risk Index for Madagascar (HRI-M) using a multi-layer benefit-of-doubt (MLBOD) approach to assess cyclone risk at the regional level, integrating hazard, exposure, and vulnerability components and highlighting differentiated risk profiles for tailored policy responses.
10. Numerical analysis for landslide deformation and its generated surge waves based on MPM and UAV tilt-photogrammetry modelling technology
Core Problem: The need for accurate and efficient numerical analysis methods to simulate complex landslide deformation and the resulting surge waves, especially for risk management near water bodies.
Key Innovation: A novel numerical analysis method combining enhanced UAV photogrammetry (using improved SIFT and IRANSAC) for high-precision 3D model construction and the Material Point Method (MPM) for simulating landslide deformation and surge waves.
11. Multi-scale structure deterioration mechanisms of the cenozoic red-bed mudstone under freeze-thaw cycles in the guide Basin, Northeastern Tibetan plateau
Core Problem: Understanding the multi-scale structural deterioration mechanisms of Cenozoic red-bed mudstone under freeze-thaw cycles, which are a primary driver of catastrophic events in the northeastern Tibetan Plateau.
Key Innovation: A comprehensive investigation using XRD, SEM, CT, and triaxial shear testing to elucidate the multi-scale deterioration process ('micropore expansion, macropore interconnection, and fracture network formation') and its impact on mechanical properties of red-bed mudstone under freeze-thaw conditions.
12. Fast landslide events in calcareous debris and weathered pyroclastic soils: a case study in Southern Italy
Core Problem: The need to understand the occurrence and characteristics of fast landslides in calcareous debris and weathered pyroclastic soils, particularly their relationship with rainfall, as previous research mainly focused on pyroclastic soils.
Key Innovation: A case study analyzing fast debris slides in Southern Italy, involving mapping, field/drone surveys, geotechnical characterization, and rainfall data analysis, revealing that moderate thickness deposits in calcareous debris can evolve into dangerous events triggered by critical rainfall.
13. Reinforced behaviors of anchored slopes with weak layer: insights into effect of weak layer and anchorage angle
Core Problem: Understanding the deformation and failure mechanisms of anchored bedding rock slopes with weak layers, and optimizing anchorage design to mitigate landslides, especially considering the influence of weak layer percentages and anchorage angles.
Key Innovation: Established a quantitative relationship between weak layer percentages and anchorage effectiveness, identified an optimal anchorage angle (25°) for mitigating rock landslides, and segmented the slope evolution into four stages based on multi-field coupling monitoring data for the first time.
14. Intelligent Identification of Evolutionary Stages of Tunnel Water and Mud Inrush Hazards Based on Multi-source Time-Series Data Fusion: A Transfer Learning and Semi-supervised Learning Approach
Core Problem: Accurate identification of evolutionary stages of tunnel water and mud inrush disasters is challenging due to reliance on single-threshold criteria, sparse monitoring data, and difficulty in label acquisition.
Key Innovation: Proposed an intelligent prediction framework integrating multi-source time-series data (Q, D, u, σ) with a hybrid 1D-CNN and GRU architecture, semi-supervised transfer learning (simulation data as source, in situ as target), and a confidence-guided pseudo-labeling strategy, achieving high accuracy in disaster stage recognition.
15. Study on Hydro-Mechanical Coupling Deformation Mechanism and Control Strategy of Mud-Gushing Tunnel with Large Deformation in Soft Rock
Core Problem: The persistent challenge of water inrush and large deformation in deep-buried tunnels in soft rock, compromising construction safety, and the inadequacy of traditional support methods.
Key Innovation: Systematically investigated hydro-mechanical coupling deformation mechanisms in a specific tunnel using integrated geophysical exploration, numerical simulation, and physical modeling; identified key influencing factors (faults, high stress, expansibility, lithologic contacts); and developed a novel, superior control strategy combining grouting with prestressed (NPR) anchor cables.
16. Probabilistic seismic hazard analysis for El Salvador
Core Problem: Lack of a comprehensive and updated probabilistic seismic hazard analysis (PSHA) for El Salvador, particularly considering specific regional seismic sources and validating ground motion prediction equations with local data.
Key Innovation: Presents a new time-independent PSHA for El Salvador, incorporating a new earthquake catalog, novel characterization of the upper-crustal volcanic chain source, a 3D subduction zone model, and validation of GMPEs with local accelerometric data to derive ASCE-compatible seismic hazard maps.
17. Consequences That Matter: Community Insights from the Southwestern Puerto Rico Seismic Sequence
Core Problem: A documented gap in the integration of social impacts into earthquake risk assessment in Puerto Rico, particularly from the perspective of emergency management stakeholders, leading to incomplete preparedness and mitigation strategies.
Key Innovation: Employed a qualitative research approach (interviews and government report analysis) to identify critical physical, social, and cascading consequences of a potential major seismic event in Puerto Rico from the perspective of emergency management stakeholders, highlighting a gap between stakeholder and government-identified social impacts and informing vulnerability assessment and mitigation strategies.
18. Seismic failure mechanism of prefabricated assembled monolithic subway station structures: investigation by spring-structure system quasi-static pushover tests
Core Problem: The seismic failure mechanisms and performance of novel prefabricated assembled monolithic (PAM) subway station structures are unclear, hindering their widespread adoption despite construction efficiency benefits.
Key Innovation: Conducted destructive comparative tests and numerical simulations on PAM and cast-in-place subway station models using a Spring-Structure System Quasi-Static Pushover Test Method, revealing that PAM structures generally meet design load-bearing capacity, have slightly superior deformation capacity, and are more sensitive to soil constraints than CIP structures.
19. Multi-task prediction framework for microseismic time series data based on hybrid signal decomposition and adaptive loss function
Core Problem: Accurate prediction of microseismic source parameters for ground pressure disaster warning in deep mining is constrained by single-task learning and lacks robust uncertainty quantification.
Key Innovation: Proposes a novel multi-task prediction framework with a dual signal decomposition strategy (VMD and CEEMD) and a Spearman correlation-aware adaptive Pinball loss function to enhance stability and quantify risk through prediction intervals, achieving high accuracy for extreme event prediction.
20. Shear failure behavior of rock under high-stress and multidirectional dynamic disturbances: Insights from true triaxial double-sided shear tests
Core Problem: The shear failure mechanism of rocks under dynamic disturbances and true triaxial stress conditions in deep excavations, leading to instability and risks, remains insufficiently investigated.
Key Innovation: Develops a true triaxial double-sided shear apparatus with dynamic disturbance functionality to conduct tests on sandstone, identifying the evolution laws of strength and deformation under various loading conditions, and proposing critical factors governing burst-type rock failure.
21. Effect of ramp-fault geometry on near-fault ground motion
Core Problem: Understanding how ramp-fault geometry influences near-fault ground motion, as simplified fault representations may underestimate regional seismic hazard.
Key Innovation: Incorporating planar, single-ramp, and double-ramp Main Himalayan Thrust (MHT) geometries into fault slip models for the 2015 Gorkha earthquake, using analytical solutions and the Spectral Element Method, demonstrating that ramp geometries better reproduce recorded near-fault velocity pulses and highlight high-amplification zones.
22. Performance of a shallow-founded building on liquefiable soil at the port of Wellington during the 2013 and 2016 New Zealand earthquakes
Core Problem: Evaluating the seismic performance of a shallow-founded building on liquefiable soil during earthquakes, specifically understanding the mechanisms of damage and settlement observed during the 2013 and 2016 New Zealand earthquakes.
Key Innovation: Using PM4Sand, PM4Silt, and UBCHyst constitutive models in FLAC for nonlinear effective stress site response and 2D nonlinear dynamic analyses, reliably capturing the observed building performance (settlement and damage) due to liquefaction during the 2013 and 2016 earthquakes, and validating design procedures for liquefaction-induced settlement.
23. Seismo‐Acoustic Evidence for Meteoric Water Modulation of Hydrothermal Fluid Discharge
Core Problem: Deciphering the signal of external processes (like meteoric water influx) on active hydrothermal dynamics is a critical challenge in understanding volcanic unrest.
Key Innovation: Revealed that meteoric water influx modulates shallow hydrothermal fluid discharge at Pisciarelli (Campi Flegrei caldera), causing persistent seismic tremor (due to steam condensation) and an inverse relationship with acoustic emissions (steam venting). This leads to pressurization cycles that may promote seismicity, demonstrating seismo-acoustic monitoring as a sensitive tool for detecting transient changes in hydrothermal dynamics, improving volcanic hazard assessments.
24. GreenPhase: A Green Learning Approach for Earthquake Phase Picking
Core Problem: Earthquake detection and seismic phase picking are challenging due to low signal-to-noise ratios and waveform variability, and current deep-learning models are computationally intensive, lacking efficiency, interpretability, and sustainability.
Key Innovation: GreenPhase, a multi-resolution, feed-forward, and mathematically interpretable model based on the Green Learning framework, eliminates backpropagation and achieves excellent performance for earthquake detection and phase picking with an 83% reduction in computational cost for inference.
25. Geographically-Weighted Weakly Supervised Bayesian High-Resolution Transformer for 200m Resolution Pan-Arctic Sea Ice Concentration Mapping and Uncertainty Estimation using Sentinel-1, RCM, and AMSR2 Data
Core Problem: High-resolution pan-Arctic sea ice concentration (SIC) mapping with reliable uncertainty is challenging due to subtle ice features, inexact labels, model uncertainty, and data heterogeneity.
Key Innovation: Proposes a novel Bayesian High-Resolution Transformer approach with: 1) a global and local module for subtle feature extraction, 2) a geographically-weighted weakly supervised loss function for inexact labels, 3) a Bayesian extension for uncertainty quantification, and 4) decision-level fusion of Sentinel-1, RCM, and AMSR2 data, achieving high accuracy and preserving SIC patterns.
26. Hybrid Belief Reinforcement Learning for Efficient Coordinated Spatial Exploration
Core Problem: Coordinating multiple autonomous agents for spatial exploration and serving heterogeneous demand requires jointly learning unknown spatial patterns and planning trajectories efficiently, a task where pure model-based or deep reinforcement learning approaches alone often fall short in terms of adaptivity or sample efficiency.
Key Innovation: A hybrid belief-reinforcement learning (HBRL) framework that combines model-based spatial belief construction (Log-Gaussian Cox Process with Pathwise Mutual Information planner) for information-driven exploration with a warm-started deep reinforcement learning agent (Soft Actor-Critic) for adaptive trajectory control, enhanced by dual-channel knowledge transfer and a variance-normalized overlap penalty for coordinated coverage.
27. Any2Any: Unified Arbitrary Modality Translation for Remote Sensing
Core Problem: Existing cross-modal translation methods for multi-modal remote sensing imagery treat each modality pair independently, leading to quadratic complexity, limited generalization, and an inability to handle frequently incomplete observations.
Key Innovation: Any2Any, a unified latent diffusion framework, performs arbitrary modality translation by projecting heterogeneous inputs into a geometrically aligned latent space with a shared backbone and lightweight residual adapters, supported by the new million-scale RST-1M dataset, significantly outperforming pairwise methods and showing strong zero-shot generalization.
28. REDNET-ML: A Multi-Sensor Machine Learning Pipeline for Harmful Algal Bloom Risk Detection Along the Omani Coast
Core Problem: The need for a reproducible and effective machine learning pipeline to detect and assess the risk of Harmful Algal Blooms (HABs) along coastal regions using multi-sensor satellite data, addressing threats to coastal infrastructure, fisheries, and water supplies.
Key Innovation: REDNET-ML, a multi-sensor machine learning pipeline that fuses Sentinel-2 optical data, MODIS ocean color/thermal indicators, and learned image evidence from object detectors into a compact decision fusion model (CatBoost) to provide calibrated probabilities of HAB risk, supporting operational exploration and demonstrating robust performance.
29. Archeological data with AI- and physics-based modeling explain typhoon-induced disasters in inland China around 3000 yr B.P.
Core Problem: The causes and impacts of extreme disasters and marked social change in inland China around 3000 yr B.P. remain unclear.
Key Innovation: Aligned paleoclimate reconstructions, archeological evidence, and AI/physics-based model simulations to demonstrate that intensified typhoon activities caused considerable impacts on climate extremes and social change in inland China around 3000 yr B.P.
30. Max-stable process framework for intensity‒duration‒frequency analysis and climate projections in monsoon Bangladesh
Core Problem: Traditional statistical approaches for extreme precipitation projections suffer from systematic biases in direct climate model precipitation outputs.
Key Innovation: Develops an innovative covariate-based max-stable process framework for intensity-duration-frequency analysis and climate projections, aiming to overcome biases in traditional extreme precipitation projection methods.
