Machine learning enhanced multiscale topology optimization with structural genome databases

Abstract

Multiscale topology optimization (MTO) offers a larger design space and enables the concurrent design of macroscopic “structure” and microscale “material”. Even after adopting the asymptotic homogenization analysis, the solution process could still be time-consuming, especially for multiscale topology optimization with multiple microstructures. To alleviate such an issue, a released structural genome database (SGD) is incorporated into the multiscale topology optimization process to replace the asymptotic homogenization analysis. The effectiveness, accuracy, and efficiency of the SGD-MTO algorithm are validated by benchmark examples, including optimization design of cellular structures with uniform and multiple microstructures, and concurrent multiscale design of uniform and multiple microstructures with connectivity constraints. It is validated that, compared with the traditional MTO algorithm with asymptotic homogenization analysis, the SGD-MTO algorithm can accelerate the solution efficiency by more than 30 times. Even for general MTO problems with optimized design desiring asymmetric microstructures, the proposed algorithm can still efficiently supply a reasonable initial design, which can be re-optimized in a few iterations using the explicit topology optimization method.