基于机器学习与元启发式特征选择的钙钛矿材料多性能预测

Abstract Extending the availability of stable perovskite materials with appealing optoelectronic characteristics is essential to surpassing the current efficiency limitations in photovoltaic absorbers. Herein, we propose a multi-properties machine learning (ML) prediction strategy to accelerate the discovery of ABX 3 and A 2 BB’X 6 perovskites . This approach evaluates key properties essential for high-performance photovoltaic materials, including formation energy ( Δ E f ), thermodynamic stability, band gap ( E g ), and the nature of the band gap. This study evaluated various feature selection methods, such as Least Absolute Shrinkage and Selection Operator, k-Best, and meta -heuristic algorithms (MHA) like Genetic Algorithm, Particle Swarm Optimization, Atom Search Optimization, Electromagnetic Field Optimization, and Multi-Verse Optimizer, to enhance model performance. The models developed in this study achieved cross-validation and testing R 2 scores consistently exceeding 0.80 for Δ E f and E g predictions, while thermodynamic stability and band gap classification models attained accuracies above 0.85. Despite relying solely on compositional features, our models demonstrate improved performance over previous studies using the same dataset. The MHA feature selection method improves the performance of ML models in predicting Δ E f , E g , and classifying the nature of the band gap. Notably, our thermodynamic stability classification model performs effectively using only compositional features.