31. Automated burned area detection using machine learning with hybrid training data generation and explainable AI: a comparative analysis of PBIA and OBIA approaches
Core Problem: The increasing frequency and severity of wildfires necessitate rapid and accurate burned area identification, but conventional methods lack scalability and generalizability.
Key Innovation: Developed a comprehensive framework for automated burned area mapping using Sentinel-2 imagery, introducing a hybrid automatic training data generation approach, evaluating four tree-based ML algorithms with optimized hyperparameters (OBIA outperforming PBIA), and using SHAP for explainable AI to identify decisive features (NBR-type indices, SWIR bands).
32. Tsunami evacuation behavior: a systematic review and proposal for a standardized research framework
Core Problem: Fragmentation of knowledge and methodological inconsistencies in tsunami evacuation research, hindering rigorous meta-analysis and effective policy development.
Key Innovation: A proposed standardized research framework, including a 'core questionnaire' and best practices, to enhance comparability across studies and support evidence-based tsunami preparedness and evacuation policies.
33. Place attachment of earthquake victims residing in Antakya (Hatay) after the Kahramanmaraş earthquakes on February 6, 2023
Core Problem: Understanding the factors that influence earthquake victims' place attachment, which can hinder relocation to safer areas and impact the creation of sustainable communities.
Key Innovation: An empirical study using ordered logistic regressions on survey data from 453 earthquake victims to identify specific factors (e.g., injury, loss of relatives, non-evacuation, place-protective behavior) that increase or decrease place attachment after a devastating earthquake.
34. Enhancing volcanic risk communication in Pucón, Chile: a science-based proposal to support decision-making
Core Problem: Challenges in inter-institutional articulation, public risk perception, and clarity of technical information transmission in volcanic risk management in Pucón, Chile.
Key Innovation: A proposed framework for volcanic risk communication across different phases, developed through mixed-method research (media analysis, interviews, focus groups), aiming to improve coordination, strengthen community bonds, and enhance territorial resilience.
35. Seismic retrofit of RC frames using self-centering rocking walls: a performance-based design method and probabilistic fragility assessment
Core Problem: Enhancing the seismic resilience of existing reinforced concrete (RC) frame buildings and developing a practical, reliable retrofit method that accounts for structural and ground motion uncertainties.
Key Innovation: Developed a practical self-centering rocking wall (SCRW) with friction devices for seismic retrofit, established a distributed parameter model to investigate SCRW influence, proposed a performance-based seismic retrofit method, and evaluated its effectiveness through deterministic analyses and probabilistic fragility assessments, demonstrating enhanced post-earthquake recovery capacity.
36. When Flood Warnings Fail: Psychological Predictors of Information Clarity and Evacuation Intention
Core Problem: Traditional flood warning models assume uniform behavioral responses, but real-world evidence shows wide variation in message effectiveness, leading to a lack of inclusive and trusted warning systems.
Key Innovation: Empirically identifies two psychological segments (Responsive and Non-responsive) in flood-prone populations, linking cognitive orientations to message responsiveness, and provides practical guidance for designing audience-segmented, more inclusive, and trusted flood warning systems.
37. Large-scale experimental study on dynamic response of bedrock and tunnel subjected to seismic and train moving loads
Core Problem: Railway bedrock and tunnels are susceptible to dynamic damage and instability under the combined effects of seismic and train moving loads, and their dynamic response in such extreme conditions is not fully understood.
Key Innovation: Conducted large-scale physical model experiments using a train-rail-shaking table test system to investigate the dynamic response of bedrock and tunnels, revealing how acceleration levels evolve with seismic intensity and train speed, and the complex interplay between seismic and train loads on vibration.
38. Momentum-enhanced stochastic gradient MCMC for composite surrogate-based inversion of nonlinear coupled hydro-mechanical problems
Core Problem: Accurately and efficiently determining multiple material parameters (Young’s modulus, Poisson’s ratio, friction angle, cohesive strength, permeability) for nonlinear coupled hydro-mechanical problems in geomechanics is computationally intensive and challenging.
Key Innovation: Introduces a novel data-driven approach combining OED and LHS for dataset construction, a high-precision composite surrogate model (blending XGBoost, CatBoost, LightGBM) to replace finite element calculations, and a Momentum-enhanced Stochastic Gradient Markov Chain Monte Carlo Method (MSGMCMC) for efficient Bayesian inversion, demonstrated on a coastal embankment application.
39. Study on the influence of retaining structures on basal heave failure of excavations in soft clay
Core Problem: Accurately assessing the stability and understanding the failure modes of basal heave in excavations in soft clay, particularly concerning the influence of retaining structure embedded depth and bending strength, is crucial for excavation engineering design.
Key Innovation: Systematically studies basal heave failure modes and stability using Discontinuity Layout Optimization (DLO), revealing staged relationships between safety factor and embedded ratio, proposing an empirical formula for critical embedded ratio, and developing a new stage-dependent simplified method for calculating the safety factor.
40. An elastoplastic analytical method for characterizing the plastic zones around twin circular tunnels excavated at shallow depth
Core Problem: Accurate determination of plastic zone distribution around twin shallow circular tunnels is essential for identifying surrounding rock instability mechanisms and optimizing tunnel design, but existing methods may lack comprehensive analytical solutions.
Key Innovation: Develops a novel elastoplastic analytical method using a new mapping function through conformal transformation to derive theoretical solutions for stress and displacement, and a differential evolution algorithm to solve for plastic zones, providing reliable theoretical guidance for design optimization and construction parameter selection in twin shallow-buried tunnel projects.
41. Sentinel‐5p Reveals Unexplained Large Wildfire Carbon Emissions in the Amazon in 2024
Core Problem: The Amazon region experienced severe wildfires in 2024, but even advanced fire emission models underestimate carbon monoxide (CO) emissions by a factor of 1.5–3, likely due to prolonged smouldering during droughts.
Key Innovation: Quantified and evaluated August-September 2024 Amazon fire carbon emissions using Sentinel-5p satellite CO observations and atmospheric transport/wildfire emission models. Estimated CO emissions at 28–62 Tg, mainly from understorey forest fires, which were about four times larger than the 2018–2023 average. The comparison revealed a significant underestimation by models, hypothesized to be from prolonged smouldering, indicating that 2024 Amazon wildfire carbon emissions are much higher than currently explained.
42. Physics-constrained symbolic regression for discovering closed-form equations of multimodal water retention curves from experimental data
Core Problem: Modeling the unsaturated behavior of porous materials with multimodal pore size distributions is challenging, as standard hydraulic models fail to capture complex characteristics, and superposition approaches lack interpretability and generalizability.
Key Innovation: A physics-constrained machine learning framework using genetic programming automatically discovers closed-form mathematical expressions for multimodal water retention curves directly from experimental data, embedding physical constraints into the loss function to ensure consistency and robustness.
43. Optimization of Cost Functions in Absolute Plate Motion Modeling
Core Problem: Existing absolute plate motion modeling techniques, such as optAPM, can be improved by refining the construction of objective functions to reduce the propagation of modeling errors and achieve more precise historical plate movement reconstructions.
Key Innovation: Modifications to the objective function, including a simpler and more intuitive hotspot cost function and pre-interpolation of hotspot trail data, improve the accuracy and reliability of optAPM outputs, leading to more precise reconstructions of historical plate movements.
44. Hierarchical Inference and Closure Learning via Adaptive Surrogates for ODEs and PDEs
Core Problem: Engineers and scientists often lack a complete understanding of system properties and governing dynamics (e.g., friction laws, damping phenomena) when calibrating physical models (ODEs/PDEs) to data, making inverse problems challenging.
Key Innovation: A principled methodology leveraging hierarchical Bayesian inference to jointly estimate individual model parameters and learn shared unknown dynamics via an ML-based closure model (neural network embedded in ODE/PDEs), combined with a bilevel optimization strategy to train adaptive surrogate forward models (FNO, PINNs) for computational efficiency.
45. On the punch-through potential of a cubic spudcan in sediments with interbedded sand-over-clay
Core Problem: Current methods for forecasting punch-through failure of spudcans primarily concentrate on generic conical spudcans, often disregarding the geometric characteristics of specific spudcans like the cubic design, leading to challenges in assessing their punch-through potential in layered sediments.
Key Innovation: Investigated the punch-through potential of a cubic spudcan using large deformation finite element analyses, demonstrating its positive effects in mitigating punch-through compared to generic spudcans, and developed predictive methods for estimating the increase in peak and post-peak resistance.
46. Response of a loaded pile in soft soils under one-dimensional nonlinear creep consolidation with non-Darcian flow and continuous drainage boundaries
Core Problem: Accurately predicting the long-term axial performance of pile foundations in soft marine clays is challenging due to complex consolidation-driven effective-stress evolution, negative skin friction, and dragload transfer, requiring an integrated framework that accounts for nonlinear creep, non-Darcian flow, and continuous drainage boundaries.
Key Innovation: Presents an integrated one-dimensional nonlinear consolidation–load transfer framework for single end-bearing piles, incorporating non-Darcian flow, an elastic visco-plastic model with a creep-strain limit, and time-dependent continuous drainage boundaries, demonstrating improved agreement with validation cases for predicting pile performance.
47. Keying process of a novel folding-plate anchor: investigation of installation depth loss
Core Problem: Existing mooring anchors require optimized design for deep-sea exploitation, and the depth loss during the keying process of novel folding-plate anchors (FPAs) can significantly affect their bearing performance, necessitating a systematic evaluation.
Key Innovation: Employed validated coupled Eulerian-Lagrangian modeling to systematically evaluate the pretension characteristics of a novel FPA, assessing depth loss, revealing strain characteristics during rotational unfolding, and detailing the soil mobilization mechanism, providing valuable design references for deep-sea anchors.
48. Winkler spring modulus for laterally loaded monopiles in layered soils
Core Problem: Accurately determining the Winkler spring modulus for laterally loaded monopiles in layered soils or weathered rocks is critical for estimating pile responses and natural frequencies of pile-supported offshore wind turbines, but existing models may not adequately address complex layered conditions.
Key Innovation: Developed a rigorous three-dimensional analytical model based on Green's function to investigate the Winkler spring modulus in layered soils, proposing a virtual homogeneous soil model and correction factors for rigid pile segments, which reliably estimates resonant frequencies of prototype OWTs.
49. An optimization study of bit stick-out in a deepwater conductor jetting project based on a pipe-string model
Core Problem: Optimizing bit stick-out is a key factor influencing the efficiency and safety of deepwater surface conductor jet installation, requiring a comprehensive theoretical model and experimental validation to provide feasible solutions for field applications.
Key Innovation: Proposed a modified model for bit stick-out based on rock-breaking theory and bit-conductor size ratio, analyzed soil damage using FEM, and conducted simulation experiments to determine the optimal bit stick-out (149.8 mm), providing design basis and theoretical guidance for jet installation.
50. Advancing operational global aerosol forecasting with machine learning
Core Problem: Traditional aerosol forecasting is complex, uncertain, and computationally expensive due to intricate interactions between aerosol physicochemical processes and atmospheric dynamics.
Key Innovation: Developed AI-GAMFS, a machine-learning-driven system combining vision transformer and U-Net, providing reliable 5-day, 3-hourly forecasts of aerosol optical components and surface concentrations in 1 minute, demonstrating improved performance over existing models for warning against events like dust storms and wildfires.
51. Wide-swath altimetry maps bank shapes and storage changes in global rivers
Core Problem: Global estimates of river water storage magnitude and variability are few and inconsistent, hindering effective water resource management and disaster mitigation due to reliance on sparse observations or incomplete hydrological models.
Key Innovation: Presents near-global-scale observations of active river channel geometry and monthly water storage changes for 126,674 river reaches worldwide, derived from the SWOT mission, revealing distinct patterns and a lower global annual storage variability than previously modelled, highlighting knowledge limitations and opportunities for improved disaster mitigation.
52. Structural vibration monitoring with diffractive optical processors
Core Problem: Current Structural Health Monitoring (SHM) solutions for civil infrastructure are constrained by cost, power consumption, scalability, and data processing complexity.
Key Innovation: A diffractive vibration monitoring system integrating a jointly optimized diffractive layer with a shallow neural network backend to remotely extract three-dimensional (3D) structural vibration spectra, offering a low-power, cost-effective, and scalable solution for SHM and disaster resilience.
53. Data-driven identification of swell potential of clayey soils for engineering surveying using genetic projection pursuit
Core Problem: Accurately classifying swell potential grades of clayey soils remains a significant challenge in geotechnical engineering due to the complex interplay of multiple governing factors.
Key Innovation: Develops a hybrid intelligent model (GAPP) integrating a Genetic Algorithm with Projection Pursuit to optimize projection directions, transforming multi-dimensional soil data into one-dimensional values for accurate, quantitative identification and grading of expansive soil swelling-shrinkage potential.
