Highly Efficient Screening of Halide Double Perovskite Optoelectronic Materials Based on Machine learning

Area of Science:

• Materials Science
• Computational Chemistry
• Solid-State Physics

Background:

• Photoelectronic properties of halide perovskites are key to their applications but are difficult to obtain experimentally.
• Traditional methods like density functional theory (DFT) are time-consuming and labor-intensive.

Purpose of the Study:

• To develop a rapid and efficient method for predicting halide perovskite properties.
• To screen a large database of halide double perovskites for stable candidates with desired band gaps.

Main Methods:

• Utilized gradient boosting tree models combined with genetic algorithms (GA) and grid search (GRID) for property prediction.
• Employed multistep prediction and SHapley Additive exPlanations (SHAP) for model validation and feature analysis.
• Screened 77,604 halide double perovskites to identify 1515 stable candidates with band gaps between 1-4 eV.

Main Results:

• Achieved high prediction accuracy with R-squared values of 0.9958 for formation energy and 0.9206 for band gap.
• Successfully validated the optimized models through DFT calculations on selected candidates.
• Identified a promising perovskite, Cs2RbBiI6, which demonstrated successful application in photodetection and photocatalysis.

Conclusions:

• The developed machine learning approach significantly accelerates the discovery of functional halide perovskites.
• This method reduces the time and experimental resources required for material screening.
• The study provides a framework for discovering novel materials tailored for specific optoelectronic applications.