Enrichment-Oriented Virtual Screening of Perovskite Solar Cells through Reinforcement Learning

Yezeyuan Zheng, Ziyang Hu, Maojun Sun. Ruijian Han, Jian Huang, Guanhua Chen, Jie LiuarXiv 2026 · arXiv

TL;DR

Researchers have developed a reinforcement learning (RL) framework to navigate the massive combinatorial "maze" of perovskite solar cell design. By using a surrogate model to guide RL agents toward "recipes" supported by high-performing experimental data, they achieved a 2.4-fold improvement in identifying promising candidates compared to traditional optimization methods.

What Problem Does This Paper Address?

Perovskite solar cells (PSCs) are the rising stars of photovoltaics, offering high efficiency and low-cost fabrication. However, designing the "perfect" cell is a nightmare of complexity. You have to choose the right absorber composition (the $ABX_3$ perovskite), select compatible electron and hole transport layers (ETL/HTL), and determine the best deposition methods.

The number of possible combinations is astronomical, making exhaustive experimental testing impossible. While "virtual screening" (using AI to predict performance before hitting the lab) helps, it suffers from a major flaw: surrogate exploitation. Optimization algorithms often find "blind spots" in predictive models—configurations where the model incorrectly predicts high efficiency—leading researchers to chase phantom results that fail in the real world.

Methods

The authors propose an enrichment-oriented virtual screening workflow. Instead of just looking for the highest predicted Power Conversion Efficiency (PCE), they look for "neighborhood support"—meaning the AI-generated recipe is similar to known, high-performing experimental devices.

Key components include:

  • Recipe Representation: Each device is encoded as a 21-dimensional vector covering materials, layer sequences, and processing flags.
  • The Surrogate (Reward Model): They used a Tabular Prior-data Fitted Network (TabPFN), a transformer-based model designed for small-data tabular tasks, to act as the "judge" for new recipes.
  • RL Agents: They trained three types of policies—Proximal Policy Optimization (PPO), Recurrent PPO (RPPO), and Trust Region Policy Optimization (TRPO)—to generate recipes that maximize the surrogate's score.
  • Neighborhood Assessment: To validate the AI's suggestions, they compared generated recipes against a reference pool of 890 measured devices using a SHAP-weighted similarity metric.
Overview of the virtual screening workflow and evaluation framework.

Datasets

The study draws from two primary sources:

  1. The Perovskite Database: A massive open-access repository of over 43,000 device records (mostly pre-2020).
  2. LLM-Extracted Literature: To include cutting-edge results, the team used a Large Language Model (DeepSeek-v4-flash) to extract 667 recent recipes from open-access papers published after 2020.
  3. High-PCE Subset: A cleaned dataset of 1,846 records with PCE $\geq$ 18% was used for training and cross-validation.

Key Insights and Ideas

  • Separating Optimization from Evidence: The core "aha" moment is treating the surrogate model as a guide for generation but using experimental neighborhoods as the validation. This prevents the RL agent from simply exploiting the model's errors.
  • Foundation Models for Tabular Data: By using TabPFN, the researchers leveraged "in-context learning" for tabular data, which proved robust for ranking devices even with moderate absolute accuracy.
  • The "Hit Rate" Shift: The paper moves the goalposts from "predicting the exact PCE" to "increasing the fraction of candidates in high-performing regions." This "enrichment" is a more practical metric for researchers who need a shortlist of recipes to actually build.

Key Results

The RL approach significantly outperformed standard optimization baselines like Bayesian Optimization (Ax BO) and Differential Evolution (L-SHADE).

  • PCE Boost: In a five-fold cross-validation, selecting the top-ten recipes identified by the model raised the mean measured PCE from 19.68% to 23.74%.
  • Superior Hit Rates: For a shortlist of 100 candidates, PPO and RPPO achieved "neighborhood-supported hit rates" of 23% and 24%, respectively. This is more than double the 10% achieved by the best non-RL baselines and an 8-fold improvement over random sampling (2.9%).
  • Compositional Agreement: 65.4% of the 9,000 RL-generated recipes matched known composition families found in recent literature, suggesting the AI learned physically plausible design rules.
Comparison of neighborhood-supported hit rates across different optimization methods.

Limitations and Future Work

Despite the success, the authors are careful to note several hurdles:

  • Physical Reality Checks: Some RL-generated recipes were assigned surrogate scores exceeding 33%, which is the theoretical Shockley-Queisser limit for single-junction cells. This confirms that even advanced models can be exploited.
  • Missing Synthesis Details: The current model doesn't account for solvents, annealing temperatures, or layer thicknesses—factors that are critical in the lab but inconsistently reported in literature.
  • The Need for "Wet Lab" Validation: All assessments in this study were computational. The next logical step is a "closed-loop" system where AI-generated recipes are fabricated and the results are fed back into the model to refine its predictions.

Summary updated

Abstract

Experimental optimization of perovskite solar cells is constrained by combinatorial diversity in absorber compositions, device layers, and processing, which makes virtual screening attractive for prioritizing recipes before fabrication. We develop a surrogate-guided reinforcement-learning framework that casts recipe generation as virtual screening, enriching candidate shortlists near high-performing experimental neighbors. A frozen Tabular Prior-data Fitted Network (TabPFN) guides recipe generation, while a reference pool of 890 measured devices enables neighborhood-based assessment. Five-fold evaluation on 1,846 devices shows top-ten selection raises mean power conversion efficiency (PCE) from 19.68\% to 23.74\%. Three learned policies are compared with six baselines under 3,000 surrogate evaluations per run. In top-100 shortlists, Proximal policy optimization (PPO) and recurrent PPO (RPPO) achieve neighborhood-supported hit rates of 23\% and 24\% (7.9- and 8.3-fold Pure Random's 2.9\%), versus 10\% (3.4-fold) for the baselines with the highest supported hit rate (L-SHADE and Ax BO, tied) at $K = 100$. Gains coexist with weak-support fractions. Of 9,000 recipes, 65.4\% match a composition family at similarity 0.80, falling to 11.7\% at 0.90. Results support enrichment-oriented prioritization but not experimental performance, seed robustness, or matched-budget advantage.
BibTeX
@misc{zheng2026enrichment,
  author = {Yezeyuan Zheng and Ziyang Hu and Maojun Sun. Ruijian Han and Jian Huang and Guanhua Chen and Jie Liu},
  title = {{Enrichment-Oriented Virtual Screening of Perovskite Solar Cells through Reinforcement Learning}},
  year = {2026},
  eprint = {2611.18043},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.18043},
}
Figure