Problem-independent machine learning-enhanced structural topology optimization of complex design domains based on isoparametric elements

Abstract

Topology optimization requires dozens or even hundreds of iterations, each requiring a complete finite element analysis (FEA). Significant computation cost limits the application of topology optimization in engineering, especially for high-resolution problems containing complex design domains. To address the issue, a Problem-Independent Machine Learning (PIML) model based on isoparametric elements is proposed. Effectively reducing the computational time of FEA, the proposed model enables efficient topology optimization and extends the solvable problem range to complex design domains. The essential idea is leveraging the substructure method and establishing a mapping from element shapes and material distribution within the substructure to its numerical shape functions through machine learning models. Both sample generation and model training are conducted offline, allowing the trained machine learning model to be directly employed during the topology optimization process. Since the shape function of the substructure is problem-independent, it requires no sample regeneration or modification of the proposed machine learning model when changing the geometry or boundary conditions of the optimization problem. Numerical examples demonstrate that the proposed machine learning model boosts the efficiency of topology optimization by one order of magnitude without parallel techniques.