Convective cooling of supercritical carbon dioxide inside tubes: heat transfer analysis through neural networks

Abstract

In a previous work, convective heating of carbon dioxide was studied with neural networks (NN), obtaining a totally heuristical heat transfer equation from the direct regression of experimental data. In the present work, the analysis focuses on the cooling process, which has a technical relevance in various applications, as for example in transcritical refrigeration cycles. Heat transfer around the critical zone presents a marked enhancement, that follows the peaks in thermophysical properties like thermal conductivity and heat capacity. Similarly, other properties like density and enthalpy, present a strong variation in narrow temperature intervals around the critical point. This constitutes then a highly non-linear phenomenon, for which it is advisable to use a very flexible function approximator like the NNs. NN models were applied both in terms of dimensionless numbers and of physical quantities, obtaining the two corresponding NN architectures. The choice of the optimal number of neurons in the NN hidden layer is discussed. The NN models are then compared with a recent correlation from literature, for which the validation results present an AAD of 27% and a bias of −26% with an evident prediction shifting. On the other hand the NN models in terms of dimensionless numbers and of physical quantities have AAD and bias of 14% and −4%, and of 7% and −2%, respectively, showing a largely better performance.