Discovering and Designing Novel Perovskite Photovoltaic Materials via Machine Learning

Perovskite semiconductors are promising materials for high-efficiency photovoltaics due to their outstanding optoelectronic properties, emerging as a sustainable energy source through solar cell applications. Perovskites with the ABX₃ composition (A, B = metal or organic cations with varying oxidation states; X = chalcogen or halogen anions) have gained interest for their excellent phase stability and compositional tunability. However, combinatorial possibilities arising from the many choices of A, B, and X site species, and their respective mixing fractions, a large number of possible ABX₃ perovskites remain undiscovered. In this work, we used machine learning (ML) methods to design new stable and synthesizable ABX3 compounds for photovoltaic (PV) applications. Hypothetical charge-neutral ABX₃ compounds were generated based on the oxidation states of the constituent ions, their associated chemical properties were used as numerical descriptors to train ML regression models for predicting stability and electronic properties. The synthesizability of the generated compounds was predicted using positive-unlabeled (PU) classification learning, referencing experimentally synthesized compounds collected from the literature using a large language model (LLM). Regression models were trained on a computational dataset of perovskite properties, namely the decomposition energy and the electronic band gap. Predictions were made for thousands of novel compositions, and screening was performed to yield many new materials with stability against decomposition to alternative phases, high probability of synthesis, and PV-suitable band gaps. To ensure the practicality of the proposed materials, future studies will involve DFT calculations of optoelectronic properties and the realization of hypothetical materials through targeted synthesis and characterization.