TerraMosaic Daily Digest: July 19, 2026
Daily Summary
Slope-hazard studies move from correlation toward failure mechanics. A multiphase debris-flow model attributes more than half of basal traction to fine solids and reproduces erosion in the 1990 Tsing Shan event; complementary work resolves finite-softening fracture in snow slabs, the destabilizing bridging effect of poorly anchored piles in reservoir landslides, and crack-controlled failure in humidified expansive soils. Susceptibility and forecasting studies add spatially resolved rainfall information from InSAR coherence, zone-specific parameter distributions in probabilistic TRIGRS, and climate controls on four decades of glacial debris-flow occurrence.
Earthquake and compound-hazard studies isolate mechanisms through controlled contrasts. In the Bedretto laboratory, hydraulic preconditioning preceded an Mw -0.54 event and a distinct aftershock sequence, whereas direct stimulation produced higher rates and a migrating seismicity front without a comparable mainshock. An open-source framework extends deformation inference to the full North American GNSS network on standard computing resources. A coupled atmosphere-ocean-wave counterfactual for Storm Babet shows that a North Sea marine heatwave increased river discharge by 12-18%, wave power by about 9%, and storm surge by about 20%. Correlation-aware PSHA, time-varying dam fragility and combined seismic-water-hammer analysis extend the same emphasis on interacting sources of risk.
The strongest AI and remote-sensing contributions focus on controlled adaptation and explicit uncertainty. Synthetic-only pretraining learns transferable spatiotemporal dynamics; lightweight wrappers adapt frozen time-series foundation models to regional drought without full retraining; and a kilometre-scale diffusion framework couples global reanalysis, a weather foundation model and regional refinement. Conformal methods, probabilistic water mapping, graph-based river forecasting and statewide or global Earth-observation datasets make calibration, spatial structure and transfer conditions visible. Product-level tests likewise show that IMERG skill varies by station, timescale and metric, while SAR-derived vegetation indices remain proxies rather than optical substitutes.
Key Trends
Five methodological shifts connect internal failure mechanics, dynamic observations, coupled counterfactuals, calibrated uncertainty and constrained foundation-model transfer.
- Slope models resolve internal force and phase interactions: Fine-solid traction, pile-mediated force transfer, crack connectivity and finite-softening weak layers are represented explicitly instead of absorbed into bulk empirical coefficients.
- Dynamic observations enter susceptibility and forecasting: InSAR coherence, continent-scale GNSS, water-level cycles, multidecadal climate indices and low-cost displacement imagery add time-varying evidence to otherwise static terrain descriptions.
- Counterfactuals separate coupled hazard drivers: Preconditioned versus direct stimulation and warm-ocean versus cool-ocean simulations quantify how antecedent state changes seismic and storm outcomes.
- Uncertainty becomes an output rather than a footnote: Probabilistic TRIGRS, conformal prediction, Monte Carlo water mapping, runoff distributions and correlation-aware PSHA report calibrated ranges or exceedance probabilities alongside point estimates.
- Foundation models are tested through constrained adaptation: Synthetic pretraining, frozen-model residual correction, source-free domain adaptation and open-vocabulary UAV segmentation target transfer under limited labels and distribution shift.
Selected Papers
The papers below trace hazard behaviour to measurable mechanisms: phase-specific bed traction, weak-layer softening, pile-mediated force transfer, hydraulic preconditioning, ocean heat and observation-dependent uncertainty.
1. Hydraulic Stimulation Experiments Attempting to Enhance Induced Seismicity for Earthquake Physics Research
Core Problem: Induced earthquakes cannot normally be observed under controlled, repeatable changes in fluid pressure, leaving the effect of injection history difficult to separate from stress transfer and environmental forcing.
Key Innovation: The Mzero experiments compare preconditioned and direct hydraulic stimulation in a densely instrumented underground testbed: preconditioning preceded an Mw -0.54 event and aftershock sequence, while direct stimulation produced higher rates and a migrating front without a comparable mainshock.
2. Implementing a unified multi-phase debris flow erosion model featuring phase-specific shear stress for internal mechanics and quantifying fine-solid effects
Core Problem: Debris-flow erosion models often collapse solid and fluid phases into one rheology, obscuring how frictional, viscous and collisional stresses entrain the bed.
Key Innovation: A unified multiphase model reproduces flume dynamics and the 1990 Tsing Shan event, predicts larger velocity and depth than a single-phase treatment, and shows that fine solids supply more than half of bed traction; halving their fraction reduces simulated erosion by 31%.
