Predicting the thermodynamic stability of perovskite oxides using machine learning models

TL;DR

This study develops machine learning models trained on DFT data to rapidly predict the thermodynamic stability of perovskite oxides, aiding materials discovery and reducing computational costs.

Contribution

The paper introduces accurate ML models for predicting perovskite stability, significantly speeding up stability screening compared to traditional DFT calculations.

Findings

Extra trees classifier achieved 93% accuracy.

Kernel ridge regression had an RMSE of 28.5 meV/atom.

Models successfully predicted stability of unseen compounds.

Abstract

Perovskite materials have become ubiquitous in many technologically relevant applications, ranging from catalysts in solid oxide fuel cells to light absorbing layers in solar photovoltaics. The thermodynamic phase stability is a key parameter that broadly governs whether the material is expected to be synthesizable, and whether it may degrade under certain operating conditions. Phase stability can be calculated using Density Functional Theory (DFT), but the significant computational cost makes such calculation potentially prohibitive when screening large numbers of possible compounds. In this work, we developed machine learning models to predict the thermodynamic phase stability of perovskite oxides using a dataset of more than 1900 DFT-calculated perovskite oxide energies. The phase stability was determined using convex hull analysis, with the energy above the convex hull (Ehull) providing a direct measure of the stability. We generated a set of 791 features based on elemental property data to correlate with the Ehull value of each perovskite compound. For classification, the extra trees algorithm achieved the best prediction accuracy of 0.93 (+/- 0.02), with an F1 score of 0.88 (+/- 0.03). For regression, leave-out 20% cross-validation tests with kernel ridge regression achieved the minimal root mean square error (RMSE) of 28.5 (+/- 7.5) meV/atom between cross-validation predicted Ehull values and DFT calculations, with the mean absolute error (MAE) in cross-validation energies of 16.7 (+/- 2.3) meV/atom. We further validated our model by predicting the stability of compounds not present in the training set and demonstrated our machine learning models are a fast and effective means of obtaining qualitatively useful guidance for a wide-range of perovskite oxide stability, potentially impacting materials design choices in a variety of technological applications.