Earthquake early warning (EEW) systems rely on rapid and accurate magnitude estimation from early P-wave signals, commonly using peak displacement amplitude (P-d). However, P-d-based magnitude prediction is strongly influenced by local site conditions, which are often neglected in conventional models. This study introduces a site-specific soil parameter, PGV/Vp(30), where PGV is the peak ground velocity measured from the first three seconds of the vertical P-wave component and Vp(30) is the average P-wave velocity in the upper 30 m of the subsurface. The parameter is integrated into linear empirical relationships for P-d and magnitude estimation and evaluated using 2317 strong-motion records from the Japanese KiK-net network. Results show that incorporating the site-specific parameter improves P-d prediction stability by 43% and reduces the uncertainty of average magnitude error by 48%. The proposed model also reduces magnitude underestimation for large events (M > 7.0). Validation against empirical models from Southern California and Taiwan confirms the robustness of the approach. These findings demonstrate that incorporating site-specific soil response significantly enhances the accuracy and reliability of EEW magnitude prediction.
Effective identification of precursors to large landslide instability is critical for hazard assessment and early warning. However, conventional displacement-based approaches are often limited by the spatiotemporal resolution of monitoring data and the inherent spatial heterogeneity of landslide deformation. Here, we propose a precursor identification framework based on the spatiotemporal evolution of landslide surface scars. This approach characterises progressive failure through temporal variations in geometric and topological features of surface morphology, capturing the transition from stable to unstable states. Multi-temporal optical imagery is used in conjunction with a vision foundation model, guided by domain knowledge, to enable automated delineation and continuous tracking of landslide scars. The extracted geometric and topological descriptors provide a quantitative basis for representing landslide evolution and detecting precursory signals. Application to representative cases demonstrates that a critical precursor signal is identified on 25 July 2018, approximately 2.5 months prior to failure of the Baige landslide. For the Sela landslide, which did not undergo system-scale failure, a transition on 7 June 2019 corresponds to externally induced localised acceleration. These results indicate that precursor signals can be identified from the perspective of surface morphological evolution, offering a complementary observational basis for landslide early warning and risk assessment.
This study investigates four uncertainty quantification-integrated deep learning techniques for developing CPTu- Vs correlations. Three techniques - na & iuml;ve conformal prediction, deep ensemble, and Monte Carlo dropout - are incorporated within a traditional deep neural network (DNN) framework. A fourth approach adopts a probabilistic framework based on Bayesian neural networks (BNNs). These techniques are applied to a multi-site seismic piezocone test (SCPTu) database comprising both onshore and offshore sites. Comparative analyses assess model accuracy, predictive uncertainty, and data coverage for individual SCPTu profiles and for the full dataset, followed by an evaluation of generalisation performance. Results indicate that deep ensemble and Monte Carlo dropout produce relatively narrow uncertainty bounds, leading to inadequate data coverage of 43% and 56%, respectively, within a 95% confidence interval. Na & iuml;ve conformal prediction attains the target coverage but yields overly conservative estimates due to its inherent assumptions. BNNs provide balanced performance, offering moderate uncertainty bounds that encompass 71% of observed measurements, thereby evidencing their robustness. For individual SCPTu profiles, higher predictive accuracy is generally associated with lower uncertainty and higher coverage. Poor predictive performance is generally attributed to specific site conditions, including rapid soil stratigraphy transitions and silty soil layers. Furthermore, BNNs demonstrate superior generalisation compared to DNN-based techniques.
