StreamLNet:结合流线可视化的流场原位压缩方法

StreamLNet:In-Situ Compression of Flow Field Combined with Streamline Visualization

为了解决大规模计算流体力学数据量巨大引起的可视化和数据存储困难,提出一种原位流线提取和速度场压缩方法.在计算流体数值模拟的原位,首先使用均匀随机方法在流场中分布和生成一定数量的流线;然后基于该组流线训练一个深度神经回归网络模型StreamLNet,该模型以流场中任意位置的速度作为输出,以该位置周围若干流线节点的相对位置关系和速度属性作为输入,学习二者间的映射关系;最后舍弃原始流场数据,仅存储该组流线和StreamLNet模型用于后续数据可视化和流场恢复.对典型流体仿真数据的实验结果表明,所提方法在原位实现流线可视化的同时,可以以2%~3%的相对误差的代价获得几十倍到几百倍的压缩比.

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.