Accelerating catalytic process optimization for water treatment via automated knowledge extraction and machine...

Traditional catalyst discovery for water treatment is inefficient due to data acquisition bottlenecks caused by manual literature review. To address this, we developed an automated pipeline integrating high-fidelity document parsing with a Large Language Model. Using Advanced Oxidation Processes (AOPs) as a representative case study, this pipeline proved ∼270 times more efficient than manual methods, enabling the construction of a comprehensive database of 3276 records from over 3000 publications. The structured database is openly available in Zenodo (https://doi.org/10.5281/zenodo.18112768). Based on this data, a multi-stage optimized XGBoost model was trained, achieving R² values of 0.6753 for the degradation rate constant (k_obs) and 0.8351 for removal efficiency (η). We objectively acknowledge that the predictive precision is influenced by the quality fluctuations and reporting heterogeneity inherent in literature-sourced data. Functioning as a "virtual laboratory," the model successfully identified three promising catalysts (Co₃O₄@BC, Fe₃O₄@GO, and CoFe₂O₄), which were subsequently validated by experiments with less than 4 % error. This work presents a data-driven methodology that shifts research from tedious data collection to automated innovation, fundamentally accelerating materials optimization in environmental catalysis.