A Machine Learning Approach for Mining Structure-Property Linkages in Magnetorheological Elastomer Composites

Since it was firstly investigated in 1996, magnetorheological elastomers (MREs), consisting of polymer and ferromagnetic particles, are one kind of smart composite materials whose mechanical properties can be altered by external magnetic field. There has been a great amount of research focusing on establishing the structure-property linkages in MREs by utilizing experimental methods or numerical approaches such as finite element methods (FEM). In recent years, data-driven methods, such as statistical continuum theories and machine learning approaches, are emerging as a new toolset in materials science, which may open new opportunities in investigating the structural-property linkage of MREs in a more efficient manner.In this thesis, a machine learning approach is employed to mine the structural-property linkages of MREs both in and out of the presence of magnetic field. Features that could well represent the microstructure of MREs are constructed and selected, following by a training and validating process utilizing a regression tree algorithm for the prediction of the strain field of MREs under a certain macroscopic load. It is demonstrated that contrast ratio of MREs has large effect on the model performance. Also, this model is proven to be able to effectively predict the localization property of MRE in both magnetic load conditions when contrast ratio is at relatively low level