Combining Variational Autoencoders and Physical Bias for Improved Microscopy Data Analysis

Abstract Electron and scanning probe microscopy produce vast amounts of data in the form images or hyperspectral data, such as Energy Loss Spectroscopy (EELS) 4D Scanning Transmission Microscope (STEM), that contain information on a wide range structural, physical, chemical properties materials. To extract valuable insights from these it is crucial to identify physically separate regions phases, ferroic variants, boundaries between them. In order derive an easily interpretable feature analysis, combining with well-defined principled unsupervised manner, here we present physics augmented machine learning method which combines capability Variational Autoencoders disentangle factors variability within driven loss function seeks minimize total length discontinuities corresponding latent representations. Our applied various materials, including NiO-LSMO, BiFeO3, graphene. The results demonstrate effectiveness our approach extracting meaningful large volumes imaging data. customized codes required functions classes develop phyVAE available at https://github.com/arpanbiswas52/phy-VAE