Data-Driven Shape-keeping Fluid Simulation Method

With the rapid development of computer science,fluid simulation is wildly applied in many areas such as movies,cartoons,video games,virtual reality,etc.An efficient and accurate fluid simulation method can greatly improve the visual experience of the simulation and remarkably optimize the productivity of related works.However,most of the existing fluid simulation algorithm is not capable of achieving best simulation result and highest performance at the same time due to the restrictions of the hardware,resulting a trade-off between good result and high performance.Euler method based fluid simulation is a very important branch of computer graphics.The key features of Euler method are high precision and good convenience to reconstruct the fluid surface.However,the Euler method is based on grids,which has high compute complexity and very time consuming when applied directly without any optimization.Convolutional neural network(CNN)is a very popular architecture among all the machinelearning architectures.Every neuron of CNN receives and responds the states of neurons locally around,so that CNN with a proper model can achieve good result of fitting and classifying problems form complex data input.Therefore CNN is wildly used in many areas such as computer vision and picture processing.The fluid simulation method in this paper is based on two level grids and CNN,which combines the advantages of these two method and reduces the computation complexity without introducing much shape precision loss.The main concept of two level grids is to divide and conquer.We divide each coarse grid in Euler method into a grid domain of several fine grids,and compute the fluid state locally in each grid domain.This method can reduce the computation scale and allows using parallelization for acceleration.Furthermore,we integrate a CNN on fluid simulation to minimize the computation error introduced by the reduction of computation scale.The CNN learns the relationship between accurate result produced by Euler method and inaccurate result produced by two level grid method.Test result shows that our method can improve the computation performance while basically keeps the detail shape of the fluid.