3. Propagation of weak layer failure in snow slab avalanche release: analytical solutions for a compliant interface with finite softening
Core Problem: Analytical snow-slab release models commonly assume a perfectly brittle weak layer or omit its pre-peak compliance, although both softening and elasticity control crack propagation.
Key Innovation: Finite-softening solutions distinguish the residual crack, process zone and intact region, recover the brittle limit, and extend the formulation to collapse-driven and mixed-mode propagation, with characteristic lengths supported by two- and three-dimensional material-point simulations.
4. Marine heatwave amplifies extreme multi-hazards of extratropical cyclone Babet
Core Problem: Storm attribution rarely quantifies how a pre-existing marine heatwave modifies rainfall, river flow, waves and surge within the same event.
Key Innovation: A kilometre-scale coupled atmosphere-ocean-wave counterfactual isolates the North Sea heatwave’s effect on Storm Babet, finding increases of 12-18% in river discharge, about 9% in wave power and about 20% in storm surge relative to a cooler-ocean simulation.
5. Deformation Laws of Coal Mining-Affected Slopes in Loess Gully Area
Core Problem: Coal extraction beneath loess gullies produces delayed, topographically variable slope deformation that cannot be inferred from the underground working-face geometry alone.
Key Innovation: Five slopes are analysed with 3DEC simulation and orthophoto-derived fractures, showing that deformation peaks above the working face, persists after extraction, and follows different pathways according to slope shape, aspect, gradient, height and overlap across working faces.
6. Mechanisms of interlayer force transmission and slip surface penetration in pile-reinforced reservoir landslides with multi-sliding zones
Core Problem: Stabilizing piles can fail in reservoir landslides with multiple slip zones because their role in transferring force between adjacent slide masses is poorly constrained.
Key Innovation: A three-dimensional fluid-solid model identifies a bridging effect in which inadequately anchored piles transfer shallow residual thrust into a deeper metastable mass, promote slip-surface penetration, and shift deformation among shallow, middle, deep and minimal-sliding modes as pile depth changes.
7. Landslide susceptibility mapping based on InSAR coherence rainfall components and deep learning: a case study of Lanping county in Southwest China
Core Problem: Kilometre-scale rainfall products cannot resolve local hydrological disturbance well enough for landslide susceptibility mapping in complex terrain.
Key Innovation: An InSAR-coherence separation model extracts a rainfall-dominated component as a finer-resolution proxy; within the same deep-learning framework it reaches 0.97 AUC, compared with 0.95 for raw coherence and 0.94 for conventional rainfall data.
8. Centrifugal model test on deformation evolution and failure mechanism of cracked expansive soil slopes under humidification
Core Problem: Conventional crack analogues do not reproduce the coupled permeability and low strength that govern humidification-driven failure of expansive-soil slopes.
Key Innovation: Centrifuge tests with permeable weak crack fills show that connectivity accelerates a sequence from shear-stress concentration to directional deformation and coordinated sliding, linking crack geometry to the transition from local deformation to global failure.
9. Displacement Prediction for Reservoir-induced Landslides under Water Level Fluctuations
Core Problem: Reservoir rim-slope screening needs a simple displacement estimate that captures repeated water-level fluctuations without a full site-specific numerical model.
Key Innovation: Physical model tests varying slope angle, fluctuation rate and cycle count are condensed into a dimensionless spread index and empirical displacement relation, then validated on three Three Gorges Reservoir landslides with errors below 10%.
10. Evaluation of PSHA logic trees considering the Pólya distribution to model the spatial correlation between observation sites
Core Problem: Standard PSHA logic-tree updating can overstate independent evidence when observations at nearby sites are spatially correlated.
Key Innovation: A Pólya-distribution formulation represents site-to-site dependence inside Bayesian logic-tree updating, allowing regional observations to revise model weights without counting correlated records as independent confirmations.
11. Integrated prediction of rainfall-induced landslides using zone parameterization and physical-based probabilistic modelling
Core Problem: Regional physically based landslide prediction is weakened when one parameter distribution is imposed across terrain with distinct geotechnical regimes.
Key Innovation: Zone-specific parameter distributions are coupled with probabilistic TRIGRS, producing 0.76 AUC and 0.69-0.71 balanced accuracy and outperforming non-zoned comparisons by roughly 19% in AUC.
