Prediction of Cavity Length Using an Interpretable Ensemble Learning Approach.

Area of Science:

• Engineering
• Computational Fluid Dynamics
• Machine Learning

Background:

• Cavity length is a critical parameter in aeration and corrosion reduction engineering.
• Calculating cavity length is challenging due to numerous influencing factors.
• Accurate prediction of cavity length is essential for optimizing engineering designs.

Purpose of the Study:

• To develop and compare accurate predictive models for cavity length.
• To evaluate the performance of ensemble learning methods against empirical approaches.
• To analyze the feature importance and interpretability of predictive models.

Main Methods:

• Utilized 10-fold cross-validation for optimal input selection.
• Applied Bayesian optimization (BO) to fine-tune hyperparameters of Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Extreme Gradient Boosting Tree (XGBOOST).
• Employed Sobol sensitivity analysis for model interpretability and feature significance assessment.

Main Results:

• XGBOOST demonstrated the highest prediction accuracy among the tested models.
• Sobol analysis revealed that models using individual feature effects (ensemble learning) outperformed those using interactive effects (SVR).
• XGBOOST's ability to capture feature significance variations aligned well with experimental observations.

Conclusions:

• Ensemble learning models, particularly XGBOOST, provide a highly accurate method for predicting cavity length.
• Model interpretability analysis using the Sobol method is crucial for understanding feature contributions under varying conditions.
• The findings support the use of advanced machine learning techniques for complex engineering calculations.