Evaluation of principal features for predicting bulk and shear modulus of inorganic solids with machine learning

Abstract

The modulus of elasticity describes the ability of a solid to resist external forces, and plays a critical role in the development of new materials to maintain structural integrity. This work identifies the ideal feature for developing a prediction model for two elastic moduli - shear modulus (G) and bulk modulus (K)- of inorganic solids. A total of 4399 were generated (4392 features using material optimal descriptor network (MODNet) featurizers and seven additional features from the previous work) for the 17,051 datapoints, and four feature selection methods were applied: Pearson's correlation coefficient-based feature selection (PFS), MODNet-based feature selection (MFS), importance-based feature selection (IFS) measured by models, and the intersection of all selected (IAS). In addition, the prediction accuracies of various machine learning algorithms and neural networks were compared. As a result, using the light gradient boosting machine (LGBM) model with the IFS method, R2 score was obtained as 0.89 ( +/- 0.02) with only 710 features for G, and 0.91 ( +/- 0.03) with 588 features for K. The latter obviously showed better performance with fewer features. Therefore, it is expected that these main features, such as XRDPowderPattern, ElementProperty, and AGNIFingerPrint, can be efficiently utilized for the prediction of G and K, in the research on the development of new inorganic materials.