Modeling for viscoelastic behaviors of magnetorheological elastomer using single hidden layer feed-forward neural network approaches

Abstract

The prediction of magnetorheological elastomer (MRE) dynamic modulus behavior is a challenging process because of the material’s highly nonlinear nature. This problem becomes apparent while considering various possible material’s fabrication parameters selection. Previously, parametric modeling techniques such as Kelvin Voigt and Maxwell's models were applied to simulate the viscoelastic behavior. Nevertheless, it required parameter identification or data fitting for each applied magnetic field which is less efficient and becomes more complex when considering various material responses. In other words, parametric modeling method’s performance was limited in the change of input-output data, especially for larger-scale cases involving vast databases. Consequently, prediction model construction using a non-parametric approach such as machine learning has gained much attention in recent years. The advantages of machine learning techniques, such as to identify complex patterns or trends, and the ability to handle multi-variety of data, allow its potential to be utilized in material science study. Therefore, this research presents a data-driven approach prediction model using machine learning techniques for predicting the dynamic viscoelastic modulus of MRE. The multiple input multiple output-dependent dynamic modulus models were formulated using two feedforward neural network approaches called backpropagation artificial neural network (BP-ANN) and extreme learning machine (ELM). In this research, the MRE samples were synthesised under various compositions to undergo dynamic testing using a rheometer for data collection purposes. For the basic model design, three inputs variables were considered which were the shear strain, magnetic flux density, and input frequency. On the output side, storage and loss modulus were the targeted material dynamic properties. Meanwhile, for extended model design, fabrication effects such as filler concentration and distribution were also considered as additional input to predict dynamic modulus. To optimize the model configuration, sensitivity analysis was conducted. Here, the hyperparameters such as a number of hidden nodes and types of activation functions were varied in the training process. Thereafter, hyperparameters for optimized model configuration were selected based on the training accuracy performance. Next, the models were evaluated by utilizing the testing data sets for generalization purposes. Evaluation results showed that the ELM model had produced higher prediction accuracy, particularly at the linear viscoelastic (LVE) region where the achieved root mean square error (RMSE) and coefficient of determination (R2) were 0.0021 MPa and 0.994 respectively. Moreover, in terms of material’s fabrication effect, the ELM model also had demonstrated promising performance in forecasting the unlearned filler concentration where a relatively small RMSE of 0.0096 MPa was recorded. It is concluded that the ELM model had shown its potential to be as an accurate, flexible, and fast prediction modeling platform. The establishment of this non-parametric approach to replace the parametric model in predicting material dynamic properties is expected to contribute towards a time-efficient and cost-effective strategy for the MRE-based device development process.