Machine Learning-Driven Band Gap Prediction/Classification and Feature Importance Analysis of Inorganic Perovskites

Abstract

Perovskites are a class of materials, known for their diverse structural, electronic, and optical properties. Band gap in perovskites is crucial in determining their suitability for applications such as solar cells, light-emitting diodes, and photodetectors. By tuning the band gap through composition and structural modifications, perovskites can be optimized for specific optoelectronic and energy-related applications, making them a versatile material in modern technology. Machine learning (ML) provides an efficient approach to predicting material band gaps by analyzing atomic and structural features, facilitating the discovery of materials with tailored electronic properties. This study employs adaptive boosting regression (ABR), random forest regression (RFR), and gradient boosting regression (GBR) for band gap prediction, alongside support vector machine (SVM), random forest classifier (RFC), and multilayer perceptron (MLP) for classifying compounds with zero and nonzero band gaps. Regression models are assessed using mean absolute error (MAE), mean squared error (MSE), and R², while classification performance is evaluated based on accuracy, precision, recall, and F1-score. ABR excels in predicting band gaps of inorganic perovskites, while RFC is the most effective model for classification. Feature analysis identifies the standard deviation of valence charges as the key predictor. This study underscores ML’s potential to accelerate perovskite discovery through accurate band gap predictions.