A Posteriori Learning for Quasi‐Geostrophic Turbulence Parametrization

Frezat, Hugo; Sommer, Julien Le; Fablet, Ronan; Balarac, Guillaume; Lguensat, Redouane

doi:10.1029/2022ms003124

Cited by 31 publications

(57 citation statements)

References 104 publications

(166 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Our open‐source framework (Appendix ) will hopefully encourage the research community to find easy‐to‐use resources for such evaluation and facilitate the development of new parameterizations that more faithfully capture the effects of subgrid‐scale processes. Data design choice: the filtering and coarse‐graining operator is key, consistent with Zanna and Bolton (2021) and Frezat et al. (2022). The online results for a given FCNN architecture are highly sensitive to filtering choice; here the best performance was obtained with a filtering that most closely follow the numerics of the model.…”

Section: Discussionsupporting

confidence: 55%

“…Data design choice: the filtering and coarse‐graining operator is key, consistent with Zanna and Bolton (2021) and Frezat et al. (2022). The online results for a given FCNN architecture are highly sensitive to filtering choice; here the best performance was obtained with a filtering that most closely follow the numerics of the model.…”

Section: Discussionsupporting

confidence: 55%

“…There are many possible directions we did not explore for NN optimization, including online learning (Dresdner et al., 2022; Frezat et al., 2022; Kochkov et al., 2021; Sirignano et al., 2020; Um et al., 2020), or training on multiple datasets (Bolton & Zanna, 2019; O’Gorman & Dwyer, 2018). New approaches to remain more faithful to the physics of the problem that could be explored as well which include non‐dimensionalizing input variables (Beucler et al., 2021), modeling subgrid‐scale organization (Shamekh et al., 2022), or finding a better latent space for our input (and eliminating spurious correlation with causal inference).…”

Section: Discussionmentioning

confidence: 99%

“…The formulation of the subgrid forcing term, either in terms of tendency or subgrid‐scale fluxes, can affect the stability of parameterized models (Yuval et al., 2021). Different filtering schemes also have significant effects on the online performance of subgrid parameterizations (Frezat et al., 2022; Piomelli et al., 1988; Zhou et al., 2019).…”

Section: Introductionmentioning

confidence: 99%

“…Recently, an increase in high‐resolution observations and simulations combined with advances in machine‐learning (ML) methods has propelled a surge in the development of data‐driven SGS parameterizations in climate models (Beucler et al., 2021; Bolton & Zanna, 2019; Frezat et al., 2022; Guan et al., 2022; Guillaumin & Zanna, 2021; Krasnopolsky et al., 2010; O’Gorman & Dwyer, 2018; Rasp et al., 2018; Subel et al., 2022; Yuval & O’Gorman, 2021; Zanna & Bolton, 2021). Directly learning from data, ML methods automatically extract relevant information from observations and high‐resolution simulations to improve coarse‐resolution models at a reduced computational cost.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Benchmarking of Machine Learning Ocean Subgrid Parameterizations in an Idealized Model

Ross

Perezhogin

Fernandez‐Granda

et al. 2023

J Adv Model Earth Syst

View full text Add to dashboard Cite

Current state-of-the-art climate models solve geophysical fluid equations on horizontal grids of size 25 km and coarser. Models at this resolution are not able to accurately and sufficiently resolve processes with physical length scales smaller than the model grid, for example, convection in the atmosphere and mesoscale eddies in the ocean. Since increases in computational power will likely not enable climate models to resolve these processes before the effects of climate change ensue (Fox-Kemper et al., 2014;Schneider et al., 2017), we must represent subgrid-scale (SGS) processes with closure models, also known as parameterizations. Yet, these SGS models are some of the largest sources of bias and uncertainties in climate simulations: for example, insufficient representations of transient eddies cause biases in modeled currents and sea surface temperature in the ocean (Griffies et al., 2015;Hewitt et al., 2020), and the precipitation pattern is strongly sensitive to the different subgrid cloud closures, thereby causing significant errors in climate projections (Stevens & Bony, 2013). Therefore, developing robust parameterizations remains an important task toward reliable climate projections.

show abstract

Section: Discussionsupporting

confidence: 55%

Section: Discussionsupporting

confidence: 55%

Section: Discussionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Benchmarking of Machine Learning Ocean Subgrid Parameterizations in an Idealized Model

