CNN based Mode Decomposition Model with Weight-Shared Decoder for Complex Flows

To perform mode decomposition of complex flow fields, such as turbulence, we develop Weight-sharing Mode Decomposition Model based on Convolutional Neural Network and Autoencoder. For decreasing the number of model layers and parameters, the Model utilizes a common decoder with weight sharing to extract multiple decomposed flow fields. The array of the latent variables mapped from the flow fields are labeled with the mode numbers. Then, the decomposed flow fields are obtained from the labeled modes using the shared decoder. The number of parameters of a shared decoder in the Model with n modes is reduced to 1/n compared to n decoders in the conventional model. In this study, we apply the Model with 4 modes to a flow around two parallel cylinders at Reynolds number, Re_D=U_∽D/v=100, as an example of complex flows. The reconstructions match the ground truth (CFD) both qualitatively and quantitatively. The decomposed flow fields have different characteristics. The 1st to 3rd modes represent similar-size structures corresponding to the Karman vortex. The 1st mode has aperiodic motion similar to the ground truth. The 2nd and 3rd modes have motion with slight periodicity. The 4th mode represents steady structures similar to the time-averaged flow fields of the ground truth.