NISSAN_TECHNICAL_REVIEW_89 (2023)
85/111

2022 JSAE Award The Outstanding Technical Paper Award1. Introduction2. Machine learning model2.1 Previous Studies & Related Studies Technical Awardsconstructed a machine learning model to learn and estimate the quantifi ed features and characteristics evaluated via CFD (fl ow velocity, pressure, and CD) for the car shape by using the Gaussian process to estimate the CD values and fl ow fi elds of the cars. They trained the model with approximately 800 examples of car shapes and CFD results. They reported that the CD and fl ow fi eld estimation results matched the CFD calculation results, which served as reference information for our research. However, it is diffi cult to learn complex shapes of actual cars using their method. Specifi cally, this method uses a polyhedron called “PolyCube” and projects the constituent points of the PolyCube onto a car shape within a range in which the topology does not vary. The coordinates and heights of each projected constituent point are used as features. It is necessary to prepare a PolyCube with a topology similar to the shape of the car to be learned. However, it is diffi cult to use the same PolyCube for all car shapes because actual shapes have different topologies; for instance, minivans, hatchbacks, and sedans have 1-, 2-, and 3-BOX topologies, respectively. Hence, a PolyCube suitable for each car model needs to be prepared for individual learning, which requires time and effort. In addition, the actual cars have complicated shapes such as engine room parts, tires, fl oor components, and door mirrors, and it is diffi cult to prepare a PolyCube with a topology that can express such complicated shape elements. Therefore, the results obtained by Umetani et al. can only be applied to relatively simple car shapes.By contrast, Guo et al. (4) used a distance function that represents the distance from an object as input data and constructed a machine learning model that estimates two- and three-dimensional fl ow fi elds (corresponding to the fl ow velocity vector and pressure) based on a convolutional neural network (CNN) (5). Because the *Integrated CAE・PLM Department  **Nissan Research Center Mobility & AI LaboratoryAbstract  In the evaluation of car aerodynamics, Computational Fluid Dynamics (CFD) are frequently used as well as a wind-tunnel. However, the CFD simulations consume a lot of resources and time. In this study, a surrogate model using machine learning was developed to reduce the amount of resource and time needed for CFD. In the proposed model, the relation between car shapes and CFD results was learned for rapid prediction of pressure, velocity, and coeffi cient of drag for aerodynamics. In this paper, we introduce the proposed model, the training dataset, the accuracy, and the computational time.In recent years, it has become necessary to further reduce aerodynamic drag in order to improve the fuel effi ciency competitiveness of cars and comply with new environmental fl uid dynamics (CFD) is often used as well as wind tunnel experiments to evaluate the aerodynamic performance of cars (1,2). However, CFD requires a signifi cant amount of time and resources. In particular, the resources required to perform CFD computations are increasing every year owing to the necessity of pre-calculations prior to experiments, the increase in the number of evaluation specifi cations, and the demand for improved calculation accuracies. Therefore, this study attempted to develop a surrogate model that estimates the CFD results (fl ow velocity, pressure, and drag coeffi cient CD) using machine learning method, by learning the relationships between the car shape and the quantities evaluated using CFD. Using this approach, aerodynamic analyses can be replaced with the surrogate model, such that the amount of CFD computations and time consumption, as well as the corresponding costs, can be reduced. This paper outlines the proposed method, the dataset, and the validation results.*Received on November 10, 2020. Presented at the Technical Sessions of the Autumn Congress of the Society of Automotive Engineers of Japan on October 23, 2020.regulations. Computational Several studies similar to the present investigation have been reported. For instance, Umetani et al. (3) NISSAN TECHNICAL REVIEW No.88 (2022)8093Surrogate Model Development for Prediction of Car Aerodynamics Using Machine LearningKei Akasaka* Fangge Chen** Takehito Teraguchi**

元のページ  ../index.html#85

このブックを見る