Interpretable and extrapolation-stable model for predicting nanofluid thermal conductivity.

Area of Science:

• Materials Science and Engineering
• Computational Physics
• Chemical Engineering

Background:

• Accurate prediction of nanofluid thermal conductivity is crucial for designing efficient thermal management systems.
• Existing models often struggle to balance predictive accuracy with physical interpretability.
• Nanofluids offer enhanced thermal properties but require precise modeling for application.

Purpose of the Study:

• To develop a physics-guided hybrid modeling framework for enhanced nanofluid thermal conductivity prediction.
• To integrate Generalized Additive Models (GAM) and Gradient Boosting Machines (GBM) for improved accuracy and interpretability.
• To overcome the limitations of traditional models in predicting thermophysical properties.

Main Methods:

• A hybrid model combining a spline-based GAM for global trends and a regularized GBM for local corrections was developed.
• A preprocessing pipeline included feature engineering, Box-Cox transformation, and Random Forest-based outlier detection.
• The model was rigorously evaluated against multiple state-of-the-art machine learning algorithms.

Main Results:

• The proposed hybrid framework achieved high predictive accuracy for nanofluid thermal conductivity.
• The model demonstrated physically consistent behavior and maintained interpretability across the studied domain.
• Leave-One-Group-Out (LOGO) validation confirmed robust generalization capabilities, even for unseen base fluids.

Conclusions:

• Physics-guided hybrid modeling offers a balanced and reliable approach for thermophysical property prediction.
• The integrated GAM-GBM model enhances interpretability without sacrificing predictive performance.
• This methodology advances the accurate design and application of advanced thermal systems utilizing nanofluids.