Application of LLM in Perovskite Solar Cell Research: An Innovative Approach Combining Large Language Models and Traditional Machine Learning

Introduce the open-source project perovskite_llm_cell_press, which combines large language models (LLMs) with traditional machine learning and applies them to the prediction and reverse engineering of perovskite solar cells. It addresses challenges in perovskite R&D such as large composition space, complex performance factors, high experimental costs, and scattered literature knowledge, accelerates the new material R&D process, and opens up new possibilities for materials science research.

Perovskite materials have become a potential direction for next-generation photovoltaic technology due to their excellent photoelectric conversion efficiency, low-cost preparation, and solution processability. However, R&D faces challenges such as a huge composition space (many ABX₃ combinations), complex performance-influencing factors (correlations between band gap, stability, etc.), high experimental costs, and scattered literature knowledge, making it an ideal scenario for AI-assisted R&D.

The project's innovation lies in the complementarity between LLM and traditional ML: LLM is responsible for literature knowledge extraction (composition, synthesis conditions, performance data, etc.), text data structuring (conversion to ML-usable formats), and hypothesis generation and interpretation; traditional ML is responsible for performance prediction (metrics like PCE, Voc), reverse design (inferring material composition from target performance), and feature importance analysis (quantifying component contributions, etc.).

The project architecture includes a data layer (integrating datasets such as open literature and material properties), a processing layer (LLM-driven text processing like PDF parsing and entity extraction, feature engineering like one-hot encoding and normalization), a model layer (hybrid modeling: random forest, gradient boosting, neural networks), and an application layer (prediction and optimization interfaces, such as rapid performance prediction and reverse design recommendations).

The value of this method includes accelerating material screening (prioritizing testing of potential candidates, narrowing the search space), guiding experimental design (focusing on key component variables, optimizing processes), and knowledge integration and discovery (identifying cross-study patterns, recognizing gaps).

Insights from the project: 1. Complementarity principle (LLM handles knowledge, traditional ML handles numerical prediction, leading to better collaborative effects); 2. Data quality first (open datasets ensure transparency, LLM-assisted cleaning improves reliability); 3. Domain knowledge integration (combining materials science principles and experimental expertise).

Current limitations: Data sparsity (insufficient data for some materials), unproven generalization ability (across component families and preparation methods), and insufficient interpretability (model black box); future directions: Multimodal data fusion (crystal structure, spectroscopy, etc.), application of generative models (new material structure generation), and automated experimental closed loop (integration with robot platforms).