Machine Learning-Accelerated Prediction of Amorphization Enthalpy in Ionic Compounds

Abstract

Amorphous ionic materials hold promise for advanced energy storage, electrocatalysis, and optical devices, yet systematically evaluating their propensity to form amorphous phases remains underdeveloped. Here, we present a machine learning framework for efficiently predicting amorphization enthalpy, a universal measure of the thermodynamic cost of converting crystalline ionic compounds to amorphous phases. By combining a standardized DFT-based melt-and-quench protocol with machine learning, we build a training set of 407 compounds and identify key descriptors linked to the amorphization enthalpy. A random forest regressor highlights relevant features but shows a limited predictive power. To overcome this, we implement e3nn, one of the state-of-art graph neural networks (GNN), and adopt a transfer-learning approach. Applying this GNN to screen 12,123 ionic compounds reveals that nitrides and sulfides typically resist amorphization, while alkali- or halogen-rich and multi-cation compositions favor it. Our data-driven approach offers a practical guide for discovering novel amorphous ionic materials, accelerating experimental development across diverse applications.