A new simulation framework has been introduced for understanding turbulent flows in fluid dynamics. Turbulence is a common phenomenon in both natural and engineered fluid flows and has many applications in various research areas. However, our current understanding and ability to predict these flows accurately are limited due to their chaotic nature and the vast range of scales they occupy. To overcome these limitations, the direct numerical simulation (DNS) method can be used, which provides a detailed representation of the flow-field without any approximations. However, DNS requires significant computational resources, making it challenging for large-scale simulations.
The new simulation framework, called “A TensorFlow simulation framework for scientific computing of fluid flows on tensor processing units,” allows for detailed large-scale simulations of turbulent flows using TPUs. By leveraging the advances in TensorFlow software and TPU hardware architecture, this framework enables researchers to push the boundaries of scientific discovery and turbulence analysis. It demonstrates excellent scalability and efficiency, outperforming other distributed computation frameworks. The software is open-source and available on GitHub.
The framework solves the variable-density Navier-Stokes equations using the TensorFlow framework on TPU architectures. It adopts a SIMD approach for parallelization and leverages TPU’s matrix multiply unit for finite difference operators. It takes advantage of the high-bandwidth interconnect between TPU accelerators and performs computations with single-precision floating-point arithmetic and optimized executable through the accelerated linear algebra compiler.
This framework is the first open-source computational fluid dynamics (CFD) framework that fully utilizes cloud accelerators like TPUs. It reduces the cost and turnaround time for large-scale CFD simulations and enables faster iteration in climate and weather research. The framework also allows for easy integration with machine learning methods, opening up new possibilities for exploring ML approaches in CFD problems.
To validate the framework, the researchers performed simulations of homogeneous isotropic turbulence (HIT), a well-studied flow benchmark. They achieved one of the largest turbulent flow simulations to date, with over eight billion degrees of freedom. The simulation results were in agreement with theoretical expectations, confirming the accuracy and reliability of the framework.
Overall, this new simulation framework provides a powerful tool for understanding turbulent flows in fluid dynamics. It allows for large-scale simulations with TPUs, offering improved scalability, efficiency, and cost-effectiveness. Researchers can now explore and analyze turbulent flows at unprecedented scale and contribute to our understanding of these complex phenomena.