AI-guided choice boosts perovskite solar-cell performance in authors’ lab tests, preprint shows

A new arXiv preprint says an evidence-grounded large language model did more than perform well on a scientific benchmark: It picked a perovskite solar-cell contact chemistry that beat a human expert-preferred alternative in the authors’ own lab tests on both efficiency and stability.

The paper, “Evidence-grounded large language models guide perovskite solar cell design and optimization,” was posted on arXiv as preprint 2611.00633, submitted Oct. 31 and revised Nov. 3. It is not peer-reviewed. The authors — YanKai Chen, Zhi Wan, ZiHao Li and Tao Jing, with affiliations including Tsinghua University and Shaanxi University of Science & Technology — describe a language-model system called OpenPerov that was adapted using structured evidence reconstructed from the published perovskite literature and then tested on a new domain benchmark.

Perovskite solar cells are a fast-moving area of solar research because they can reach very high lab efficiencies, but long-term durability remains a major barrier. Against that backdrop, the paper combines two claims: improved question-answering and reasoning on perovskite science, and a real materials-selection result that held up in device experiments.

To evaluate scientific reasoning, the researchers built PSM-Bench, an 800-question benchmark derived from 446 studies in perovskite photovoltaics. They also assembled 4,532 curated article evidence records, with 4,035 used for training or adaptation and 497 held out. On PSM-Bench, the authors say OpenPerov Flash ranked first among 13 closed-book systems with a Scientific Mastery score of 87.6, ahead of the strongest cited comparator, GLM-5.2, at 83.3. A retrieval- and evidence-conditioned version, OpenPerov Pro, raised that score to 89.1.

The more notable result came in an experimental design task centered on a self-assembled monolayer, or SAM — a single-molecule-thick contact layer — at the interface between indium tin oxide, or ITO, and the perovskite absorber. In that head-to-head choice, the model prioritized BCPCz-TMPS-K, a potassium-counterion version of the contact material. Human experts in the study preferred the sodium analogue, BCPCz-TMPS-Na.

In the authors’ lab tests, the model-selected potassium version produced the better devices. The best small-area cell using BCPCz-TMPS-K reached 27.05% power conversion efficiency on a reverse scan through a 0.09 square centimeter aperture. The same device showed 26.8% stabilized power-output efficiency over about 600 seconds. Across 16 devices per SAM, the paper reports an average efficiency of 26.69% plus or minus 0.27% for BCPCz-TMPS-K, versus 25.24% plus or minus 0.30% for BCPCz-TMPS-Na.

The abstract summarizes the result this way: “Devices incorporating the selected contact reached a power conversion efficiency of 27.05% and retained 96.74% of their initial efficiency after 1,000 h of dark thermal ageing at 85 °C.”

Those numbers are notable, but the caveats matter. The top result came from the authors’ own lab, not an independent certification body, and the paper does not report third-party certification of the device. The headline efficiency was measured on a small-area device and reported on a reverse scan, while the stabilized efficiency — generally the more conservative figure — was slightly lower at 26.8%. And while roughly 27% is near the leading edge of current lab-scale single-junction perovskite results, it is not clearly beyond the best research-cell performance already shown on NREL charts in 2025 and 2026.

The stability data are also central to the paper’s claim. The model-selected potassium contact retained 96.74% of its initial efficiency after 1,000 hours of dark thermal aging at 85 degrees Celsius, the authors report. It also retained 94.84% under continuous one-sun maximum-power-point tracking.

The paper offers some transparency, but not full reproducibility from the outside. In its code availability section, the authors write: “Code for model adaptation, inference, evidence retrieval, local revision and evaluation is available at https://github.com/Yan-Kai-Chen/OpenPerov.” But the underlying literature-derived training and retrieval corpora, source-support passages and evidence indexes are not publicly distributed. That means the benchmark and lab result can be examined as a preprint, but key parts of the evidence pipeline are not yet open for full outside inspection.