Machine Learning Techniques to Model Highly Nonlinear Multi-field Dynamics

Modelling the dynamics of the membrane displacement in a micromachined beam fixed at both ends for different applied voltages is important for real applications. The strong nonlinearities involved and the interaction between multiple physical fields make this task challenging for classical modelling and model reduction approaches. In this work we search for a simplified, yet accurate, data-driven models, based on different recurrent neural network architectures, using only peripheral input-output information of the original system. The main goal is to find the most suitable neural network architecture having the smallest number of hidden units that provides low error of the minimum gap dynamics for different applied voltages. We show that these black-box models, with only 4 hidden units, are able to accurately reproduce the original system’s response to a variety of different stimuli, and a strategy to make them parameter aware is proposed.