Prediction for magnetostriction magnetorheological foam using machine learning method

Abstract

Magnetorheological (MR) foam is a magnetic polymer composite (MPC) that can be used for soft sensors and actuators in soft robotics. Modeling mechanical properties and magnetostriction behavior of MR foam is critical to developing into MR foam devices. This study uses extreme learning machines (ELM) and artificial neural networks (ANN) to predict magnetostriction behavior. These models describe the nonlinear relationship between different carbonyl iron particle compositions, magnetic field, strain, and normal force. The model's hyperparameters (learning algorithms and activation functions) are varied. For ANN, RMSProp, and ADAM learning algorithms were used with sigmoid and ReLU activation functions. The ELM model considered the Hard limit, ReLU, and sigmoid activation function. The model was then evaluated for both training and testing data. Based on the results, ANN RMSProp Sigmoid, ELM with activation function ReLU, and Hard limit are more accurate than other models. However, the correlation analysis and comparison between prediction and experimental data show ELM Hard limit are more generalized in predicting strain and normal force with R2$$ {\mathrm{R}}<^>2 $$, 0.999, and RMSE less than 0.002. In conclusion, the ELM Hard limit model accurately predicts the magnetostriction behavior of MR foam, paving the way for future MR foam device development.