12. Extreme climate effects on the occurrence and frequency of glacial debris flows: evidence from the Parlung Tsangpo Basin, Southeastern Tibetan Plateau
Core Problem: The climatic controls on glacial debris-flow occurrence and recurrence remain difficult to separate because inventories are short and precipitation and heat extremes act on different stages of the process.
Key Innovation: A 1985-2024 inventory in the Parlung Tsangpo Basin links warm-spell duration to predisposition and repeated 5-day precipitation extremes to triggering frequency, separating thermal conditioning from rainfall initiation.
13. A low-cost image-based monitoring system for automatic rock displacement measurement using YOLO
Core Problem: Continuous rock-displacement monitoring is often too expensive for small sites and difficult to automate under changing outdoor illumination.
Key Innovation: A roughly EUR 200 system combines a wildlife camera, crackmeter target and YOLO detection to recover displacement at 0.01 mm resolution, providing a field-oriented alternative to specialized imaging hardware.
14. Effects of stochastic and natural seismic noise on the performance of waveform cross-correlation used to recover low-magnitude seismicity prior to the July 29, 2025, Kamchatka earthquake
Core Problem: Waveform cross-correlation can recover weak seismicity, but stochastic and natural noise may create apparent changes in event rate before a major earthquake.
Key Innovation: Controlled noise tests around the 29 July 2025 Kamchatka earthquake quantify when correlation lowers the detection threshold by about one magnitude and distinguish improved completeness from precursor-like artefacts that require cautious interpretation.
15. An Agentic Interface for End-to-End Probabilistic Seismic Hazard and Risk Analysis
Core Problem: Probabilistic seismic hazard and risk tools expose complex inputs and outputs that are difficult to connect safely through a conversational interface.
Key Innovation: An agentic layer exposes 24 typed OpenQuake operations through the Model Context Protocol and evaluates provenance-preserving calculations across 73 cities, with median results within 5% of reference workflows.
16. Learning Spatio-Temporal Foundation Models from Pure Synthetic Data
Core Problem: Spatiotemporal foundation models usually require large observational archives whose coverage and biases constrain the dynamics they can learn.
Key Innovation: NeoST pretrains entirely on generated dynamical systems and refines latent trajectories at inference, testing whether synthetic variation can supply transferable priors under observational distribution shift.
17. Towards Open Science: Monitoring Crustal Deformations in North America
Core Problem: Crustal-deformation studies often analyse local or subsampled GNSS networks because full-network inference is computationally expensive, potentially obscuring long-wavelength signals and intraplate strain.
Key Innovation: An open-source framework performs inference across the full North American GNSS network on standard computing resources while retaining performance comparable to existing approaches and recovering established tectonic trends.
18. Spectral-Morphological Attention U-Net: An Efficient Network for Active Wildfire Detection
Core Problem: Active-fire segmentation must separate small, irregular burn fronts from cloud, smoke and bright land cover without a computationally heavy architecture.
Key Innovation: Spectral-morphological attention combines band relationships with shape-sensitive features in a compact U-Net, reaching 75.16% IoU on TS-SatFire and improving transfer to Sen2Fire.
19. Apeliotes: A Diffusion-Based Modeling Framework for km-scale Multi-Level Atmospheric Fields
Core Problem: Global weather models do not directly resolve kilometre-scale, multilevel atmospheric structure needed for regional extremes.
Key Innovation: Apeliotes couples global reanalysis, a weather foundation model and conditional diffusion to generate kilometre-scale fields, reporting correlations of 0.91 for wind and 0.99 for temperature and vertical-wind errors below 3% in the evaluated cases.
20. Residual-Guided Multi-Resolution Refinement of Foundation Models: A Case Study in Drought Forecasting
Core Problem: Frozen time-series foundation models can miss regional drought structure at multiple temporal scales, while full fine-tuning is costly and data hungry.
Key Innovation: Residual-guided multi-resolution refinement corrects a frozen model across timescales and reduces drought-forecast MSE by as much as 18.9% in the reported comparisons.
21. Lightweight Wrappers for Adapting Time Series Foundation Models to Regional Drought Forecasting
Core Problem: Regional drought forecasting needs adaptation strategies that work when the underlying time-series foundation model is available only as a black box.
Key Innovation: Lightweight input-output wrappers adapt frozen forecasters without modifying their weights and reduce MSE by up to 26% across the evaluated regional drought tasks.
