Technical Awards：2022 JSAE Award The Outstanding Technical Paper Award - Surrogate Model Development for Prediction of Car Aerodynamics Using Machine LearningFig.9 Comparison of pressure between CFD and proposed model in central cross-sectionFig.10 Comparison of velocity magnitude between CFD and proposed model in central cross-sectionNISSAN TECHNICAL REVIEW No.88 (2022)Fig.11 Comparison of velocity magnitude between Guo’s model and proposed model in central cross-section3.3 Effect of the residual blocks3.4 CD validation resultsFig.12 Histograms of error in CDThe pressure distribution at the central cross-section of the car is shown in Fig. 9, where the pressure is converted into a dimensionless value by using the dynamic pressure. In each case, the pressure distribution estimated by the proposed model reproduces the tendency of the CFD results.The velocity magnitudes in the central cross-section of the car are shown in Fig. 10. The velocity was made dimensionless by using the car speed. The proposed model qualitatively reproduces the CFD results in terms of the size of the rear stream region behind the car and the tendency of the velocity magnitude, which are important, especially in aerodynamic evaluations.According to the comparison results shown in Fig. 9 and Fig. 10, it was confi rmed that the proposed model can reproduce not only the qualitative tendency of the fl ow fi eld even in case D, in which the error between pressure and velocity is relatively large, but also the characteristics of the fl ow fi elds of different car types. The network structure proposed by Guo et al., in which a fully connected layer was adopted for the connection between the encoder and decoder, and the network structure proposed in this study, in which residual blocks were adopted instead of a fully connected layer, are compared in this section. The velocity magnitudes at the central cross-section of the car are shown in Fig. 11. The results of the CFD calculations and training using the residual blocks are presented in Figs. 11 (a) and (b), respectively. Here, the results of the proposed model are similar to the CFD calculation results. By contrast, in the results of training using the fully connected layer, shown in Fig. 11 (c), the tendency of the velocity magnitude is different from that in the CFD results. Owing to the gradient explosion, the weight optimization failed during training, and the value of the loss function of the velocity vector was 0.1. This value is larger than that of the loss function of the proposed method, as shown in Table 3, and it did not decrease even when the number of epochs was increased. According to these results, the residual blocks were considered to be effective.The errors in the CD values estimated using the proposed model are depicted in Fig. 12; the errors obtained in the training and testing cases are shown in (a) and (b), respectively. Table 4 shows the results of evaluating the error using the mean absolute percentage error (MAPE), described in equation (2), and the standard deviation:8497
元のページ ../index.html#89