Study on the extrapolability of artificial neural network for predicting convective heat transfer of supercritical fluid based on a small number of samples

Artificial neural networks are promising for predicting highly nonlinear characteristics of convective heat transfer. Dimensionless and parametric networks are two strategies for supercritical fluid heat transfer prediction, yet their relative accuracy and extrapolation performance with few samples remain unclear. This study obtained a training sample dataset containing 410 sets of data and an extrapolation test dataset containing 940 sets of data of n-decane under supercritical pressure via experiments and calculations. Two corresponding networks were constructed and trained using the sample data, and wall temperature from the dimensionless network was obtained by a surface-intersection method. Results show that in the training sample dataset, both networks show similar errors, while the dimensionless network better captures the characteristics of heat-transfer deterioration. In the extrapolation test dataset, the dimensionless network demonstrates higher accuracy, while the parametric network yields unreasonable predictions in which the predicted wall temperature is lower than the bulk fluid temperature. The reason is that the dimensionless network leverages prior knowledge about the correlation of dimensionless number groups. By directly learning dimensionless variables strongly associated with thermophysical properties, it reduces the degree of nonlinearity in its structure. In contrast, although the parametric network has a simpler structure, it conceals the nonlinear relationships of thermophysical properties and the implicit relationships between wall temperature and other variables. This makes it difficult to extract sufficient information from a small number of samples. This research provides insights into the differences in the extrapolation capability of different artificial neural networks when faced with a limited sample size.