22. Community-scale assessment of flood-related public health vulnerability using multi-criteria AHP in Northwestern Bangladesh
Core Problem: Flood-health vulnerability varies within communities, but district-scale indicators obscure household differences in water, sanitation, healthcare, relief and adaptation capacity.
Key Innovation: An AHP assessment of 315 households across six unions in Bangladesh identifies flood intensity as the largest contribution and reveals high-vulnerability areas driven by low socioeconomic status and limited relief access despite similar physical exposure.
23. Statewide Forest Monitoring Data for California
Core Problem: California lacks one statewide, analysis-ready record that aligns forest structure, fuels and optical-radar observations at operational mapping scales.
Key Innovation: The dataset harmonizes airborne LiDAR, Sentinel-1, Sentinel-2 and PlanetScope into statewide 10 m and 3 m products, providing open training and validation material for forest and wildfire-fuel monitoring.
24. GEM-Forest: A Global satellite EMbedding–based map of forests and tree crops for 2020
Core Problem: Global forest products often merge natural forest with tree crops and lose local detail when transferred across biomes.
Key Innovation: GEM-Forest uses Alpha Earth satellite embeddings to map forests and tree crops globally at 10 m for 2020, reporting overall accuracies of 88-92% across evaluation settings.
25. The complex effect of climate change and urbanization on streamflow in small-medium eastern Mediterranean catchments
Core Problem: Climate and urbanization can push runoff in opposite directions, so treating either driver alone can misstate future flood response in small Mediterranean catchments.
Key Innovation: Scenario modelling finds urbanization alone raises peak flow by 43% and volume by 41%, climate alone lowers them by 21% and 30%, and their combination produces net increases of about 13%.
26. Enhancing Flood Susceptibility Mapping Through High-Resolution Earth Observation: A Data-Driven Comparative Analysis
Core Problem: Flood-susceptibility maps in rapidly changing terrain are limited by coarse predictors and by untested choices among conventional bivariate weighting methods.
Key Innovation: A 12.5 m Earth-observation analysis compares frequency ratio and weights of evidence; the latter reaches 0.945 AUC and maps substantial built-up area into high and very high susceptibility classes.
27. Mapping Thermokarst Lakes Using Sentinel-2 Imagery in the Qinghai–Tibet Engineering Corridor in 2020
Core Problem: Thermokarst-lake inventories along infrastructure corridors need object boundaries that remain reliable across spectrally heterogeneous permafrost terrain.
Key Innovation: A Sentinel-2 convolutional workflow maps 2020 lakes along the Qinghai-Tibet Engineering Corridor with 98.04% overall accuracy, 97.18% F1 and 0.79 polygon IoU.
28. Hybrid Deep Learning–Monte Carlo-Based MNDWI Ensemble for Probabilistic Water Extent Mapping Under Dynamic Spectral Flow Conditions in Large Waterbodies
Core Problem: Single-threshold water indices provide no spatial uncertainty and are unstable where flow, turbidity and mixed pixels change the spectral boundary of large waterbodies.
Key Innovation: A U-ResNet and 1,000 Monte Carlo MNDWI realizations produce probabilistic water extent, achieving 0.92-0.99 accuracy and below 2% false detection while identifying persistent flood-prone hotspots.
29. Atmospheric state along the European coasts during high frequency (T < 2 h) sea level extremes: a view from above
Core Problem: High-frequency coastal sea-level extremes are difficult to attribute because atmospheric pressure, winds, waves and local geometry interact over periods shorter than two hours.
Key Innovation: A coast-wide European analysis characterizes the atmospheric states associated with these short-period extremes, supplying an event-scale basis for distinguishing common and region-specific forcing patterns.
30. Surface subsidence and stability analysis on reconstructed salt cavern for underground gas storage
Core Problem: Reconstructed salt caverns can remain mechanically uncertain after conversion to gas storage because subsurface geometry and damage are incompletely observed.
Key Innovation: Surface-subsidence observations are coupled with stability analysis to invert cavern behaviour and diagnose whether reconstruction has restored an acceptable mechanical state for underground storage.
31. CO2 Fracturing in Pre-conditioned Reservoir Rocks Maximizes Geothermal Extraction Potential While Minimizing Induced Seismicity
Core Problem: Geothermal stimulation must create connected fractures without producing concentrated microseismic release, and the interaction between fluid choice and prior conditioning is uncertain.
Key Innovation: Laboratory comparison shows that CO2 injection into preconditioned rock creates the most extensive fracture network while producing the lowest peak microseismicity among the tested treatments.
