An inorganic ABX3 perovskite materials dataset for target property prediction and classification using machine learning

TL;DR

This paper introduces a new, validated dataset of inorganic ABX3 perovskite materials for machine learning, enabling accurate prediction and classification of key properties to accelerate materials discovery.

Contribution

The study provides a high-quality, preprocessed dataset from OQMD with 16,323 samples and demonstrates its effectiveness through ML models predicting properties like formation energy, band gap, and crystal system.

Findings

Best MAE for formation energy: 0.013 eV/atom

Best MAE for band gap: 0.216 eV

Crystal system classification accuracy: 85%

Abstract

The reliability with Machine Learning (ML) techniques in novel materials discovery often depend on the quality of the dataset, in addition to the relevant features used in describing the material. In this regard, the current study presents and validates a newly processed materials dataset that can be utilized for benchmark ML analysis, as it relates to the prediction and classification of deterministic target properties. Originally, the dataset was extracted from the Open Quantum Materials Database (OQMD) and contains a robust 16,323 samples of ABX3 inorganic perovskite structures. The dataset is tabular in form and is preprocessed to include sixty-one generalized input features that broadly describes the physicochemical, stability/geometrical, and Density Functional Theory (DFT) target properties associated with the elemental ionic sites in a three-dimensional ABX3 polyhedral. For validation, four different ML models are employed to predict three distinctive target properties, namely: formation energy, energy band gap, and crystal system. On experimentation, the best accuracy measurements are reported at 0.013 eV/atom MAE, 0.216 eV MAE, and 85% F1, corresponding to the formation energy prediction, band gap prediction and crystal system multi-classification, respectively. Moreover, the realized results are compared with previous literature and as such, affirms the resourcefulness of the current dataset for future benchmark materials analysis via ML techniques. The preprocessed dataset and source codes are openly available to download from github.com/chenebuah/ML_abx3_dataset.