Developing 3D stratigraphic models for a target site has been a long-standing challenge due to limited boreholes within the target site, a situation often encountered in engineering practice. Although several machine learning methods have been developed recently for 3D stratigraphic modelling by incorporating prior geological knowledge to supplement existing borehole data, they still face considerable challenges when the number of boreholes within a target site is extremely limited (e.g. less than three). In engineering practice, boreholes may be available outside the target site, which can offer valuable stratigraphic information to supplement borehole data within the target site. In this study, a data-driven generative framework is proposed for constructing 3D stratigraphic models using boreholes outside the target site. A recently developed machine learning method, called multi-scale generative adversarial network (MS-GAN), is employed to model subsurface stratigraphy and quantify uncertainty. An enhancement is made to MS-GAN for effectively handling irregularly distributed boreholes. The proposed framework is validated through a set of boreholes in Hong Kong. The results show that the proposed method successfully utilises boreholes outside the target site to generate reliable 3D stratigraphic models with quantified uncertainty.
The prediction of rainfall-induced shallow landslides depends not only on rainfall characteristics but also critically on geoenvironmental conditions. However, current empirical rainfall thresholds are typically developed neglecting spatial variations in geoenvironmental settings and are applied uniformly across heterogeneous terrain, limiting their predictive capability. To this end, we propose a novel framework that adaptively partitions the study area into sub-regions where homogeneous geoenvironmental conditions give rise to consistent landslide patterns, and derives localised, pattern-specific rainfall thresholds. The framework first evaluates geoenvironmental conditions at the slope unit scale and leverages Gaussian mixture modelling to segment the area based on similarity. For each sub-region, the separability between triggering and non-triggering rainfall events is quantified to assess the consistency of rainfall-landslide response patterns, validating the segmentation. Guided by this assessment, an adaptive procedure optimises the predictors and segment number, yielding sub-regions exhibiting consistent landslide patterns and associated probabilistic rainfall thresholds. The framework is applied in Zhuji, China. Results show that it improves the prediction of historical landslides, increasing the AUC from 0.936 to 0.966 over a 16-year rainfall record. A real-time early-warning case further demonstrates a reduction in false alarms by over 20% while maintaining a high hit rate, outperforming conventional thresholds.
Physically based models like TRIGRS are utilized to assess regional shallow landslide hazards. However, TRIGRS accuracy is limited by spatial heterogeneity in geotechnical properties, especially colluvial thickness and shear strength. To address this limitation, we propose an enhanced Physico-Data TRIGRS model. First, a physico-data coupled soil zoning optimization integrating frequency ratio analysis and Random Forest (RF)-based dominant factor identification for improved soil zoning. Then, a physically constrained estimation of colluvial thickness by combining field data with an improved slope-thickness function, enhancing the model's applicability in complex terrains. Finally, Monte Carlo simulation is incorporated to represent the spatial variability of shear strength parameters and to transform deterministic factors of safety into slope failure probabilities under four rainfall scenarios: no rainfall, 10-, 50- and 100-year return periods. Pingyang County is selected as a case study. Results indicate that the proposed model improves the accuracy of landslide hazard by 10%, increasing the AUC from 0.708 to 0.843. Monte Carlo simulation effectively captures parameter uncertainty, enabling robust dynamic probabilistic hazard evaluation under different rainfall conditions. Overall, the proposed framework improves both the physical interpretability and predictive precision of TRIGRS, providing a more flexible and accurate approach for landslide hazard assessment.
Shear-wave velocity (V-s) is an essential parameter in evaluating geotechnical hazards and in foundation design. Conventional techniques to measure V-s are time-intensive, costly, and infeasible to perform at high resolution across large sites such as offshore wind farms. However, predicting V-s, particularly the time-averaged V-s in the uppermost 30 m (V-s30), using readily available geospatial proxies has significant technical and economic benefits. This study employs geospatial data and machine learning techniques to advance V-s30 mapping in the North Sea, leveraging onshore geotechnical data from the United Kingdom, Norway, and the Netherlands. To encourage reliable extrapolation from onshore to offshore environments, the workflow prioritises domain-limited predictions and independent verification against offshore datasets. The outcomes of this modelling effort include a proof-of-concept simplified site classification map in the North Sea and a scalable methodologic framework adaptable to other offshore regions and geotechnical parameters. Importantly, this study identifies critical data gaps that constrain current model performance and that should be prioritised to advance offshore V-s mapping for hazard assessment, siting, and infrastructure design.
