Large language model-driven decision-making for single-atom Fenton-like catalysts toward wastewater treatment

Tailoring catalysts to diverse environmental conditions for wastewater treatment demands rational material design, yet such decisions remain largely empirical due to fragmented literature knowledge. Here we establish an LLM-powered framework designed to underpin data-driven decision-making in Fenton catalysis. Using optimized few-shot Chain-of-Thought prompting, our framework processed over 1000 domain publications and patents on Fenton-like single-atom catalysts (SACs), systematically extracting knowledge including synthesis, properties, and performance metrics with over 87 % accuracy and 90 % completeness. This process generated a high-fidelity knowledge baseline, revealing key synthesis-structure-property relationships governing SAC performance. Building on these structured insights, we developed an interactive decision-support platform that integrates machine learning-based performance prediction with literature-anchored synthesis recommendations, guiding researchers toward optimal catalyst configurations based on reaction condition. Experimental validation confirmed that platform-recommended catalysts achieved predicted performance under realistic conditions. This work demonstrates a transferable paradigm for intelligent decision-making, accelerating the development of robust catalysts for advanced water purification.