Updated multi-layer perceptron algorithm for predicting solute transport parameters and processes in karst conduits with variable flow rates

Abstract

This study pioneers the application of a Bayesian-optimized multilayer perceptron (MLP) framework to predict the complete breakthrough curve (BTC) in two conduits under various flow conditions, unlike prior research that predicted only partial BTC. MLP shows significant advances in BTC prediction accuracy compared with Random Forest and Support Vector Regression. The transient storage model then simulates predicted BTCs to derive parameters: dispersion coefficient (D), cross-sectional area of main channel (A), cross-sectional area of storage zone (A_s), and exchange coefficient (α). Fifty-four MLP models are developed and trained, with an incremental increase in training data and input variables across scenarios S1-S3. S2 and S3 notably improve BTC prediction accuracy over S1, with R² ≥ 0.9. S2 and S3 predict A with <6.6 % error, and A_s and α with <20 % and <50 % errors respectively. These results prove MLP's effectiveness in predicting solute transport parameters in karst conduits with variable discharges.