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

The overwhelming volume of unstructured scientific literature presents a fundamental bottleneck to materials discovery, where critical data on synthesis and properties remain locked in text. Here, a closed-loop framework that integrates automated knowledge extraction with interpretable machine learning and targeted experimental validation is presented. This approach is centered on a novel data extraction pipeline, which combines a prompt-engineered large language model with a model ensemble strategy, systematically optimized to interpret complex materials science narratives. When deployed to construct a database for defect-engineered carbon nitride photocatalysts, the system achieved 90% accuracy and recall for key parameters. Analysis of the high-fidelity dataset enabled reliable machine learning models to identify specific surface area (170 m² g^-1) and bandgap (≈2.31 eV) as dominant performance parameters. Crucially, SHapley Additive exPlanations analysis elucidated a non-monotonic relationship for bandgap, identifying an optimal range of 2.2-2.4 eV and quantifying the fundamental trade-off between light absorption and charge recombination. These data-driven insights guided the synthesis of representative materials, with experimental hydrogen evolution rates deviating by less than 5% from predictions. This work establishes a scalable and transferable paradigm, transforming fragmented literature into actionable intelligence and offering a powerful strategy for accelerating the development of functional materials.