Physics-Optimized Design of 3D Shapes with Part-Based Control

Abstract

We introduce PhysiOPart, a computational approach for rapid generative design of 3D objects optimized for physical integrity. PhysiOPart enables users to edit and combine object parts to explore a vast design space. To model continuous surfaces of arbitrary resolution without topology restrictions, we parametrize parts with neural implicit representations. However, when parts are assembled to form an object, the resulting geometry is not guaranteed to be functional. Existing generative modeling approaches use task-specific neural predictors to approximate physical behaviors with limited accuracy. We propose an end-to-end differentiable physics simulation pipeline that performs linear static analysis to optimize for user-specified objectives, leveraging learned geometry priors. Our part-based formulation with finite element method is highly customizable, allowing for user-defined per-part materials, loads, and boundary conditions. The optimized designs exhibit improved physical behavior and can be fabricated.