LEAP: A closed-loop framework for perovskite precursor additive discovery

2026-05-22 · Source: cs.LG updates on arXiv.org · Field: Science & Research — Physical Sciences & Chemistry, Engineering & Applied Sciences, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

LEAP (LLM-driven Exploration via Active Learning for Perovskites) is a new expert-in-the-loop closed framework designed to accelerate the discovery of precursor additives for perovskite solar cells (PSCs). This framework integrates a domain-specialized large language model (LLM), Perovskite-RL, with active learning and Bayesian optimization. Perovskite-RL extracts mechanism-relevant knowledge from perovskite additive literature and represents candidate molecules using interpretable descriptors. These descriptors are then fed into a Bayesian optimization workflow for uncertainty-aware prioritization, especially useful in low-data conditions. Benchmarking showed Perovskite-RL achieved 78.1% accuracy on unseen literature, outperforming general-purpose models in mechanistic reasoning. Experimental validation across three screening rounds demonstrated improved additive prioritization, leading to average device PCEs of 20.13% for 6-CDQ-treated devices and 20.87% for 2-CNA-treated devices, compared to 19.25% for the control, with a champion PCE of 21.32%. This provides preliminary evidence for the effectiveness of literature-grounded mechanistic descriptors combined with Bayesian optimization and expert review.

Key takeaway

For research scientists developing new materials in data-limited fields, adopting a closed-loop framework like LEAP can significantly accelerate discovery. You should consider integrating domain-specialized LLMs for mechanistic reasoning and interpretable descriptor generation. This approach, combined with Bayesian optimization and expert review, enables efficient, uncertainty-aware candidate prioritization. It allows you to translate literature knowledge into experimentally actionable hypotheses, improving validation success and reducing resource intensity.

Key insights

Coupling a domain-specialized LLM with active learning and expert review accelerates perovskite additive discovery under low-data conditions.

Principles

• Domain-specialized LLMs outperform general models in mechanistic reasoning.
• Hybrid descriptors combine molecular features with mechanism-level effects.
• Bayesian optimization balances exploitation and exploration for candidate prioritization.

Method

LEAP trains Perovskite-RL on literature to generate mechanism-aware descriptors, integrates them with molecular features into a Gaussian process for Bayesian optimization, and refines selection with expert review and experimental feedback.

In practice

• Use Perovskite-RL for mechanism-consistent reasoning in additive design.
• Integrate soft mechanistic descriptors with hard molecular features.
• Apply Bayesian optimization for uncertainty-aware candidate prioritization.

Topics

• Perovskite Solar Cells
• Precursor Additive Discovery
• Large Language Models
• Active Learning
• Bayesian Optimization
• Materials Informatics

Best for: AI Scientist, Research Scientist

Related on AIssential

• Finetuning-Free Diffusion Model with Adaptive Constraint Guidance for Inorganic Crystal Structure Generation — Adaptive constraint guidance in diffusion models enables finetuning-free generation of inorganic crystal structures with targeted properties.
• Battery-Sim-Agent: Leveraging LLM-Agent for Inverse Battery Parameter Estimation — LLM agents can reframe inverse scientific problems as reasoning tasks, outperforming traditional black-box optimization.
• A unified predictive and generative solution for liquid electrolyte formulation — A unified AI framework accelerates liquid electrolyte design by combining predictive and generative models.
• From Blind Guess to Informed Judgment: Teaching LLMs to Evaluate Materials by Building Knowledge-Augmented Preference Signals — Teaching LLMs to evaluate materials reliably involves pairing expert-rule-informed judgments with blind guesses as preference signals.
• De novo molecular generation with optical property preconditioning at the token level — Token-conditioned autoregressive models can generate OLED molecules with targeted optical properties, but control varies by chemical motif.
• Coupling Language Models with Physics-based Simulation for Synthesis of Inorganic Materials — Coupling LLMs with physics-based simulations offers more viable inorganic material synthesis routes than classical algorithms, leveraging LLMs' implicit priors.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.