"Predicting properties of structurally and chemically complex materials using physics-informed statistical learning"
To apply statistics and data science tools to aid computational designs of materials is under fast development. There are two unique aspects of the applications of these tools in materials science. First, the training sets are usually small. Second, physical mechanisms of material properties can be applied to facilitate the constructions of descriptors and statistics learning methods. In this talk, I will give three examples to address these two issues. The first example is to use machine learning to predict density and elastic moduli of SiO2-based glasses. Our machine learning approach relies on a training set generated by high-throughput atomistic simulations and a set of elaborately constructed descriptors with the fundamental physics of interatomic bonding. The predictions of our model are comprehensively compared and validated with a large amount of both simulation and experimental data. In the second example, a general linear correlation can be found between two descriptors of local electronic structures at defects in pure metals and the solute-defect interaction energies in binary alloys of refractory metals with transition-metal substitutional solutes. This linear correlation plus a residual-corrected regression model provides quantitative and efficient predictions on the solute-defect interactions in alloys. In addition, with these local/global electronic descriptors and a simple bond-counting model, we developed regression models to accurately and efficiently predict the unstable stacking fault energy (γusf) and surface energy (γsurf) for refractory multicomponent alloys. Using the regression models, we performed a systematic screening of γusf, γsurf, and their ratio in the high-order multicomponent systems to search for alloy candidates that may have enhanced strength-ductile synergies. First-principles calculations also confirmed search results.