32. Integrating UAV image time-series with multi-source environmental features for sequential wildfire spread prediction
Core Problem: Wildfire-spread models trained on scarce field sequences struggle to transfer from idealized simulations to noisy UAV observations.
Key Innovation: A synthetic-to-UAV sequential model fuses image time series with environmental drivers, reaches 0.921 Jaccard similarity, reduces parameters by 70%, and runs in 47.8 ms per prediction in the reported tests.
33. Learning geometry-aware streamflow forecast models from a large-scale river graph dataset
Core Problem: River forecasts are usually trained basin by basin, limiting transfer across large, connected drainage networks with irregular geometry.
Key Innovation: A geometry-aware graph model is trained on more than 500 basins, about 100,000 river nodes and roughly 6 billion observations, attaining Kling-Gupta efficiency above 0.8 at nodes and above 0.96 at outlets.
34. Time-variant seismic fragility analysis framework for arch dams considering material deterioration and stochastic ground motions
Core Problem: Arch-dam fragility changes as materials deteriorate, but conventional assessments hold structural properties fixed over the service life.
Key Innovation: A time-variant framework propagates material ageing and stochastic ground motions into evolving seismic fragility, separating present-day safety from deterioration-driven changes in exceedance probability.
35. SkyVLaM: Multimodal Large Language Model for UAV Video Understanding in Remote Sensing
Core Problem: UAV video understanding lacks benchmarks that join temporal reasoning, fine-grained object recognition and natural-language questions over remote-sensing scenes.
Key Innovation: SkyVLaM contributes 101 videos, 33,600 frames and 1.53 million object instances with multimodal annotations, then benchmarks vision-language models on temporal and spatial UAV-video understanding.
36. LFM: Leveraging Foundation Models for Source-Free Universal Domain Adaptation
Core Problem: Source-free universal domain adaptation must handle both covariate and label shift without revisiting source data, including target classes that were absent during source training.
Key Innovation: LFM combines vision-language similarities, LLM-generated names for unknown classes, mixture-model detection and source-target consensus to refine pseudo-labels across multiple label-shift settings.
37. Adaptive Mamba Neural Operators
Core Problem: Neural PDE solvers lose accuracy when one representation must span point clouds, structured meshes and irregular domains.
Key Innovation: Adaptive Mamba Neural Operators replace a fixed state-space integral with reproducing kernels built from Takenaka-Malmquist systems and report lower relative L2 error than the compared solvers across fluid, solid and finance PDE benchmarks.
38. Isotonic Conformal Prediction
Core Problem: Marginally calibrated regression can remain biased conditional on the predicted value, while exact self-calibrating conformal methods are costly for continuous outcomes.
Key Innovation: Isotonic Conformal Prediction fits one monotone recalibration map and forms intervals within prediction strata; split and transductive variants trade asymptotic versus finite-sample self-calibration while retaining prediction-conditional coverage at far lower cost than repeated refitting.
39. Do Maps Still Matter for Machines: Revisiting the Role of Choropleth Maps in Foundation Model Spatial Understanding
Core Problem: It is unclear whether foundation models gain useful spatial structure from a map when the same geographic values are already available as machine-readable data.
Key Innovation: ChoroplethMap-Bench tests 22 models on 2,400 maps and 12,000 questions under data-only, map-only and combined inputs; the combined condition performs best, particularly for higher-order pattern reasoning.
40. Time-Aware Prior Fitted Networks for Zero-Shot Forecasting with Exogenous Variables
Core Problem: Most time-series foundation models ignore known external drivers even when weather, prices or calendar effects cause abrupt regime changes.
Key Innovation: ApolloPFN builds exogenous dependence and temporal structure into its synthetic pretraining prior and architecture, then outperforms the evaluated zero-shot baselines on four forecasting benchmarks with external covariates.
41. SemDINO: DINOv3-Guided Cross-Temporal Semantic Alignment Network for Remote Sensing Change Detection
Core Problem: Foundation-model features for remote-sensing change detection can be spatially misaligned and sensitive to the order of the two observation dates.
Key Innovation: SemDINO fuses DINOv3 semantics with convolutional detail, uses bidirectional temporal interaction to reduce order bias, and separates real transitions from pseudo-change; it leads the reported semantic and binary results across five benchmarks.
42. Spectral Adaptive Conformal Prediction for Structured Non-Exchangeable Data
Core Problem: Classical conformal intervals lose their finite-sample guarantee when seasonal regimes and frequencies make calibration and test observations non-exchangeable.
