Data-driven modeling for complex contacting phenomena via improved neural networks considering link switches

Recent years saw tremendous developments of data-driven modeling in various engineering fields. As for the contact modeling between complex surfaces, the utilization of neural networks successfully eliminates the limitations encountered by the traditional physics-based contact modeling strategy. However, contrary to its increasingly extensive applications, very little attention has been paid to the role of network hyper-parameters in reducing the model redundancy and improving its training efficiency. In this work, a novel neural network considering link switches has been presented for the data-driven modeling of complex contact phenomena. In order to further boost its prediction performance, genetic algorithm (GA) is employed for the optimal settings of relevant hyper-parameters. An indoor experimental setup is utilized to demonstrate the effectiveness of the presented methodology. Comprehensive comparisons with the base models indicate the superiorities of the established locally-connected-neural-network-based contact force model for complex geometries.