Turbulence closure with small, local neural networks: Forced two-dimensional and $\beta$-plane flows

We parameterize sub-grid scale (SGS) fluxes in sinusoidally forced two-dimensional turbulence on the $\beta$-plane at high Reynolds numbers (Re$\sim$25000) using simple 2-layer Convolutional Neural Networks (CNN) having only O(1000)parameters, two orders of magnitude smaller than recent studies employing deeper CNNs with 8-10 layers; we obtain stable, accurate, and long-term online or a posteriori solutions at 16X downscaling factors. Our methodology significantly improves training efficiency and speed of online Large Eddy Simulations (LES) runs, while offering insights into the physics of closure in such turbulent flows. Our approach benefits from extensive hyperparameter searching in learning rate and weight decay coefficient space, as well as the use of cyclical learning rate annealing, which leads to more robust and accurate online solutions compared to fixed learning rates. Our CNNs use either the coarse velocity or the vorticity and strain fields as inputs, and output the two components of the deviatoric stress tensor. We minimize a loss between the SGS vorticity flux divergence (computed from the high-resolution solver) and that obtained from the CNN-modeled deviatoric stress tensor, without requiring energy or enstrophy preserving constraints. The success of shallow CNNs in accurately parameterizing this class of turbulent flows implies that the SGS stresses have a weak non-local dependence on coarse fields; it also aligns with our physical conception that small-scales are locally controlled by larger scales such as vortices and their strained filaments. Furthermore, 2-layer CNN-parameterizations are more likely to be interpretable and generalizable because of their intrinsic low dimensionality.

[1]  L. Zanna,et al.  Benchmarking of Machine Learning Ocean Subgrid Parameterizations in an Idealized Model , 2022, Journal of Advances in Modeling Earth Systems.

[2]  A. Chattopadhyay,et al.  Explaining the physics of transfer learning a data-driven subgrid-scale closure to a different turbulent flow , 2022, ArXiv.

[3]  R. Keisler Forecasting Global Weather with Graph Neural Networks , 2022, ArXiv.

[4]  Li-Wei Chen,et al.  Learned turbulence modelling with differentiable fluid solvers: physics-based loss functions and optimisation horizons , 2022, Journal of Fluid Mechanics.

[5]  T. Pfaff,et al.  Learned Coarse Models for Efficient Turbulence Simulation , 2021, ArXiv.

[6]  J. McWilliams,et al.  Stochastic rectification of fast oscillations on slow manifold closures , 2021, Proceedings of the National Academy of Sciences.

[7]  A. Chattopadhyay,et al.  Stable a posteriori LES of 2D turbulence using convolutional neural networks: Backscattering analysis and generalization to higher Re via transfer learning , 2021, J. Comput. Phys..

[8]  Stephan Hoyer,et al.  Machine learning–accelerated computational fluid dynamics , 2021, Proceedings of the National Academy of Sciences.

[9]  Jeremy McGibbon,et al.  Machine Learning Climate Model Dynamics: Offline versus Online Performance. , 2020, 2011.03081.

[10]  C. Hill,et al.  Use of Neural Networks for Stable, Accurate and Physically Consistent Parameterization of Subgrid Atmospheric Processes With Good Performance at Reduced Precision , 2020, Geophysical Research Letters.

[11]  Andre N. Souza,et al.  Uncertainty Quantification of Ocean Parameterizations: Application to the K‐Profile‐Parameterization for Penetrative Convection , 2020, Journal of Advances in Modeling Earth Systems.

[12]  Janni Yuval,et al.  Stable machine-learning parameterization of subgrid processes for climate modeling at a range of resolutions , 2020, Nature Communications.

[13]  G. He,et al.  Subgrid-scale model for large-eddy simulation of isotropic turbulent flows using an artificial neural network , 2019, Computers & Fluids.

[14]  J. McWilliams,et al.  Variational Approach to Closure of Nonlinear Dynamical Systems: Autonomous Case , 2019, Journal of Statistical Physics.

[15]  S. Rasp Coupled online learning as a way to tackle instabilities and biases in neural network parameterizations: general algorithms and Lorenz 96 case study (v1.0) , 2019, Geoscientific Model Development.

[16]  E Weinan,et al.  Model Reduction with Memory and the Machine Learning of Dynamical Systems , 2018, Communications in Computational Physics.

[17]  Prakash Vedula,et al.  Subgrid modelling for two-dimensional turbulence using neural networks , 2018, Journal of Fluid Mechanics.

[18]  Ding-Xuan Zhou,et al.  Universality of Deep Convolutional Neural Networks , 2018, Applied and Computational Harmonic Analysis.

