A Critical Examination Of Robustness And Generalizability Of Machine Learning Prediction Of Materials Properties

This article discusses the use of graph neural network (GNN) models for predicting formation energy, a fundamental property that dictates the phase stability of a material. The Atomistic LIne Graph Neural Network (ALIGNN) model is used as a representative GNN model and is pretrained on the Materials Project 2018.06.01 version (MP18). The MP18-pretrained ALIGNN model (ALIGNN-MP18) is then used to predict the formation energies of the new structures in the latest (2021.11.10 version) Materials Project database (MP21). The article then focuses on the alloys of interest (AoI) which are defined as the space formed by the first 34 metallic elements (from Li to Ba) and the alloys formed exclusively by these elements.