Towards convolutional neural networks for finite element modeling of the tongue

Muscular hydrostats such as the tongue are modeled using the finite-element (FE) method, which is computationally expensive. To reduce the computational cost, approximate techniques based on dimensionality reduction (DR) and deep neural networks (DNN) have been proposed. On the one hand, linear DR techniques are not expressive enough, while on the other hand fully-connected DNNs are highly over-parameterized. The large number of parameters require a large number of training samples that are expensive to obtain. Therefore, DNNs are impractical for modeling complex systems such as a high-resolution FE tongue. Given the lumped-mass assumption and local connectivity of the nodes, we propose convolutional neural networks (CNN) for FE model approximation. We applied the proposed CNN architecture to a muscle actuated beam and FE tongue model. The CNN architecture resulted in better fidelity to the FE models compared to DNNs with an order of magnitude fewer parameters.