Machine Learning-Aided Design of Materials with Target Elastic Properties

A set of universal descriptors which combines atomic properties with crystal fingerprint are presented to build interpretable models for elastic property prediction. Using the well-performed model, 100 materials with large predicted elastic moduli are screened out and then validated by the first-principles calculations. When performing projection analysis, we find that compounds with large and small elastic moduli are clearly divided into two parts by the average value of volume and atomization enthalpy (ΔH atomic), and the relation between them is given by two discriminant equations, suggesting that compounds composed of elements with large ΔH atomic are potential large elastic moduli materials. Following this rule, we design several new stable materials like ReTcB4 and ReB which have high elastic moduli. This method is valuable for high-throughput screening and material design.