Key Innovation: Spectral adaptive conformal prediction weights residuals by local frequency similarity and adjusts miscoverage online, with bounds that separate spectral mismatch from effective sample size and a safeguard that detects the observed low-sample failure case.
43. The DTU25 mean sea surface: from and for SWOT
Core Problem: Conventional mean-sea-surface models smooth short-wavelength and coastal geodetic features that SWOT now observes directly.
Key Innovation: DTU25 combines long conventional-altimetry records with almost two years of SWOT, uses 250 m data within 40 km of coasts, and increases resolved spatial detail by about 30% while reducing geodetic leakage into ocean signals.
44. The Greenland GNSS Network (GNET): geodetic grade GNSS measurements of Greenland's 3D bedrock displacement from 1995-2025
Core Problem: Greenland bedrock-motion studies have relied on partial releases from a remote GNSS network whose station histories and outages complicate long-term analysis.
Key Innovation: GNET releases daily RINEX data, processed east-north-up series and installation metadata for 71 bedrock stations spanning 1995-2025, creating a traceable record of three-dimensional crustal response around the ice sheet.
45. Accurate and robust geometric algorithms for regridding on the sphere
Core Problem: Spherical regridding pipelines combine many geometric predicates and clipping operations whose floating-point failure modes are rarely documented as one system.
Key Innovation: The study organizes non-conservative and conservative regridding into a common set of spherical kernels, supplies stable formulas and error characterizations, and identifies where error-free transformations are needed for additional precision.
46. State-Space-Based Mamba and Convolutional Neural Network Integration for Anthropogenic Geomorphic Feature Extraction from Land Surface Parameters
Core Problem: Claims that state-space architectures improve terrain segmentation have rarely been tested against tuned convolutional baselines across different anthropogenic landforms.
Key Innovation: A three-task comparison finds the hybrid Mamba-CNN model most consistent for broad spatial continuity, but its gains are modest and task dependent; CNN variants retain the sharpest boundaries and sometimes the best overall result.
47. Reliability Analysis of Agricultural Foundation Models Under Distribution Shift
Core Problem: Operational foundation-model monitoring often assumes that a large representation shift implies a large prediction error.
Key Innovation: A Prithvi-based crop-yield pipeline measures input- and latent-space drift across normal and anomalous seasons; latent distances separate drift most clearly, but accurate out-of-distribution and inaccurate in-distribution cases show that drift score alone is not an uncertainty estimate.
48. Unified UAV Open-Vocabulary Semantic Segmentation: Benchmark Construction and LLM-Guided Text–Visual Enhancement
Core Problem: Open-vocabulary segmentation transfers poorly to UAV imagery because category wording and ground-level visual priors do not match small aerial objects and viewpoints.
Key Innovation: A harmonized cross-dataset benchmark pairs LLM-expanded UAV descriptions with DINO geometric features; UAV-OVSeg reaches 65.5% and 64.5% mIoU in two transfer settings, 2.6 points above CAT-Seg in each.
49. Seismic Inversion Based on Mamba and ResUNet++
Core Problem: Convolutional seismic inversion has limited reach along traces and can miss broad subsurface trends without a low-frequency starting model.
Key Innovation: DMamUNet inserts Mamba sequence blocks and deformable convolutions into ResUNet++, adds multitrace supervision and reconstruction loss, and lowers blind-well MSE by 78.22% versus FCRN and 14% versus TransUNet in the post-stack field test.
50. Atmospheric Dust as an Air Quality Hazard to the World Population
Core Problem: The global population exposed to fine atmospheric mineral dust is poorly constrained when dust size and health-based concentration thresholds are not resolved consistently.
Key Innovation: Satellite and population analysis estimates that about 33.5% of people, roughly 2.5 billion, experience submicrometre dust concentrations above the referenced PM2.5 recommendation.
51. Construction of relationships among acceleration, velocity and displacement response spectra using machine learning
Core Problem: Engineering applications often provide only acceleration, velocity or displacement response spectra, while existing conversions among them rely on simplified regressions.
Key Innovation: Bayesian-optimized LSTM relations trained on 16,660 records from 338 stations map among all three spectra and deliver more accurate, stable estimates than the compared conversion models across four site classes.
52. A new method to extract strong velocity pulse in a pulse-like ground motion
Core Problem: Wavelet-based extraction can mislocate or distort the strong velocity pulse that controls near-fault structural demand.
