Evolutionary Discrete Multi-Material Topology Optimization Using CNN-Based Simulation Without Labeled Training Data

Multi-material topology optimization problem under total mass constraint is a challenging problem owning to the incompressibility constraint on the summation of the usage of the total materials. A novel optimization approach is proposed here that utilizes the wide search space of the genetic algorithm, and greatly reduced computational effects achieved from the direct structure-performance mapping. The former optimization is carefully designed based on our recent theoretical insights, while the latter simulation is derived via a novel convolutional neural network based simulation which does not rely on any labeled simulation data but is instead designed based on a physics-informed loss function. As compared with results obtained using latest approach based on density interpolation, structures of better compliances are observed under acceptable computational costs, as demonstrated by our numerical examples.