Cross-scale covariance for material property prediction

Correlations between fundamental small-scale properties computable from first principles, which we term "canonical properties," and complex large-scale quantities of interest (QoIs) provide an avenue to predictive materials discovery. We propose that such correlations can be discovered through simulations with efficient approximate interatomic potentials (IPs), which serve as an ensemble of "synthetic materials." In a manner similar to statistical studies in public health, we analyze correlations of QoI with canonical properties, identify the best predictor properties, and build a cross-scale QOI-on-predictors regression model, which can be used to estimate regression errors over the statistical pool of IPs. We demonstrate this with two QoI examples: (1) symmetric tilt grain boundary (GB) energies computed for 234 IPs; and (2) plastic flow strength computed for 178 IPs through large-scale (~100 million atom) molecular dynamics simulations. We further postulate that IP and first principles density functional theory (DFT) predictions belong to the same statistical pool, allowing the regression model obtained from the IPs to be used with DFT values. We are able to confirm this hypothesis for the GB energy example for which DFT values are available for the QoIs.