Machine Learning Approach To Predict Aerodynamics Performance of Under-Body Drag Enablers 2020-01-0684
Implementing stringent emission norm and fuel economy requirement in the coming decade will be very challenging to the whole automotive industry. Aerodynamic losses contributes upto 13% to overall fuel economy and aerodynamicists will be challenged to have optimum content on the vehicle to reduce this loss. Improving Aerodynamic performance of ground vehicles has already reached its peak and the industry is moving towards active mechanisms to improve performance. Calibrating or simulating these active mechanisms in the tunnel or in Computational Fluid Dynamics (CFD) would be very challenging as the model complexity increases. Computationally expensive CFD models are required to predict the transient behaviors of model complexity. To balance these complexities and to reduce cost, objective of this piece of work is to explore feasibility of statistical data analytics and machine learning methods to come up with good predictive meta-models with least data, which can help to make technical decisions.
Machine Learning (ML) is the scientific study of algorithms and statistical models that computer systems use to perform a specific task without using explicit instructions, relying on patterns and inference instead. ML algorithms build a mathematical model based on sample data, in order to make predictions or decisions without being explicitly programmed to perform the task. ML algorithms are used in a wide variety of applications, where it is difficult or infeasible to develop a conventional algorithm for effectively performing the task.
Training data is generated through full factorial Design of Experiments (DOE) using CFD simulations. In this paper, various ML algorithms are used to generate models for predicting the aerodynamic drag and cooling flow with three continuous variables (Airdam, TireDam and Bellypan) and three discrete variables (Duckbill, Lower Bumper Stiffener and Engine Panel). Accuracy of various machine learning algorithms like, decision trees, linear regression, random forest, neural network etc, are discussed in this paper. The ML models can be used to predict a drag and cooling flow for any small changes in experimental space. Effect of all the enablers and interactions between different aero enablers are also discussed.