Heat transfer prediction of supercritical water with artificial neural networks

Supercritical fluids have been under intensive investigation due to their broad applications in the domain of energy conversion. They are able to significantly increase the efficiency of thermal cycles. However, abrupt changes of thermophysical properties of the supercritical fluids have been observed near the critical point, causing heat transfer deterioration that is challenging to predict. This largely thwarts the technology development with the supercritical fluids, making accurate prediction of heat transfer the fundamental issue to address. In this paper, we propose to train an artificial neural network (ANN) based on 5280 data points collected from published experimental results, for the heat transfer prediction of supercritical water. Validation (strictly separated from training) shows that the mean error percentage and its standard deviation are both below 0.5%. Furthermore, a series of tests, including operational conditions out of the training and validation data, are performed in comparison with four well-established correlations. The results demonstrate that the performance of the ANN is considerably better than the correlations. Training of the ANN takes less than an hour on a regular computer, and the prediction takes several milliseconds. This is the first time that ANNs are trained for general heat transfer prediction of supercritical water. This paper should open further opportunities in the supercritical fluids research to be pursued jointly by the fluid dynamics community and the computer science community.