Evidence-trained AI helps design a better contact for perovskite solar cells

Scientific papers contain more than findings: they document controls, competing explanations and the conditions under which an idea works. OpenPerov aims to learn those relationships, then use retrieved evidence to revise its answers. Its developers report gains on scientific reasoning tasks and a laboratory demonstration involving potassium-based molecular contacts.

Learning how evidence supports an explanation

A language model that recalls materials research does not necessarily know which experiment would distinguish competing explanations. In an arXiv research preprint, Chen YanKai and colleagues present OpenPerov, a framework intended to connect observations, mechanisms and experimental decisions in perovskite solar-cell research. Its reported results are preliminary research findings, not established evidence of general scientific superiority.

The team assembled 4,532 articles and reconstructed their scientific arguments into expert-calibrated records. These connect research questions with experimental strategies, controls, observations and interpretations, while preserving relevant conditions and relationships to other studies. Language models assisted with text processing; domain experts constructed and checked the scientific content.

OpenPerov Flash was adapted from Qwen3.6-27B using records from 4,035 articles, followed by scientific instruction training and answer-style tuning. The central idea is to teach relationships between evidence and explanations, rather than simply train on questions and answers extracted from papers.

Better answers on a specialist benchmark

The researchers also developed PSM-Bench: 800 open-ended questions drawn from 446 studies. Tasks cover mechanism diagnosis, stability-constrained design, failure analysis and problems requiring several constraints to be satisfied together. Its Scientific Mastery score combines answer adequacy, coverage of scientific criteria, supported claims and experimental decision quality.

Flash scored 87.6 out of 100, leading the 13 systems tested without retrieval. The strongest comparator, GLM-5.2, scored 83.3. A double-blind expert assessment of answers from four systems supported Flash’s leading position.

A training comparison used the same source articles and backbone. After identical subsequent training, the structured-evidence route scored 87.57, versus 84.07 for conventional question–answer adaptation, despite using fewer tokens during initial adaptation. However, the initial training objectives also differed, so this comparison evaluates two learning routes rather than isolating data structure alone.

Revising selectively instead of starting over

OpenPerov Pro adds literature retrieval to Flash. Rather than regenerate an entire answer, it makes evidence-linked additions or replacements in selected passages.

On PSM-Bench, Pro raised Scientific Mastery to 89.1. With the same evidence packets and answer-model checkpoint, full-answer regeneration improved the aggregate score by only 0.22 points, compared with 1.49 points for Pro. Regeneration caused declines of at least five points on 24.38% of questions; Pro did so on 0.50%. Within this experiment, selective revision better preserved useful existing reasoning.

A separate evaluation used ten studies published after the literature cutoff, supplying 40 questions in each of two information settings. Both models scored above expert-authored answers. The experts could consult cutoff-limited online resources and use AI for wording and organization. These small, rubric-based comparisons do not establish that the models outperform researchers generally.

A molecular choice tested in the laboratory

The experimental demonstration concerned a self-assembled monolayer: a molecular contact at the interface between indium tin oxide and the perovskite. The design combined an oxide-facing silane anchoring group with carboxylate groups intended to interact with undercoordinated lead on the perovskite side.

Researchers supplied the design constraints and chemical evidence. OpenPerov selected potassium as the counterion, while experts preferred sodium. Both versions were synthesized and tested under matched fabrication conditions.

Across 16 devices per group, the potassium version averaged 26.69% power conversion efficiency, versus 25.24% for sodium—a gain of 1.45 percentage points. A small-area potassium-contact cell reached 27.05% in a reverse voltage scan; a separate fixed-bias measurement gave stabilized output of 26.8% over approximately 600 seconds.

The authors propose that anchoring, lead coordination and counterion-dependent interactions jointly improve the contact. The performance measurements support the design choice, but do not by themselves establish every proposed molecular mechanism.

Stability results need their reference points

Encapsulated potassium-contact cells retained 96.74% of their reference efficiency after a reported 1,000-hour interval of dark ageing at 85 °C. Crucially, that reference was taken 650 hours into the original record, not at the beginning. The result therefore describes retention over a later interval, not retention of initial efficiency.

Under continuous illumination with maximum-power-point tracking, retention was 94.84% over 994 hours, again measured from a later reference point. These laboratory protocols do not establish outdoor service life or large-module performance.

Other limitations affect interpretation. PSM-Bench includes development questions and items selected using model performance. Foundation-model pretraining exposure is unknown, and Pro can retrieve held-out source articles. Literature-derived training and retrieval corpora are not publicly distributed, limiting independent reconstruction.

The preprint offers a concrete example of evidence-organized training and targeted revision informing a successful laboratory choice—not proof of autonomous discovery or broad replacement of scientific expertise.