Exploring Data Augmentation and Dimension Reduction Opportunities for Predicting the Bandgap of Inorganic Perovskite through Anion Site Optimization

Significant focus has been directed towards inorganic perovskite solar cells because of their notable capabilities in converting sunlight to electricity effectively, their efficient light absorption, and their suitability for conventional semiconductor manufacturing methods. The identification of the composition of perovskite materials is an ongoing challenge to achieve high performing solar cells. Conventional methods of trial and error frequently prove insufficient, especially when confronted with a multitude of potential candidates. In response to this challenge, the suggestion is to employ a machine-learning strategy for more precise and efficient prediction of the characteristics of new inorganic perovskite materials. This work utilized a dataset sourced from the Materials Project database, consisting of 1528 ABX3 materials with varying halide elements (X = F, Cl, Br, Se) and information regarding their bandgap characteristics, including whether they are direct or indirect. By leveraging data augmentation and machine learning (ML) techniques along with a collection of established bandgap values and structural attributes, our proposed model can accurately and rapidly predict the bandgap of novel materials, while also identifying the key elements that contribute to this property. This information can be used to guide the discovery of new organic perovskite materials with desirable properties. Six different machine learning algorithms, including Logistic Regression (LR), Multi-layer Perceptron (MLP), Decision Tree (DT), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Random Forest (RF), were used to predict the direct bandgap of potential perovskite materials for this study. RF yielded the best experimental outcomes according to the following metrics: F1-score, Recall, and Precision, attaining scores of 86%, 85%, and 86%, respectively. This result demonstrates that ML has great potential in accelerating organic perovskites material discovery.