Evaluation of Fluid-Driven Debris Impacts in a High-Performance Multi-GPU Material Point Method

Abstract

Tsunamis, storm surges, landslides, lahars, and avalanches driving groups of debris (e.g. cars, trees, boats, collapsed structures) are complex coinciding hazards. Yet, in recent 3D animated films these events are rendered to an incredible visual benchmark at computational scales exceeding what most engineers can match. Disney’s Frozen and Moana, in an effort to simulate snow and sand media (history-dependent, topology changing, large-deformation, and nonlinear), champions an interesting although expensive numerical tool: The Material Point Method (MPM). Their software optimization bypassed a computational barrier that had limited engineering use. Just as animators adopt engineering techniques to improve physical accuracy, we argue engineers should adopt the optimized codes of computer graphics professionals. Doing so can accelerate engineering simulations 10x - 100x, grow them by 2x - 1000x, and amplify their public impact. We introduce innovations by Disney, Pixar, Tencent, DreamWorks, and their research contemporaries into our own project with minimal overhead. Our open-source numerical tool, a high-performance multi-physics coupling of the MPM and finite elements (FEA), scales across Multi-GPUs, devices typically used for video games and machine learning. In doing so, we broach the doorstep of exa-scale computation-- Where billion particle simulations are common. Prioritizing flexibility, speed, and simplicity lead us to the most basic algorithm approach for debris-fluid-structure interaction (DFSI) identified yet, which we show to be a novel solution to fluid-driven debris-field loads on structures. This work develops an augmented MPM framework suitable for scaling on Multi-GPU infrastructure and designed for natural hazard simulations. We validate against analytical cases, alternative numerical methods, and experimental results. Five digital twins are developed for wave and lahar flumes located at USGS sites, Oregon State University, Waseda University, and the University of Washington. Structural loads from organized and random groups of debris at multiple scales are numerically replicated with novel accuracy. Preliminary steps are taken into a performance-based engineering framework and a generalized debris-field characterization scheme for multi-hazard events.