Gravity currents in open channels are important transport mechanisms that influence the propagation of saline plumes in rivers, reservoirs and waterways. Predicting the evolution of the current front in channels with varying geometry and bed roughness conditions remains a challenge due to the non-linear interactions between geometric confinement, buoyancy and hydraulic resistance. In the present study, an explainable machine learning (ML) framework is developed to predict the front propagation of saline gravity currents in a composite trapezoidal open-channel configuration. Eight ML algorithms were employed, combined with a group-aware validation procedure to ensure generalization. Model performance was assessed utilizing standard regression metrics. Among the tested ML models, the CatBoost algorithm achieved the highest predictive accuracy. Interpretation of the model was carried out with the Shapley Additive Explanation (SHAP) approach to quantify the contribution of governing variables including time, initial water depth, density difference and bed condition. The SHAP analysis reveals that the initial depth in the channel has a stronger impact on the front propagation than the density difference, reflecting the combined effects of buoyancy, geometric confinement and bed roughness. Bed roughness is also a contributing factor to propagation dynamics by modifying hydraulic resistance. The proposed ML-SHAP framework provides a robust and interpretable tool for gravity current evolution prediction in channels with complex geometry and varying bed roughness. It may further aid in rapid assessment of transport processes in hydraulic and environmental settings.
Polychronopoulos et al. (Fri,) studied this question.