Symmetry-constrained machine learning on 3D-ΔPDF to identify short range correlations

The 3D difference pair distribution function (3D-ΔPDF) is a powerful way to understand structural inhomogeneity without the need for simulation. Although 3D-ΔPDF maps are often intuitive to interpret and provide a guide to the appropriate disorder models without computationally expensive modeling, a detailed analysis is hindered by the difficulty of visualizing high dimensional data, the loss of phase information and the uncertainty of embedded local disorder orientations. To address these challenges, we have developed a symmetry-constrained machine learning architecture to identify short-range structural distortion modes from experimental 3D-ΔPDF maps. Use cases of this approach to predict local structural distortions will be discussed.