Deep learning accelerated efficient framework for topology optimization

Abstract

In this work we propose a novel deep learning assisted topology optimization model that predicts the final optimization result (which is usually obtained through hundreds of redesign/trial-error iterations) from only a few initial redesign iterations and few fine tuning iterations. The broad idea is to sandwich a deep learning model between the initial redesign iterations and the fine tuning iterations. The deep learning model acts as a surrogate model and allows us to learn the evolution law in an efficient manner. Although the proposed framework can accommodate any deep learning model, we have used the state-of-the-art flow based generative model. The flow based model takes the topological configurations from the initial redesign steps as input and yields the optimal configuration which is then further fine tuned by running the topology optimization algorithm for a few iterations. The final fine tuning topology optimization iteration assures structural soundness and integrity of the optimized configuration. The total computation time of this model is seen to be significantly less than the standard topology optimization routines. The generalization capability of this framework is demonstrated by having promising result in design domains with different boundary and loading conditions (dynamic and multiple point-loads). More interestingly, the proposed model generalizes to nonlinear systems without any retraining. Overall the proposed framework is robust and can be used as an alternative tool in conceptual design optimization process.