Physics-Informed Neural Network for Scalable Soft Multi-Actuator Systems

Soft actuators, distinguished by their complex nonlinear behavior, are difficult to model analytically and cumbersome to prototype. Finite element (FE) models allow for more efficient behavioral prediction, but often require onerous setup, especially for large systems. We present a physics-informed neural network model formed by combining a low fidelity analytical model and input-convex neural networks to learn an underlying energy potential for the actuator from experimental and finite element simulation data. In doing this, the neural network can provide sufficiently accurate predictions about systems made up of multiple units, essentially scaling the model from a single unit to an assembly of many. To test this concept, we compare predictions of the deformation of a 5-actuator system from an FE model and from the physics-informed neural network. The neural network, which provides a prediction similar in accuracy to the FE equivalent, can more easily be adjusted to execute systems of greater quantities of units without drastic increases in computational consumption. In this way, we can scale our predictive understanding with adequate accuracy without compounding resources.