Rapid market changes are requiring automakers to reduce development time through agile development processes. To meet this demand, it is important to make clear the feasible design space by utilizing simulation in the early stages of vehicle development. However, it is still challenging to use Computational Fluid Dynamics (CFD) in the early stage of vehicle development because of its computational costs. Therefore, a simple alternative to CFD is required. The purpose of this paper is to propose a method to construct a surrogate model which can predict three-dimensional flow filed composed of velocity magnitude and temperature. In this paper, we propose a surrogate modeling of CFD simulation for heat damage evaluation using supervised and unsupervised learning. In the proposed method, surrogate model was created by combining model feature extraction process and regression process and this method was applied to CFD model for heat damage evaluation in a vehicle engine compartment. In the feature extraction process, unsupervised learning was applied in order to extract features from the tensor type training data. Then, supervised learning was introduced in the regression process to learn the relationship between each feature and the design parameter. This paper describes the results of applying this proposed method to CFD model for heat damage evaluation and verifying its technical effectiveness. As the result, the prediction accuracy was higher than that of conventional method, especially in the strongly nonlinear regions. In addition, the computational cost was significantly reduced compared to CFD. These results confirmed the technical usefulness of this proposed method. A limitation for this study is that the number of samples is not large. Due to the limitation of computing resources, only 99 samples could be prepared. Using more data can provide more accurate validation results. We presented a method to construct a surrogate model which can predict three-dimensional flow filed of velocity magnitude and temperature. The proposed method combined unsupervised learning and supervised learning. Unsupervised learning was used to extract features from the training data and supervised learning was introduced into regression model. In numerical experiments, we applied this method to CFD model for heat damage evaluation in the engine compartment. As a result, we obtained higher accuracy compared to the previous surrogate model. The proposed method can be used in early stage of vehicle development to make development more effective.
Ms. Haruna Kawai, technical staff, Toyota automoble company