Bayesian-optimized machine learning for flow boiling heat transfer prediction in microchannel heat sinks

Flow boiling is an efficient thermal management technique for high-power-density electronic devices. Though, use of conventional methods to simulate flow boiling is computationally expensive and often fails to capture complex microscale phenomena inherent to two-phase flows. To alleviate this drawback, this work presents a surrogate modeling approach that leverages Bayesian-optimized machine learning algorithms to enable accurate predictions of flow boiling performance. Gaussian Process Regression (GPR) and Random Forest (RF), each optimized using Bayesian hyperparameter tuning, were used to improve predictive accuracy. Experimental tests were conducted using three different coolants, i.e., ethanol, acetone, and Novec-7000, at different heat fluxes and inlet volumetric flow rates. The developed model showed excellent performance, with GPR outperforming RF on generalization and error metric scores. Bayesian-optimized GPR (BOGPR) model achieved a regression coefficient (R 2 ) of 0.9995 and a Mean Squared Error (MSE) of 0.0025, compared to 0.9613 and 14.2466 for the RF, respectively. Also, the variance outputs of GPR were used to estimate within 95% confidence intervals, thus enhancing the model interpretability and improving uncertainty quantification. The proposed model is a powerful, data-driven tool for precise prediction of heat transfer in microchannel heat sinks under flow boiling conditions.