Feature Blending: An Approach toward Generalized Machine Learning Models for Property Prediction

From studying the atomic structure and chemical behavior to thediscovery of new materials and investigating properties of existingmaterials, machine learning (ML) has been employed in realms thatare arduous to probe experimentally. While numerous highly accuratemodels, specifically for property prediction, have been reported inthe literature, there has been a lack of a generalized framework.Herein we propose a novel feature selection approach that enablesthe development of a unified ML model for property prediction forseveral classes of materials. It involves an ingenious blending ofselected features from various classes of data such that the resultantfeature set equips the model with global data descriptors capturingboth class-specific as well as global traits. We took accurate bandgaps of three distinct classes of 2D materials as our target propertyto develop the proposed feature blending approach. Using Gaussianprocess regression (GPR) with the blended features, the ML model developedhere resulted in an average root-mean-squared error of 0.12 eV forunseen data belonging to any of the participating classes. The featureblending approach proposed here can be extended to additional classesof materials and also to predict other properties.