Technical Awards:2022 JSAE Award The Outstanding Technical Paper Award - Surrogate Model Development for Prediction of Car Aerodynamics Using Machine LearningTable 4 Error in CD of train and testFig.13 Error history in CD with increase of train casesFig.14 Comparison of coeffi cient of aerodynamic drag (CD) between CFD and proposed modelTable 5 Comparison of computational time between CFD and 3.5 Estimated time required for machine learning 85984. Summary and conclusionproposed modelmodelIn equation (2), ai and yi represent the CFD result and the estimation result of the proposed model, respectively, and n represents the number of training cases or the number of testing cases.Because the standard deviation and MAPE of the training case (a) shown in Fig. 12 and Table 4 are suffi ciently small, it is evident that the dataset has been trained appropriately. For the testing case (b), however, the error is larger than that for the training case, and there exists a problem with the generalization performance.The error trend with an increase in training cases is shown in Fig. 13. The horizontal axis represents the number of training cases, and the vertical axis represents the error (standard deviation). As the number of training cases increases, the error of the testing cases decreases. Therefore, it is considered effective to increase the number of training cases in order to improve the generalization performance.The CD values and CFD results of the test cases estimated by the proposed model are compared using a scatter plot, as shown in Fig. 14. The vertical axis represents the CD values estimated by the proposed model, and the horizontal axis represents the values calculated using CFD. The testing case includes six car types, and the CD values estimated by the proposed model for each car type are generally within ±0.012 of the CFD results.In the initial phase of car development, aerodynamic evaluation is performed on multiple car design proposals; however, various car shapes have been proposed, and CD also varies considerably more than the error of the proposed model. Therefore, in the initial phase, the proposed model can be used to evaluate the superiority or inferiority of CD for each proposed design, and it is expected that the frequency of CFD usage and the amount of calculations can be reduced.The times required to predict the velocity, pressure, and CD in Case A described above when using the proposed model are shown in Table 5 and compared with the times required when the commercial CFD software is used.In the case of CFD, approximately 24 h was required for mesh preparation and fl ow calculation (1). This calculation time was recorded when using Intel® Xeon® (approximately 80 million meshes and 256 CPU cores). For the proposed model, a notebook PC (CPU: Intel® Core-i5™) was used to compute the distance function; this required approximately 11 minutes, whereas the predictions of the velocity, pressure, and CD only required 3 seconds. Therefore, the proposed model is deemed useful for checking the qualitative fl ow fi eld and CD for various shape designs within a short period of time.In this study, a practical surrogate model that can predict the fl ow fi eld and CD around a car with a complicated three-dimensional shape was developed.1 We developed a model using residual blocks, instead of the fully connected layers used in the network structure of Guo et al., to effectively avoid gradient explosion and accurately predict the fl ow fi eld around the car shape.2 It was confi rmed that the fl ow fi eld estimated by the proposed model could qualitatively reproduce the No.88 (2022) NISSAN TECHNICAL REVIEW
元のページ ../index.html#90