Reconstruction of three-dimensional structures from two-dimensional sectional flow fields of cavitating turbulent flows using machine-learning-based super-resolution

A method is proposed for reconstructing spanwise cross-sectional data of a cavitating turbulent flow field around a 2D hydrofoil to a 3D flow field using supervised machine learning (ML). One of the variable of interest was the cross-sectional pseudo-measurement data of particle image velocimetry. Therefore, the spanwise component w of velocity was not obtained, that is, it cannot be input into neural networks (NN). To compensate for this, we introduced a mass conservation error into the loss function of machine learning. The computational fluid dynamics (CFD) data for the training were obtained using large-eddy simulation with a homogeneous fluid model for cavitation. The super-resolution convolutional neural network (SRCNN) was adopted for the supervised ML framework. As a result, under attached sheet cavitation mode, which corresponds to a quasi-steady condition, prediction was relatively simple. In contrast, under sheet/cloud cavitation mode, which is a strongly unsteady condition, the learning curve exhibited oscillations, suggesting that learning is difficult. Physical variables with a uniform distribution in the homogeneous spanwise direction can be reconstructed with high accuracy even under unsteady conditions, whereas the rotational components of streamwise vortices, which are non-uniform in the spanwise direction and have high spatio-temporal frequencies, fluctuate even under quasi-steady conditions, making them extremely difficult to reconstruct. When the loss function is augmented with a mass conservation error, the accuracy of w reconstruction is improved as the NN attempts to satisfy the conservation of mass.