ISSN Online: 2377-424X
International Heat Transfer Conference 12
Forced convective heat transfer to supercritical carbon dioxide inside tubes. Analysis through neural networks
Sinopsis
In literature a variety of correlations has been proposed to predict the coefficients for heat transfer to fluids in the
near-critical region, but discrepancies are reported. In fact in the near-critical region the thermophysical
properties such as density, heat capacities, enthalpy, viscosity, thermal conductivity, etc. present intense
variations for limited changes of temperature or pressure. As it is usual, the correlations have been developed
assuming an initial approximate model in which successive modifications are introduced to correct discrepancies
with a "trial and error" like procedure. The present work aims instead at developing an analytical expression
directly from experimental evidences regressing organized experimental data.
Neural networks have been assumed for this task, because they are a very versatile and powerful function approximator tool. A neural network has been trained on a limited amount of data covering homogeneously the operative conditions range. Once the network has been successfully trained, it is able to represent the behaviour of the whole data set. Besides, the extrapolation beyond the training range proves to be very regular.
Three different correlation architectures are proposed for the neural networks, based both on dimensionless groups as variables of the neural network function and on directly accessible physical quantities. In all the three architectures the definition of the optimal functional form for the correlation is obtained through a completely heuristic procedure only based on experimental data, and the reached accuracy is comparable with the claimed experimental uncertainties.
Neural networks have been assumed for this task, because they are a very versatile and powerful function approximator tool. A neural network has been trained on a limited amount of data covering homogeneously the operative conditions range. Once the network has been successfully trained, it is able to represent the behaviour of the whole data set. Besides, the extrapolation beyond the training range proves to be very regular.
Three different correlation architectures are proposed for the neural networks, based both on dimensionless groups as variables of the neural network function and on directly accessible physical quantities. In all the three architectures the definition of the optimal functional form for the correlation is obtained through a completely heuristic procedure only based on experimental data, and the reached accuracy is comparable with the claimed experimental uncertainties.