Predicting efficient and stable inorganic photovoltaic materials using interpretable machine learning combined with DFT calculations based on band edge orbital engineering

Abstract

The discovery of new materials with high power conversion efficiency (PCE) is critical for the advancement of solar energy technologies. In this study, we combine traditional machine learning (ML) methods and deep learning methods to predict and categorize the inorganic photovoltaic materials, which are more stable than the organic counterpart. By employing gradient boosting trees, extremely randomized trees, random forests, backpropagation neural network, and convolutional neural network, we classify materials by bandgap (zero or non-zero) and predict their values with regression model. The stability, bandgaps, optical absorption, and PCE of these materials are validated through density functional theory (DFT) calculations, leading to the identification of promising candidates: Li2Bi4Se7, Na2Bi4Se7, and Mo2Ba5N7. The Shapley Additive Explanation method is applied to analyze feature interactions, intuitively establishing the relationship between band-edge orbitals and material properties, while uncovering hidden connections between structural and electronic properties. The results reveal that the delocalization of valence electrons, along with variations in atomic coordination environments, modifies the charge density distribution at the band edges, affecting the transition probabilities and atomic orbital connectivity. Among these, Li2Bi4Se7 and Na2Bi4Se7 stand out as the most promising materials for solar cell applications. Our findings provide a novel framework for accelerating the discovery of efficient photovoltaic materials using ML and DFT methods.