Interpretable machine learning-based prediction and analysis of supercritical fluid heat transfer characteristics at different boundary conditions

The study of the heat transfer characteristics of supercritical fluids is crucial for the safe and economical operation of supercritical energy systems. This study presents a new method to study the heat transfer characteristics of supercritical fluids, employing interpretable machine learning. Two back propagation neural network models are proposed for the prediction of heat transfer to supercritical fluid for the cases with given heat flux and given wall temperature, respectively. The particle swarm optimization algorithm was employed to search for the optimal hyperparameters, and the resulting accuracy was evaluated in comparison with that of the traditional empirical correlation and a number of machine learning models. Combining the SHAP interpretable algorithm with prediction models, the supercritical heat transfer mechanism is explored from the perspective of global and local prediction behaviors based on explainable models. The results demonstrate that the average error of two established neural network models on the test set is 0.47 % and 0.69 %. Furthermore, for vertical upward flow, the buoyancy and acceleration effects are of greater feature importance in heat transfer deterioration. They are the primary factors contributing to heat transfer deterioration behavior. The research method proposed in this study has certain reference significance for further study of the heat transfer characteristics of supercritical fluids. • The supercritical heat transfer mechanism is explored based on interpretable algorithm. • Two prediction models are proposed at two different boundary conditions. • The optimization algorithm is used to determine the network hyperparameters.