Materials physics from microscopy: statistical and machine learning methods for tackling inverse problems

Materials imaging has undergone a dramatic transformation in the previous two decades, caused by the proliferation of new modalities in scanning probe microscopy that enable better functional property measurements at high resolution, along with advances in electron microscopy, that now make atomic-scale imaging nearly routine for a wide variety of material systems. This explosion of new imaging data brings forth challenges related to data compression, analysis, and ultimately, physics knowledge extraction.  

In this talk, I will highlight our recent advances at both accelerating knowledge acquisition through Bayesian optimization methods, as well as providing better insight into materials physics via Bayesian inference, Bayesian model selection, and statistical physics-based methods. Generative models that combine both local and global characterization data allow ‘stochastic’ model matching by comparing statistics of between target configurations and model predictions at the atomic and mesoscopic scale, and promise the ability to autonomously guide sample characterization towards proving or disproving available models describing the system and its dynamics. These improvements suggest that the next decade of microscopy will be greatly enriched by a close coupling between analytics at the edge, computational modeling, and machine learning methods. This work was conducted at and supported (RKV) at the Center for Nanophase Materials Sciences at Oak Ridge National Laboratory, a US DOE Office of Science User Facility.