LEAP: AI-driven framework for perovskite additive discovery
A team of researchers has introduced LEAP (LLM-driven Exploration via Active Learning for Perovskites), a closed-loop system that merges a specialized large language model with active learning techniques to hasten the identification of precursor additives for perovskite solar cells. This LLM draws mechanism-related insights from existing literature and depicts molecules through interpretable descriptors, which are incorporated into Bayesian optimization for prioritizing under low-data scenarios with awareness of uncertainty. Benchmark evaluations indicate that the specialized model surpasses general-purpose models in reasoning consistent with mechanisms. Currently, experimental validation is in progress as part of a proof-of-concept involving experts.
Key facts
- LEAP stands for LLM-driven Exploration via Active Learning for Perovskites.
- The framework couples a domain-specialized LLM with active learning.
- It is designed for iterative additive prioritization in perovskite solar cells.
- The LLM extracts mechanism-relevant knowledge from perovskite additive literature.
- Candidate molecules are represented through interpretable descriptors.
- Bayesian optimization enables uncertainty-aware prioritization under low-data conditions.
- Benchmark results show the domain-specialized model outperforms general-purpose models.
- Experimental validation is conducted in an expert-in-the-loop proof-of-concept.
Entities
—