Computational Evaluation of Perovskite Materials through Machine Learning-Driven Prediction and Screening of Photovoltaic Cell Candidates

Abstract

While current solar technology mainly uses silicon, perovskites — crystalline compounds with a chemical formula of ABX3 — are emerging technologies due to their high energy conversion efficiency. However, experimentally discovering new perovskite compositions is both costly and time-intensive. This study introduces Random Forest models to predict photovoltaic-relevant properties of perovskites and identify promising candidates for solar cell applications. A dataset from the Open Quantum Materials Database was reconstructed due to inconsistencies and used to predict the band gap, stability energy, and formation energy of the compounds. Further, ablation studies and Shapley-value analyses indicated greater dependence on crystal-structure inputs. Finally, a Photovoltaic Suitability Score ranked candidate materials based on these three properties relevant to solar-cell performance, revealing that rare-earth vanadate compounds carry strong prospects for future investigations. Based on these findings, future efforts in engineering photovoltaic materials can consider crystallographic features alongside elemental properties. Broadly, this analysis of band gap, stability energy, and formation energy serves as a preliminary step to identify candidates for photovoltaic cell engineering and inverse design pipelines.