Figure 1
From: A scalable convolutional neural network approach to fluid flow prediction in complex environments

Architectures of the deep learning model. (A) Architecture of Gated Residual U-net, which consists of one encoder, one decoder, and multiple skip connector units. The encoder incorporates a convolution with stride 2 down-scaling, and the decoder employs a deconvolution layer with stride 2 for up-scaling. (B) (Left) Building block of Residual U-net, in which the 1D convolution enables residual learning. (Middle) Building block of Gated Residual U-net. (Right) Building block of Recurrent Residual U-net. (C) Nine different feature channels of input dataset. Inlet velocity is directed from left to right.