Discovery of direct band gap perovskites for light harvesting by using machine learning

• Published in: Computational materials science 2022-07, Vol.210, p.111476, Article 111476
• Main Authors: Rath, Smarak, Sudha Priyanga, G., Nagappan, N., Thomas, Tiju
• Format: Article
• Language: English
• Subjects: Machine learning; Matminer; Nature of band gap; Perovskite; XGBOOST
• ISSN: 0927-0256; 1879-0801
• DOI: 10.1016/j.commatsci.2022.111476

[Display omitted] •XGBOOST (eXtreme Gradient BOOST) classifier predicts direct band gap ABX3 perovskites with 81% precision.•Model can be used for the prediction of the nature of band gap of non-oxide perovskites.•SHAP (SHapley Additive exPlanations) based elucidation of factors influencing nature of band gap is done.•Absence of transition metals and IIIA-VIIIA elements (Z > 20) increases the likelihood of direct band gap. An approach that would allow quick determination of compositions that are most likely to be direct band gap materials would significantly accelerate research on light-harvesting materials. Inorganic perovskites are attractive for this purpose since they afford compositional flexibility, while also offering stability. Here, ABX3 inorganic perovskites (A and B are cations and X is an anion) are classified into direct band gap and indirect band gap materials by using the XGBOOST (eXtreme Gradient BOOST) classifier. We use a dataset containing 1528 ABX3 compounds (X = O, F, Cl, Br, I, S, Se, Te, N, or P) along with information on the nature of their band gap (direct or indirect). All the data is taken from the Materials Project database. Descriptors for these materials are generated using the Matminer python package. Ten-fold cross-validation with the XGBOOST classifier is used on the dataset and the average accuracy is found to be 72.8%. To generate a confusion matrix, the dataset is once again split into a training set and a testing set after cross-validation. Subsequently, the confusion matrix is generated for that particular test set. It is found that the precision for the prediction of direct band gap materials is 81% i.e., 81% of the materials predicted to be direct band gap materials are actually direct band gap materials. Thus, machine learning can be an effective tool for discovering novel direct band gap perovskites. Finally, SHAP (SHapley Additive exPlanations) analysis is performed to determine the most important descriptors. One key insight gained from the SHAP analysis is that the absence of transition metals and elements belonging to groups IIIA to VIIIA with atomic number greater than 20 increases the probability of the perovskite having a direct band gap.