Technical Summary: From Literature to Lab: Closed-Loop Advancement of Perovskite Solar Cells via Domain Knowledge Guided LLM

1. Problem Statement

Perovskite solar cells (PSCs) represent a disruptive next-generation photovoltaic technology with efficiencies exceeding 27%. However, their advancement is hindered by a vast, high-dimensional design space encompassing precursor compositions, solvent engineering, and processing parameters. Traditional researcher-driven trial-and-error approaches are laborious and inefficient, often failing to identify global optima within this complex combinatorial landscape.

While Large Language Models (LLMs) offer potential for leveraging textual scientific knowledge, existing models (e.g., GPT Series, ChemCrow, Coscientist) face critical limitations in the specialized domain of PSCs:

1. Semantic Misalignment: They struggle to align general semantic reasoning with precise, quantitative perovskite domain knowledge, often generating scientifically invalid recipes.
2. Inefficiency in High-Dimensional Spaces: They lack the ability to perform quantitative extrapolation in multi-parametric design spaces, making them ineffective for orchestrating the closed-loop optimization required for device breakthroughs.

2. Methodology: The PVK-LLM Framework

The authors propose PVK-LLM, a domain-knowledge-guided framework designed to bridge the gap between unstructured scientific literature and precise materials recipes. The methodology consists of two primary components: a specialized LLM trained via curriculum learning and a hierarchical Bayesian Optimization workflow (PVK-BO).

A. PVK-LLM: Three-Stage Curriculum Learning

Built upon the Qwen2.5-32B backbone, the model undergoes a three-stage training strategy to transform a general LLM into a domain expert:

1. Stage I: Knowledge Injection (PVK-Sci): The model is fine-tuned on a dataset of 55,104 question-answer pairs covering seven research themes (e.g., Device Structure, Interface Engineering, Defects). This stage internalizes core domain vocabulary and semantic patterns.
2. Stage II: Instruction Alignment:
   • PVK-Cite: Fine-tuning on 22,916 QA pairs to ground responses in specific literature citations, ensuring factual accuracy.
   • PVK-Exp: Fine-tuning on 10,648 QA pairs derived from experimental tables to enable the interpretation of quantitative performance metrics and efficiency mechanisms.
3. Stage III: Knowledge Retrieval-Augmented Generation (RAG): The model is connected to a Perovskite Knowledge Graph (PVK-KG) containing 23,789 entities and 22,272 triples. This graph is dynamically updated via an automated pipeline, allowing the model to access the latest scientific discoveries without frequent retraining.

B. PVK-BO: Hierarchical Bayesian Optimization

To navigate the high-dimensional design space, the framework employs PVK-BO, an optimization engine that integrates the LLM's generative capabilities with Bayesian Optimization. The workflow operates in a closed loop:

1. Warmstarting: PVK-LLM generates high-quality initial candidates based on domain knowledge, addressing the "cold-start" problem.
2. Recipe Design & Candidate Sampling: The LLM acts as a generative proposal distribution, suggesting new recipes conditioned on historical feedback and target performance.
3. Surrogate Modelling: The LLM functions as a probabilistic surrogate to predict the performance of candidates before physical verification.
4. Acquisition Function: Instead of rigid statistical functions, PVK-LLM itself acts as an intelligent acquisition function, weighing potential performance gains against exploration value to select the next optimal recipe.
5. Verification: Recipes are validated via physics-based simulators (SCAPS-1D) or wet-lab experiments, with results fed back to refine the strategy.

3. Key Results

A. Benchmark Performance

• Domain Knowledge: On the PVK-MCQ benchmark (1,103 multiple-choice questions), PVK-LLM achieved a state-of-the-art accuracy of 87.25%, outperforming general LLMs like Qwen2.5-72B and GPT-4o.
• Qualitative Assessment: In open-ended generation tasks (PVK-QA), PVK-LLM scored highest across accuracy, completeness, relevance, and clarity.
• Human Evaluation: In pairwise comparisons with five domain experts, PVK-LLM achieved dominant win rates: 69.5% against Qwen2.5-32B, 65.5% against GPT-4, and 68.0% against Deepseek-R1.
• Semantic Structure: t-SNE visualization confirmed that PVK-LLM's embedding space forms distinct, semantically meaningful clusters for perovskite materials, unlike the random distribution of the base model.

B. Simulator Experiments

Using SCAPS-1D, the framework was tested on Band Alignment and Doping Optimization tasks:

• Initialization Efficiency: PVK-BO achieved a higher initial mean Power Conversion Efficiency (PCE) than random search, Sobol sequences, and Latin Hypercube sampling, effectively bypassing the initial exploration phase.
• Convergence: In Band Alignment, PVK-BO reached a maximum PCE of 26.52%. In Doping Optimization, it reached 25.44%, outperforming standard Bayesian Optimization (BO), HEBO, and TuRBO baselines with lower variance and faster convergence.

C. Wet-Lab Experiment

In an autonomous closed-loop wet-lab experiment targeting p-i-n PSCs:

• Discovery: PVK-LLM identified interface passivation as the bottleneck and autonomously designed a novel, unreported four-component synergistic passivation system: 3MTPAI, PDAI2, EDAI2, and PipDI.
• Optimization: Through three iterative epochs, the model refined the mixture ratios.
  • Epoch 0 (Baseline): 23.68% PCE.
  • Epoch 1: 25.07% PCE.
  • Epoch 2: 25.23% PCE.
  • Epoch 3 (Champion): 26.00% PCE ($V_{OC}=1.155$ V, $J_{SC}=26.35$ mA/cm², $FF=85.44\%$).
• This result approaches world records typically achieved through extensive expert trial-and-error.

4. Significance and Claims

The paper claims that PVK-LLM successfully demonstrates that domain-knowledge-guided AI can:

1. Internalize Scientific Expertise: By bridging the semantic gap between natural language and quantitative material science, the model can comprehend complex fabrication parameters and experimental records.
2. Accelerate Material Discovery: The framework effectively navigates high-dimensional design spaces, overcoming cold-start problems and reducing the iteration cycle from theoretical design to laboratory validation.
3. Enable Autonomous Optimization: The integration of LLMs with Bayesian Optimization allows for the autonomous proposal of novel, high-performance recipes (such as the 26.0% PCE device) without human guidance in the decision-making loop.

The authors note that while the current physical execution relies on a "human-in-the-loop" for wet-lab execution, the framework is designed for extensibility. By integrating standard API interfaces for laboratory automation, the system aims to evolve into a foundational agent for self-driving laboratories, applicable to other material science challenges such as battery electrolyte optimization and organic photovoltaic screening.

Limitations Acknowledged:

• The model's predictive accuracy is bounded by the quality, noise, and reporting inconsistencies of existing open-source academic literature.
• Current wet-lab throughput is limited by manual intervention, though the decision-making loop is fully closed.