54. Social memory of Queensland floods: a perspective of 262 newspapers
Core Problem: The process of constructing social memory, despite its recognized role in building community resilience against flood risk, remains systemically unexamined.
Key Innovation: Develops a structured 6W analytical framework to systematically examine how social memory of floods is constructed by newspapers across temporal and spatial scales, identifying shifts in reporting focus, stakeholder roles, narrative themes, and sentiments, which can assist in enhancing community resilience.
55. Spatiotemporal evaluation of downscaled and native high-resolution satellite soil moisture products
Core Problem: Differences in the ability of downscaled and native high-resolution satellite soil moisture (SM) products to represent true spatial features are poorly understood due to limitations in current evaluation strategies.
Key Innovation: Applies the Point-Scale Downsampling (PSD) framework to six high-resolution satellite SM products (Sentinel-1 SAR and downscaled/data fusion products) for a coordinated spatiotemporal evaluation, revealing that enhanced products generally show stronger spatiotemporal agreement with in situ measurements despite lower spatial variability, providing a viable framework for comparative assessment of SM product quality.
56. Quantifying the hydrological services of wetlands with respect to the Strahler order of their river segments
Core Problem: The hydrological services of wetlands, especially their role in mitigating extreme flows and maintaining stability, are critical but not fully quantified, particularly concerning their spatial distribution relative to river segments (Strahler order).
Key Innovation: Uses a semi-distributed hydrological model across 20 diverse subwatersheds to quantify wetland hydrological services, revealing that upstream wetlands (Strahler order 1) disproportionately influence hydrological regulation and that wetland loss amplifies high flows linearly while reducing low flow resilience exponentially.
57. Millennial forearc erosion driven by upper plate faulting, Southeastern North Island, Aotearoa New Zealand
Core Problem: The contributions of various drivers (coseismic/interseismic deformation, forearc faults, subduction interface) to uplift and erosion in subduction zone forearcs over geologic time remain unknown.
Key Innovation: Found high rates of erosion and uplift along the Palliser–Kidnappers coastline, particularly in the south near Cape Palliser, correlating with channel steepness and Late Pleistocene–Holocene uplift rates. This indicates that the highest forearc uplift rates occur above the narrow, locked portion of the Hikurangi subduction zone, influenced by offshore oblique contractional deformation on the Palliser–Kaiwhata fault.
58. Filling Streamflow Data Gaps in Indian Catchments Using Machine Learning and K‐Means Clustering
Core Problem: Reliable and continuous water level and streamflow records are essential for hydrological modeling and water resource management, but observations often suffer from substantial data gaps, limiting the applicability of existing gap-filling methods to large-scale networks.
Key Innovation: Developed a robust framework integrating geomorphological, meteorological, and hydrological parameters with Quantile Regression Forests and K-means clustering to fill daily streamflow data gaps at 343 stations across Peninsular India (1961–2021), achieving high accuracy (NSE > 0.8 for 72-90% of stations) and providing a robust dataset for hydrological modeling.
59. Prediction of Extreme Events in Multiscale Simulations of Geophysical Turbulence using Reinforcement Learning
Core Problem: Accurate subgrid-scale closures are essential for weather/climate models to predict extreme events, but traditional closures have structural errors, and existing AI methods for closure modeling struggle with data requirements, stability, or scalability.
Key Innovation: SMARL, a reinforcement learning approach, develops closures for atmospheric/oceanic turbulence using only the enstrophy spectrum as a reward, enabling stable simulations with significantly fewer degrees of freedom that accurately reproduce high-fidelity statistics and capture extreme events.
60. mHC-HSI: Clustering-Guided Hyper-Connection Mamba for Hyperspectral Image Classification
Core Problem: Existing manifold-constrained hyper-connection (mHC) approaches are not tailor-designed for hyperspectral image (HSI) classification, which requires effective spatial-spectral feature learning and handling complex, heterogeneous data.
Key Innovation: Introduction of mHC-HSI, a clustering-guided mHC Mamba model for enhanced HSI classification. It features a novel clustering-guided Mamba module for spatial-spectral learning, a new residual matrix implementation for soft cluster membership maps (improving explainability), and leverages physically-meaningful spectral band grouping for enhanced interpretability and accuracy.
61. Hazard-Aware Traffic Scene Graph Generation
Core Problem: Maintaining situational awareness in complex driving scenarios is challenging because existing scene understanding methods lack the ability to assess safety-relevance and model traffic-specific relations between prominent hazards and the ego vehicle.
Key Innovation: A novel task, Traffic Scene Graph Generation, and a framework that explicitly uses traffic accident data and depth cues to generate intuitive scene graphs. These graphs stress prominent hazards by color-coding severity, effect mechanism, and relative location to the ego vehicle, enabling hazard-aware ego-centric reasoning.
62. Graph Negative Feedback Bias Correction Framework for Adaptive Heterophily Modeling
Core Problem: Conventional Graph Neural Networks (GNNs) are inherently limited by the homophily assumption, leading to degraded performance on heterophilic graphs where nodes with different labels are connected.
Key Innovation: Graph Negative Feedback Bias Correction (GNFBC), a framework that leverages a negative feedback mechanism and a negative feedback loss to correct bias introduced by label autocorrelation in GNNs, improving performance on heterophilic graphs by incorporating graph-agnostic model output as a feedback term.
63. DAGE: Dual-Stream Architecture for Efficient and Fine-Grained Geometry Estimation
Core Problem: Estimating accurate, view-consistent geometry and camera poses from uncalibrated multi-view/video inputs remains challenging, especially at high spatial resolutions and over long sequences.
Key Innovation: DAGE, a dual-stream transformer, disentangles global coherence from fine detail using a low-resolution stream for view-consistent representation and camera estimation, and a high-resolution stream for sharp boundaries, fused by a lightweight adapter, achieving state-of-the-art results for video geometry estimation and multi-view reconstruction.
64. LiDAR Prompted Spatio-Temporal Multi-View Stereo for Autonomous Driving
Core Problem: Current multi-view stereo approaches for autonomous driving struggle with achieving high metric accuracy, multi-view and temporal consistency, and cross-domain generalization for depth estimation.
Key Innovation: Introduces DriveMVS, a multi-view stereo framework that uses sparse, metrically accurate LiDAR observations as geometric prompts (hard prior and soft feature-wise guidance) and employs a spatio-temporal decoder to ensure consistency across frames, achieving state-of-the-art performance in metric accuracy and temporal stability.
65. Small Object Detection in Complex Backgrounds with Multi-Scale Attention and Global Relation Modeling
Core Problem: Small object detection in complex backgrounds remains challenging due to severe feature degradation, weak semantic representation, and inaccurate localization caused by downsampling operations and background interference.
Key Innovation: Proposes a multi-level feature enhancement and global relation modeling framework, introducing a Residual Haar Wavelet Downsampling module to preserve fine-grained details, a Global Relation Modeling module for long-range dependencies, a Cross-Scale Hybrid Attention module for multi-scale feature fusion, and a Center-Assisted Loss for improved localization.
66. Bridging Human Evaluation to Infrared and Visible Image Fusion
Core Problem: Current Infrared and Visible Image Fusion (IVIF) methods primarily optimize handcrafted losses and objective metrics, often resulting in fusion outcomes that do not align with human visual preferences, limiting their effectiveness in real-world perceptual environments.
Key Innovation: A feedback reinforcement framework is proposed that bridges human evaluation to IVIF, introducing the first large-scale human feedback dataset for IVIF, a domain-specific reward function, and a reward model to quantify perceptual quality, guiding the fine-tuning of fusion networks for human-aesthetics-aligned results.
67. Point Cloud Feature Coding for Object Detection over an Error-Prone Cloud-Edge Collaborative System
Core Problem: Efficient and reliable transmission of point cloud features between edge and cloud devices for object detection is challenging, particularly in meeting low-latency and low-power requirements over error-prone wireless channels.
Key Innovation: A task-driven point cloud compression and reliable transmission framework is proposed, based on source and channel coding, featuring a lightweight feature compaction module, SNR-adaptive channel encoding/decoding, LDPC encoding/decoding, and diffusion-based feature upsampling for robust multi-scale feature reconstruction, achieving significant feature size reduction with high accuracy.
68. Architecture and evaluation protocol for transformer-based visual object tracking in UAV applications
Core Problem: Object tracking from Unmanned Aerial Vehicles (UAVs) is challenged by platform dynamics, camera motion, and limited onboard resources, leading to a lack of robustness or high computational demands for real-time embedded use.
Key Innovation: MATA, a Modular Asynchronous Tracking Architecture combining a transformer-based tracker with an Extended Kalman Filter and ego-motion compensation, along with a new hardware-independent evaluation protocol and metric (NT2F) for quantifying real-time embedded performance.
69. DISC: Dense Integrated Semantic Context for Large-Scale Open-Set Semantic Mapping
Core Problem: Current instance-centric approaches for open-set semantic mapping are bottlenecked by context-depriving and computationally expensive crop-based feature extraction, limiting language-driven robotic perception in large-scale environments.
Key Innovation: DISC (Dense Integrated Semantic Context), a fully GPU-accelerated architecture featuring a novel single-pass, distance-weighted extraction mechanism that derives high-fidelity CLIP embeddings directly from vision transformer intermediate layers, enabling real-time, voxel-level instance refinement for large-scale continuous semantic mapping.
70. GeoSeg: Training-Free Reasoning-Driven Segmentation in Remote Sensing Imagery
Core Problem: Remote sensing lacks a generalizable, reasoning-driven segmentation solution due to the prohibitive cost of reasoning-oriented data and domain-specific challenges like overhead viewpoints.
Key Innovation: Presents GeoSeg, a zero-shot, training-free framework that couples MLLM reasoning with precise localization via bias-aware coordinate refinement and a dual-route prompting mechanism, outperforming baselines on a new diagnostic benchmark.
71. Long-Term Visual Localization in Dynamic Benthic Environments: A Dataset, Footprint-Based Ground Truth, and Visual Place Recognition Benchmark
Core Problem: Long-term visual localization in dynamic benthic environments is understudied due to a lack of curated datasets and precise ground-truthing methods for near-nadir underwater imagery, hindering cost reduction and mapping quality in AUV monitoring.
Key Innovation: Presented a curated dataset for long-term visual localization in benthic environments, a novel footprint-based ground-truthing method for accurate ground-truth links, and a benchmark of VPR methods, advancing AUV-based environmental monitoring.
72. Efficient Point Cloud Processing with High-Dimensional Positional Encoding and Non-Local MLPs
Core Problem: Existing MLP-based point cloud processing models suffer from complex architectures that obscure their strengths and time-consuming local MLP operations, limiting their efficiency and effectiveness.
Key Innovation: HPENets, a suite of MLP networks, introduce a High-dimensional Positional Encoding (HPE) module and replace local MLPs with efficient non-local MLPs, achieving a strong balance between efficiency and effectiveness in point cloud processing across various tasks and datasets.
73. Accurate and Efficient Hybrid-Ensemble Atmospheric Data Assimilation in Latent Space with Uncertainty Quantification
Core Problem: Existing traditional and machine-learning data assimilation (DA) methods for atmospheric state estimation struggle to simultaneously achieve accuracy, efficiency, and uncertainty quantification.
Key Innovation: Proposes HLOBA, a three-dimensional hybrid-ensemble DA method operating in a latent space, which maps forecasts and observations into this space and fuses them via Bayesian update. It matches dynamically constrained 4D DA methods in skill, achieves inference-level efficiency, and provides element-wise uncertainty estimates.
74. FlowCLAS: Enhancing Normalizing Flow Via Contrastive Learning For Anomaly Segmentation
Core Problem: Normalizing Flows (NFs) for anomaly segmentation struggle in dynamic scenes with complex, multi-modal data distributions, failing to efficiently identify out-of-distribution samples and leaving a performance gap to leading discriminative methods.
Key Innovation: Introduced FlowCLAS, a hybrid framework that enhances NFs' maximum likelihood objective with a discriminative, contrastive loss. This explicitly enforces separation between normal and anomalous features in the latent space, achieving new state-of-the-art performance across multiple challenging anomaly segmentation benchmarks.
75. FastLSQ: A Framework for One-Shot PDE Solving
Core Problem: Traditional PDE solvers, including iterative PINN solvers, can be slow and less accurate, especially for complex or high-dimensional problems, hindering rapid simulation and inverse problem solving.
Key Innovation: FastLSQ, a framework for one-shot PDE solving and inverse problems built on sinusoidal random Fourier features with exact analytical derivatives, enabling graph-free operator assembly and achieving orders of magnitude faster and more accurate solutions than iterative PINN solvers.
