UCLA

The challenge of efficiently and plausibly simulating deformable solids and fluids remains significant in the domains of Computer Graphics and Scientific Computing. This dissertation presents an in-depth exploration of physics-based simulation, with an emphasis on the Material Point Method (MPM) --- a dominant technique in this arena. Our research aims to extend the capabilities of MPM, focusing on enhancing its performance, scalability, range of applications, and integration with emerging AI technologies. We first summarize our development of optimized MPM leveraging GPU architectures. This advancement accelerates scenarios involving hundreds of millions of particles in multi-GPU computational environments. Furthermore, the thesis introduces a device-agnostic and distributed MPM framework. This system is adept at dynamically allocating workloads across multiple computing ranks, thus enabling simulations at unprecedented particle-count scales. Additionally, the dissertation examines the application of physics-based simulation, specifically MPM, in real-time contexts. It also integrates simulation with generative AI tasks. This exploration includes developing unified frameworks for simulations, image rendering, and natural language processing, showcasing the versatile applicability of MPM in tackling contemporary computational challenges.