Application of artificial neural networks and support vector regression modeling in prediction of magnetorheological fluid rheometery

Abstract

In this study, rheological properties of Magnetorheological Fluid (MRF) have been investigated by Artificial Neural Network (ANN) and Support Vector Regression (SVR) methods The effects of temperature and magnetic field strength on rheological properties have been modeled. The experimental results (about 600 data) have been estimated about 3 percent average error in interpolations and about 5 percent average error in extrapolation of models. Low accuracy of data at low shear stresses and inability to determine the reliable values of dynamic yield stress are challenging problems in using analytical models for MRF rheological properties prediction. The shear stress has been predicted by using ANN and SVR models at low range of shear rates (to 10−5s−1) and the results showed that SVR is the most reliable model in predicting shear stress as well as dynamic yield stress (at low shear rate values). Therefore, the dynamic yield stress has been obtained with log extrapolation prediction of experimental data, Herschel-Bulkley model, ANN and SVR models. It was revealed that Herschel-Bulkley model was unable to predict the dynamic yield stress in a wide range of temperature and magnetic field strength. The SVR model showed the most reliable predicting results both in interpolations as well as extrapolations when trying to predict the dynamic yield stress. The SVR model had an appropriate estimation of shear stress at low shear rates, also a good coincidence with existing experimental data.