Reconstruction and analysis of negatively buoyant jets with interpretable machine learning

Area of Science:

• Environmental Engineering
• Fluid Dynamics
• Computational Science

Background:

• Negatively inclined buoyant jets occur during wastewater discharge, e.g., from desalination.
• Assessing and minimizing their environmental impact requires understanding jet geometry.
• Traditional methods involving experiments and simulations are time-consuming.

Purpose of the Study:

• To apply machine learning (ML) for predicting geometrical characteristics of inclined buoyant jets.
• To evaluate the performance of various ML models in this application.
• To identify key parameters influencing jet behavior using feature interpretation.

Main Methods:

• Dataset generation using OpenFOAM simulations, validated with experimental data.
• Training and evaluation of multiple ML models: Support Vector Regression, Artificial Neural Networks, Random Forests, XGBoost, CatBoost, and LightGBM.
• Utilizing the SHAP (SHapley Additive exPlanations) method for feature interpretation.

Main Results:

• ML models achieved an average prediction accuracy (R ²) of 0.94 ± 0.05.
• Artificial Neural Networks demonstrated superior performance with R ² = 0.98.
• SHAP analysis provided insights into input parameter influence on jet geometry.

Conclusions:

• ML models, particularly ANNs, are effective tools for predicting inclined buoyant jet characteristics.
• This approach can significantly reduce the need for extensive simulations and experiments.
• The findings support better environmental management of industrial wastewater discharges.