76. Wind shear enhances soil moisture influence on rapid thunderstorm growth
Core Problem: Predicting the initiation and rapid growth of individual convective storms remains a fundamental challenge, despite understanding large-scale environmental factors.
Key Innovation: Demonstrates that the most extreme thunderstorm initiations are significantly enhanced over soil moisture contrasts when interacting with wind shear, particularly when soil moisture-driven circulations oppose shear-induced cloud displacement, providing a new source of predictability for rapidly developing thunderstorms.
77. Geostatistical modeling of meteorological condition for a decision support system in wildfire resilience management
Core Problem: The need for a systemic method to investigate and model parameters determining wildfire risk and its effects on infrastructure resilience, to optimize disaster preparedness and risk management.
Key Innovation: A methodology employing advanced geostatistical modeling of territorial and temporal parameters and computational fluid dynamics simulations to predict fire risk scenarios, thereby improving understanding of wildfire development and supporting effective risk management strategies.
78. Strength properties of a composite geomaterial for soft soil stabilization
Core Problem: Developing and characterizing an effective composite geomaterial for stabilizing soft soils, which is crucial for improving ground stability in engineering applications.
Key Innovation: Experimentally investigated the unconfined compressive strength (UCS), California Bearing Ratio (CBR), and flexural strength of a composite geomaterial (river dredged soils, GGBFS, waste fishing net), demonstrating that waste fishing net significantly improves flexural strength and changes brittle behavior into ductile response, and established an empirical relationship between UCS, CBR, and secant modulus.
79. Quantifying the Social Costs of Power Outages and Restoration Disparities Across Four U.S. Hurricanes
Core Problem: The multifaceted nature of disaster impact, particularly power outages from hurricanes, leads to disproportionate suffering in sparsely populated regions and a lack of a standardized, welfare-based framework to quantify societal costs and equity in restoration.
Key Innovation: Develops a transferable framework to quantify societal costs of power outages and equity in restoration, providing comparable cross-event evidence linking restoration dynamics to social losses, and actionable spatial analyses for equity-informed prioritization and resilience investments.
80. Integrated Data Imputation of National Inventories for Bridging Information Gaps and Outcome Risk Uncertainty in Community Resilience Modeling
Core Problem: Reliable natural hazard resilience modeling is hindered by data scarcity and incomplete records in public building inventories, leading to uncertainty in quantitative risk analysis.
Key Innovation: Proposes an integrated data imputation approach using stratified Monte Carlo sampling to impute missing exposure attributes in national inventories, generating multiple complete building-inventory realizations to propagate uncertainty into community-level risk metrics, demonstrated for seismic risk, allowing quantification of risk uncertainty due to data incompleteness.
81. Last Interglacial shoreline successions in southeastern Australia: A framework for identifying a waning mantle upwelling, neotectonic movements and sea-level change
Core Problem: Accurately inferring recent tectonism, surface displacement, and past sea-level changes from relict shoreline successions, which are critical for understanding regional geodynamics.
Key Innovation: Reviews and analyzes Last Interglacial shoreline successions across southeastern Australia to infer surface displacement, identifying a waning Cosgrove mantle upwelling, neotectonic movements (e.g., fault displacement, tilting), and refining paleosea level estimates, providing a framework for defining subtle geodetic changes.
82. Differential deformation of small- to medium-scale strike-slip faults and its control on carbonate fracture–cavity reservoir and hydrocarbon accumulation in the ultra-deep Tarim Basin
Core Problem: Understanding how small- to medium-scale strike-slip faults, with their complex deformation patterns and reactivation histories, control the development and heterogeneity of carbonate fracture-cavity reservoirs and hydrocarbon accumulation in the ultra-deep Tarim Basin.
Key Innovation: Provides a detailed analysis of the differential deformation of strike-slip faults in the Tarim Basin, characterizing their segmentation, stratified deformation, and multiple reactivation events, and demonstrating their fundamental control on reservoir evolution, distribution, and hydrocarbon migration and enrichment.
83. KGBDCNet: keyword-guided building damage captioning network for bi-temporal remote sensing images
Core Problem: Current models for building damage captioning struggle to accurately identify damage features and generate comprehensive captions due to the complexity of disaster-affected environments and diverse damage manifestations.
Key Innovation: Introduces the BD-CC dataset (4300 bi-temporal images of conflict-damaged buildings) and proposes KGBDCNet, a keyword-guided captioning model that uses LoRA-fine-tuned Remote-CLIP for multi-modal feature extraction, a cross-modal attention guided module (CAGM) for keyword-enhanced visual features, and a spatial-temporal memory-enhanced module (STMEM) for improved change detection and temporal evolution capture, achieving superior results in building damage captioning and zero-shot damage assessment.
84. A simplified solution to predict surface uplift induced by a point non-isothermal well leakage
Core Problem: The need for rapid and efficient tools to predict and analyze coupled thermo-hydro-mechanical (THM) behavior and ground surface displacements induced by non-isothermal fluid injections in subsurface operations, as numerical simulations are computationally expensive.
Key Innovation: Development of a simplified analytical solution to estimate ground surface displacements due to point non-isothermal injections, combining existing hydraulic deformation solutions with a novel temperature-induced deformation solution, verified against numerical models and field measurements.
85. High Resolution Microscopy and Raman Spectroscopic Studies on the Freshest Mukundpura Meteorite, Rajasthan, India: Presence of Nanodiamond
Core Problem: Understanding the precise composition and geological significance of rare carbonaceous chondrite meteorites, such as the Mukundpura meteorite, requires detailed material characterization.
Key Innovation: High-resolution scanning and transmission electron microscopy combined with Raman spectroscopy confirmed the presence of nanocrystalline diamond and graphitic carbon in the Mukundpura meteorite, linking its high iridium content to impact-related iridium anomalies and mass extinctions.
86. Graph Hopfield Networks: Energy-Based Node Classification with Associative Memory
Core Problem: Developing robust and effective node classification methods for graph-structured data, particularly in scenarios with sparse information or feature masking, and for handling heterophilous graphs.
Key Innovation: Introduction of Graph Hopfield Networks, which combine associative memory retrieval with graph Laplacian smoothing in an energy function for node classification. This approach improves robustness under feature masking and outperforms standard baselines, with tuning enabling graph sharpening for heterophilous benchmarks.
87. Phys4D: Fine-Grained Physics-Consistent 4D Modeling from Video Diffusion
Core Problem: Recent video diffusion models, despite impressive generative capabilities, often struggle with fine-grained physical consistency, exhibiting physically implausible dynamics over time.
Key Innovation: Introduces Phys4D, a three-stage training pipeline that progressively lifts appearance-driven video diffusion models into physics-consistent 4D world representations through pseudo-supervised pretraining, physics-grounded supervised fine-tuning, and simulation-grounded reinforcement learning. Also introduces 4D world consistency evaluation metrics.
88. PhyPrompt: RL-based Prompt Refinement for Physically Plausible Text-to-Video Generation
Core Problem: State-of-the-art text-to-video (T2V) generators frequently violate physical laws despite high visual quality, stemming from insufficient physical constraints in prompts that are difficult to manually add.
Key Innovation: Presents PhyPrompt, a two-stage reinforcement learning framework that automatically refines prompts for physically realistic generation. It fine-tunes an LLM on a physics-focused Chain-of-Thought dataset and applies Group Relative Policy Optimization with a dynamic reward curriculum to achieve synergistic optimization of physical commonsense and semantic adherence.
89. Role-Aware Conditional Inference for Spatiotemporal Ecosystem Carbon Flux Prediction
Core Problem: Accurate prediction of terrestrial ecosystem carbon fluxes is challenging due to strong spatiotemporal heterogeneity, as existing learning approaches implicitly assume a global response function, leading to brittle generalization.
Key Innovation: Proposing Role-Aware Conditional Inference (RACI), a process-informed learning framework that disentangles slow regime conditioners from fast dynamic drivers and incorporates role-aware spatial retrieval, improving accuracy and spatial generalization across diverse environmental regimes.
90. Adaptive Sensing of Continuous Physical Systems for Machine Learning
Core Problem: Traditional methods of learning from physical dynamical systems do not optimize how to measure the system to extract the most useful information for a given machine learning task.
Key Innovation: A general computing framework with a trainable attention module that adaptively learns both where to probe a physical system's state and how to combine these measurements to optimize prediction performance, demonstrated on spatiotemporal fields.
91. QD-PCQA: Quality-Aware Domain Adaptation for Point Cloud Quality Assessment
Core Problem: No-Reference Point Cloud Quality Assessment (NR-PCQA) struggles with generalization due to scarce annotated point cloud datasets, and existing Unsupervised Domain Adaptation (UDA) methods overlook key perceptual quality characteristics.
Key Innovation: QD-PCQA, a Quality-aware Domain adaptation framework for PCQA, comprising Rank-weighted Conditional Alignment (RCA) to align features under consistent quality levels and Quality-guided Feature Augmentation (QFA) to enhance perceptual feature alignment and improve generalization.
92. Harmonic Dataset Distillation for Time Series Forecasting
Core Problem: Time Series Forecasting (TSF) faces significant computational and storage costs due to massive data, and conventional Dataset Distillation (DD) methods are not tailored for time series, suffering from architectural overfitting and limited scalability.
Key Innovation: Proposes Harmonic Dataset Distillation for Time Series Forecasting (HDT), which decomposes time series into sinusoidal bases via FFT and aligns core periodic structures by Harmonic Matching in the frequency domain, achieving strong cross-architecture generalization and scalability.
93. Universal Pansharpening Foundation Model
Core Problem: Existing pansharpening methods are satellite-specific and scene-dependent, severely limiting their generalization across heterogeneous sensors and varied scenes, which reduces their real-world practicality.
Key Innovation: FoundPS, a universal pansharpening foundation model, introduces a modality-interleaved transformer, a latent diffusion bridge model, and infinite-dimensional pixel-to-latent interaction mechanisms to achieve satellite-agnostic and scene-robust image fusion, along with a new large-scale benchmark (PSBench).
94. PatchDecomp: Interpretable Patch-Based Time Series Forecasting
Core Problem: Neural network models for time series forecasting often achieve high accuracy but lack interpretability, limiting human understanding of their predictions.
Key Innovation: PatchDecomp, a neural network-based time series forecasting method that divides input series into patches and aggregates their contributions, providing both high accuracy and clear, visualizable interpretability of patch-wise contributions.
95. Cross-Modal Mapping and Dual-Branch Reconstruction for 2D-3D Multimodal Industrial Anomaly Detection
Core Problem: Existing unsupervised multimodal industrial anomaly detection methods often rely on fragile fusion schemes, memory banks, or teacher-student architectures, limiting robustness under noisy depth, weak texture, or missing modalities.
Key Innovation: CMDR-IAD, a lightweight and modality-flexible unsupervised framework combining bidirectional 2D↔3D cross-modal mapping to model appearance-geometry consistency with dual-branch reconstruction for independent texture and geometric structure capture, integrated via reliability-gated mapping anomaly and confidence-weighted reconstruction anomaly fusion strategies.
96. UniRain: Unified Image Deraining with RAG-based Dataset Distillation and Multi-objective Reweighted Optimization
Core Problem: Most existing image deraining methods are developed for specific types of rain degradation and fail to generalize across diverse real-world rainy scenes (e.g., rain streaks, raindrops, daytime/nighttime conditions).
Key Innovation: Proposes UniRain, a unified image deraining framework that employs a RAG-based dataset distillation pipeline for selecting high-quality training samples and a multi-objective reweighted optimization strategy within an asymmetric mixture-of-experts architecture, achieving state-of-the-art performance across diverse scenes.
97. NOVA3R: Non-pixel-aligned Visual Transformer for Amodal 3D Reconstruction
Core Problem: Limitations of pixel-aligned 3D reconstruction methods, specifically their inability to recover complete scene representations (visible and invisible points) and issues with duplicated structures in overlapping regions, especially from unposed images.
Key Innovation: NOVA3R, a non-pixel-aligned visual transformer that learns a global, view-agnostic scene representation using a scene-token mechanism and a diffusion-based 3D decoder, leading to improved reconstruction accuracy and completeness for both visible and invisible points.
98. Robust Unscented Kalman Filtering via Recurrent Meta-Adaptation of Sigma-Point Weights
Core Problem: The Unscented Kalman Filter (UKF) is limited by the static parameterization of its Unscented Transform, failing to adapt to time-varying dynamics or heavy-tailed measurement noise.
Key Innovation: Introduces the Meta-Adaptive UKF (MA-UKF), which reformulates sigma-point weight synthesis as a hyperparameter optimization problem addressed via memory-augmented meta-learning, dynamically adapting weights to maximize tracking accuracy and maintain consistency.
