LEAP Framework Speeds Perovskite Solar Cell Additive Discovery
A closed-loop language model and active learning system identified precursor additives that raised perovskite solar cell efficiency.
Updated on Nov. 4, 2026 in Chemistry
Researchers introduced LEAP, a closed-loop framework combining a domain-specialized language model with active learning, for discovering perovskite precursor additives. Two later-round additives increased average device power conversion efficiencies to 20.13% and 20.87%.
Why it matters
Conventional trial-and-error screening is inefficient because the chemical space is large, so a method that prioritizes candidates under low-data conditions could shorten the search for better perovskite solar cell materials. The approach also converts literature knowledge into interpretable molecular descriptors.
The two later-round additives lifted average power conversion efficiency from a 19.25% control to 20.13% and 20.87%, with the best device reaching 21.32%. Only a limited number of candidates were tested prospectively, so broader performance is not yet established.
The details
LEAP extracts mechanism-relevant knowledge from the literature using a domain-specialized language model and converts it into molecular descriptors used by uncertainty-aware Bayesian optimization. On an evaluation drawn from unseen literature, the specialized model gave more mechanism-consistent answers than the tested general-purpose models.
Timeline
• Early publication as accepted peer-reviewed research: November 4, 2026
The Big Picture
LEAP extends Bayesian optimization, an active-learning method long used to search large materials spaces, by adding a literature-trained language model to guide which candidates are tested first. The result is a closed-loop approach aimed at reducing trial-and-error screening of perovskite additives.
If the approach holds up beyond the limited set of prospectively tested additives, it could help researchers reach higher-efficiency perovskite solar cells with fewer experiments. Practical commercialization timelines have not been established in the study.
The takeaway
Literature-trained language models paired with active learning may narrow the search for new materials when experimental data is scarce. Results from a small set of prospective tests should be read as early evidence rather than a settled benchmark.
Further reading
Explore more on Chemistry to see how molecular discovery methods like this one are developed.
Source note: This article includes information reported by Nature.
