A Research Of Deep Learning-based Fluid Animation Acceleration Methods

Computer fluid animation is widely used in production and life,such as 3D games,3D animation,movies,advertisements,etc.The grid-based Euler method is one of the physics-based simulation methods,but the large demand for computing resources limits its application.Deep learning based on multi-layer neural networks,which has emerged in recent years,has strong fitting ability and fast calculation speed,so it can be used to accelerate physics-based fluid simulation.Since projection step calculation occupies most of the computational resources of grid-based methods,this paper studies the use of deep learning to accelerate the projection step of Euler fluid animation.Existing methods that use deep learning to accelerate fluid animation have differences in training datasets,input and output,network types,etc.,and each method declares that it has achieved some good results,so it is not clear that how to configure the input and output,network type and other aspects of accelerated fluid animation deep learning is still unclear.This paper analyzes and experiments the loss function and the input feature vector,and finds that the loss function should constrain the pressure and velocity divergence,and optionally the velocity and pressure gradient.The input feature vector should include the geometric symbol,the previous frame pressure and speed or speed divergence.Based on the analysis of the network input,this paper jumps the geometric marker data into each layer of the network,which improves the effect of generating fluid in many ways.Based on the analysis of the projection step Poisson's equation,this paper introduces dilated convolution into the acceleration of fluid animation to make the output layer of the network have a larger receptive field,solving the problem that the smoke generated by the existing neural network extends too little to both sides.This paper finds that the spatial resolution of fluid scenes that can be handled by a fully convolutional network is limited,so the applicable resolution of the reference network and the network proposed in this paper is studied,and the characteristics of several networks at different spatial resolutions are found.