Title: Disentangling ferroelectric domain wall geometries and pathways in dynamic piezoresponse force microscopy via unsupervised machine learning

Abstract: Domain switching pathways in ferroelectric materials visualized via dynamic Piezoresponse Force Microscopy are explored via rotationally invariant variational autoencoders (rVAE). rVAEs simplify the elements of the observed domain structure, crucially allowing for rotational invariance, thereby reducing the variability of local polarization distributions to a small number of latent variables. Importantly, these latent variable representations generally enable characteristic aspects of materials behavior to be disentangled. For small window sizes the latent space is degenerate, and variability is observed only in the direction of a single latent variable that can be identified with the presence of domain wall. In this case, the rVAE effectively becomes an edge filter allowing for arbitrary rotations, and analysis of the latent space distributions yields insight into wall angle distributions. For larger window sizes, the latent space is 2D, and the disentangled latent variables can be generally interpreted as the degree of switching and complexity of domain structure. Applied to multiple consecutive PFM images acquired while monitoring domain switching, the polarization switching mechanism can thus be visualized in the latent space, providing insight into domain evolution mechanisms and their correlation with the microstructure. This approach is furthermore widely tunable via a modification of VAE's loss function, as well as the priors in the latent space. This enables broad applicability for investigating a wide range of physical and biological phenomena, not just with scanning probes but also for electron, optical, and x-ray microscopies. The interactive Jupyter notebook that reproduces this paper's results is available at this https URL