Research On CNN-based Improving Method Of Smoke Simulation's Resolution

Fluid simulation technology has always been a hot research topic in the field of simulation.After over 30 years of research and development,its theoretical basis and research methods have been fairly systematic,which can be divided into non-physical simulation and physical simulation.Existing fluid technology has been applied to simulate real,natural,detailed and delicate fluid,in which physical simulation is indispensable.But physical simulation is very time-consuming,so acceleration techniques or detail enhancement techniques(up-res techniques)for physical simulation emerged.At present,these technologies still have some problems,such as low fitness to fluid,long time-consuming or large storage-consuming,so fluid simulation has not made great progress in the real-time field.In order to avoid these problems,and considering that convolutional neural network(CNN)can fit the multi-dimensional structure of data and has a strong fitting ability,a new idea is proposed,which is converting low-resolution smoke into high-resolution smoke based on CNN.Then a novel CNN converter is proposed in this paper.The converter is completely formed by locally connected layers,this makes it possible to learn the small training data set efficiently.Its input is velocity and density fields of 2D low-resolution smoke in the current frame,and its output is density field of 2D high-resolution smoke.In order to train the converter,a two-dimensional smoke fluid data set is generated,which has only 800 pairs of high-and low-resolution smoke samples.In addition,three cost functions that are suitable for calculating the error of fluid density field are proposed to adjust the structure of the converter or improve its efficiency.After training,the high-resolution smoke generated by the converter is visually trustable.