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

TL;DR

This work introduces PVK-LLM, a domain-knowledge–guided LLM tailored for perovskite solar cell research, and couples it with PVK-BO, a Bayesian optimization loop that leverages domain knowledge and experimental feedback to navigate a high-dimensional material-design space. The framework is trained via a three-stage curriculum (PVK-Sci, PVK-Cite, PVK-Exp) and grounded through PVK-KG, enabling retrieval-augmented generation and robust knowledge grounding. Benchmark results show state-of-the-art domain understanding (PVK-MCQ accuracy $87.25\%$) and superior QA performance, while simulator and wet-lab experiments demonstrate accelerated closed-loop optimization, achieving a champion PCE of $26.00\%$ in a four-component passivation system. The work demonstrates that integrating structured domain knowledge with LLMs can rapidly advance real-world PSC development and offers a scalable blueprint for autonomous experimentation in other high-dimensional material systems.

Abstract

Perovskite solar cells (PSCs) have been considered as a next-generation disruptive photovoltaic technology, yet their advancement is constrained by the complexity of perovskite recipe with high-dimensional material and process design space. Despite the impressive general reasoning of Large Language Models (LLMs), they struggle with two limitations for application in PSCs: an inability to align general semantics with the perovskite domain knowledge, and an inefficiency in navigating high-dimensional perovskite material and recipe design spaces. To address these limitations, we introduce a domain-knowledge-guided framework PVK-LLM, a specialized model to serve as an expert to bridge general semantics with perovskite domain knowledge. By integrating this domain knowledge into a hierarchical Bayesian Optimization workflow, our approach efficiently navigates the high-dimension design space on a solar cell simulator platform. The domain knowledge resolves cold-start problems while dynamically adapting to simulator feedback. Moreover, in an individual wet-lab experiment aimed at maximizing power conversion efficiency (PCE), our framework autonomously proposes a novel synergistic four-component recipe comprising specialized organic passivation recipe (3MTPAI, PDAI2, EDAI2, and PipDI) which has not been reported in existing literature. This AI-designed recipe effectively achieves a champion PCE value of over 26.0 %, approaching world records achieved through extensive expert trial-and-error. Our approach can effectively enable LLM comprehend the domain knowledge, which can efficiently navigate in a high-dimensional, capable to accelerate the advancement in real-world perovskite as well as other material science development.

Figures (10)

• Figure 1: Architecture and implementation of the PVK-LLM framework for autonomous scientific advancement. (a) Overview of the closed-loop active learning workflow to integrate domain knowledge. Literal knowledge extracted from scientific literature constructs a knowledge base (KB) to fine-tune the model through three phases. This enables PVK-LLM to drive the design, verification, and optimization cycles in wet-lab and simulator (sim.) environments. (b) The model architecture and learning objective for PVK-LLM. GT means ground truth. (c) Composition of the three-stage curriculum learning datasets. Stage I (PVK-Sci) covers domain knowledge; Stage II integrates knowledge grounding (PVK-Cite) and experiment analysis data (PVK-Exp); Stage III establishes a structured knowledge graph (PVK-KG). Performance Enhancement and Defect & Recombination are abbreviated as Perf. Enhan. and Def. & Rec., respectively. (d) The distribution of the perovskite literature corpus, spanning from $2018$ to $2025$ July. (e-g) Representative examples of the instruction-tuning datasets: (e) PVK-Sci addresses scientific QA; (f) PVK-Cite incorporates retrieval-based evidence for knowledge grounding; (g) PVK-Exp utilizes structured experimental records to facilitate detailed mechanism analysis and performance attribution.
• Figure 2: Benchmark results in domain-specific evaluation. (a) Objective evaluation on the PVK-MCQ dataset. PVK-LLM achieves a state-of-the-art score of $87.25 \%$, surpassing powerful general LLMs, demonstrating the efficacy of domain knowledge injection. All experiments are conducted with five independent runs. (b) Subjective Human Expert evaluation. Expert reviewers rate PVK-LLM higher than other models. (c) Detailed metric breakdown using LLM-as-a-Judge. Our PVK-LLM outperforms baselines across accuracy, completeness, relevance, and clarity. All experiments are conducted with five independent runs. (d) t-SNE visualization of embedding spaces. Comparison between the base model (Qwen2.5-32B) and PVK-LLM reveals that the fine-tuned model forms more distinct, semantically meaningful clusters of perovskite materials. The legend abbreviations correspond to specific material categories: Back (Back Contact Layer), Device (Complete Device Structure), ETL/HTL (Electron/Hole Transport Layers) and their additives (Add.), A/B/C-site (Perovskite Cations), Pero. Add./Abbr. (Perovskite Additives/Abbreviations), and Substrate.
• Figure 3: Framework and performance of simulator experiments. (a) Workflow of the PVK-LLM-based Bayesian Optimization (PVK-BO). (b) Schematics of the physical optimization tasks. Task #1 optimizes Band Alignment to minimize energy barriers at interfaces, while Task #2 optimizes Doping concentrations to align Fermi levels and reduce defects. (c) For Band Alignment and Doping Optimization, PVK-BO achieves maximum PCEs of $26.52 \%$ and $25.44 \%$, respectively. All experiments are conducted with five independent runs.
• Figure 4: Wet-lab experimental validation and optimization of perovskite solar cells empowered by PVK-LLM. (a) The human-in-the-loop active learning workflow. The process proceeds through three stages: Initialization of the base recipe, Screening of recipe based on experimental feedback, and Optimization of the mixture ratio using PVK-BO. (b) Evolution of the PCE distribution across iterative experimental epochs. The results demonstrate a clear upward trajectory in device performance from Epoch 0 to Epoch 3. (c) PCE distribution for each recipe from Epoch 0 to Epoch 3. We let our frame work to recommend seven recipes in each epoch. (d) Current density-voltage (J-V) characteristics of the champion device ("Target") compared to the control group. The optimized device achieves a performance boost with a peak PCE of $26.00 \%$.
• Figure 5: Other analysis of our pipeline. (a) and (b) are conducted with five independent runs. (a) Model ablation on experiment analysis. Comparing the PVK-LLM (Fine-tuned) with Qwen2.5-32B (Base) shows gains. (b) Evaluation of knowledge grounding. The PVK-LLM (Fine-tuned) demonstrates improvement. (c) Assessment of PVK-KG quality. PVK-KG achieves higher metrics compared to other automatic methods. (d) Case study on KG construction. Our pipeline identifies SnO$_2$ as the ETL layer while other methods fail. (e) The correlation of simulator experiment parameters.