Geomechanical assessment using coupled reservoir-geomechanical simulation is becoming increasingly important for analyzing the potential geomechanical risks in subsurface geological developments. However, a robust and efficient geomechanical upscaling technique for heterogeneous geological reservoirs is lacking to advance the applications of three-dimensional (3D) reservoir-scale geomechanical simulation considering detailed geological heterogeneities. Here, we develop convolutional neural network (CNN) proxies that reproduce the anisotropic nonlinear geomechanical response caused by lithological heterogeneity, and compute upscaled geomechanical properties from CNN proxies. The CNN proxies are trained using a large dataset of randomly generated spatially correlated sand-shale realizations as inputs and simulation results of their macroscopic geomechanical response as outputs. The trained CNN models can provide the upscaled shear strength (R-2 > 0.949), stress-strain behavior (R-2 > 0.925), and volumetric strain changes (R-2 > 0.958) that highly agree with the numerical simulation results while saving over two orders of magnitude of computational time. This is a major advantage in computing the upscaled geomechanical properties directly from geological realizations without the need to perform local numerical simulations to obtain the geomechanical response. The proposed CNN proxy-based upscaling technique has the ability to (1) bridge the gap between the fine-scale geocellular models considering geological uncertainties and computationally efficient geomechanical models used to assess the geomechanical risks of large-scale subsurface development, and (2) improve the efficiency of numerical upscaling techniques that rely on local numerical simulations, leading to significantly increased computational time for uncertainty quantification using numerous geological realizations. (c) 2024 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/ licenses/by-nc-nd/4.0/).

Landslides are a prevalent natural hazard in West Bengal, India, particularly in Darjeeling and Kurseong, resulting in substantial socio-economic and physical consequences. This study aims to develop a hybrid model, integrating a Genetic-based Random Forest (GA-RF) and a novel Self-Attention based Convolutional Neural Network and Long Short-term Memory (SA-CNN-LSTM), for accurate landslide susceptibility mapping (LSM) and generate landslide vulnerability-building map in these regions. To achieve this, we compiled a database with 1830 historical data points, incorporating a landslide inventory as the dependent variable and 32 geoenvironmental parameters from Remote Sensing (RS) and Geographic Information Systems (GIS) layers as independent variables. These parameters include features like topography, climate, hydrology, soil properties, terrain distribution, radar features, and anthropogenic influences. Our hybrid model exhibited superior performance with an AUC of 0.92 and RMSE of 0.28, outperforming standalone SA-CNN-LSTM, GA-RF, RF, MLP, and TreeBagger models. Notably, slope, Global Human Modification (gHM), Combined Polarization Index (CPI), distances to streams and roads, and soil erosion emerged as key layers for LSM in the region. Our findings identified around 30% of the study area as having high to very high landslide susceptibility, 20% as moderate, and 50% as low to very low. The vulnerability-building map for 244,552 building footprints indicated varying landslide risk levels, with a significant proportion (27.74%) at high to very high risk. Our model highlighted high-risk zones along roads in the northeastern and southern areas. These insights can enhance landslide risk management in Darjeeling and Kurseong, guiding sustainable strategies for future damage qualification.