Enhancing topology optimization with adaptive deep learning

Abstract

Topology optimization (TO) is a pivotal technique for generative design of high-performance structures. Practical designs often face complex boundary conditions and require non-gradient optimizers for solving TO with thousands of design variables or more. This paper presents the Adaptive Deep Learning (ADL) which supports both gradient-based topology optimization (GTO) and non-gradient-based topology optimization (NGTO). The ADL roots in convolutional neural network to link material layouts with structural compliance. A small number of training data is generated dynamically based on the ADL’s prediction of the optimum. The ADL explores the region of interest in a probabilistic setup and evolves with increased data. The presented ADL has been evaluated on four cases including beam design, heat dissipation structure design, three-dimensional machine tool column design and heat transfer enhancement optimization. The ADL achieved 0.04 % to 4.08 % increasement of structural performance compared to GTO algorithm, and 0.88 % to 81.98 % increasement compared to NGTO algorithms.