A supercritical carbon dioxide cooling heat transfer machine learning prediction model based on direct numerical simulation

Abstract

The severe thermo-physical properties variations of supercritical fluids in the vicinity of the critical point lead to difficulty in heat transfer prediction. In this paper, a novel prediction model for supercritical CO₂ (sCO₂) cooling heat transfer is proposed, integrating a Direct Numerical Simulation (DNS) database with the CatBoost algorithm. A high-precision heat transfer prediction database was established based on DNS data (8192 data points in total). The feature parameters were screened utilizing the random forest feature importance method. More importantly, a newly dimensionless parameter, ${Re}^{-0.9}{\pi }_A$, was selected as one of the feature parameters. ${Re}^{-0.9}{\pi }_A$ represents the impact of near-wall acceleration caused by buoyancy, and it demonstrated the highest feature importance during training and screening processes. Based on the selected ten characteristic parameters, a robust data-driven heat transfer prediction model for sCO₂ was developed. The CatBoost algorithm outperformed the other three widely used machine learning algorithm s across training sets, testing sets, and actual predictions, achieving a mean absolute percentage error reduction of up to 22.69 %. Through comparison with 6 traditional heat transfer correlations, the results showed that the CatBoost-based sCO2 cooling heat transfer prediction model exhibits superior training speed and predictive accuracy, with a maximum relative error of merely 6.57 %. Moreover, when validated through experimental data with large Reynolds number, this model still has the highest accuracy, with 91.3 % of the data sets having a prediction accuracy within ±30 % for the Nusselt number.