Machine learning-accelerated computational fluid dynamics.

Area of Science:

• Computational fluid dynamics
• Turbulence modeling
• Scientific machine learning

Background:

• Numerical simulation of fluids is crucial for modeling phenomena like weather and aerodynamics.
• Solving Navier-Stokes equations is computationally expensive, limiting accuracy and tractability.
• Current methods face trade-offs between simulation accuracy and computational cost.

Purpose of the Study:

• To enhance fluid dynamics approximations using end-to-end deep learning.
• To improve the modeling of two-dimensional turbulent flows.
• To achieve faster and more accurate fluid simulations.

Main Methods:

• Implemented an end-to-end deep learning framework within computational fluid dynamics.
• Applied the method to direct numerical simulation and large-eddy simulation of turbulence.
• Validated the approach on two-dimensional turbulent flow models.

Main Results:

• Achieved accuracy comparable to baseline solvers with 8-10x finer resolution.
• Realized computational speedups of 40- to 80-fold.
• Demonstrated stability during long simulations and generalization to unseen parameters.

Conclusions:

• Deep learning significantly improves the efficiency and accuracy of fluid dynamics simulations.
• The developed method offers substantial computational speedups without sacrificing accuracy.
• This approach showcases the potential of integrating machine learning into scientific computing for enhanced simulations.