Dmitrii Kochkov,Jamie A. Smith,Ayya Alieva,Qing Wang,Michael P. Brenner,Stephan Hoyer
Abstract:Numerical simulation of fluids plays an essential role in modeling many physical phenomena, such as weather, climate, aerodynamics and plasma physics. Fluids are well described by the Navier-Stokes equations, but solving these equations at scale remains daunting, limited by the computational cost of resolving the smallest spatiotemporal features. This leads to unfavorable trade-offs between accuracy and tractability. Here we use end-to-end deep learning to improve approximations inside computational fluid dynamics for modeling two-dimensional turbulent flows. For both direct numerical simulation of turbulence and large eddy simulation, our results are as accurate as baseline solvers with 8-10x finer resolution in each spatial dimension, resulting in 40-80x fold computational speedups. Our method remains stable during long simulations, and generalizes to forcing functions and Reynolds numbers outside of the flows where it is trained, in contrast to black box machine learning approaches. Our approach exemplifies how scientific computing can leverage machine learning and hardware accelerators to improve simulations without sacrificing accuracy or generalization.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is that in hydrodynamic calculations, when dealing with complex physical phenomena (such as weather, climate, aerodynamics, and plasma physics, etc.), direct numerical simulation (DNS) and large - eddy simulation (LES) require the resolution of the smallest spatio - temporal features, resulting in excessively high computational costs, and thus there is an unfavorable trade - off between accuracy and computational feasibility. Specifically, traditional computational fluid dynamics (CFD) methods, such as the finite - difference method, the finite - volume method, the finite - element method, and the pseudo - spectral method, when resolving complex flows such as turbulence, must use very fine grids to ensure convergence, which causes the computational cost to increase with the cube of the Reynolds number. Therefore, direct numerical simulation for large - scale problems such as climate and weather becomes impossible to achieve.
To solve this problem, the paper proposes an end - to - end deep - learning method for improving the modeling of two - dimensional turbulent flows in computational fluid dynamics. This method can use a resolution that is 8 - 10 times coarser in each spatial dimension while maintaining the same accuracy as the baseline solver, thereby achieving a 40 - 80 - fold computational acceleration. In addition, this method remains stable in long - term simulations and can be generalized to forcing functions and Reynolds numbers outside the training data, overcoming the limitations of black - box machine - learning methods.
By introducing machine - learning - based interpolation and correction techniques to replace the parts in the traditional solver that are most affected by resolution loss, the paper accurately solves nonlinear partial differential equations on coarser grids. This method not only improves computational efficiency but also maintains the accuracy and generalization ability of the model, providing new ideas for how scientific computing can use machine learning and hardware accelerators to improve simulations.