Prompt Engineering Accelerates the Data-Driven Discovery of Photocatalysts via an LLM-Based Model Ensemble Strategy

Area of Science:

• Materials Science
• Artificial Intelligence
• Chemical Engineering

Background:

• Scientific literature contains vast, unstructured data crucial for materials discovery.
• Extracting synthesis and property data from text is a significant bottleneck.
• Automated methods are needed to unlock this information.

Purpose of the Study:

• To develop a closed-loop framework integrating automated knowledge extraction and machine learning for materials discovery.
• To systematically extract and analyze data on defect-engineered carbon nitride photocatalysts.
• To accelerate the development of functional materials through data-driven insights.

Main Methods:

• A novel data extraction pipeline using a prompt-engineered large language model and model ensemble strategy.
• Construction of a high-fidelity dataset for carbon nitride photocatalysts.
• Application of interpretable machine learning (SHapley Additive exPlanations) to identify key performance parameters and relationships.
• Experimental validation of data-driven predictions.

Main Results:

• The data extraction system achieved 90% accuracy and recall for key parameters.
• Specific surface area and bandgap were identified as dominant performance parameters for carbon nitride photocatalysts.
• A non-monotonic relationship for bandgap was elucidated, revealing an optimal range (2.2-2.4 eV).
• Experimental hydrogen evolution rates showed less than 5% deviation from model predictions.

Conclusions:

• The developed framework effectively transforms fragmented literature into actionable intelligence.
• This approach accelerates the discovery and development of functional materials.
• The scalable and transferable paradigm offers a powerful strategy for materials science research.