Creation of a structured solar cell material dataset and performance prediction using large language models.

Area of Science:

• Materials Science
• Artificial Intelligence
• Natural Language Processing

Background:

• Materials scientists rely on experimental data for material prediction and improvement.
• Utilizing unstructured scientific literature to update structured datasets remains a significant challenge.
• Bridging the gap between text-based knowledge and structured data is crucial for applied science.

Purpose of the Study:

• Introduce a novel natural language processing task: structured information inference (SII).
• Develop an end-to-end approach to extract and structure device-level information from scientific literature.
• Enhance existing materials datasets for improved data analysis and predictive modeling.

Main Methods:

• Proposed a structured information inference (SII) task using natural language processing.
• Developed an end-to-end deep learning framework to process multi-layered, device-level information.
• Fine-tuned the LLaMA model, achieving an 87.14% F1 score for data extraction and structuring.

Main Results:

• Successfully updated a perovskite solar cell dataset with newly published research using the fine-tuned LLaMA model.
• Enabled direct use of updated structured data for subsequent machine learning analyses.
• Developed regression models for predicting solar cell electrical performance, showing competitive results against traditional methods.

Conclusions:

• Large language models show significant potential for scientific knowledge acquisition.
• The proposed SII approach effectively transforms unstructured literature into valuable structured data.
• This methodology accelerates material development and enhances the predictive capabilities in solar cell research.