Predictive machine learning approaches for perovskites properties using their chemical formula: towards the discovery of stable solar cells materials

• Published in: Neural computing & applications 2024-09, Vol.36 (26), p.16319-16329
• Main Authors: Touati, Soundous, Benghia, Ali, Hebboul, Zoulikha, Lefkaier, Ibn Khaldoun, Kanoun, Mohammed Benali, Goumri-Said, Souraya
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Classification; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Crystal structure; Data Mining and Knowledge Discovery; Datasets; Density functional theory; Energy conversion efficiency; Energy gap; Energy of formation; Error analysis; Free energy; Halides; Heat of formation; Image Processing and Computer Vision; Machine learning; Materials science; Original Article; Perovskites; Photovoltaic cells; Prediction models; Probability and Statistics in Computer Science; Root-mean-square errors; Solar cells
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-024-09992-5

In recent years, notable progress in computational density functional theory (DFT) has facilitated the collection of extensive datasets in the field of materials science. Machine learning is a crucial technique for effectively processing and analyzing these large datasets and accelerating the creation of new compounds. In this work, we employ the Extreme Gradient Boosting (XGBoost) classification algorithm to predict the crystal structure of 381 halides and oxide perovskites using 78 features. We achieved a classification accuracy of 76.62% for these materials. Subsequently, we utilized the Random Forest and XGBoost algorithms to investigate a dataset comprising 761 perovskite materials from the material project database to predict the band gap energy (best accuracy = 84.78%, mean absolute error(MAE) = 0.410 eV, root mean square error (RMSE) = 0.594 eV) and formation energy (best accuracy = 94.81%, MAE = 0.083 eV/atom, RMSE = 0.157 eV/atom), all of these prediction results based on the chemical formulas. Notably, the feature importance analysis shows the significant influence of the number of d electrons in the d orbit of the B atom on both band gap (E g ) and formation energy (E F ) properties. Finally, we applied our prediction model to identify stable perovskite solar cells, achieving an accuracy of 85.18%. These findings provide valuable guidelines for the discovery of stable perovskite solar cells and help researchers tailor the composition of perovskite materials to optimize their light-harvesting capabilities, leading to higher-power conversion efficiencies.