Title: MPMWorlds: Leveraging Material-Point-Method Simulations to Infer and Extrapolate Physical Dynamics

Abstract

To investigate the capacity for deducing physical dynamics from video footage and projecting them into the future, we have curated a comprehensive dataset comprising 2D Material Point Method (MPM) simulations. This collection encompasses a wide array of physical behaviors, including fluids, deformable bodies, kinetic entities, and emitters.

Our research evaluates both code generation and video diffusion techniques using this dataset, assessing their respective merits and limitations by manipulating the volume of physically relevant auxiliary information provided. The findings indicate that while the code generation model successfully demonstrates the automatic synthesis of MPM simulations, it faces significant challenges in extracting physical parameters directly from visual data. However, when compared to video diffusion, it yields extrapolations that are superior in both temporal and physical stability. Conversely, the video diffusion model excels at recognizing geometric attributes from visual inputs but tends to generate extrapolations that lack physical plausibility.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC