Abstract:Many material properties are manifested in the morphologicalappearanceand characterized using microscopic images, such as scanning electronmicroscopy (SEM). Polymer miscibility is a key physical quantity ofpolymer materials and is commonly and intuitively judged using SEMimages. However, human observation and judgment of the images is time-consuming,labor-intensive, and hard to be quantified. Computer image recognitionwith machine learning methods can make up for the defects of artificialjudging, giving accurate and quantitative judgment. We achieve automaticmiscibility recognition utilizing a convolutional neural network andtransfer learning methods, and the model obtains up to 94% accuracy.We also put forward a quantitative criterion for polymer miscibilitywith this model. The proposed method can be widely applied to thequantitative characterization of the microstructure and propertiesof various materials.

Automatically Predicting Material Properties with Microscopic Images: Polymer Miscibility As an Example