Mechanism-embedded graph neural network for predicting and interpreting methane and sulfide production in sewers

Understanding and predicting methane (CH₄) and sulfide (H₂S) production in sewers remains significantly challenging due to the complex influence of dynamic water quality on the production pathways intensities. In this study, we develop a Mechanism-Embedded Graph Neural Network (ME-GNN) that explicitly incorporates established CH₄ and H₂S production pathways into its architecture. ME-GNN achieves superior predictive accuracy, with R² of 0.78 ± 0.03 for CH₄ production and 0.85 ± 0.11 for H₂S production, significantly outperforming both mechanistic and machine learning models. By structurally embedding mechanistic knowledge, the model improves performance and reduces dependence on large training datasets. Interpretability analysis identifies that acetate (S_ac) and sulfate (SO₄²⁻) concentrations are key water quality factors that dynamically regulate production pathway intensities. Elevated SO₄²⁻ concentration shifts the dominant pathway from CH₄ production (72 % to 11 %) to H₂S production (28 % to 89 %), whereas higher S_ac concentration reverse this trend. This framework bridges the gap between data-driven modeling and biochemical principles, offering a powerful tool for predicting CH₄ and H₂S production and quantifying dynamic production pathway intensities in sewer.