Analysis of forced convection heat transfer to supercritical carbon dioxide inside tubes using neural networks

The modeling of forced convection heat transfer for carbon dioxide flowing inside a heated tube at supercritical conditions was studied. The conventional models in the literature tend to modify a constant property correlation by including thermodynamic property terms that follow the heat flux trends. An innovative heuristic method is assumed here for the first time to draw the case-specific heat transfer coefficient correlation from the experimental data on said quantity alone. Neural networks were used since they constitute a general, powerful function-approximator tool proving able to represent a conventional heat transfer surface precisely in the present case. Four different correlation architectures were considered for the neural network function, alternatively based on dimensionless groups and on directly accessible physical quantities as independent variables. In all these architectures, the optimal functional form of the correlation was obtained using a completely heuristic procedure based exclusively on experimental data, reaching an accuracy comparable with the experimental uncertainties declared. An improved performance of the present model was found with respect to conventional correlations. On all the data sets, the third architecture reaches an AAD of 3.98% against 4.09% for the conventional equation and the fourth architecture an AAD of 2.67% against 4.30% for the conventional equation. Besides both these NN architectures present Bias values very close to 0, whereas the conventional equation has a Bias considerably greater.