Surrogate Modeling of a Lumped-Mass Multibody Structure Using Hamiltonian Neural Networks

Abstract

The complexity of highly flexible structures restricts their use in real-time simulations. To address this challenge, we investigate the use of Hamiltonian neural networks (HNNs) as an alternative method for modeling a highly flexible cantilever beam. We derived the reference structural model using a lumped-mass rigid multibody method considering the Hamiltonian formalism and used it to generate a dataset consisting of generalized coordinates and momenta as inputs and their respective time derivatives as outputs. The trained neural networks are used as surrogate models to simulate the cantilever beam under free and forced conditions. Preliminary findings indicate that HNNs create accurate and efficient surrogate models whilst learning conservation laws. For forced-response simulations, our approach requires analytical calculation of external forces, offsetting the computational Efficiency gains of our surrogate models. The outcomes of this study give initial perspectives and limitations of the use of surrogate models based on HNNs as a means to efficient simulations of highly flexible structures.