Upscaling micro-mechanical heterogeneity to predict macroscopic failure in coal using deep learning and grain-based DEM

Quantitative characterisation of multiscale mechanical failure in heterogeneous coal is critical for mitigating subsurface engineering risks, such as borehole instability and hydraulic fracture containment failure during geoenergy extraction. However, accurate predictive modelling is hindered by the intrinsic disconnect between microscopic sedimentary textures and macroscopic engineering responses. To delineate the controls governing multiscale mechanical behaviour, we employed an integrated methodology combining non-destructive geophysical evaluation, advanced experimental mechanics, and grain-based discrete element modelling (GBM-DEM). Specifically, ultrasonic testing and high-fidelity in situ X-ray computed tomography (CT) under varying triaxial confining pressures were used to capture the dynamic evolution of damage and stress-induced anisotropy. This workflow is complemented by a novel deep-learning segmentation approach (MSR-EnlightenGAN-DeepLabV3+) to resolve sub-resolution natural fracture networks, enabling the reconstruction of high-precision digital rock models. By explicitly mapping the micro-mechanical contrast between brittle vitrinite matrices and stiff inertinite inclusions via nanoindentation and a dual-channel neural network (ResNet34-MLP), we revealed a hierarchical control system governing damage evolution. We observed a distinct “mechanical invasion” phenomenon, in which stiff inertinite grains act as stress concentrators, inducing localised shear bands that propagate into the softer vitrinite matrix. Crucially, our quantitative results demonstrate that primary sedimentary fabrics—specifically the vitrinite-to-inertinite ratio—override post-depositional thermal maturity in dictating the fundamental mechanical strength and failure modes. Furthermore, the anisotropy induced by historical stress states and natural fractures significantly dictates the structural damage pathways. This study established a mechanism-based upscaling framework that translates microstructural geological attributes into macroscopic failure criteria and proposes that simple macrolithotypes serve as a robust, cost-effective geological index for rapid stability assessment in underground engineering projects.