Using a deep convolutional network to predict the longitudinal dispersion coefficient.

Area of Science:

• Hydraulic engineering and water quality modeling.
• Machine learning applications in environmental science.

Background:

• Accurate prediction of the Longitudinal Dispersion Coefficient (D_x) is crucial for hydraulic and water quality modeling.
• Existing empirical and machine learning models have limitations in predicting D_x, especially at extreme values.

Purpose of the Study:

• To propose and evaluate a Deep Convolutional Network (DCN) for improved prediction of the Longitudinal Dispersion Coefficient (D_x).
• To compare the DCN model's performance against empirical, Artificial Neural Network (ANN), Support Vector Machine (SVM), and other machine learning (ML) models.

Main Methods:

• Developed a DCN architecture comprising a 1D CNN for feature extraction and fully connected layers for D_x estimation.
• Utilized dimensionless parameters (Width/Depth, Velocity/Shear Velocity, D_x/(Depth * Shear Velocity)) for model training.
• Performed statistical analysis and five-fold cross-validation to assess model accuracy, sensitivity, and robustness.

Main Results:

• The DCN model significantly outperformed all tested empirical, ANN, SVM, and ML models in predicting D_x.
• Cross-validation confirmed that dataset selection did not significantly impact the DCN model's accuracy.
• The DCN demonstrated excellent accuracy across the full range of D_x values, including extreme upper and lower ranges, with low errors.

Conclusions:

• The proposed DCN model offers a superior approach for predicting the Longitudinal Dispersion Coefficient (D_x) in streams.
• DCN's ability to learn high-level features incrementally provides a significant advantage over existing methods for environmental modeling.
• The DCN model provides a reliable tool for accurate D_x estimation using river geometry and hydraulic data.