ABSTRACT This paper presents a proof‐of‐concept for an inter‐material transfer learning framework to accelerate the prediction of nanofluid convective heat transfer in laminar internal pipe flow. While machine learning (ML) surrogate models can reduce the computational cost of CFD simulations, they typically require large, material‐specific datasets. Our work addresses this by demonstrating that an XGBoost model, pre‐trained on 200 CFD simulations of a source material (Al 2 O 3 ‐water), can be fine‐tuned to accurately predict the Nusselt number for a different target material (a hybrid Al 2 O 3 –TiO 2 nanofluid) using only 20 data points. The transfer learning model achieved a high coefficient of determination (R 2 = 0.980, 95% CI 0.951, 0.988), representing an estimated 80% reduction in data requirements compared to a model trained from scratch. The framework also incorporates uncertainty quantification via bootstrap ensembles and advanced explainability (XAI) to ensure the model's predictions are both reliable and physically interpretable. This study establishes a robust methodology for developing accurate, data‐efficient, and trustworthy ML models for thermal engineering applications.
Salah et al. (Thu,) studied this question.