Towards data-driven modeling for complex contact phenomena via self-optimized artificial neural network methodology

In recent years, various physics-based contact force models have been proposed for describing contact/impact phenomena in different engineering domains. However, difficulties emerged in obtaining the critical dynamical parameters as well as suitable constitutive relations for complex contacting surfaces, become the limiting factors of the traditional contact modeling strategy. Meanwhile, data-driven methodology gradually flourishes and develops as a fast and reliable alternative to the traditional method. The accuracy and reliability of artificial neural network (ANN) rely heavily on the settings of its architecture and initial weights and biases. But less attention has been paid to the selection of these hyper-parameters for the moment. In this work, we aim to present a self-optimized ANN methodology, as an endeavor to discover more accurate and robust models which can simulate the interaction process between complex geometries. Genetic algorithm combined with sequential quadratic programming (GA-SQP), is utilized to boost the consistent performance of the network. An indoor experiment rig is employed to validate the presented strategy, of which the results show great performance enhancement and excellent robustness of the neural-network-based contact force model.