99. ZipMap: Linear-Time Stateful 3D Reconstruction with Test-Time Training
Core Problem: State-of-the-art 3D reconstruction methods have a computational cost that scales quadratically with the number of input images, making them inefficient for large collections, while sequential approaches sacrifice reconstruction quality.
Key Innovation: Introduces ZipMap, a stateful feed-forward model that achieves linear-time, bidirectional 3D reconstruction by zipping an entire image collection into a compact hidden scene state via test-time training layers, enabling significantly faster and accurate reconstruction.
100. Real-time loosely coupled GNSS and IMU integration via Factor Graph Optimization
Core Problem: Achieving accurate and available positioning in challenging environments (e.g., urban canyons) using GNSS alone is difficult, and traditional FGO methods are often computationally intensive for real-time use.
Key Innovation: Proposes a real-time loosely coupled GNSS and IMU integration architecture using FGO, demonstrating increased service availability in urban canyons compared to batch FGO, while analyzing the trade-offs between accuracy, availability, and computational efficiency.
101. Real-time tightly coupled GNSS and IMU integration via Factor Graph Optimization
Core Problem: Reliable positioning in dense urban environments is challenging due to GNSS signal degradation, and most robust FGO-based GNSS-IMU fusion methods are offline.
Key Innovation: Presents a real-time tightly coupled GNSS-IMU FGO method that enables causal state estimation via incremental optimization with fixed-lag marginalization, evaluating its performance in highly urbanized, GNSS-degraded environments.
102. HBRB-BoW: A Retrained Bag-of-Words Vocabulary for ORB-SLAM via Hierarchical BRB-KMeans
Core Problem: The binary vocabulary in ORB-SLAM, trained with k-majority-based bag-of-words, suffers from precision loss and degradation of visual words due to conventional binary clustering, especially as errors propagate through its hierarchical structure.
Key Innovation: Proposed HBRB-BoW, a refined hierarchical binary vocabulary training algorithm that integrates a global real-valued flow within the clustering process. This preserves high-fidelity descriptor information until final binarization, yielding a more discriminative vocabulary expected to improve ORB-SLAM performance in loop closing and relocalization.
103. FireANTs: Adaptive Riemannian Optimization for Multi-Scale Diffeomorphic Matching
Core Problem: State-of-the-art diffeomorphic image matching methods are slow due to inefficient implementations and ill-conditioned optimization, while deep learning methods lack generalization across diverse image modalities and require extensive training/memory.
Key Innovation: Proposes FireANTs, a training-free, GPU-accelerated multi-scale Adaptive Riemannian Optimization algorithm for fast and accurate dense diffeomorphic image matching, achieving significant speedup and robustness across various matching problems without domain-specific training.
104. Toward Reasoning on the Boundary: A Mixup-based Approach for Graph Anomaly Detection
Core Problem: GNN-based anomaly detection methods struggle with 'boundary anomalies' (subtly camouflaged nodes) due to their reliance on easy negatives in standard Graph Contrastive Learning, leading to simplistic decision boundaries.
Key Innovation: Proposed ANOMIX, a framework that synthesizes informative hard negatives by linearly interpolating normal and abnormal subgraph representations. This graph mixup strategy populates the decision boundary with hard-to-detect samples, enhancing GNN's reasoning capacity for robust graph anomaly detection.
105. Track Anything Behind Everything: Zero-Shot Amodal Video Object Segmentation
Core Problem: Existing amodal completion methods require pretrained class labels, limiting zero-shot inference, and there is a lack of accurate ground truth and specialized evaluation for amodal video object segmentation, especially for completely occluded objects.
Key Innovation: Presented Track Anything Behind Everything (TABE), a novel dataset, pipeline, and evaluation framework for zero-shot amodal completion from visible masks. It enables flexible, zero-shot inference using a single query mask and provides highly accurate ground truth for objects, even when completely occluded.
106. FSMLP: Modelling Channel Dependencies With Simplex Theory Based Multi-Layer Perceptions In Frequency Domain
Core Problem: Multi-Layer Perceptrons (MLPs) used for time series forecasting are prone to overfitting when modeling inter-channel dependencies, especially with extreme values, due to high Rademacher complexity.
Key Innovation: Proposed FSMLP, a novel framework for time series forecasting that introduces a Simplex-MLP layer where weights are constrained within a standard simplex. This encourages learning simpler patterns, reducing overfitting, and demonstrating significant improvements in forecasting accuracy, efficiency, and scalability.
107. Building a Mind Palace: Structuring Environment-Grounded Semantic Graphs for Effective Long Video Analysis with LLMs
Core Problem: Long-form video understanding with Large Vision Language Models (LVLMs) is challenged by the need to analyze temporally dispersed yet spatially concentrated key moments within limited context windows.
Key Innovation: Introduced VideoMindPalace, a framework inspired by the 'Mind Palace,' which organizes critical video moments into a topologically structured semantic graph. This incorporates hand-object tracking, clustered activity zones, and environment layout mapping, enabling LLMs to provide grounded spatio-temporal and 3D contextual insights for long video analysis.
108. Token Adaptation via Side Graph Convolution for Efficient Fine-tuning of 3D Point Cloud Transformers
Core Problem: Existing parameter-efficient fine-tuning (PEFT) methods for 3D point cloud Transformers suffer from high temporal and spatial computational costs during fine-tuning, despite minimizing tunable parameters.
Key Innovation: Proposes STAG (Side Token Adaptation on a neighborhood Graph), a novel PEFT algorithm that uses a graph convolutional side network in parallel with a frozen backbone Transformer to adapt tokens, significantly reducing computation time and memory consumption while maintaining accuracy. Also introduces PCC13 benchmark.
109. Unsupervised Representation Learning - an Invariant Risk Minimization Perspective
Core Problem: Traditional Invariant Risk Minimization (IRM) relies on labeled data, limiting its application in settings where labels are unavailable for learning representations robust to distributional shifts.
Key Innovation: Proposes a novel unsupervised IRM framework that redefines invariance through feature distribution alignment, introducing PICA and VIAE methods to extract invariant directions and separate latent factors from unlabeled data, demonstrating effectiveness in capturing invariant structure and generalizing across environments.
110. TSPulse: Tiny Pre-Trained Models with Disentangled Representations for Rapid Time-Series Analysis
Core Problem: Existing pre-trained time-series models entangle heterogeneous signals into single large embeddings, limiting transferability and zero-shot usability, and suffer from mask-induced bias.
Key Innovation: Proposes TSPulse, a family of ultra-light pre-trained models with disentangled representations (temporal, spectral, semantic views) learned through augmented masked reconstruction and explicit disentanglement, achieving state-of-the-art zero-shot performance and efficient fine-tuning across various time-series diagnostic tasks.
111. Extremely Simple Multimodal Outlier Synthesis for Out-of-Distribution Detection and Segmentation
Core Problem: Lack of supervision signals from unknown data in multimodal Out-of-Distribution (OOD) detection and segmentation leads to overconfident predictions on OOD samples, and existing methods primarily focus on unimodal data.
Key Innovation: Proposes "Feature Mixing," an extremely simple and fast method for multimodal outlier synthesis, which is modality-agnostic and applicable to various modality combinations, achieving state-of-the-art performance with significant speedup on OOD detection and segmentation benchmarks, and introduces the CARLA-OOD dataset.
112. Robust Adversarial Quantification via Conflict-Aware Evidential Deep Learning
Core Problem: Evidential Deep Learning (EDL) models, while efficient for uncertainty quantification, are vulnerable to adversarial and out-of-distribution (OOD) inputs, leading to overconfident and unreliable predictions in high-stakes applications.
Key Innovation: Proposes C-EDL, a lightweight post-hoc uncertainty quantification approach that enhances adversarial and OOD robustness by generating diverse, task-preserving transformations and quantifying representational disagreement, significantly reducing coverage for OOD and adversarial data.
113. Effective Sample Size and Generalization Bounds for Temporal Networks
Core Problem: Standard evaluation protocols for learning from time series conflate raw sequence length with statistical information, leading to misleading generalization assessments due to temporal dependence.
Key Innovation: A dependence-aware evaluation methodology that controls for effective sample size ($N_{\text{eff}}$) rather than raw length, providing end-to-end generalization guarantees for Temporal Convolutional Networks (TCNs) on $\beta$-mixing sequences, and demonstrating that stronger temporal dependence can reduce generalization gaps when $N_{\text{eff}}$ is controlled.
114. Raw-JPEG Adapter: Efficient Raw Image Compression with JPEG
Core Problem: Raw image data requires large storage, making it impractical in constrained scenarios, while JPEG is efficient but not well-suited for raw storage, leading to a trade-off between storage and fidelity.
Key Innovation: Presents RawJPEG Adapter, a lightweight, learnable, and invertible preprocessing pipeline that adapts raw images for standard JPEG compression using spatial and optional frequency-domain transforms, enabling accurate raw reconstruction with higher fidelity than direct JPEG storage.
115. A Geometry-Based View of Mahalanobis OOD Detection
Core Problem: The performance of Mahalanobis-based Out-of-Distribution (OOD) detectors varies widely and unpredictably across modern pretrained representations, lacking a clear understanding of which feature space properties drive this variability.
Key Innovation: Links OOD detection variability to in-distribution geometry (within-class spectral structure and local intrinsic dimensionality), and introduces radially scaled "l"_2 normalization as a geometric control mechanism to improve OOD detection performance by adjusting feature radii based on ID-only geometry signals.
116. BumpNet: A Sparse MLP Framework for Learning PDE Solutions
Core Problem: Traditional methods for solving Partial Differential Equations (PDEs) can be computationally intensive or lack flexibility, and existing neural network approaches may not fully leverage modern MLP training techniques for efficient and accurate solutions.
Key Innovation: Introduces BumpNet, a sparse MLP framework that uses trainable basis functions constructed from sigmoid activations, enabling efficient and accurate numerical solutions for PDEs and operator learning, and proving universal approximation capabilities, with direct applicability to physical models in geohazard simulation.
117. Measurement-Consistent Langevin Corrector for Stabilizing Latent Diffusion Inverse Problem Solvers
Core Problem: Latent diffusion model (LDM)-based solvers for inverse problems frequently suffer from instability due to a discrepancy between the solver dynamics and the stable reverse diffusion dynamics learned by the diffusion model.
Key Innovation: Introduces Measurement-Consistent Langevin Corrector (MCLC), a theoretically grounded plug-and-play stabilization module that remedies LDM-based inverse problem solvers through measurement-consistent Langevin updates, providing a principled stabilization mechanism in latent space, highly relevant for remote sensing data interpretation in geohazard monitoring.
118. Dr.Occ: Depth- and Region-Guided 3D Occupancy from Surround-View Cameras for Autonomous Driving
Core Problem: Existing 3D semantic occupancy prediction methods for autonomous driving struggle with geometric misalignment in view transformation (due to inaccurate pixel-level depth estimation) and severe spatial class imbalance.
Key Innovation: Dr.Occ, a depth- and region-guided occupancy prediction framework. It introduces a D$^2$-VFormer (depth-guided 2D-to-3D View Transformer) for precise geometric alignment using dense depth cues and an R/R$^2$-EFormer (region-guided Expert Transformer) that adaptively allocates region-specific experts to address spatial semantic variations.
119. TIGeR: Tool-Integrated Geometric Reasoning in Vision-Language Models for Robotics
Core Problem: Vision-Language Models (VLMs) are limited to qualitative spatial reasoning and lack the computational precision (centimeter-level accuracy) required for real-world robotic manipulation, failing to leverage metric cues from depth sensors and camera calibration.
Key Innovation: TIGeR (Tool-Integrated Geometric Reasoning), a novel framework that transforms VLMs into geometric computers by enabling them to generate and execute precise geometric computations through external tools, supported by the TIGeR-300K dataset and a two-stage training pipeline, achieving SOTA performance and centimeter-level precision in robotic manipulation.
120. Synthetic Augmentation in Imbalanced Learning: When It Helps, When It Hurts, and How Much to Add
Core Problem: Imbalanced classification causes standard training procedures to prioritize the majority class, performing poorly on rare but important minority cases, and it is unclear when synthetic augmentation helps and how much to add.
Key Innovation: A unified statistical framework showing that synthetic data is not always beneficial, identifying 'local symmetry' and 'local asymmetry' regimes, and proposing Validation-Tuned Synthetic Size (VTSS) to select the optimal synthetic size for improved performance.
121. Rheological and infiltration properties of wheat straw powder-containing slurry for slurry shield tunneling
Core Problem: Difficulty in forming effective slurry membranes in highly permeable formations during slurry shield tunneling, which can impact ground stability.
Key Innovation: Development and characterization of wheat straw powder (WSP) as an environmentally friendly slurry viscosity enhancer, demonstrating improved membrane-forming performance, reduced bentonite use, and detailed rheological and infiltration properties.
