GPGPU-Parallelized Data-Driven Hierarchical Multiscale 3D FDEM for Rock Meso-macro-numerical Simulation

A hierarchical multiscale three-dimensional (3D) FDEM approach is employed to investigate the meso–macro-mechanical response of rock. The core of the multiscale approach involves a hierarchical coupling of meso-scale and macro-scale 3D FDEM. Specifically, the temporal convolutional network with mixture density network (TCN-MDN) is specially developed for efficiently training and predicting meso-scale 3D FDEM simulation data of upscale finite and crack elements assembly (UFEA and UCEA). Then, UFEA and UCEA driven by TCN-MDN serve as equivalent elements, replacing phenomenological constitutive relationships in macro-scale 3D FDEM. Further, general purpose graphic processing unit (GPGPU) parallel computing is implemented within 3D FDEM and TCN-MDN to further enhance computational efficiency. Upon constructing GPGPU-parallelized DHM-3DFDEM, uniaxial compression and Brazilian disk tests confirm its consistency with reference solutions and laboratory test results. The ability of the proposed method to reproduce the mechanical behavior of rock under complex 3D loading conditions is further verified through cyclic loading–unloading and true triaxial compression tests. Results demonstrate that TCN-MDN has the advantage of one-to-many mapping and high parallelism, capturing rock heterogeneity and efficiently processing large 3D sequential strain–stress data sets. Finally, the acceleration of GPGPU parallel computing on DHM-3DFDEM is verified, achieving a maximum speed-up ratio of 20.09.