Abstract: In order to solve the difficulties of visualization and data storage caused by the huge volume of com-putational fluid dynamics data,an in-situ streamline extraction and velocity field compression method is pro-posed.Firstly,a uniform random method is used in situ to distribute and generate a certain number of streamlines in the computational fluid dynamics.Then,a deep neural regression network model StreamLNet is trained to build the mapping between this set of streamlines and velocity of any position in the flow flied.Finally,the origi-nal flow field data is discarded and only the streamlines and the StreamLNet model are stored for subsequent data visualization and flow field recovery.Experimental results on typical fluid simulation data show that proposed method can achieve a compression ratio of tens to hundreds of times for large-scale flow field at the cost of a relative error of 2%to 3%,in addition to streamline visualization of the filed at the same time.