122. Enabling fast greenhouse gas emissions inference from satellites with GATES: a Graph-Neural-Network Atmospheric Transport Emulation System
Core Problem: Scaling well-established inverse modeling techniques, particularly those relying on Lagrangian Particle Dispersion Models (LPDM), to modern satellite datasets for greenhouse gas flux estimates faces significant computational challenges.
Key Innovation: Introduced GATES (Graph-Neural-Network Atmospheric Transport Emulation System), a data-driven LPDM emulator that outputs source-receptor relationships approximately 1000 times faster than an LPDM, accelerating satellite-based greenhouse gas emissions estimation.
123. Data-driven discovery and model reduction methods for the atmospheric effects of high altitude emissions
Core Problem: Chemistry transport models are computationally expensive and require specialized expertise, making them impractical for applications needing efficient surrogate models for evaluating atmospheric effects of anthropogenic emissions.
Key Innovation: Investigated data-driven discovery and reduced-order modeling methods (optimized DMD, bagging optimized DMD) to create computationally efficient surrogate models for chemistry transport models, demonstrating their ability to reconstruct and forecast full-atmospheric ozone responses with significantly reduced computational and storage requirements.
124. Do wet or dry soils trigger thunderstorms? It depends on how the wind blows
Core Problem: The complex relationship between soil moisture and thunderstorm initiation, and how this can be leveraged for improved forecasting.
Key Innovation: Satellite data analysis revealing that wind conditions significantly modulate the connection between soil moisture and thunderstorms, providing insights for better weather forecasting.
125. Daily briefing: Galileo’s notes discovered in the margins of an ancient book
Core Problem: Understanding the precise timeline of the initiation of tectonic plate movement on Earth.
Key Innovation: A brief mention of new findings suggesting that tectonic plates might have started moving earlier than previously thought, contributing to fundamental Earth science knowledge.
126. Spatio-temporal analysis of climate-driven drought dynamics in Pakistan using geo-spatial and machine learning approaches
Core Problem: The need for a comprehensive spatio-temporal analysis of climate-driven drought dynamics and its impact on agricultural productivity, moving beyond single indicators.
Key Innovation: A two-tiered analytical framework developing a Composite Drought Index (CDI) using Principal Component Analysis from multiple remote sensing indices, and then applying Kernel Regularized Least Squares (KRLS) machine learning to quantify CDI's impact on crop yields.
127. A comprehensive risk assessment model for wind hazards on railway safety: a case study of China’s high-speed rail
Core Problem: Limitations in existing meteorological disaster prevention systems for railways, lacking precise risk assessment methodologies and reliable wind hazard forecasting capabilities.
Key Innovation: An innovative and comprehensive risk assessment model for railway systems exposed to wind hazards, integrating meteorological, geographical, and operational factors to provide a dynamic spatiotemporal assessment framework.
128. Bayesian Calibration of TBM Cutter Wear Under Geological Uncertainty
Core Problem: Predicting TBM cutter wear accurately and making scientific maintenance decisions under significant uncertainties in rock mass parameters and incomplete geological information.
Key Innovation: Developed a disc cutter wear prediction model based on sliding distance and geotechnical parameters, quantified spatial variability using Monte Carlo simulations, introduced Bayesian updating to dynamically correct model bias with field data, and proposed a machine learning-based rock mass classification model (MLP, XGBoost, weighted ensemble) using TBM operational parameters.
129. A probabilistic framework for clogging assessment in EPB tunnelling
Core Problem: Clogging of fine-grained soils in EPB tunneling is a major source of uncertainty, causing operational disruptions and cost increases, with existing empirical charts not fully accounting for interaction forces or uncertainty.
Key Innovation: Introduced a probabilistic framework for quantifying clogging potential by integrating laboratory pull-out test-soil state correlations with empirical susceptibility criteria using a Bayesian framework; reformulated correlations with Weibull likelihood; defined a joint state-force exceedance contribution; and investigated strain-rate dependency of pull-out resistance, proposing a normalization model.
130. Seismic evaluation of existing low- to mid-rise RC buildings strengthened by a new external steel frame connected with an innovative EWSF joint device
Core Problem: Existing low- to mid-rise RC buildings often lack sufficient seismic resistance, requiring effective retrofitting techniques that enhance lateral load resistance, improve constructability, and allow for retrofitting while the building is in use.
Key Innovation: Proposes and validates a new seismic retrofitting technique using an external welded steel frame (EWSF) with an innovative joint device, demonstrating enhanced load-carrying capacity and effective seismic performance under design basis and maximum considered earthquakes through cyclic loading tests, pseudo-dynamic tests, and finite element analyses.
131. A probabilistic performance-based framework for heat vulnerability and risk assessment of buildings
Core Problem: Existing heat stress assessment for buildings often uses simplified code-based approaches, lacking a unified probabilistic framework that estimates multi-domain consequences and accounts for multi-factor uncertainties.
Key Innovation: Proposes a probabilistic performance-based framework for heat vulnerability and risk assessment of buildings, integrating hazard analysis, building performance evaluation, fragility modeling, and loss estimation with Monte Carlo to quantify multi-domain consequences and propagate uncertainty, enabling informed design decisions and multi-hazard resilience planning.
132. SuperSTF: A latent diffusion model for cloud-free spatiotemporal remote sensing image fusion
Core Problem: Optical remote sensing images are often degraded by clouds and suffer from a trade-off between temporal and spatial resolutions, and existing methods treating cloud removal and spatiotemporal fusion separately can introduce cumulative errors.
Key Innovation: Proposes SuperSTF, an all-in-one latent diffusion model framework that simultaneously reconstructs cloud-free, fine-resolution image series by jointly modeling cloud removal and spatiotemporal fusion, adaptively exploiting their intrinsic correlations and enhancing performance through a Swin Transformer-based autoencoder, cross-attention, cloud location encoding, and acquisition date modulation.
133. Advanced prediction of mine airflow based on WOA-optimized VMD and a deep learning hybrid model
Core Problem: The calculation of required mine airflow is intricate, time-intensive, and often lags behind, posing a threat to mine safety and efficient air supply.
Key Innovation: Proposes a hybrid deep learning model (WOA-optimized VMD, CNN, BiLSTM) for advanced and accurate real-time prediction of mine airflow, achieving greater prediction accuracy and robustness compared to contrast models, thereby addressing delayed calculations and inaccurate air volume predictions.
134. Mechanistic interpretation of microwave-induced rock fracturing: an analytical perspective on thermally-driven stresses
Core Problem: Insufficient understanding of the triggering mechanisms underlying characteristic fracturing behaviors in microwave fracturing of rocks due to the difficulty in isolating individual physical contributions.
Key Innovation: Develops an analytical model for microwave-driven stresses to elucidate the effects of geometric size, free-surface, and external confinement on rock fracturing, providing theoretical support for enhancing and controlling microwave-assisted rock breakage.
135. DeepDiscover: towards autonomous discovery of bucket-type conceptual models – a proof of concept applied to hydrology
Core Problem: Traditional conceptual hydrological models rely on expert-defined structures and equations, limiting scalability, structural diversity, and systematic exploration of alternative process representations.
Key Innovation: Introduced DeepDiscover, a physics-embedded machine learning framework with a modular neural architecture to autonomously infer bucket-type conceptual hydrological models from data. Outperformed benchmark models in streamflow prediction, recovered hydrologically meaningful internal dynamics, and demonstrated physically coherent responses to perturbations, showing feasibility for data-driven process discovery.
136. Radar Specularity Content Indicates a Strong Geothermal Heat Flow Gradient in Antarctica's South Pole Basin
Core Problem: Geothermal heat flow (GHF) is one of the least constrained boundary conditions for the Antarctic Ice Sheet, making accurate estimates critical for predicting basal melting and identifying stable ice sites. Existing GHF models fail to capture observed gradients in basal conditions.
Key Innovation: Evaluated nine published Antarctic GHF models against radar-derived specularity content in the South Pole Basin. Found that no existing GHF map captures the observed gradient in basal conditions better than a uniform field. Radar observations require a spatial GHF gradient aligned with a major ice-sheet and geomorphological boundary, suggesting that shallow geology controls heat at smaller scales than continent-wide products predict.
137. Knowledge Graph and Hypergraph Transformers with Repository-Attention and Journey-Based Role Transport
Core Problem: Integrating knowledge and language representations in a joint training framework while keeping them separable, and effectively encoding structured data (knowledge graphs/hypergraphs) for language transformers, remains a challenge.
Key Innovation: Developed a concise architecture that treats knowledge graphs and hypergraphs as structured instances, encoding them into a key-value repository for a language transformer to attend over, conditioned by journey-based role transport, enabling explicit and inspectable separation between linguistic context and structured knowledge with tight alignment through cross-attention.
138. Beyond Pixel Histories: World Models with Persistent 3D State
Core Problem: Existing interactive world models lack a persistent 3D representation of the environment, leading to poor 3D consistency, limited spatial memory, and unrealistic user experiences.
Key Innovation: Introduces PERSIST, a new world model paradigm that simulates the evolution of a latent 3D scene (environment, camera, renderer) to synthesize new frames with persistent spatial memory and consistent geometry, enabling coherent, evolving 3D worlds and fine-grained, geometry-aware control.
139. Parallax to Align Them All: An OmniParallax Attention Mechanism for Distributed Multi-View Image Compression
Core Problem: Existing distributed multi-view image compression (DMIC) methods treat all images equally, overlooking varying inter-view correlations during decoding, leading to suboptimal coding performance for 3D applications.
Key Innovation: Proposes the OmniParallax Attention Mechanism (OPAM) to explicitly model correlations and align features between arbitrary information sources, and integrates it into a Parallax Multi Information Fusion Module (PMIFM) to create ParaHydra, the first DMIC method to significantly surpass state-of-the-art MIC codecs with high efficiency gains.
140. Extending Neural Operators: Robust Handling of Functions Beyond the Training Set
Core Problem: Neural operators struggle to robustly handle input functions that are out-of-distribution from their training set, limiting their generalization and reliability for complex scientific problems like solving PDEs.
Key Innovation: Develops a rigorous framework for extending neural operators using kernel approximation techniques and Reproducing Kernel Hilbert Spaces (RKHSs), providing theoretical guarantees for reliable extensions and approximation accuracy, and establishing formal relationships between kernel choices and Sobolev Native Spaces to capture function values and derivatives, empirically validated on elliptic PDEs.
141. Local Shapley: Model-Induced Locality and Optimal Reuse in Data Valuation
Core Problem: Exact computation of Shapley values for data valuation is computationally intractable due to the exponential coalition space, and existing accelerations remain global, ignoring model-induced locality.
Key Innovation: Formalizing model-induced locality for Shapley computation and proposing LSMR (Local Shapley via Model Reuse), an optimal subset-centric algorithm that trains each influential subset exactly once, and LSMR-A, a reuse-aware Monte Carlo estimator, significantly reducing retraining operations and speeding up data valuation.
142. EvoPrune: Early-Stage Visual Token Pruning for Efficient MLLMs
Core Problem: The inference efficiency of Multimodal Large Language Models (MLLMs) is severely limited by the exponential growth of visual tokens, especially in complex scenarios, with existing pruning methods overlooking substantial computational costs during visual encoding.
Key Innovation: EvoPrune, an early-stage visual token pruning method for MLLMs that performs pruning directly during visual encoding using a layer-wise strategy guided by token similarity, diversity, and attention-based importance, achieving significant inference speedup with minimal performance degradation.
143. Seeing as Experts Do: A Knowledge-Augmented Agent for Open-Set Fine-Grained Visual Understanding
Core Problem: Fine-grained visual understanding is limited by closed-set taxonomies and single-label prediction, leading to degradation in open-set or context-dependent conditions, and existing agents treat retrieval and reasoning as independent processes.
Key Innovation: Presents KFRA, a Knowledge-Augmented Fine-Grained Reasoning Agent, which uses a three-stage closed reasoning loop (open-vocabulary detection, discriminative region localization, multimodal evidence integration) and establishes a retrieval-grounding coupling to convert retrieved knowledge into spatially grounded evidence for factual, interpretable, and task-agnostic reasoning.
144. Adaptive Enhancement and Dual-Pooling Sequential Attention for Lightweight Underwater Object Detection with YOLOv10
Core Problem: Underwater object detection faces significant challenges due to pronounced visual impairments from light absorption, scattering, and diminished contrast, requiring robust and lightweight solutions for marine surveillance and autonomous systems.
Key Innovation: Proposes a streamlined YOLOv10-based framework integrating a Multi-Stage Adaptive Enhancement module for image quality, a Dual-Pooling Sequential Attention (DPSA) mechanism for multi-scale feature representation, and a Focal Generalized IoU Objectness (FGIoU) loss for improved accuracy and objectness prediction in resource-constrained underwater settings.
