Application of artificial neural networks for prediction of crushing performance of a Quadricycle’s S-rail

Heavy electric quadricycles (L7e) typically have the problem of crash incompatibility because their size and
weight are smaller and lighter than other vehicles. The S-rails, located in front of the automotive underbody, are the main components for absorbing the impact energy of the L7e. The design of the L7e's S-rail for crashworthiness is challenging because the design space is relatively limited. High computational resources are commonly required for crash simulation due to a vast amount of data from nonlinear explicit dynamic analyses. Machine learning techniques can more quicky provide and extract accurate crash predictions from a combination of feasible solutions. This work employs a finite element analysis via LS-DYNA to obtain discrete results of the peak crushing force and the specific energy absorption for S-rail impact. The artificial neural network (ANN) module in MATLAB and response surface methodology (RSM) were then applied to predict the crushing-performance surrogate models based on the S-rail's thickness and geometry. To conduct an accurate learning process of the ANN, 81 data points for the crashworthiness simulation were adopted as training inputs with ten test data points. Furthermore, the number of neurons and activation function are examined to acquire the most accurate results for predicting the dynamic crushing of an S-rail. The investigation shows that ANN is more accurate than RSM in predicting the S-rail's peak crash force and specific energy absorption, especially when applying ReLU activation functions. With its high accuracy and capability to predict multiple targets, the ANN-ReLU is recommended to perform surrogate-based model optimization.