Abstract The proliferation of antibiotics in the environment poses a severe threat to public health and ecosystems. While biochar-based catalytic systems offer a promising remediation strategy, their design is complicated by nonlinear interactions between material properties and reaction conditions. To overcome this, we present an interpretable machine learning framework for the accurate prediction of reaction rate constants ( k ) for antibiotic degradation. A comprehensive dataset was compiled from the literature, encompassing 16 features across three domains including biochar catalyst properties, elemental composition, and reaction conditions. Among the six machine learning (ML) algorithms evaluated, the transformer-based deep learning model TabPFN (Tabular Prior-data Fitted Network) demonstrated superior performance (Test R 2 = 0.91, RMSE = 0.021), outperforming tree-based, kernel-based, and neural network models. Model interpretability analyses revealed that catalyst properties contributed the largest share of predictive capability (59.3%), followed by reaction conditions (25.9%) and elemental composition (14.8%). Persistent free radicals (formed at 450–550 °C) and high total pore volume (> 0.23 cm 3 g − 1 ) were identified as key drivers of reaction kinetics, along with optimal oxidant (0.5–5.5 mg L − 1 ) and pollutant concentrations (< 22 mg L − 1 ). These insights were embedded into a user-friendly web-based GUI, enabling rapid k prediction for new biochar catalysts with prediction errors below 20% on external validation samples. This work provides both a robust predictive tool and a generalizable, data-driven methodology for understanding and optimizing complex environmental catalytic processes. Graphical Abstract
Latif et al. (Fri,) studied this question.