DeepCFD: Efficient Steady-State Laminar Flow Approximation with Deep Convolutional Neural Networks The "DeepCFD.py" script will run an example on how to train the DeepCFD model using a toy dataset, which is provided together with the code. The toy dataset has the following format: dataX.pkl: Input Geometry Information (SDF and flow region multi-class channel) dataY.pkl: Ground-truth steady-state CFD data (Ux, Uy, p) Both dataX and dataY have the same dimensions (Ns, Nc, Nx, Ny), in which the first axis is the number of samples (Ns), the second axis is the number of channels (Nc), and third and fourth axes are the number of elements in x and y (Nx and Ny). Regarding the input dataX, the first channel is the SDF calculated from the obstacle's surface, the second channel is the multi-label flow region channel, and the third channel is the SDF from the top/bottom surfaces (not used in the current implementation). For the output dataY file, the first channel is the Ux horizontal velocity component, the second channel is the Uy vertical velocity component, and the third channel is the pressure field p.