Enhancing Structural Topology Optimization with Deep Learning: A Stress Field-Based Approach

Abstract

Traditional structural topology optimization process depends on series of finite element analysis (FEA) which is a computationally expensive process. As an attempt to minimize this, in this work a deep-learning (DL) based approach for structural topology optimization is proposed. In addition to increased computational efficiency the approach is designed to scale across various load and boundary condition (BC) combinations. The DL topology predictor model which is built using the Attention-Res-U-Net architecture is trained on the data generated by employing the automated FEA data generation technique. Incorporating an input image encoding technique, that utilizes the stress field of the structure to encode both load and BC, the model is capable of predicting the optimal structure of different load cases directly. The approach is validated by implementing it in a cantilevered beam problem. The result exhibited prediction accuracy above 96%. When compared against conventional FEA methods the proposed approach was able to predict solutions within 5% of VF error while outperforming in terms of computational efficiency by taking 1/10^th of their total time.

Cite as

Sourav Parida, & Rati Ranjan Dash. (2023). Enhancing Structural Topology Optimization with Deep Learning: A Stress Field-Based Approach. Advancement in Mechanical Engineering and Technology, 6(2), 24–37. https://doi.org/10.5281/zenodo.8003313