Integrated machine learning and density functional theory to design C₃N₅ piezo-catalysts toward precise control of selective reactive oxygen species and efficient water purification.

The precise modulation of reactive oxygen species (ROS) generation pathways is crucial for enhancing the selectivity and efficiency of piezo-catalytic water purification. Herein, we report a mechanism-informed machine learning (ML) strategy integrating density functional theory (DFT) to rationally design non-metal-doped C₃N₅ piezo-catalysts. Using dipole moment as a key descriptor, our optimized ML model achieved high predictive accuracy and identified boron (B), sulfur (S), and phosphorus (P) as optimal dopants. Experimental validation confirmed that S-doping significantly enhanced polarization and charge separation, leading to generation of •OH and •O₂- and superior degradation performance of sulfamethoxazole (SMX). In contrast, B-doping favored a hole-driven •OH pathway. Crucially, S-doped catalyst maintained high efficiency and stability across a wide pH range and in realistic water matrices. This work demonstrates the power of ML-guided design for developing advanced piezo-catalysts with tailored ROS pathways for efficient environmental remediation.