Harnessing interpretable and unsupervised machine learning to address big data from modern X-ray diffraction

Recent instrumental advances now enable large volumes of X-ray diffraction to be collected with high efficiency at synchrotron sources. This article shows that machine learning can produce an unbiased and comprehensive analysis of such data that uniquely combines both long-range and short-range structural correlations as a function of temperature. In Cd₂Re₂O₇, machine learning characterizes both the critical behavior of the primary order parameter and the Goldstone mode fluctuations that drive symmetry breaking at a lower temperature. The approach results from a synergy between computer scientists and physicists, producing a machine learning strategy that is interpretable within the established framework of physics and adaptable to other "big data" problems in materials science and engineering.