LEAP framework boosts perovskite solar cells to 21.32% efficiency using LLMs

A team led by Xin-De Wang introduced LEAP (LLM-driven Exploration via Active Learning for Perovskites), a closed-loop framework that marries a domain-specialized large language model with Bayesian optimization for rapid additive discovery in perovskite photovoltaics. The LLM extracts mechanistic knowledge from literature and represents molecules via interpretable descriptors, feeding into an uncertainty-aware prioritization workflow that works efficiently under low-data conditions, dramatically cutting the time needed compared to conventional trial-and-error screening.

In a proof-of-concept expert-in-the-loop study, LEAP prioritized additives across three rounds, yielding average device power conversion efficiencies (PCE) of 20.13% and 20.87% for 6-CDQ- and 2-CNA-treated cells, respectively, versus 19.25% for controls. The champion device reached 21.32%. The domain-specialized LLM outperformed general-purpose models on mechanism-consistent reasoning, providing preliminary evidence that literature-grounded descriptors coupled with Bayesian optimization can deliver mechanism-aware additive prioritization for perovskite solar cells.