Evidence-grounded large language models guide perovskite solar cell design and optimization

YanKai Chen, Zhi Wan, ZiHao Li, Tao Jing

TL;DR

Researchers have developed OpenPerov, a large language model (LLM) framework that masters scientific reasoning in the complex field of perovskite photovoltaics. By training on "structured-evidence" records rather than raw text, the model successfully designed a novel solar cell interface that achieved a 27.05% efficiency and outperformed human-expert recommendations in both performance and long-term stability.

What Problem Does This Paper Address?

In materials science, the sheer volume of literature makes it nearly impossible for researchers to keep up. While standard LLMs are great at summarizing papers, they often fail at scientific reasoning—the ability to link a specific experimental observation (like a voltage drop) to a physical mechanism (like ion migration) and then propose a discriminating experiment to prove it.

Perovskite solar cells are a "high-stakes" test case for this. They are incredibly efficient but notoriously unstable. Designing them requires balancing "coupled constraints": a chemical additive that improves efficiency might inadvertently accelerate degradation. The authors tackle the gap between simple information retrieval and the deep, causal reasoning required for genuine materials discovery.

Methods

The authors introduced two versions of their model based on the Qwen architecture:

  • OpenPerov Flash: This model underwent Domain-Adaptive Pretraining (DAPT) using a unique "structured-evidence" approach. Instead of feeding the model raw PDFs, experts reconstructed 78,299 records from 4,532 articles. Each record explicitly links a research question to its experimental strategy, controls, observations, and mechanistic interpretations.
  • OpenPerov Pro: This version adds a sophisticated retrieval-augmented generation (RAG) layer. Unlike standard RAG, which often regenerates an entire answer (and can introduce new errors), OpenPerov Pro uses local revision. It identifies specific "unresolved scientific requirements" in the initial answer and patches only those sections with retrieved evidence.

Datasets

  • PSM-Bench: A new, expert-developed benchmark consisting of 800 open-ended questions drawn from 446 studies. It evaluates models on four task families: mechanism diagnosis, stability-constrained design, stability-failure analysis, and cross-constraint synthesis.
  • OpenPerov Corpus: A curated collection of 4,532 articles (published 2013–2026) used for training and retrieval.
  • Perovskite-R1: An external, independent benchmark used to validate the model's general domain knowledge.

Key Insights and Ideas

  • Structure Trumps Scale: The "structured-evidence" training route used only 19.33 million tokens but significantly outperformed a conventional training route using 63.39 million tokens. This suggests that how scientific data is presented to a model is more important than the raw volume of text.
  • The "Local Revision" Advantage: Full-answer regeneration (standard RAG) often led to "frequent large declines" in answer quality. By focusing on local patches, OpenPerov Pro achieved a much better "gain-harm balance," preserving the logic of the original answer while adding precise evidence.
  • AI vs. Expert Intuition: In a head-to-head design task for self-assembled monolayers (SAMs), human experts favored a Sodium (Na) analogue based on traditional intuition. OpenPerov chose a Potassium (K) analogue, arguing that its larger ionic radius would better tune the competition between alkali-metal association and lead (Pb) coordination. Experimental testing proved the AI was right.

Key Results

  • Leaderboard Dominance: OpenPerov Flash ranked first among 13 closed-book systems (including GLM-5.2 and DeepSeek variants) on the PSM-Bench.
  • Outperforming Humans: On questions derived from studies published after the model's training cutoff, both Flash and Pro outscored expert-authored answers.
  • World-Class Efficiency: The AI-designed BCPCz–TMPS–K contact enabled solar cells to reach a power conversion efficiency (PCE) of 27.05%.
  • Exceptional Stability: The AI-designed devices retained 96.74% of their efficiency after 1,000 hours of dark thermal ageing at 85 °C, significantly outperforming the industry-standard Me-4PACz contact.

Limitations and Future Work

While the results are impressive, the authors note that the model's performance is heavily dependent on the quality of the expert-calibrated evidence records. Reconstructing these records is labor-intensive.

Future work may focus on:

  1. Automating Evidence Reconstruction: Using LLMs to build these structured records with minimal human supervision.
  2. Cross-Domain Transfer: Applying the "structured-evidence" framework to other complex fields like battery chemistry or catalyst design.
  3. Closed-Loop Discovery: Integrating OpenPerov directly with automated fabrication labs to enable fully autonomous materials discovery.

Summary updated

Abstract

Using large language models to interpret experimental evidence, design discriminating experiments and formulate well-supported mechanistic explanations remains a central challenge in materials research. Here we establish PSM-Bench, an 800-question benchmark developed by perovskite researchers to evaluate scientific reasoning in perovskite photovoltaics. We also introduce OpenPerov, an expert-informed language-model framework that learns from expert-calibrated evidence records reconstructed from 4,532 articles; these records link experimental strategies, controls and observations to mechanistic interpretations. Structured-evidence adaptation outperformed conventional question-answer training, and OpenPerov Flash ranked first among 13 closed-book systems. OpenPerov Pro raised Scientific Mastery further, from 87.6 to 89.1, by revising selected passages with retrieved evidence. On questions derived from subsequent studies, both models also outscored expert-authored answers. In an experimental test of self-assembled monolayer design, OpenPerov selected a K analogue that outperformed the Na analogue preferred by experts. Devices incorporating the selected contact reached a power conversion efficiency of 27.05% and retained 96.74% of their steady-state reference efficiency after 1,000 h of dark thermal ageing at 85 $^\circ$C. These results show that reconstructed scientific reasoning provides an effective basis for building domain models that interpret experiments and inform materials design.
BibTeX
@misc{chen2026evidence,
  author = {YanKai Chen and Zhi Wan and ZiHao Li and Tao Jing},
  title = {{Evidence-grounded large language models guide perovskite solar cell design and optimization}},
  year = {2026},
  eprint = {2611.00633},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.00633},
}
Figure