Predicting Finite Element Simulation output using Machine Learning: A study to understand the potential of Graph Neural Networks in predicting vehicle occupant pre-crash simulation kinematics

Abstract

Each year around 1.35 million people die in traffic crashes and many more get injured. Human Body Models have become a popular tool to understand and prevent injuries and fatalities as they simulates the human body response during crash sce narios. With improvements in the accuracy and bio-fidelity of Human Body Models, the demand for computational resources is increased up to requiring days for a pre single crash simulation. Therefore, Dimensionality Reduction methods can be used to create a surrogate that is a smaller representation of the model. The reduced representation translated to a latent space that still retains the complexity and the features of the simulation. To build the surrogate, supervised learning tasks have been implemented on the reduced dimension to map the simulation to its parameters. To create a compressed version of the input simulation, Graph Neural Networks have been considered since they incorporate the geometrical structure of the model. The Graph Neural Network has been compared to an equivalent Convolutional Neu ral Network architecture and to a Principal Component Analysis. Random Forest, Gradient Boosting, and XGBoost were chosen and compared to build the surrogate model. Due to computational limitations, and since pre-crash simulations are al ready time-consuming, a simplified Hybrid III dummy model was chosen in place of a full Human Body Model. For all Dimensionality Reduction methods, the accuracy of the reconstruction im proves by increasing both the number of samples and the latent space size. Principal Component Analysis shows a better performance in terms of lower errors compared to Graph Neural Network. Moreover, the Graph Neural Network structure is compa rable to an equivalent Convolutional Neural Network architecture in term of perfor mance. Random Forest and Gradient boosting proved to be better than XGBoost. Principal Component Analysis requires less computational time and resources. Although Graph Neural Network was outperformed in this study, further improve ments in the development of the method may still have potential.