Key Innovation: Shock-waveform decomposition isolates single and double pulses as the first one or two components and derives pulse period directly from component frequency; tests on 91 benchmark motions and 26 records from the 2023 Turkey earthquakes agree with established period estimates.
53. A probabilistic prediction model based on uncertainty quantification and interpretable deep learning: Application to runoff prediction in the Yellow River mainstream
Core Problem: Deterministic runoff forecasts cannot express the data and model uncertainty produced by non-stationary climate and human regulation along a connected river.
Key Innovation: A topology-aware graph network combines predictive variance, Gaussian likelihood, Monte Carlo dropout and kernel density estimation, improving point-error metrics by at least 11.46-15.45% and probabilistic metrics by at least 7.32-10.42% on the Yellow River.
54. Analytical solution for dynamic response of deep fluid-filled tunnels under combined effects of seismic waves and water hammer loading
Core Problem: Hydraulic tunnels can experience earthquake excitation and valve-induced water hammer together, yet their transient response is usually assessed as separate load cases.
Key Innovation: A time-domain analytical solution superposes seismic scattering and method-of-characteristics water-hammer pressure, matches ABAQUS simulations, and exposes sensitivity to flow velocity, closure time and ground-liner flexibility.
55. FlashPDE: A Drop-in Fused Triton Operator Library for Neural PDE Solvers
Core Problem: Grid-based physics-informed learning is constrained by automatic-differentiation memory and fragmented execution of finite-difference operators.
Key Innovation: FlashPDE fuses 14 differentiable PDE operators and analytic discrete adjoints in Triton kernels, reducing peak memory by up to 37 times and delivering up to 2.30 times end-to-end and 19.2 times kernel-level acceleration with numerical agreement to PyTorch references.
56. Potential failures of physics-informed machine learning in traffic flow modeling: theoretical and experimental analysis
Core Problem: Adding a PDE residual does not guarantee better learning when observations are too coarse to reconstruct the state from which that residual is computed.
Key Innovation: Traffic-flow analysis traces failure to misaligned data and physics gradients under low-resolution sampling and derives residual-error lower bounds that explain why the simpler LWR physics can outperform higher-order ARZ constraints.
57. WINO: A Weak-Form Physics Informed Neural Operator for Hyperelasticity on Variable Domains
Core Problem: Supervised neural operators for nonlinear mechanics require many converged solutions and usually assume a fixed, body-fitted domain.
Key Innovation: WINO minimizes weak-form residuals on level-set-defined variable domains without paired labels, trains in about 15-70% of the time reported for supervised phi-FEM-FNO baselines, and supplies warm starts that reduce nonlinear solver iterations.
58. Interpretation and representation in geomodels: the POKIMON ontology for formalizing geomodelling knowledge
Core Problem: Three-dimensional geological models rarely record the assumptions, inference steps and uncertainties that shaped their interpreted geometry.
Key Innovation: The POKIMON ontology formalizes geological concepts and modelling decisions so that expert reasoning can be inspected, transferred and used by knowledge-driven geomodelling systems.
59. Scattering Center Prior-Guided Diffusion for Unknown-Azimuth SAR Image Generation
Core Problem: Sparse viewing angles make SAR synthesis unstable because image-domain interpolation does not preserve the target’s angle-dependent scattering structure.
Key Innovation: A conditional diffusion model extracts point-spread-function-constrained scattering centres at known angles and fuses them into an unseen-angle prior, reducing local artefacts and preserving dominant scattering regions better than the reported GAN and interpolation baselines.
60. Revisiting Deep Learning-Based Semantic Segmentation on Large-Scale Hydraulic-Structure LiDAR Point Clouds: A Spatial Surrogate Modeling Perspective
Core Problem: Large LiDAR scenes are sampled into manageable blocks, leaving many points to receive labels from simplistic nearest-neighbour propagation that can distort measured structures.
Key Innovation: Inverse-distance weighting treats label completion as a transparent spatial-surrogate problem; bridge and hydraulic-structure tests show modest, class-dependent gains and quantify how neighbourhood size and distance decay affect segmentation.
61. C2Fusion: Collaborative Conditional Diffusion Model for Infrared and Visible Remote Sensing Image Fusion
Core Problem: Infrared-visible fusion methods designed for natural images lose small structures and sparse textures in remote-sensing scenes.
