Physical simulation and deep learning using particle-based representations

RWTH Publications (RWTH Aachen)

Abstract

Points are simple and versatile primitives for representing geometry in computer graphics, with applications including physical simulation and geometric deep learning. Current approaches in these directions are already powerful, but they are still lacking in specific areas. For one, implicit particle-based fluid simulation is an active area of research that has shown improvements for a variety of effects, but, at the time of publication of the relevant work, had not been successfully explored for surface tension. This is an important force for realistic fluid motion. In addition, particle simulation has not been quantifiably established as a viable alternative to more traditional grid-based simulation methods, particularly for complex melt-driven processes. Finally, reconstructing closed surfaces from surface particles is an ill-posed problem without a clear or simple solution. Existing methods, particularly in deep learning, often fail to generate satisfying surfaces using direct inference and may rely on additional optimization passes at test-time to create a better fit. To address these shortcomings, this thesis investigates how the fidelity and efficiency of using particles can be improved in two primary directions. First, for particle-based fluid simulation with Smoothed Particle Hydrodynamics (SPH), methods are developed to address challenging thermodynamic, melt-driven scenarios. An implicit cohesion-based surface tension model is introduced and strongly coupled with implicit viscosity to improve stability, especially in high-tension regimes. In addition, implicit formulations for melting and solidification, together with accurate particle-to-grid transfer, enable quantitative studies of welding and thermal spraying. These studies show good agreement between the SPH framework and established grid-based approaches while providing efficient high-resolution simulations for free-surface dynamics. Second, for point-based deep learning, an efficient hybrid point-grid architecture is presented to reconstruct neural distance fields from point-cloud inputs. The method combines point processing and latent grid representations in a single forward-pass encoder-decoder design, enabling accurate and efficient shape reconstruction of closed surfaces. It is shown that this architecture is fast to evaluate and train, while outperforming other baselines on the reconstruction task. Taken together, the presented works improve point-based methods both in terms of physical accuracy in simulation and efficient geometric representation learning.

Authors 1

1. Stefan Rhys Jeske corresponding Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen