Reduced Order Modeling of CFD Model Using Machine Learning and an Application for Heat Damage Evaluation

Abstract

The purpose of this paper is to propose a method to construct a surrogate model which can predict flow filed of velocity and temperature aiming at decrease the computational cost of CFD. In the proposed method, training data are corrected from CFD simulation based on a Design of Experiments (DOE). Then, after performing missing value interpolation on the training data, Tucker decomposition is applied to training data to extract features from the tensor type training data. For regression model, Gaussian process is introduced to construct surrogate models. The feasibility of the proposed method is illustrated by an application for CFD model using for the heat damage design problem.