Machine learning-driven prediction and interpretability analysis of nucleate pool boiling heat transfer coefficient for R141B-BASED ternary nanofluids

Abstract

This study presents a machine learning framework for predicting the nucleate pool boiling heat transfer coefficient of R141b-based ternary nanofluids containing Al2O3, SiO2, and multi-walled carbon nanotubes (MWCNT) with SDBS surfactant. Although nanofluids have been investigated extensively for boiling enhancement, the manner in which multiple nanoparticle species interact on the boiling surface remains poorly quantified, and no prior study has isolated these synergies or verified whether data-driven models can generalize to unseen ternary compositions. A dataset of 334 experimental data points spanning 34 formulation-pressure conditions (1, 2, and 3 bar) was used to train five ML models under leave-one-condition-out cross-validation, which rigorously tests extrapolation to unseen nanofluid compositions. XGBoost achieved R2 = 0.924 +/- 0.002 and MAPE =11.0 +/- 0.2% across 10 random seeds, with three independent model families converging to R2 = 0.91-0.92. To benchmark ML against established theory, the exact functional forms of four classical correlations were preserved, and their tunable constants re-fitted via the Differential Evolution algorithm; all achieved R2 <= 0.83, revealing a structural limitation due to the absence of nanoparticle concentration terms rather than coefficient miscalibration. SHAP interaction analysis points to a possible SiO2-MWCNT interaction in the ternary formulations; because the two species occur together only at a fixed Al2O3:SiO2 ratio, it is reported as a hypothesis for dedicated binary experiments. An extended power-law correlation incorporating individual nanoparticle terms (R2 = 0.882) is proposed for engineering use. GPR uncertainty analysis identified intermediate concentrations (0.04-0.06 vol%) as regions of highest predictive uncertainty, providing data-driven guidance for future experiments.