Predicting suspended floc size in estuarine waters using self-adaptive parameterized physics-informed neural networks

Area of Science:

• Environmental Science
• Fluid Dynamics
• Machine Learning

Background:

• Suspended floc size dynamics are crucial for sediment transport and estuarine ecology.
• Traditional flocculation models struggle with accuracy in complex hydrodynamic conditions due to fixed parameters.

Purpose of the Study:

• To develop a novel self-adaptive parameterized physics-informed neural networks (SAP-PINNs) model for accurate floc size prediction.
• To enhance the adaptability and physical consistency of flocculation dynamic models.

Main Methods:

• Implemented a SAP-PINNs model that dynamically optimizes aggregation, breakage, and erosion parameters.
• Integrated data-driven machine learning with physical constraints for improved modeling.
• Validated the model using laboratory experiments under varying shear stress and field data.

Main Results:

• The SAP-PINNs model demonstrated robustness across variable hydrodynamic regimes, accurately predicting floc size.
• Field validation showed an 88.31% increase in accuracy, R² of 0.99, and MAE of 0.78 compared to traditional models.
• SHAP analysis identified shear stress and salinity as key drivers, with suspended sediment concentration having an optimal range effect.

Conclusions:

• SAP-PINNs effectively combines physics and machine learning for superior accuracy, interpretability, and generalizability in floc dynamics.
• The model offers significant potential for applications in complex hydrodynamic systems, aiding sediment transport and water quality management.
• This approach advances predictive capabilities for estuarine and coastal environments.