Temperature Assimilation for Convective Flows by Means of Convolutional Neural Networks

A convolutional encoder-decoder network trained on instantaneous velocity fields is used to assimilate corresponding temperature fields of convective flows. In particular, synthetic data of Rayleigh-Bénard convection in a cubic sample at \(Ra = 10^8, Pr = 0.7\) and \(Ra = 10^{10}, Pr = 6.9\) is studied and the shape and size of the windowed input and output of the network are varied to determine a favorable domain. Additionally, the amount of training data is also varied to determine it’s extrapolation potential. This approach proves to predict the temperature fields well for all parameter variations considered. Particularly good correlations between the predictions and the ground truth are achieved for horizontal planar domains and large amounts of training data.