LEAP: AI-Driven Closed-Loop Discovery Boosts Perovskite Solar Cell Efficiency to 21.32%

| Property | Details | | :--- | :--- | | Paper | LEAP: A closed-loop framework for perovskite precursor additive discovery | | Authors | Xin-De Wang, Zhi-Rui Chen, Ze-Feng Gao, Peng-Jie Guo, Cheng Mu, Zhong-Yi Lu | | arXiv ID | 2605.20242 | | Submitted | May 18, 2026 | | Categories | cs.LG; cond-mat.mtrl-sci; cs.AI; physics.chem-ph | | Key contribution | Expert-in-the-loop AI framework combining a domain-specific LLM with Bayesian optimization; discovered novel perovskite additives raising champion power conversion efficiency to 21.32%, a ~2 point gain over the 19.25% control | | Keywords | perovskite solar cells, AI materials discovery, Bayesian optimization, expert-in-the-loop, closed-loop framework, active learning |

The Problem: Trial and Error in a Vast Chemical Space

Perovskite solar cells are attractive because they can be solution-processed at room temperature, far cheaper than silicon. But pristine perovskite films contain defects—grain boundaries, vacancies, impurities—that trap photogenerated carriers. Additive engineering addresses this: tiny amounts of organic molecules passivate defects. The catch is knowing *which* molecule to add, among tens of millions of known organic small molecules. The traditional approach—a senior chemist picks candidates by intuition, synthesizes, tests, repeats—can take decades.

How LEAP Works

LEAP closes the loop between literature and the lab:

1. AI recommendation. A domain-specific LLM trained on perovskite additive literature proposes candidates with human-readable mechanistic rationales (e.g., "this molecule has an electron-withdrawing group that may passivate positively charged iodine vacancies"). A Bayesian optimization model tracks each candidate's uncertainty—what hasn't been tested yet.

2. Expert review. Human chemists filter the AI's suggestions for synthetic feasibility and chemical soundness, removing at least half.

3. Real experiments. Selected molecules are synthesized, films spin-coated, electrodes deposited, and current–voltage curves measured on real devices—not simulations.

4. Feedback. New efficiency data update both the LLM and the Bayesian model, sharpening the next round.

A deliberate design choice: the LLM expresses molecular features in natural-language descriptors rather than opaque vectors, because the framework is expert-in-the-loop—humans must be able to judge the AI's reasoning. Benchmark tests showed the domain model significantly outperforms general-purpose LLMs (like GPT) on mechanistically consistent reasoning.

Results: 19.25% → 21.32%

| Round | Control | Best additive | Average efficiency | Champion efficiency | | :--- | :--- | :--- | :--- | :--- | | Control | 19.25% | — | 19.25% | — | | Round 2 | — | 6-CDQ | 20.13% | — | | Round 3 | — | 2-CNA | 20.87% | 21.32% |

Neither 6-CDQ nor 2-CNA had ever been reported as a perovskite additive. Only a few dozen molecules were tested in total across three rounds—highlighting Bayesian optimization's ability to prioritize "uncertain but promising" experiments where each measurement is expensive.

Honest Limitations

• Three rounds ≠ chemical space. Better additives likely remain unexplored; these are the best of a few dozen tested, not global optima.
• 21.32% is a champion device, not the average; the 20.87% mean is the more honest metric (still well above control).
• Synthesis cost is unreported—a molecule needing a 20-step route has no practical value regardless of efficiency gains.
• Expert filtering cuts both ways: it removes implausible suggestions but may also discard unconventional-but-effective ones (false negatives that can't be estimated).
• Three rounds in one lab is a proof of concept; generalization to other compositions, labs, and humidity conditions is unverified.

Why It Matters Beyond Solar Cells

The LEAP formula—AI extracts mechanistic knowledge from literature → proposes candidates → experts filter → experiments validate → data feed back—transfers directly to catalyst design, drug discovery, and alloy development. Its deeper contribution is demonstrating that expert-in-the-loop AI discovery works in the real physical world: real labs, real chemicals, real devices. AI didn't replace the chemist; it decided what the chemist should try next.

With commercialization of perovskite cells pegged around 25% efficiency, the remaining ~4-point gap from 21.32% may be shorter if each AI screening round keeps squeezing out half a point to a full point.

---

*Via zhichai.net — AI research frontline.*