Interpretable machine learning on a curated dataset identifies chemical descriptors governing 2D perovskite solar cell performance

Two-dimensional (2D) hybrid perovskites are promising candidates for stable and efficient solar energy conversion, yet their complex chemistry and device architecture pose challenges for rational design. In this study, we compile a curated dataset—from database and literature sources—of 792 2D perovskite solar cell devices, including detailed chemical descriptors for large organic spacer cations, and investigate how these descriptors relate to photovoltaic performance. Using an interpretable machine learning model, we evaluate the relative importance of chemical and device parameters in predicting key performance metrics. Our analysis reveals that chemical descriptors, particularly the large organic-to-metal ratio and small organic cation type, have a greater influence on device efficiency than stack architecture or deposition method. The large organic-to-metal ratio emerges as the most influential feature in our model, accounting for 19.5 % of the total feature importance. The results highlight the importance of chemically informed data representation and demonstrate how structured datasets can uncover key structure–performance relationships in complex hybrid materials.

Duke Scholars

Author David Mitzi Thomas Lord Department of Mechanical Engineering and Materia ...