Fluid simulation utilizing 3D convolutional networks

In this work, we propose a data driven method of fluid simulation which approximates the classic Lagrangian method - position based fluids (PBF) [1]. Our method is based on subpixel convolutional networks which serve the solution to image and video superresolution problems. In out method we use subpixel convolutional layers to upsamle velocity correction field. Our final solution supports real-time interaction with a scene. Depending on the accuracy of approximation and ration between domain volume and the number of particles our method runs up to 200 times faster than supervisor method (PBF).