[19]  Frank O. Bryan,et al.  Evaluation of scale-aware subgrid mesoscale eddy models in a global eddy-rich model , 2017 .

[20]  James C. McWilliams,et al.  The emergence of fast oscillations in a reduced primitive equation model and its implications for closure theories , 2017 .

[21]  Baylor Fox-Kemper,et al.  A scale-aware subgrid model for quasi-geostrophic turbulence: SUBGRID MODEL FOR QG TURBULENCE , 2017 .

[22]  K. Duraisamy,et al.  Non-Markovian Closure Models for Large Eddy Simulations using the Mori-Zwanzig Formalism , 2016, 1611.03311.

[23]  I. Held,et al.  Parameterizing subgrid-scale eddy effects using energetically consistent backscatter , 2014 .

[24]  P. Ioannou,et al.  S3T stability of the homogeneous state of barotropic beta-plane turbulence , 2014, 1407.3354.

[25]  Nathan E. Glatt-Holtz,et al.  Invariant Measures for Dissipative Dynamical Systems: Abstract Results and Applications , 2011, 1110.4354.

[26]  Yoshua Bengio,et al.  On the Expressive Power of Deep Architectures , 2011, ALT.

[27]  Nsf A Theory of Hypoellipticity and Unique Ergodicity for Semilinear Stochastic PDEs , 2011 .

[28]  Dirk P. Kroese,et al.  Kernel density estimation via diffusion , 2010, 1011.2602.

[29]  G. Eyink,et al.  Localness of energy cascade in hydrodynamic turbulence. I. Smooth coarse graining , 2009, 0909.2386.

[30]  G. Eyink,et al.  Physical mechanism of the inverse energy cascade of two-dimensional turbulence: a numerical investigation , 2009, Journal of Fluid Mechanics.

[31]  Martin Hairer,et al.  A Theory of Hypoellipticity and Unique Ergodicity for Semilinear Stochastic PDEs , 2008, 0808.1361.

[32]  P. Ioannou,et al.  Structure and Spacing of Jets in Barotropic Turbulence , 2007 .

[33]  T. Schneider,et al.  Statistics of an Unstable Barotropic Jet from a Cumulant Expansion , 2007, 0705.0011.

[34]  A. Chorin,et al.  Stochastic Tools in Mathematics and Science , 2005 .

[35]  G. Eyink Locality of turbulent cascades , 2005 .

[36]  A. Stuart,et al.  Extracting macroscopic dynamics: model problems and algorithms , 2004 .

[37]  Jonathan C. Mattingly,et al.  Ergodicity of the 2D Navier-Stokes equations with degenerate stochastic forcing , 2004, math/0406087.

[38]  W. Large,et al.  Oceanic vertical mixing: a review and a model with a nonlocal boundary layer parameterization , 1994 .

[39]  S. Lele Compact finite difference schemes with spectral-like resolution , 1992 .

[40]  Claude Basdevant,et al.  Nonlinear galerkin method and subgrid-scale model for two-dimensional turbulent flows , 1992 .

[41]  Y. Pomeau,et al.  Rates, pathways, and end states of nonlinear evolution in decaying two‐dimensional turbulence: Scaling theory versus selective decay , 1992 .

[42]  P. Moin,et al.  Subgrid-scale backscatter in turbulent and transitional flows , 1991 .

[43]  McWilliams,et al.  Evolution of vortex statistics in two-dimensional turbulence. , 1991, Physical review letters.

[44]  R. Temam,et al.  Approximate inertial manifolds and effective viscosity in turbulent flows , 1991 .

[45]  Kurt Hornik,et al.  Multilayer feedforward networks are universal approximators , 1989, Neural Networks.

[46]  Marc Brachet,et al.  The dynamics of freely decaying two-dimensional turbulence , 1988, Journal of Fluid Mechanics.

[47]  James C. McWilliams,et al.  The emergence of isolated coherent vortices in turbulent flow , 1984, Journal of Fluid Mechanics.

[48]  R. Kraichnan Inertial-range transfer in two- and three-dimensional turbulence , 1971, Journal of Fluid Mechanics.

[49]  L. Zanna,et al.  Data‐Driven Equation Discovery of Ocean Mesoscale Closures , 2020, Geophysical Research Letters.

[50]  Nitish Srivastava,et al.  Dropout: a simple way to prevent neural networks from overfitting , 2014, J. Mach. Learn. Res..

[51]  M Israeli,et al.  Numerical Simulation of Viscous Incompressible Flows , 1974 .