Modeling the butterfly behavior of SMA actuators using neural networks

Shape memory alloy (SMA) actuators are an important application of smart materials for robotics. However, the nonlinear behavior of SMA leads to difficulties in real-time simulations using numerical methods. Artificial Intelligence can be used to bypass this problem. In this paper, we study several neural networks (NNs) to model the superelastic or pseudo-elasticity effect (SEE) as well as the shape memory effect (SME) used in SMA. Focusing on antagonistic actuating, we first model a single wire to train the best NN with the proper characteristics that fit the behavior of SEE. Then, we model the SME of two linear antagonistic SMA wires used as an actuator. In both systems, single and antagonistic wires, we train the networks to obtain the stress-strain diagrams representing the behavior. The network type and training algorithm are key factors and are evaluated depending on the RMSE values. As a result, we find that the long short-term memory NN, used with a regression layer on standardized data sets, models the butterfly-shaped behavior of the actuator system with less RMSE value.