The peroxymonosulfate (PMS) activation was influenced strongly by active sites on a catalyst surface. To some extent, biochar characters, such as structure, elemental composition, and specific surface area (SSA), can reflect the composition of active sites. However, the understanding of active centers lacked accuracy in the structure–activity relationships of biochar-based catalysts (BBC) due to the consideration of one single feature. In this study, XGBoost (XGB), random forest regression (RF), and supporting vector regression (SVR) models were employed to construct descriptors to predict non-radical (NR) contribution in BBC/PMS systems. Consequently, the XGB model provided the best prediction with an R2 value of 0.81. Thereby, activation preference (γ), as a characteristic descriptor of BBC, was developed using the XGB model. Interestingly, the descriptor fully considered carbon content (C%), oxygen content (O%), defects (ID/IG), and SSA. Besides, O% played the most critical role in predicting NR contribution. Noticeably, the NR oxidation pathway could be promoted within the range of 50–60 % for C%, 30–50 % for O%, 0.4–1.2 for ID/IG, and 579–1259 m2/g for SSA in the BBC/PMS system. Herein, machine learning (ML) was applied to develop a comprehensive descriptor, further providing an innovative strategy for high-throughput screening of BBCs to activate PMS efficiently. The comprehensive descriptor offered a new perspective for elucidating the structure–activity relationship of BBC, facilitating the application of BBC/PMS systems in water treatment.