Supervised learning method for the physical field reconstruction in a nanofluid heat transfer problem

This paper presents a supervised learning method for the physical field reconstruction in a specific heat transfer problem. The deep convolutional neural network (CNN) is applied to predict fields from a few measurable information, while heat transfer characteristics of interest can be then easily inferred from the fields. This data-driven method can establish an end to end mapping from low-dimensional measurable information to full physical fields. Two modes of measurable information are considered as inputs of the network. When the measurable information is an accurate structure or work condition parameters, this method is equivalent as an efficient surrogate model instead of computational fluid dynamics (CFD) simulation. This network can also reconstruct the full-field from local information with several measuring points as inputs. To our best knowledge, this is the first time a CNN based model has been used as a high-fidelity field predicator for the flow heat transfer. To validate this method, the fields of Al₂O₃-water nanofluid laminar flow in a grooved microchannel are employed to be reconstructed from a set of reduced parameters. It indicates that the reconstruction model enables accurate results for all the temperature, velocity and pressure fields. Meanwhile, the characteristics concerned in a heat transfer process, such as Nu and f, can also be extracted from the reconstructed fields with high precision. Furthermore, the reconstruction performance and stability are verified from several perspectives, including the loss function, train-data size, measuring noise and points layout. At last, the comparison of computational costs shows that a well-trained CNN model has three orders of magnitude faster than CFD solver. The proposed approach can provide an efficient analysis tool with acceptable accuracy for heat transfer research.