The 6th Machine Learning in Geotechnics Dialogue (6MLIGD), held on 24 August 2025, in Oslo, was a pivotal event under the International Symposium on Geotechnical Safety and Risk (ISGSR 2025). The dialogue brought together leading experts from academia, industry and policymakers to explore the transformative potential of machine learning (ML) in geotechnical engineering. The event featured a structured format, including invited talks, panel discussions and group sessions, fostering collaboration and knowledge exchange. Key topics included advancements in ML-driven site investigations, tunnelling, geohazard monitoring and infrastructure resilience. The dialogue also addressed critical challenges such as data sharing, standardisation, the integration of emerging technologies like large language models (LLMs) and geo-data governance. The 6MLIGD emphasised the importance of interdisciplinary collaboration, real-world case studies, and the development of benchmark datasets to accelerate research and practice. The outcomes of the dialogue underscored the need for innovative ML solutions to address geotechnical challenges, paving the way for smarter, more resilient infrastructure systems.
Regional landslide risk assessment is essential for disaster prevention, yet traditional methods are limited by inadequate characterisation of conditioning factor heterogeneity in landslide susceptibility prediction (LSP) and spatial scale mismatch between administrative population data and slope units. This study develops an integrated framework combining automated slope unit division, heterogeneity quantification, and machine learning. Using Multi-Scale Segmentation in Anyuan County, 148,608 slope units were divided. Intra-unit heterogeneity for LSP was quantified through statistical features of conditioning factors. The Heterogeneous Slope-RF model incorporating heterogeneity achieved higher predictive accuracy (AUC = 0.941) than models ignoring it. For population density, an upper threshold of 120 persons/km & sup2; was applied, and a slope unit-scale RF model produced precise estimates (R & sup2; = 0.8157, MAE = 1.7 persons/km & sup2;). This refined distribution reduces overestimation of risk in urban centres and identifies dispersed settlements in high-susceptibility mountainous areas. By integrating this population distribution, landslide susceptibility, and AHP-based vulnerability, a spatial map of potential population loss risk was generated for Anyuan County. The study confirms that incorporating spatial heterogeneity and refined population modelling improves landslide risk assessment accuracy, supporting identification of genuine high-risk zones.
Metro tunnel linings are highly susceptible to deterioration, particularly water leakage, while manual inspection remains inefficient and the limited availability of data constrains deep learning-based detection. To address this challenge, a novel two-stage framework is proposed for water-leakage detection under extreme small-sample conditions, introducing diffusion-based inpainting for synthetic data augmentation and an enhanced segmentation architecture, WLSNet. WLSNet builds upon U-Net by incorporating a Global Perception Module (GPM) for large-scale context modelling and Multi-Scale Attention Augmentation (MSAA) for fine-grained feature refinement, effectively reducing false positives and improving boundary accuracy. Extensive experiments demonstrate that the proposed approach significantly outperforms existing methods in mIoU and precision, achieving notable gains even with as few as 10 original training images. This study represents the first application of diffusion models to tunnel leakage segmentation and establishes an effective strategy for mitigating data scarcity. The proposed SD-WLSNet framework offers a accurate solution for metro tunnel inspection, supporting proactive maintenance and improved operational safety.
Monopile foundations significantly disturb local hydrodynamics and are prone to severe local scour under steady currents. This study proposes a novel active scour protection concept, termed the spoiler pipe, installed upstream of the monopile to modify the approaching flow field. A three-dimensional CFD model coupled with sediment transport is developed and validated against classical laboratory experiments. The results show that the spoiler pipe attenuates the incoming flow, weakens erosive vortex structures around the pile, and induces wake-related sediment deposition, leading to effective scour mitigation. Parametric simulations indicate that the maximum scour depth can be reduced by up to 49%, depending on the spoiler pipe diameter and its upstream stand-off distance. The stand-off distance primarily controls the maximum scour depth, while the pipe diameter governs sediment replenishment and scour pit extent. The spoiler pipe provides a simple, low-disturbance active scour protection solution, particularly suited to current-dominated or mixed-flow offshore environments.
