CAEサロゲートモデル構築における学習データの効率的なサンプリング手法の検討

Abstract

In order to construct a surrogate model for a finite element analysis (FEA) model using machine learning, we investigated which sampling method should be used to construct a highly accurate surrogate model with less training data. Random sampling and Latin hypercube sampling were used as methods for predetermining sampling points, and Voronoi sampling and active learning were used as methods for adaptively adding sampling points. In active learning, a Gaussian process regressor (GPR) was used as the learner model, and the variance of the probability distribution output from the GPR was used as the acquisition function. Also, in the other sampling methods, the same GPR kernel configuration used in the active learning was utilized to create surrogate models. Using these four sampling methods, we created surrogate model for the FEA model which is used to predict the strength of rupture disc for air-conditioning compressor and clarified that active learning could create a highly accurate surrogate model with a small number of training data.