A high-performance multiscale space-time approach to high cycle fatigue simulation based on hybrid CPU/GPU computing

A multiscale space/time computational framework for high cycle fatigue (HCF) life predictions is established by integrating the extended space-time finite element method (XTFEM) with a multiscale progressive damage model. While the robustness of the multiscale space/time method has been previously demonstrated, the associated high computational cost remains a critical barrier for practical applications. In this work, a novel hybrid iterative/direct linear system solver is first proposed with a unique preconditioner. Computational efficiency is further improved by taking advantage of the high-performance computing platform featuring hierarchy of the distributed- and the shared-memory parallelisms using CPUs and GPUs. Robustness of the accelerated framework is demonstrated through benchmark problems. It is shown that the serial version of the hybrid solver is at least 1-2 orders of magnitude faster in computing time and cheaper in memory consumption than the conventional sparse direct or iterative solver, while the parallel version efficiently handles XTFEM stiffness matrix equations with over 100 million unknowns using 64 CPU cores. Optimal speedups are achieved in the parallel implementations of the multiscale progressive damage model using either CPUs or GPUs. HCF simulations on 3D specimens are performed to quantify key effects due to mean stress and multiaxial load conditions.