Differentiation and high-performance infrastructure for multiphysics simulation
RWTH Publications (RWTH Aachen)
Abstract
Modern physical simulation is expected to handle increasingly complex requirements: heterogeneous object representations, nonlinear materials, robust contact and friction, differentiability, and runtime efficiency. Optimization-based simulation provides a common language for many of these problems by formulating the balance of forces as the minimization of a global objective, which can be robustly solved with Newton-type methods. However, the path from a compact mathematical model to a rich, robust, fast, and differentiable program remains difficult. This thesis presents the work developed during my doctoral studies on simulation infrastructure to bridge this gap along two connected threads. The first builds on symbolic representation, differentiation, assembly, and solution of variational multiphysics problems. The second concerns high-performance data structures and algorithms for simulation, with emphasis on concurrency, vectorization, pipelining, and task decomposition. Together, these threads support the design of simulation systems that meet demanding modern requirements while remaining approachable, robust, and efficient. The thesis first introduces SymX, a simulation-native symbolic framework that turns concise mechanical model definitions into compiled kernels and assembled global derivatives for Newton-type solvers. This infrastructure supports follow-up work, including a strongly coupled simulator for rigid and deformable bodies with frictional contact, a selective strategy for stabilizing Newton solves through local Hessian projection, and differentiable pipelines for inverse simulation. Applications in robotics and digital humans accompany the presentation and validate the methods. Finally, the thesis studies high-performance proximity detection in the context of fluid simulation.Together, these contributions show that expressive, robust, differentiable, and efficient simulation emerges not from any single technique, but from the careful integration of modeling, optimization, differentiation, and high-performance infrastructure.
Authors 1
-
Affiliation as printed
RWTH Aachen