Interpretable and extrapolation-stable model for predicting nanofluid thermal conductivity

Abstract

Accurate prediction of nanofluid thermal conductivity is essential for the design of advanced thermal systems; however, existing approaches often face a trade-off between predictive accuracy and physical interpretability. In this study, a physics-guided hybrid modeling framework is proposed by integrating a Generalized Additive Model (GAM) with a Gradient Boosting Machine (GBM) to address this limitation. The proposed methodology combines a spline-based GAM to capture global thermophysical trends with a regularized boosting model to learn localized nonlinear corrections. A comprehensive preprocessing pipeline is implemented, including feature engineering, Box-Cox transformation for variance stabilization, and a Random Forest-based outlier detection strategy. The hybrid model is evaluated against several state-of-the-art machine learning methods, including Random Forest, Support Vector Regression, Gaussian Processes, Neural Networks, and Decision Trees. The results demonstrate that the proposed framework achieves high predictive accuracy (RMSE [Formula: see text] W/mK) while maintaining physically consistent behavior across the studied domain. The model's generalization capability is further validated using a Leave-One-Group-Out (LOGO) protocol, where it exhibits stable performance when predicting unseen base fluids. Additionally, the hybrid approach produces smooth and interpretable response profiles that align with known thermodynamic trends. These findings indicate that physics-guided hybrid modeling provides a balanced and reliable approach for thermophysical property prediction, offering improved interpretability without compromising predictive performance.