Key Innovation: C2Fusion separates thermal structure and visible texture in two conditional-diffusion branches, refines them jointly, and adds frequency- and edge-consistency losses to improve structural clarity and detail retention in the evaluated imagery.
62. Evaluation of IMERG V07 Precipitation Datasets at Hourly and Daily Scales in Texas, USA
Core Problem: Satellite precipitation products can change rank across stations, temporal resolutions and rainfall metrics, so a calibrated final product is not automatically the best input for every hazard application.
Key Innovation: A two-year comparison at four Texas stations finds that IMERG V07 detects 53-75% of hourly events and 71-87% of daily events, with Early, Late and Final products alternating as the strongest performer by site and metric.
63. On the Use of SAR Images for Predicting Vegetation Indices: Challenges and Limitations
Core Problem: Cloud-robust SAR is attractive for filling optical gaps, but reconstructing vegetation-sensitive spectral information from radar alone may erase details needed for biophysical interpretation.
Key Innovation: Three models reconstruct Sentinel-2 bands from Sentinel-1 before deriving multiple indices; Efficient-UNet performs best and approaches index-specific models, yet fine spectral and vegetation detail remains incomplete, making the outputs proxies rather than optical substitutes.
64. Displacement response spectra based on 2DOF system for the isolated design of railway bridges
Core Problem: Single-degree-of-freedom displacement spectra cannot separately represent pier and bearing demand in seismically isolated railway bridges.
Key Innovation: A two-degree-of-freedom formulation provides peak pier, peak bearing and residual bearing spectra; parameter tests separate the influence of mass, pier period, damping and bearing nonlinearity and extend displacement-based bridge design.
65. Machine learning-aided optimization framework for friction dampers in traditional Chinese timber structures and case study evaluation
Core Problem: Friction-damper design for traditional timber buildings is a costly multi-objective problem because reducing drift can increase acceleration and joint stiffness is highly sensitive to damper settings.
Key Innovation: An MLP surrogate, NSGA-II search and entropy-weighted ranking identify a retrofit that cuts maximum drift by 46.36% with an 11.72% acceleration increase, while accelerating the optimization by about 57.7 times.
66. Derivative of canopy reflectance with respect to leaf density: A probabilistic rendering framework
Core Problem: Remote-sensing inversion needs the derivative of canopy reflectance with respect to leaf density, but finite differences and importance sampling can be noisy.
Key Innovation: A Bernoulli-thinning rendering formulation yields a closed-form expected radiance derivative and an unbiased Monte Carlo estimator with accuracy comparable to existing methods and lower derivative noise.
67. Quantifying the effect of lateral seepage on the vertical saturated hydraulic conductivity of loess determined by the double-ring infiltration tests
Core Problem: Double-ring infiltration tests treat lateral seepage as an imprecise nuisance even though it can dominate flow and inflate vertical saturated hydraulic conductivity.
Key Innovation: Field tests and simulations show lateral seepage commonly supplies more than 50%, and sometimes 90%, of steady flow; when its share falls below 55%, the measured conductivity approaches the one-dimensional value, providing a criterion for ring design.
68. FusionNet: Physics-Aware Representation Learning for Multi-Spectral and Thermal Data via Trainable Signal-Processing Priors
Core Problem: Thermal-only mapping of cement plants confuses kilns with other heat sources and misses persistent spectral changes in surrounding material.
Key Innovation: FusionNet combines thermal imagery with a geological SWIR ratio and trainable signal-processing priors, reaches 90.6% accuracy, and shows that generic ImageNet pretraining can degrade both thermal and SWIR performance.
69. Seismic performance and design method of novel InorgBam beam-column bolted connections utilizing glued-in rods in the column
Core Problem: New inorganic-bonded bamboo beam-column connections need cyclic evidence and a design rule that avoids brittle fracture of glued-in column rods.
Key Innovation: Six connection tests distinguish three failure modes and show higher ductility when bolt yielding and bamboo splitting precede rod fracture; a capacity ratio of 1.2 is proposed to steer the joint toward the safer mechanism.
70. Seismic performance of prefabricated bridge pier with comb-shape transverse reinforcement
Core Problem: Open-ended transverse reinforcement can simplify precast bridge construction, but its confinement and cyclic behaviour differ from conventional closed hoops.
Key Innovation: Axial tests on 11 columns and cyclic tests on two full-scale piers show that section-penetrating comb reinforcement improves capacity and ductility; a modified stress-strain model and OpenSees simulations reproduce the measured response.