StressD: 2D Stress estimation using denoising diffusion model

Abstract

Finite element analysis (FEA), a common approach for simulating stress distribution for a given geometry, is generally associated with high computational cost, especially when high mesh resolution is required. Furthermore, the non-adaptive nature of FEA requires the entire model to be solved even for minor geometric variations creating a bottleneck during iterative design optimization. This necessitates a framework that can efficiently predict stress distribution in geometries based on given boundary and loading conditions. In this paper, we present StressD, a framework for predicting von Mises stress fields based on the denoising diffusion model. The StressD framework involves two models, a U-net-based denoising diffusion model and an auxiliary network to generate and predict stress distribution in structures. The denoising diffusion model generates a normalized stress map based on the given geometry, boundary conditions and loading condition, while the auxiliary network is used to determine the scaling information needed to un-normalize the generated stress map. We evaluate the StressD framework on cantilever structures and show that it is able to accurately predict von Mises stress fields while significantly reducing computational cost compared to traditional FEA.