Machine Learning Models for Efficient Property Prediction of ABX₃ Materials: A High-Throughput Approach

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• ABX3 materials are crucial for photovoltaics, catalysis, and optoelectronics.
• High experimental costs and time-consuming DFT calculations hinder rapid discovery.
• Machine learning offers a potential solution to accelerate materials exploration.

Purpose of the Study:

• To utilize the extreme gradient boosting (XGBoost) algorithm for accelerated discovery and characterization of ABX3 compounds.
• To predict key material properties including space group, volume, formation energy, and band gap energy.
• To identify new ABX3 formulas and predict their properties for enhanced performance and functionality.

Main Methods:

• Employed the XGBoost algorithm on large datasets generated by DFT calculations.
• Predicted space groups for 13947 oxides and halides using elemental features and the Open Quantum Materials Database.
• Utilized XGBoost regression to predict volume, formation energy, and band gap energy, achieving high accuracies.

Main Results:

• Achieved high classification accuracies (82.39%–99.14%) for space group prediction.
• Obtained optimal prediction accuracies of 98.41% for volume, 97.36% for formation energy, and 87.00% for band gap energy.
• Successfully identified possible space groups and predicted properties for 1252 new ABX3 formulas.

Conclusions:

• XGBoost algorithm effectively accelerates the discovery and characterization of ABX3 materials.
• The predictive models enable rapid exploration of new materials with desired properties.
• Machine learning significantly enhances the efficiency of materials science research.