A TensorFlow-based new high-performance computational framework for CFD
- PDF / 3,918,050 Bytes
- 12 Pages / 595.22 x 842 pts (A4) Page_size
- 36 Downloads / 282 Views
Available online at https://link.springer.com/journal/42241 http://www.jhydrodynamics.com Journal of Hydrodynamics, 2020, 32(4): 735-746 https://doi.org/10.1007/s42241-020-0050-0
A TensorFlow-based new high-performance computational framework for CFD * Xi-zeng Zhao1, 2, Tian-yu Xu1, Zhou-teng Ye1, Wei-jie Liu1 1. Ocean College, Zhejiang University, Zhoushan 316021, China 2. The Engineering Research Center of Oceanic Sensing Technology and Equipment, Ministry of Education, Zhejiang University, Zhoushan 316021, China (Received March 14, 2019, Revised June 23, 2019, Accepted July 22, 2019, Published online August 26, 2020) ©China Ship Scientific Research Center 2020 Abstract: In this study, a computational framework in the field of artificial intelligence was applied in computational fluid dynamics (CFD) field. This Framework, which was initially proposed by Google AI department, is called “TensorFlow”. An improved CFD model based on this framework was developed with a high-order difference method, which is a constrained interpolation profile (CIP) scheme for the base flow solver of the advection term in the Navier-Stokes equations, and preconditioned conjugate gradient (PCG) method was implemented in the model to solve the Poisson equation. Some new features including the convolution, vectorization, and graphics processing unit (GPU) acceleration were implemented to raise the computational efficiency. The model was tested with several benchmark cases and shows good performance. Compared with our former CIP-based model, the present TensorFlow-based model also shows significantly higher computational efficiency in large-scale computation. The results indicate TensorFlow could be a promising framework for CFD models due to its ability in the computational acceleration and convenience for programming. Key words: TensorFlow, vectorization, Navier-Stokes equations, graphics processing unit (GPU) acceleration, constrained interpolation profile (CIP) method, preconditioned conjugate gradient (PCG) method
Introduction Computational fluid dynamics (CFD) models directly solving Navier-Stokes (N-S) equations are convenient tools to study problems involving water flows. It is well known that the computational efficiency is the main limitation of a CFD model when it comes to complex physical problems . Therefore, researchers have devoted their efforts to accelerating the computation of CFD models. Among various computational acceleration methods, graphics processing unit (GPU) acceleration is one of the useful methods. Despite the rapid development of central processing units (CPUs), GPUs usually have the advantage of higher computational efficiency especially when the grid number is very large. A number of researchers have successfully applied GPU acceleration for CFD model calculations[1-2]. However, in order to take advantage of GPUs, researchers are supposed to get knowledge of how to program based * Project supported by the National Natural Science Foundation of China (Grant No. 51679212, 51979245). Biography: Xi-zeng Zha
Data Loading...
