NISSAN_TECHNICAL_REVIEW_89 (2023)
86/111

Technical Awards:2022 JSAE Award The Outstanding Technical Paper Award - Surrogate Model Development for Prediction of Car Aerodynamics Using Machine LearningFig.1 CNN structure of prediction model for fl ow fi led proposed 2.2 Outline of estimation model for estimating flow fieldFig.2 Structure of proposed model for fl ow fi eld estimationFig.3 Details of encoder, residual blocks and decoder of proposed model for fl ow fi eld estimationdistance function can be applied to complicated shapes and can also be extended to three dimensions relatively easily, the model structure proposed by Guo et al. was adopted for the fl ow fi eld (i.e., fl ow velocity and pressure) estimation.However, because CD cannot be estimated using the model of Guo et al., a separate surrogate model for estimating CD is required. Therefore, we proposed a surrogate model that estimates CD (6) based on the input of voxel data. Because the estimation accuracy of CD is satisfactory, the surrogate model structure of CD was adopted in this study as well.In related studies, a single model that can learn both the fl ow fi eld and CD has not been proposed thus far. Therefore, in this research, two different machine learning models based on related studies were developed to estimate the fl ow fi eld and CD.The basic confi guration of the fl ow fi eld estimation model adopts the network structure proposed by Guo et al., as shown in Fig. 1. This model consists of an encoder and a decoder connected by a fully connected layer. When applying the model practically, such as in this study, it is necessary to increase the size of the distance function to be input according to the degree of complexity of the shape. In addition, the intermediate layers between the encoder and decoder need to be multilayered to improve the estimation performance. However, the use of multilayered intermediate layers easily causes gradient disappearance or gradient explosion problems during training, making it diffi cult to proceed with training. Therefore, to solve the gradient disappearance and explosion problems, in this study, residual blocks (7) were used instead of the fully connected layer in the model structure of Guo et al. The model proposed herein is illustrated in Figs. 2 and 3. This model inputs a distance function and outputs (a) the three components of the fl ow velocity vector and (b) the pressure. For the encoder and instance decoder, normalization (9) are used to suppress the disappearance of the gradient.normalization batch (8) and 2.3 Outline of estimation model for estimating CD The estimation model for estimating CD is described in this section. We previously proposed an estimation model that estimates CD based on the input of voxel data (6). Although the model is based on Voxnet, proposed by Maturana (10), global average pooling, instead of a fully connected layer, is used according to the research of Lin et al. (11) to reduce the amount of calculation when converting the intermediate layer into one dimension. Because the estimation accuracy of CD is satisfactory, the estimation model structure based on our previous study was adopted in this study as well.8194by Guo (4)No.88 (2022) NISSAN TECHNICAL REVIEW

元のページ  ../index.html#86

このブックを見る