Using Deep Learning to Improve Structural Topology Optimization: A Stress Field-Based Method

Abstract

The conventional method of optimizing structural topology relies on a series of computationally costly finite element analyses (FEAs). This article proposes a deep-learning (DL) based structural topology optimization method in an effort to reduce this. The method is intended to scale across different load and boundary condition (BC) combinations in addition to offering enhanced computational efficiency. The data produced by using the automated FEA data creation technique is used to train the DL topology prediction model, which is constructed using the Attention-Res-U-Net architecture. The model can directly forecast the ideal structure of various load scenarios by incorporating an input picture encoding technique that uses the structure's stress field to encode both load and BC.