Surrogate modeling of vehicle dynamics using deep learning

Abstract

Nowadays, Computer Aided Engineering (CAE) is indispensable for vehicle design; however, its computation time is a heavy burden when tuning design parameters. In this respect, some studies in several fields have been carried out for replacing CAE with machine learning methods. In this paper, we propose neural sequence modeling with Recurrent Neural Networks (RNNs), which are neural networks that treat sequences such as temporal sequences to imitate CAE as a computationally-efficient neural surrogate model. Our target task is the NCAP Fishhook test to evaluate vehicle dynamics of the rollover propensity. Our model reduced the error in approximately 10% for NCAP Fishhook test dataset compared to that of the baseline hierarchical neural network model. Furthermore, our model has task-specific characteristics: (1) skip connection, (2) hybrid loss and (3) scheduled sampling to improve performance and stability. We confirmed that the skip connection reduced errors in the additional ablation study. Our experiments showed that the sequential model is effective as a surrogate model for CAE.