Point-based Diffusion Model for Predicting 2D Spatio-Temporal and 3D Large-Scale Physical Systems with Shape Variations

Traditional diffusion-based approaches for physical system prediction typically depend on grid-based, snapshot-level representations, which restrict their applicability to irregular domains and varying geometries. In this work, we propose a point-wise conditional diffusion framework that enables flexible and efficient prediction of complex physical phenomena across diverse and irregular geometries without requiring fixed grids. The proposed method is evaluated on 2D spatio-temporal systems in both Eulerian and Lagrangian settings, as well as on 3D large-scale aerodynamic problems, demonstrating strong generalization across varying geometric complexities.

Korean Society of Mechanical Engineers (KSME 2025)