Deciphering peroxydisulfate vs. peroxymonosulfate activation by biochar catalysts via explainable machine learning

Biochar (BC) activated persulfate advanced oxidation processes (PS-AOPs), employing peroxymonosulfate (PMS) and peroxydisulfate (PDS), have emerged as a promising approach for eliminating emerging contaminants from water. Nevertheless, a systematic comparison of catalytic behavior and pollutant removal performance between PMS and PDS over biochar is still lacking, which greatly restricts the further optimization of biochar catalyst design and operational parameters in relevant persulfate activation systems. Herein, machine learning (ML) techniques were systematically applied to investigate PMS and PDS activation by non-metal-doped BC, with distinct ML models constructed for each oxidant. SHAP analysis revealed that PMS activation is predominantly governed by experimental conditions, which collectively accounted for 71.3% of the total feature importance. In contrast, PDS activation is more strongly influenced by BC properties, contributing 70.1% to the overall importance. Further SHAP-based interpretation, conducted separately for two systems, identified the most influential BC-related features and their optimal ranges from three perspectives: physico-porous structural properties, surface element composition, and preparation conditions. Additionally, ionization potential was employed to classify contaminants according to degradability, and a "divide-and-conquer" strategy was adopted to build contaminant-directed sub-models, enabling more in-depth mechanistic analysis. For different oxidant-contaminant combinations, priority should be given to tuning key parameters such as biomass precursor, specific surface area, and defect degree to achieve an optimized BC design strategy. Collectively, these findings provide a mechanistic foundation for designing efficient, sustainable, and tailored PS-AOPs using BC catalysts, and offer strategic guidance for pollutant-specific treatment in complex water matrices.