Performance Study of Surrogate Models for Large-Scale Optimization

As the role of simulation-based optimization grows in prominence for the design of advanced machines, its computation cost becomes a significant concern. Machine-learning-based surrogate models are expected to be a solution to solve the problem. Surrogate models need a certain amount of training, which requires simulation, such as Finite Element Analysis, to gain enough accuracy, which can be costly. The question is whether the use of surrogate models is cost-effective or not. To answer this question, we conducted numerical experiments. The study observed the initial acceleration of the optimization thanks to the surrogate model. However, the acceleration effect cannot be seen in the later stage, and the surrogate model misleads the optimization, preventing the gain of feasible solutions. This negative impact appears even in cases with a large amount of training.