超解像機械学習による疎な疑似圧力センサ点データからの二次元圧力場推定とキャビテーション場の予測

抄録

The two-dimensional pressure field around a hydrofoil is estimated from the pressure values of sparse sensors flush-mounted on the hydrofoil. The sparse data is expanded to pressure field data using a combination of multi-layer perceptron (MLP) and super-resolution convolutional neural network (SRCNN) techniques, where MLP is employed to temporarily increase the size of the sparse data to that of the output two-dimensional field data. The training dataset for these neural networks (NNs) is CFD data of the cavitating flow field computed using a transport-equation-based cavitation model. Therefore, the prediction accuracy depends on the accuracy of the CFD of the training dataset. The input is the pressure values at 25 points on the hydrofoil surface in a certain spanwise section, and the NNs are trained with the corresponding pressure fields in the spanwise section as the label data. As a result of the training, the pressure field for one cycle of sheet/cloud cavitation is qualitatively estimated. However, the accuracy is relatively low in the vicinity of the trailing edge of the hydrofoil, where cavitation grows rapidly. When the pressure wave causes due to the disappearance of the cloud cavitation, the accuracy is also poor for most of the wake. The homogeneous fluid density field is predicted from the estimated pressure field and the barotropic cavitation model. The boundary of the cavitation region and liquid phase region are blurred. This may be due to the use of estimated pressure based on the transport equation model, which considers non-equilibrium effect and advection.