145. From Misclassifications to Outliers: Joint Reliability Assessment in Classification
Core Problem: Most prior work treats out-of-distribution (OOD) detection and in-distribution error prediction in classifiers as separate problems, overlooking their closed connection and hindering overall system reliability.
Key Innovation: A unified evaluation framework with new metrics (DS-F1, DS-AURC) for jointly assessing OOD detection and failure prediction, and SURE+, a new approach that significantly improves classifier reliability across diverse scenarios.
146. BD-Merging: Bias-Aware Dynamic Model Merging with Evidence-Guided Contrastive Learning
Core Problem: Existing model merging (MM) methods for multi-task learning often assume clean, distributionally aligned test data, leading to biased predictions and degraded generalization under real-world distribution shifts.
Key Innovation: BD-Merging, a bias-aware unsupervised model merging framework that explicitly models uncertainty using a joint evidential head, quantifies evidential alignment with an Adjacency Discrepancy Score (ADS), and refines representations via discrepancy-aware contrastive learning to train a debiased router for adaptive weight allocation.
147. Towards Generalized Multimodal Homography Estimation
Core Problem: Supervised and unsupervised homography estimation methods suffer substantial performance deterioration when applied to unseen image modalities.
Key Innovation: Proposes a training data synthesis method that generates unaligned image pairs with ground-truth offsets from a single input image, rendering diverse textures and colors while preserving structural information, leading to improved generalization across various domains.
148. Scaling Dense Event-Stream Pretraining from Visual Foundation Models
Core Problem: Learning versatile, fine-grained representations from irregular event streams is challenging due to heavy annotation requirements, limiting scalability, and existing distillation paradigms often lead to semantic collapse.
Key Innovation: Introduces a novel self-supervised pretraining method that distills visual foundation models (VFMs) to scale event representation learning, curating an extensive synchronized image-event collection and proposing a structure-aware distillation loss for higher-quality image-event correspondences.
149. Two-Stage Photovoltaic Forecasting: Separating Weather Prediction from Plant-Characteristics
Core Problem: Existing photovoltaic forecasting methods often omit error-distribution details and do not adequately analyze the source of prediction error when using weather forecasts as input.
Key Innovation: Decomposes photovoltaic forecasting into a weather forecast model (for environmental parameters like solar irradiance and temperature) and a plant characteristic model (for site-specific parameters). It uses satellite-based weather observation as an intermediate layer and analyzes error distributions, showing significant impact of weather forecast errors.
150. PlaneCycle: Training-Free 2D-to-3D Lifting of Foundation Models Without Adapters
Core Problem: Extending powerful 2D foundation models to effectively process 3D volumetric data without requiring extensive retraining, adapters, or architectural redesign.
Key Innovation: PlaneCycle, a training-free, adapter-free operator that reuses pretrained 2D backbones by cyclically distributing spatial aggregation across orthogonal planes, enabling progressive 3D fusion while preserving pretrained inductive biases. It shows intrinsic 3D fusion capability and competitive performance.
151. Scalable Evaluation of the Realism of Synthetic Environmental Augmentations in Images
Core Problem: Evaluating AI systems, especially for rare or safety-critical conditions, requires synthetic test cases, but the realism of generative AI-produced images (e.g., adding environmental conditions like fog, rain, snow, nighttime) needs to be reliably assessed to ensure meaningful evaluation.
Key Innovation: A scalable framework is presented for assessing the realism of synthetic image-editing methods, applied to adding environmental conditions to car-mounted camera images. It uses a vision-language model (VLM) jury for perceptual realism and embedding-based distributional analysis for similarity to genuine adverse-condition imagery, demonstrating that generative AI methods significantly outperform rule-based approaches in generating realistic adverse conditions for evaluation pipelines.
152. Out-of-distribution transfer of PDE foundation models to material dynamics under extreme loading
Core Problem: The utility of PDE foundation models, primarily trained on fluid dynamics, for material dynamics under extreme loading conditions (e.g., shocks, fracture) with highly non-smooth fields is unclear.
Key Innovation: Benchmarks the out-of-distribution transfer of pretrained PDE foundation models (POSEIDON, MORPH) to shock-driven multi-material interface dynamics and dynamic fracture/failure evolution, demonstrating their sample efficiency for terminal-state prediction under distribution shift.
153. Non-Invasive Reconstruction of Cardiac Activation Dynamics Using Physics-Informed Neural Networks
Core Problem: Complex electromechanical interactions governing cardiac arrhythmogenesis are not directly observable in vivo, making it difficult to reconstruct 3D activation dynamics non-invasively.
Key Innovation: A physics-informed neural network (PINN) framework that integrates nonlinear anisotropic constitutive modeling, heterogeneous fiber orientation, weak formulations, and finite-element-based loss functions to accurately reconstruct spatiotemporal cardiac activation dynamics from measurable deformation data, even with noise and reduced resolution.
154. Stable and Steerable Sparse Autoencoders with Weight Regularization
Core Problem: Learned features in Sparse Autoencoders (SAEs) can vary substantially across different training runs (random seeds, choices), impacting their stability and reliability for extracting human-interpretable features.
Key Innovation: Demonstrated that L2 weight regularization, especially when combined with tied initialization and unit-norm decoder constraints, dramatically increases cross-seed feature consistency and steering success rates in SAEs, making the extracted features more stable and functionally controllable.
155. Beyond Mixtures and Products for Ensemble Aggregation: A Likelihood Perspective on Generalized Means
Core Problem: The optimal choice of aggregation method for combining predictions from Deep Ensembles (e.g., linear vs. geometric pooling) remains an open question in machine learning.
Key Innovation: Studied normalized generalized means of order 'r' for ensemble aggregation through a log-likelihood lens, demonstrating that only the range r ∈ [0,1] ensures systematic improvements relative to individual distributions, thereby providing a principled justification for linear and geometric pooling.
156. Semi-Supervised Generative Learning via Latent Space Distribution Matching
Core Problem: Semi-supervised generative modeling of conditional distributions often relies heavily on scarce paired data, limiting its effectiveness and geometric fidelity in generated outputs.
Key Innovation: Introduced Latent Space Distribution Matching (LSDM), a two-stage framework that learns a low-dimensional latent space from both paired and unpaired data, then performs joint distribution matching using only paired data. This approach reduces reliance on scarce paired samples, enables fast one-step generation, and provides theoretical insights into Latent Diffusion Models.
157. Agentics 2.0: Logical Transduction Algebra for Agentic Data Workflows
Core Problem: Agentic AI deployments require robust software quality attributes like reliability, scalability, and observability beyond merely plausible text generation, which current research prototypes often lack.
Key Innovation: Presented Agentics 2.0, a lightweight, Python-native framework based on logical transduction algebra, which formalizes Large Language Model inference as typed semantic transformations. This framework provides semantic reliability through strong typing, observability through evidence tracing, and scalability through stateless parallel execution for agentic data workflows.
158. Catch Me If You Can Describe Me: Open-Vocabulary Camouflaged Instance Segmentation with Diffusion
Core Problem: Existing diffusion-based text-to-image models struggle with learning features of camouflaged individuals due to significant blending between visual boundaries and surroundings, making open-vocabulary camouflaged instance segmentation challenging.
Key Innovation: Proposes a diffusion-based method for Open-Vocabulary Camouflaged Instance Segmentation (OVCIS) that leverages text-image models to learn multi-scale textual-visual features, enabling effective segmentation of camouflaged and novel objects by fusing cross-domain features.
159. Beyond Accuracy: What Matters in Designing Well-Behaved Image Classification Models?
Core Problem: While deep neural networks excel in predictive performance for image classification, they often fall short in other critical quality dimensions like robustness, calibration, or fairness, and a general understanding of 'well-behavedness' is lacking.
Key Innovation: Conducts a large-scale study analyzing nine quality dimensions across 326 backbone models and various training paradigms, revealing insights into factors like vision-language models, self-supervised learning, and dataset size. Introduces the QUBA score (Quality Understanding Beyond Accuracy) to rank models across multiple quality dimensions.
160. Do We Need All the Synthetic Data? Targeted Image Augmentation via Diffusion Models
Core Problem: Existing synthetic data augmentation methods using diffusion models increase dataset size substantially and struggle with diversity, leading to high computational overhead without always optimizing generalization effectively.
Key Innovation: Introduces TADA, a principled framework that selectively augments only the training examples not learned early, using faithful synthetic images. This approach improves generalization by up to 2.8% while augmenting only 30-40% of the data, reducing computational cost.
161. Fast Equivariant Imaging: Acceleration for Unsupervised Learning via Augmented Lagrangian and Auxiliary PnP Denoisers
Core Problem: The inefficiency and performance limitations of existing unsupervised learning frameworks for training deep imaging networks, particularly in the context of Equivariant Imaging, and the lack of efficient test-time adaptation.
Key Innovation: Fast Equivariant Imaging (FEI), a novel unsupervised learning framework that reformulates the Equivariant Imaging problem using an augmented Lagrangian and plug-and-play denoisers, achieving an order-of-magnitude (10x) acceleration and improved generalization for tasks like X-ray CT reconstruction and image inpainting.
162. Bridging Computational Social Science and Deep Learning: Cultural Dissemination-Inspired Graph Neural Networks
Core Problem: Graph Neural Networks (GNNs) face challenges including feature oversmoothing in deep architectures, poor handling of heterogeneous relationships, and monolithic feature aggregation.
Key Innovation: Introduces AxelGNN, a novel GNN architecture based on Axelrod's cultural dissemination model, incorporating similarity-gated interactions, segment-wise feature copying, and global polarization to handle both homophilic and heterophilic graphs and prevent oversmoothing.
163. Talking Trees: Reasoning-Assisted Induction of Decision Trees for Tabular Data
Core Problem: Tabular foundation models, while performing well on low-resource problems, are black boxes that are difficult to interpret and costly for inference. There is a need for more lightweight, controllable, and interpretable alternatives.
Key Innovation: Proposes an agentic setup using reasoning-capable LLMs to induce lightweight, interpretable decision trees for small tabular datasets. This approach combines LLM prior knowledge with user constraints and data learning, achieving competitive performance with state-of-the-art black-box models while providing human-readable reasoning traces and incorporating fairness/monotonicity constraints.
164. Scalable Second-order Riemannian Optimization for $K$-means Clustering
Core Problem: K-means clustering is a hard discrete optimization problem, and existing relaxation algorithms struggle to balance constraint feasibility and objective optimality, making it challenging to compute second-order critical points with rigorous guarantees.
Key Innovation: Provides a new formulation of the K-means problem as a smooth unconstrained optimization over a submanifold, characterizing its Riemannian structures. It introduces a second-order cubic-regularized Riemannian Newton algorithm that solves each Newton subproblem in linear time, achieving significantly faster convergence and optimal statistical accuracy compared to state-of-the-art first-order methods.
165. Topological Alignment of Shared Vision-Language Embedding Space
Core Problem: Multilingual Vision-Language Models (VLMs) are biased towards English and neglect the global geometry of the shared embedding space, leading to suboptimal cross-modal alignment.
Key Innovation: Introduces ToMCLIP, a topology-aware framework that uses persistent homology to define a topological alignment loss, enhancing structural coherence of multilingual representations and improving zero-shot accuracy and retrieval performance.
166. Automatic Map Density Selection for Locally-Performant Visual Place Recognition
Core Problem: Ensuring local performance requirements for Visual Place Recognition (VPR) in long-term deployment is challenging, with the critical factor of reference mapping database density largely neglected.
Key Innovation: A dynamic VPR mapping approach uses pairs of reference traverses to automatically select an appropriate map density, satisfying user-defined requirements for target Local Recall@1 and Recall Achievement Rate (RAR), consistently achieving specified local recall levels and avoiding unnecessary over-densification.
167. Turbulence-induced anti-Stokes flow: experiments and theory
Core Problem: Understanding and quantifying the Eulerian-mean flow generated by the interaction of surface waves and ambient sub-surface turbulence, which partly cancels the Stokes drift, and its implications for ocean transport.
Key Innovation: Experimental evidence and supporting theory for a turbulence-induced anti-Stokes flow, showing that waves encountering ambient turbulence lead to vertical redistribution of Eulerian-mean momentum, with the near-surface ratio of mean current gradient to Stokes drift gradient relating to Reynolds normal stresses.
168. Category-Level Object Shape and Pose Estimation in Less Than a Millisecond
Core Problem: The need for fast and accurate category-level object shape and pose estimation for robotics tasks (manipulation, scene understanding, navigation) that also provides efficient certificates of global optimality.
Key Innovation: A fast local solver for shape and pose estimation that uses a learned front-end for sparse, category-level semantic keypoints and solves a maximum a posteriori optimization problem efficiently with self-consistent field iteration, achieving sub-millisecond speeds and a global optimality certificate.
