Master Thesis [Doc]
In this work, we design a new neural network, where the global information of 3D data and the implicit information of filter are extracted. It then restores the small scales of a flow by combining the current large eddy simulation field with the aforementioned information. For evaluating the performance of our model, we convert the well-performed models on 2D image reconstruction task to 3D models as our baseline models.
We train our model on two kinds of filtered data, singularly filtered data and variously filtered data. For singularly filtered data set, there is only one kind of large eddy simulation field filtered from the direct numerical simulation field with a Gaussian filter kernel. For variously filtered data set, the large eddy simulation fields are filtered from a direct numerical simulation field with different Gaussian filter kernels. Our model outperforms the baseline models for both training sets and obtains significantly better results for different metrics.