Predicting Vehicle Aerodynamics Using a Machine Learning Model Based on Physics

Abstract

In this research, we construct a machine learning model that can predict the drag coefficient based on an existing physics-based machine learning model. The base model can predict velocity and pressure fields accurately thanks to the physical knowledge embedded. Our novelty is to add a model that computes the drag coefficient inside it rather than postprocessing for more accurate results. Also, we generated a dataset using aerodynamic simulation with various shapes generated based on the DrivAer model. The model shows high accuracy with an error of 0.0025 in the drag coefficient for the considered dataset.