Ross

Perezhogin

Fernandez‐Granda

et al. 2023

J Adv Model Earth Syst

View full text Add to dashboard Cite

show abstract

Symbiotic Ocean Modeling using Physics-Controlled Echo State Networks

Mulder

Baars

Wubs

et al. 2022

Preprint

View full text Add to dashboard Cite

One of the most important spatial scales in the ocean circulation is the internal Rossby radius of deformation L D ; it ranges from 50 to 100 km at mid-latitudes to a few km in the polar regions (Hallberg, 2013). At this scale, perturbations are amplified on mean flows through mixed barotropic/baroclinic instability, giving rise to ocean eddies. Interactions between these eddies and the mean flow can lead to up-gradient momentum transport affecting the strength and separation of ocean western boundary currents such as the Kuroshio and Agulhas (Chassignet et al., 2020).Most climate models, in particularly those used in CMIP5 and CMIP6, do not resolve ocean processes at the scale L D as the spatial grid size used is too large; typically 1° (Eyring et al., 2016). The main reason is computational costs, as doubling the horizontal resolution increases these costs roughly by a factor 10. Effects of subgrid-scale processes are hence parameterized in these models. For example, the effect of ocean eddies on tracer transport is represented by the Gent- McWilliams (Gent et al., 1995) scheme, but such a scheme cannot capture, for example, the up-gradient momentum transport. Hence, western boundary flows are too weak and diffuse, and do not separate at the correct location (Chassignet et al., 2020).Over the last few years, first simulations have been performed with global climate models, where the ocean model component has a resolution of 0.1°, which is smaller than L D for many locations on the globe (Chang Abstract We introduce a "symbiotic" ocean modeling strategy that leverages data-driven and machine learning methods to allow high-and low-resolution dynamical models to mutually benefit from each other. In this work we mainly focus on how a low-resolution model can be enhanced within a symbiotic model configuration. The broader aim is to enhance the representation of unresolved processes in low-resolution models, while simultaneously improving the efficiency of high-resolution models. To achieve this, we use a grid-switching approach together with hybrid modeling techniques that combine linear regression-based methods with nonlinear echo state networks. The approach is applied to both the Kuramoto-Sivashinsky equation and a single-layer quasi-geostrophic ocean model, and shown to simulate short-term and long-term behavior better than either purely data-based methods or low-resolution models. By maintaining key flow characteristics, the hybrid modeling techniques are also able to provide higher quality initial conditions for high-resolution models, thereby improving their efficiency.Plain Language Summary Models of the ocean vary in complexity. Some are very detailed and manage to show oceanic vortices, whereas others are very efficient but coarse, and unable to compute such vortices. The idea in this paper is to let these different model types work together and benefit from each other, as if in a symbiosis. With knowledge of differences between the detailed and coarse model we can use machine learning techniques to improv...

show abstract

Online model error correction with neural networks in the incremental 4D-Var framework

Farchi

Chrust

Bocquet

et al. 2022

Preprint

View full text Add to dashboard Cite

Recent studies have demonstrated that it is possible to combine machine learning with data assimilation to reconstruct the dynamics of a physical model partially and imperfectly observed. Data assimilation is used to estimate the system state from the observations, while machine learning computes a surrogate model of the dynamical system based on those estimated states. The surrogate model can be defined as an hybrid combination where a physical model based on prior knowledge is enhanced with a statistical model estimated by a neural network. The training of the neural network is typically done offline, once a large enough dataset of model state estimates is available. By contrast, with online approaches the surrogate model is improved each time a new system state estimate is computed. Online approaches naturally fit the sequential framework encountered in geosciences where new observations become available with time. In a recent methodology paper, we have developed a new weak-constraint 4D-Var formulation which can be used to train a neural network for online model error correction. In the present article, we develop a simplified version of that method, in the incremental 4D-Var framework adopted by most operational weather centres. The simplified method is implemented in the ECMWF Object-Oriented Prediction System, with the help of a newly developed Fortran neural network library, and tested with a two-layer two-dimensional quasi geostrophic model. The results confirm that online learning is effective and yields a more accurate model error correction than offline learning. Finally, the simplified method is compatible with future applications to state-of-the-art models such as the ECMWF Integrated Forecasting System.

show abstract

A Posteriori Learning for Quasi‐Geostrophic Turbulence Parametrization

Cited by 31 publications

References 104 publications

Benchmarking of Machine Learning Ocean Subgrid Parameterizations in an Idealized Model

Benchmarking of Machine Learning Ocean Subgrid Parameterizations in an Idealized Model

Symbiotic Ocean Modeling using Physics-Controlled Echo State Networks

Online model error correction with neural networks in the incremental 4D-Var framework

Contact Info

Product

Resources

About