Optimizing novel device configurations for perovskite solar cells: Enhancing stability and efficiency through machine learning on a large dataset

Abstract

Perovskite solar cells (PSCs) have emerged as promising, cost-effective, and efficient alternatives to silicon-based solar cells, yet achieving both high stability and efficiency remains challenging. To address these challenges, we developed Random Forest and Extreme Gradient Boosting models to optimize the stability and power conversion efficiency (PCE) of PSCs, using a large dataset from the Perovskite Database. Our models demonstrated strong predictive performance, achieving an accuracy of 0.848 in stability classification and an R² of 0.751 for PCE prediction on the test set. Stability prediction used a classification approach, labeling devices as stable if they retained at least 80 % of their initial PCE after 1,000 h, a threshold that allows the inclusion of both T₈₀ and E_1000h data. Using the trained models, we do high-throughput screening of 29,016 new device configurations with varied cell architectures, electron transport layers, hole transport layers, and perovskite ion compositions. Among these, we identified 100 top-performing, predicted stable lead-based PSCs configurations with potential PCEs reaching up to 26.06 %, surpassing the highest stable device in the Perovskite Database, which has a PCE of 22.3 %. This study demonstrates that machine learning-driven approaches can effectively guide PSCs optimization, surpassing the performance of previously reported configurations.