Correlative and causal machine learning in scanning probe and electron microscopy

M. Ziatdinov
Oak Ridge National Laboratory,
United States

Keywords: machine learning, AI, electron microscopy


Classical deep and machine learning have recently emerged as powerful tools for the analysis of mesoscopic and atomically resolved images and spectroscopy in electron and scanning probe microscopy experiments. Here, I will start by illustrating the successful examples of classical machine learning applications to the imaging/spectroscopy of nano-materials, including the applications of deep convolutional neural networks for atom and defect identification in scanning transmission electron microscopy from 2D materials and utilization of Gaussian processes for the reconstruction and resolution enhancement of 3D and 4D spectroscopic data and uncertainty-guided sample exploration in hyperspectral scanning probe microscopy on ferroelectric materials. I will then discuss one of the fundamental limitations of the classical machine learning methods, namely their correlative nature, which makes them very susceptible to confounding factors and observational biases abundant in any real experiment and talk about our recent work on causal learning from structural scanning transmission electron microscopy observations on several material systems that attempts to address this limitation.