Landslide susceptibility mapping (LSM) is essential for regional geohazard assessment. However, conventional data-driven models often fail to capture slope-failure mechanisms and produce scientifically inconsistent predictions, particularly in regions with complex geology or limited data. This study proposes a novel physics-informed machine learning (PIML) framework that formulates LSM as a multi-objective optimisation problem. A positive-unlabelled (PU) bagging strategy identifies reliable non-landslide samples, and model training jointly minimises three complementary losses: a supervised loss ensuring predictive accuracy, a physical-consistency loss constraining monotonic relationships with the factor of safety (FoS), and a risk-consistency loss constraining monotonic relationships with the probability of failure (PoF). The FoS and PoF, derived from the simplified transient infiltration model (STIM) and the first-order reliability method (FORM), are incorporated into model training to enforce scientific consistency. A case study of rainfall-induced landslides in Gansu Province, China, shows that the proposed framework achieves Pareto-optimal trade-offs between accuracy and scientific consistency. Compared with the baseline, the optimal PIML model improved average AUC (0.882 vs. 0.870) under spatial cross-validation, reduced inconsistency by 73%, and outperformed the physically based probabilistic model. These results highlight that embedding geotechnical knowledge into ML produces susceptibility maps that are both accurate and scientifically meaningful.
Tailings dams are loose due to sequential accumulation and prone to failure under high-intensity rainfall. Particle gradation determines the mechanical, seepage, and rheological properties of tailings and is a key factor in dam scouring damage, while rainfall-induced gully progression mechanisms for different gradations remain underexplored. This study adopted laboratory physical model tests combined with Three-dimensional Light Detection and Ranging (3D LiDAR) and high-definition photography to dynamically capture 3D surface evolution of tailings dams with three gradation types. Quantitative analysis revealed the spatiotemporal evolution of scouring depth, deposition height, and surface shape. Results show scouring damage initiates at high-energy zones near dam toes and propagates upward through four phases, including embryonic gully formation, gully progression, stable gully growth, and collapse deformation. Scouring and deposition interactions govern gully progression through scouring dominance, deposition regulation, and dynamic equilibrium, via particle-selective transport and self-inhibition deposition. An optimised time-dependent gully depth predictive model incorporating effective rainfall intensity, gradation parameters, and permeability quantifies their synergistic effects on scouring kinetics with 95% accuracy. This study reveals gradation-dependent gully evolution in tailings dams under high-intensity rainfall and develops a predictive model to quantify these processes, providing a theoretical basis and quantitative tool for forecasting rainfall-induced dam failure.
Water-sand mixture inrush (WSMI) is a severe geological hazard in deep mining, challenging conventional control methods due to its complexity and concealment. Digital Twin (DT) technology, with advanced data fusion and interactive functions, offers an innovative approach for mine hazard governance. Based on the Cuihongshan iron-polymetallic mine, this study integrated the IoT -enabled sensor networks, finite element analysis, and deep learning algorithms to construct a multidimensional DT framework and apply it in practice. First, the connotation and key features of the DT model were defined, clarifying the interaction between physical entities and their digital counterparts within a comprehensive architecture. Second, a perception layer with sensor networks and dynamic updating was established, while the modelling layer adopted multiple approaches to construct three categories of geometric models and a semantic model. Finite-element simulations enabled multiphysics coupling of grouting, accurately revealing grout diffusion. A novel multi-activation adaptive multilayer perceptron (MMLP) was proposed for high-precision grouting prediction. Furthermore, in the application layer, an operational platform established closed-loop management, comprising perception, cognition, decision, and application. Results show that the DT framework greatly improves governance efficiency and accuracy, offering new theoretical foundations and technological pathways for mine hazard prevention.
