Interpretable machine learning for toluene degradation in biological trickling filters: prediction, parameter optimization.

Biological trickling filters (BTFs) are established sustainable technologies for volatile organic compound (VOC) abatement, yet their performance is often constrained by complex, non-linear interactions among operational parameters. This study proposes a robust and interpretable machine learning (ML) framework to accurately predict and optimize toluene removal efficiency (RE) in BTFs. A comprehensive dataset was systematically acquired from long-term, phased experiments (189 days) by systematically manipulating empty bed residence time (EBRT), nutrient spraying velocity (NSV), and inlet gas concentration (IGC). Among six evaluated models, the Back-Propagation/Multi-Layer Perceptron (BP/MLP) model, optimized via the Sparrow Search Algorithm (SSA), demonstrated superior predictive accuracy and generalization capability. SHAP (SHapley Additive exPlanations) analysis identified EBRT as the dominant factor, followed by IGC, while revealing the nuanced, non-monotonic impact of NSV. To quantify model certainty, predictions were augmented with 95% confidence intervals, achieving a Prediction Interval Coverage Probability (PICP) of 94.4%, which quantitatively validates the model’s reliability. Furthermore, ML-based functional mapping visualized the operational landscape and identified an optimal window: EBRT of 120 s, NSV of 1.5 L/h, and IGC of 250–300 mg/m3N. Independent experimental validation under these conditions yielded an average toluene RE of 91.1%, closely aligning with model predictions. Complementing these engineering metrics, microbial community analysis revealed a significant shift towards Ralstonia dominance (77.3%) and the enrichment of key metabolic enzymes, providing a biological basis for the superior performance. This work demonstrates the integration of interpretable AI with process modeling, offering both a precise predictive tool for complex biosystems and data-driven insights for intelligent operational control.