Feasibility Study on Boiling Heat Flux Prediction in A Nanofluid using Physics and Correlation-Informed Machine Learning

Abstract

A PIML (Physics-Informed Machine Learning) approach is proposed for predicting pool boiling heat flux of Graphene Quantum Dot (GQD) nanofluids. GQD nanofluids have gained attention as next- generation heat transfer fluids due to their unique advantages including low toxicity, high surface area, and excellent dispersibility. However, accurate prediction of their heat transfer characteristics through conventional physical and mathematical models has been challenging due to numerous factors affecting nanofluid heat transfer characteristics. In this study, a novel prediction methodology combining physi- cal knowledge with data-driven learning was developed to overcome these limitations. The Rohsenow correlation was utilized as a physical prior model, and property prediction models for GQD nanoflu- ids were investigated and employed to construct a property dataset. This was integrated with machine learning based on Feedforward Neural Networks (FFNN) and Random Forests (RF). Extensive hyper- parameter exploration was conducted for model optimization. The developed NN-based PIML model demonstrated excellent performance with rRMSE 0.1881, R² 0.9752, and MAPE 14.6107% on test data. Notably, compared to conventional machine learning models including Random Forest, Support Vector Machine (SVR), and Feedforward Neural Network, stable prediction performance was maintained even in high heat flux regions (above 1200 kW/m²). This research demonstrates that complex heat trans- fer phenomena can be effectively predicted even in limited data environments through the synergy of physical laws and data-driven learning.