LEAP
Expert-in-the-loop closed-loop framework that couples a domain-specialized LLM with Bayesian optimization for iterative perovskite precursor additive discovery.
| Affiliation | Renmin University of China (Ze-Feng Gao, Peng-Jie Guo, Cheng Mu, Zhong-Yi Lu) |
| First introduced | 2026-05 (arXiv:2605.20242) |
| Lifecycle stages | Multi-stage (literature mining + descriptor generation + experiment selection + analysis with expert feasibility review) |
| Autonomy level | Semi-autonomous — model-prioritized candidates are reviewed by domain experts before validation |
| Domain focus | Materials science / chemistry — perovskite solar cell precursor additives |
| Availability | Unknown (no public code link in the preprint) |
Approach
LEAP (LLM-driven Exploration via Active Learning for Perovskites) builds a domain-specialized LLM, Perovskite-RL, by supervised fine-tuning and reinforcement learning on the perovskite-additive literature. Perovskite-RL evaluates candidate molecules and emits soft mechanistic descriptors along five interpretable axes — binding mode, interfacial shielding, hydrophobic protection, ion-interaction potential, and electronic modulation — derived through repeated probabilistic reasoning and aggregation. These soft descriptors are concatenated with deterministic hard molecular features to form a hybrid representation that feeds a Gaussian-process surrogate model with expected-improvement acquisition.
Each iteration runs four steps: Perovskite-RL re-generates soft descriptors under an updated reasoning context that includes the latest experimental observations; the GP surrogate is retrained on accumulated ∆PCE data; EI scores are recomputed; and prioritized candidates pass through expert review for chemical plausibility, precursor-solution compatibility, safety, cost, and device-fabrication feasibility before wet-lab validation. Device-level PCE, film quality, and defect-suppression readouts flow back into the dataset for the next round.
Validation
On a 32-question mechanism-consistency benchmark drawn from 16 unseen additive papers, Perovskite-RL scored 78.1% (25/32), outperforming general-purpose baselines (gemini-3-flash 50.0%, DeepSeek-V3.2 46.9%, GPT-5 43.8%, Qwen3-32B 37.5%, llama-3.3-70b-instruct 28.1%) with Holm-Bonferroni-adjusted McNemar p-values from <0.001 to <0.05. A retrospective representation ablation on 36 hot-start additives showed the hybrid LEAP descriptor space lifted Spearman correlation from 0.274 (hard-only) to 0.394 and top-20% overlap from 0.125 to 0.375.
Three rounds of expert-in-the-loop wet-lab validation in inverted ITO/NiOx/4PADCB/perovskite/C60/BCP/Ag devices tested one LEAP-prioritized additive per round (Boc-DCPy, 6-CDQ, 2-CNA), with 24 devices fabricated per condition.
Notable results
- Mean device PCE 20.13 ± 0.25% (6-CDQ) and 20.87 ± 0.25% (2-CNA) vs 19.25 ± 0.28% for control; champion PCE 21.32% (2-CNA, VOC 1.128 V, JSC 23.92 mA/cm², FF 0.790).
- First-round candidate Boc-DCPy underperformed (16.76 ± 0.58%); the workflow recovered in subsequent rounds after feeding the negative result back into the surrogate and prompting context.
- Mechanism-consistency accuracy 78.1% vs ≤50% for general-purpose LLM baselines on unseen perovskite-additive papers.
Primary paper
Other references
None yet.
Code
Not released as of the preprint.