169. Generalized non-exponential Gaussian splatting
Core Problem: The standard 3D Gaussian splatting (3DGS) relies on exponential transmittance, which can lead to a high number of overdraws and computational cost in complex scenes during radiance field rendering and reconstruction.
Key Innovation: This work generalizes 3DGS to a wider family of physically-based alpha-blending operators, introducing non-exponential variants (e.g., quadratic transmittance) that significantly reduce the number of overdraws and achieve up to 4x speed-ups in rendering complex real-world captures while maintaining similar quality.
170. A robust insnavigation system for underwater vehicles using an adaptive <math xmlns:mml="http://www.w3.org/1998/Math/MathML" altimg="si3.svg" class="math"><msub><mi mathvariant="script">H</mi><mi>∞</mi></msub></math> filter based on a variational Bayesian approach: Development and experimental evaluation
Core Problem: Existing adaptive H∞ methods lack automatic adaptive noise estimation, while pure Variational Bayesian approaches lack worst-case guarantees, making robust and adaptive navigation for AUVs challenging, especially with diverse sensor anomalies.
Key Innovation: Proposes H∞-enhanced Variational Bayesian multiplicative error-state Kalman filter (H∞-VBMAESKF), an adaptive robust navigation framework that synergistically integrates H∞ minimax optimization with Variational Bayesian inference, providing deterministic worst-case robustness bounds and automatic adaptive noise estimation, validated to significantly improve position RMSE and stability under sensor contamination.
171. Heterogeneous agent proximal policy optimization for UAVs-USV formation navigation with multi-modal perception
Core Problem: Existing research on UAVs-USV formation collaborative navigation decision-making (FCND) often decomposes the problem into isolated sub-tasks, lacking a unified framework for simultaneous formation maintenance, multi-agent cooperation, and collision avoidance.
Key Innovation: Proposes a novel FCND strategy based on the Heterogeneous Agent Proximal Policy Optimization (HAPPO) algorithm, integrating an advantage decomposition mechanism, symmetric policy update, a joint state space with multimodal feature fusion, and a comprehensive reward function to achieve robust and collaborative navigation.
172. Event-triggered global uniform asymptotic stabilization of under-actuated surface vessels with input saturation and actuator faults
Core Problem: Existing full-state stabilization control strategies for under-actuated surface vessels (USVs) are limited by their inability to simultaneously handle input saturation in both surge force and yaw moment channels and integrate fault tolerance.
Key Innovation: Presents a novel event-triggered adaptive control scheme for USVs that achieves global uniform asymptotic stabilization despite simultaneous input saturation and actuator faults, using coordinate transformation and online parameter estimation to reduce computational complexity and enhance robustness.
173. Non-planar crack propagation prediction under mixed-mode fatigue in steel jacket structures using machine learning
Core Problem: Detailed fracture mechanics analyses of mixed-mode fatigue crack propagation in offshore steel jacket structures are computationally demanding, hindering routine design and maintenance.
Key Innovation: A multi-fidelity finite element (FE) model validated for crack propagation, used to compute mixed-mode stress intensity factors (SIFs) for training machine learning surrogates (Deep Neural Network achieved highest accuracy), which then predicts crack propagation and fatigue life with high accuracy and significantly reduced computational time.
174. Time-domain fully coupled aero-hydro-servo-elastic analysis of the IEA 15-MW VolturnUS-S semi-submersible FOWT with an in-loop shell-element flexible platform
Core Problem: As floating offshore wind turbines (FOWTs) become larger, the structural flexibility of semi-submersible platforms can bias coupled motion and fatigue predictions, leading to unreliable design assessments.
Key Innovation: A fully coupled aero-hydro-servo-elastic model of a FOWT, embedding a shell-element finite-element platform within a multibody-dynamics framework to enable two-way interactions and accurately capture the impact of platform flexibility on dynamic response and fatigue loads, providing a reference for load-mitigation strategies.
175. A Climate Intervention Dynamical Emulator (CIDER) for scenario space exploration
Core Problem: The computational cost of fully coupled climate model simulations constrains the comprehensive exploration of the vast scenario space for Stratospheric Aerosol Injection (SAI) implementations.
Key Innovation: Developed and evaluated CIDER (Climate Intervention Dynamical EmulatoR), a climate emulator trained on Earth System Model simulations, capable of quickly simulating regional and global responses to novel, out-of-sample SAI deployments at a small fraction of the cost of full ESM simulations.
176. Technical note: Transit times of reactive tracers under time-variable hydrologic conditions
Core Problem: The actual transit times of tracers can differ from water TTDs due to physical processes and tracer input patterns, which is often unexplored, hindering improved understanding of water quality dynamics and solute circulation.
Key Innovation: Derivation of analytical solutions and numerical implementation for TTDs of reactive tracers (sorption, degradation, evapotranspiration) in randomly sampled systems under time-variable flow, demonstrating how reactive transport parameters impact tracer TTDs and highlighting the importance of distinguishing tracer TTDs from water TTDs.
177. Merlin: a computed tomography vision–language foundation model and dataset
Core Problem: The need for automated medical image analysis tools for abdominal CT scans due to high volume and radiologist shortage, with existing vision-language models limited to 2D images and short reports.
Key Innovation: Introduces Merlin, a 3D vision-language foundation model trained on volumetric CT scans, electronic health record data, and radiology reports, demonstrating strong performance across various diagnostic and prognostic tasks and external sites.
178. Radiology AI makes consistent diagnoses using 3D images from different health centres
Core Problem: The growing demand for radiology AI tools and the challenge of achieving consistent and reliable diagnoses from 3D medical images across diverse clinical settings.
Key Innovation: Development of Merlin, a 3D vision-language model, which demonstrated superior and consistent diagnostic performance on abdominal CT scans across multiple hospital sites compared to other models, indicating its potential for broader clinical adoption.
179. Earth’s oldest crystals suggest an early start for plate tectonics
Core Problem: Determining the timing of the onset of plate tectonics in Earth's early history.
Key Innovation: Analysis of Earth's oldest crystals provides evidence suggesting that the planet's crust could have been actively churning due to plate tectonics as early as 3.3 billion years ago.
180. Hydrogeochemical evolution and driving factors of groundwater quality: based on SOM neural networks and spatial analysis in the mining area, Northern Ordos, China
Core Problem: Understanding the mechanisms of groundwater quality deterioration and associated health risks in a coal mining area affected by acid mine drainage and other anthropogenic activities.
Key Innovation: Applied a combination of Self-Organizing Maps (SOM), K-means clustering, hydrogeochemical analysis, and GIS spatial analysis to identify dominant hydrochemical types, categorize groundwater samples into distinct groups, and elucidate the multiple interacting factors controlling regional hydrochemical processes in a mining region.
181. Retardation mechanism and migration of Cr(VI) by the active porous material in modified soil-bentonite barriers
Core Problem: Enhancing the effectiveness of soil-bentonite barriers in retarding the migration of hexavalent chromium (Cr(VI)) from contaminated sites to protect surrounding areas.
Key Innovation: Demonstrated that incorporating active porous materials (active carbon, zeolite) into soil-bentonite barriers significantly improves their adsorption capacity for Cr(VI) and prolongs breakthrough time, with active carbon being more effective, and elucidated the microscopic mechanism of enhanced retardation.
182. Cutting, Localization and Damage in Soft Rocks
Core Problem: Understanding the response of soft rocks to severe plastic deformations during cutting, including localization, fracture propagation, and material damage, to optimize cutting parameters in tunneling and excavations.
Key Innovation: Detailed study of initiation and propagation of localization/fractures using PIV, quantification of damage using grain size distribution, and establishing relationships between cutting depth/speed, damage, and specific energy, providing force estimations from a brittle tensile model.
183. Parametric study on the seismic performance of SRRC columns under strong-axis and weak-axis loadings: simulation and theoretical analysis
Core Problem: Understanding the differential seismic performance of steel reinforced recycled aggregate concrete (SRRC) columns under strong-axis and weak-axis loadings and the influence of various material and geometric parameters on their bearing capacity and ductility.
Key Innovation: Conducted a comprehensive parametric study using developed software to simulate SRRC column seismic performance, identifying significant differences in bearing capacity and ductility under strong-axis vs. weak-axis loading, and proposing an empirical formula for the weak-to-strong-axis bearing capacity ratio.
184. A Metric to Support Higher-Resolution Equity-Based Infrastructure Prioritization
Core Problem: Existing equity metrics for infrastructure prioritization primarily provide single community-wide terms, lacking the resolution needed to accommodate equity-based prioritization at higher, more granular infrastructure divisions.
Key Innovation: Develops an equity-based prioritization framework with a metric derived from individual Theil’s T, enabling higher-resolution evaluation for infrastructure divisions, and demonstrates its ability to support equity-based infrastructure prioritization, showing significant divergence from conventional methods.
185. A novel deep learning framework for fault propagation analysis in nonlinear and dynamic processes
Core Problem: Challenging fault propagation analysis in nonlinear and dynamic industrial processes due to a lack of prior knowledge and complex correlations between variables.
Key Innovation: A novel deep learning framework combining a window-level reconstruction-based contribution method for fault localization and an adversarial learning-based method for dynamic Bayesian network structure learning for causal inference.
186. Label-free mangrove mapping from temporally consistent PlanetScope imagery with interpretable deep unfolding network
Core Problem: Existing mangrove mapping methods struggle with insufficient spatial/temporal resolution, low accuracy, limited generalization, and high demand for labeled samples, while deep learning methods lack interpretability.
Key Innovation: Proposes a label-free sample annotation strategy and a novel interpretable deep unfolding network for mangrove mapping using low-tide, temporally consistent PlanetScope imagery, achieving high accuracy (93.51% to 97.63%) and providing detailed, complete national-scale mapping results for China.
187. Spatiotemporal CNN framework for quantifying crop-specific salinity damage in coastal agriculture
Core Problem: Accurately and timely assessing the vulnerability of coastal agriculture to saltwater intrusion (SWI) at a regional scale, especially quantifying early-stage impacts under crop cover and differential responses across crop types, remains challenging.
Key Innovation: Proposes a machine learning-driven spatiotemporal CNN framework using satellite imagery for crop-specific, monthly mapping of SWI impacts on coastal agriculture, achieving 82.4% to 92.5% identification accuracy in damaged areas and enabling early detection of vulnerable areas and support for timely management strategies.
188. Young water fractions in spring discharge
Core Problem: Quantifying the magnitude and variability of young water fractions (Fyw) in spring discharge, as little is known compared to rivers, to better understand catchment storage, transport, and release of water.
Key Innovation: Quantified Fyw in 469 Austrian springs, finding generally low Fyw (mean 0.06) with variability controlled by aquifer structure (karst springs highest, fracture springs lowest), and distinct responses to hydrologic forcing, providing systematic insight into young water contributions to spring discharge for water resource management and contamination risk assessments.
189. Global shifts in rainfall drought relationship: weakening association in tropics
Core Problem: The traditional association between drought and rainfall deficit is weakening, and there's a need for accurate quantification of drought intensity and understanding of changing precipitation-drought linkages under climate change.
Key Innovation: Revealed a sixfold increase in global drought frequency (1951-2016), primarily due to rainfall deficit, evaporative losses, and rainfall variability. Found the traditional link between drought and rainfall deficit has weakened, with drought likelihood rising by 60% during surplus rainfall years. Attributed tropical droughts increasingly to precipitation variability rather than mean rainfall deficit.
190. Evaluation of adaptation to water scarcity in Farmers’ communities using a co-evolutionary network of agents
Core Problem: Previous socio-hydrology studies often overlook the simultaneous dynamics of collective behavior and individual decision-making in response to hydrological variation, making it difficult to evaluate local communities' adaptability to water scarcity.
Key Innovation: Develops a novel socio-hydrology framework that establishes a two-way bridge between micro-level stakeholder decision-making (ABM, SNA) and macro-level water resources management (MODFLOW), enabling simultaneous analysis of dynamics at both levels and their mutual feedback to evaluate and enhance adaptability to water scarcity.
191. Crystalline swelling of GMZ bentonite under controlled relative humidity: Mineralogical basis and cation effects
Core Problem: A comprehensive understanding of the mineralogical basis and the specific effects of interlayer cations on the crystalline swelling behavior of GMZ bentonite under controlled relative humidity is needed to accurately predict its macroscopic swelling performance in engineered barriers.
Key Innovation: Provided an integrated mineralogical and crystallographic characterization of GMZ bentonite, developed an ex-situ RH-controlled XRD protocol to quantify cation-dependent crystalline swelling, and established a basis for linking crystallographic behavior to macroscopic swelling performance of bentonite barriers.