Abstract
Maximum stress minimization is an active research topic in topology optimization. In general, gradient-based topology optimization has been used to solve such a problem, but it fundamentally faces problematic aspects. For example, singularity problems require relaxation treatments. They transform an exact minimax problem, such as the maximum stress minimization problem, into any differentiable pseudo-problem. Therefore, gradient-based topology optimization can be regarded as a low-fidelity optimization that deals with pseudo-models. Our interest is that there is room for improvement regarding the exact maximum stress in the optimized designs of the conventional method. To clarify, we focus on data-driven multifidelity topology design (MFTD) that optimizes topology without gradient information, even under a high degree of design freedom. Its basic idea is based on evolutionary algorithms (EAs). In the optimization process, the initial design candidates are generated by solving low-fidelity optimization problems. Then, they are iteratively updated a deep generative model by conducting high-fidelity forward analyses. This paper tackles a bi-objective problem of the exact maximum stress and volume minimization by data-driven MFTD incorporating initial solutions composed of the optimized designs derived by solving the gradient-based topology optimization using the p-norm stress measure. The numerical results indicate that the optimized designs by data-driven MFTD completely dominate the initial solutions and achieve a volume reduction of up to 10% under the same maximum stress value.
