Implementation and Evaluation of a Machine Learned Mesoscale Eddy Parameterization Into a Numerical Ocean Circulation Model

Journal of Advances in Modeling Earth Systems, Год журнала: 2023, Номер 15(10)

Опубликована: Окт. 1, 2023

Abstract We address the question of how to use a machine learned (ML) parameterization in general circulation model (GCM), and assess its performance both computationally physically. take one particular ML (Guillaumin & Zanna, 2021, https://doi.org/10.1002/essoar.10506419.1 ) evaluate online different from which it was previously tested. This is deep convolutional network that predicts parameters for stochastic subgrid momentum forcing by mesoscale eddies. treat as we would conventional once implemented numerical model. includes trying flow regime trained, at spatial resolutions, with other differences, all test generalization. whether tuning possible, common practice GCM development. find parameterization, without modification or special treatment, be stable action diminishing resolution refined. also some limitations learning implementation: (a) outputs various depths necessary; (b) near boundaries not predicted well open ocean; (c) cost prohibitively high on central processing units. discuss these limitations, present solutions problems, conclude this does inject energy, improve backscatter, intended but might need further refinement before can production mode contemporary climate models.

Язык: Английский

---

Opinion: Optimizing climate models with process knowledge, resolution, and artificial intelligence

Atmospheric chemistry and physics, Год журнала: 2024, Номер 24(12), С. 7041 - 7062

Опубликована: Июнь 19, 2024

Abstract. Accelerated progress in climate modeling is urgently needed for proactive and effective change adaptation. The central challenge lies accurately representing processes that are small scale yet climatically important, such as turbulence cloud formation. These will not be explicitly resolvable the foreseeable future, necessitating use of parameterizations. We propose a balanced approach leverages strengths traditional process-based parameterizations contemporary artificial intelligence (AI)-based methods to model subgrid-scale processes. This strategy employs AI derive data-driven closure functions from both observational simulated data, integrated within encode system knowledge conservation laws. In addition, increasing resolution resolve larger fraction small-scale can aid toward improved interpretable predictions outside observed distribution. However, currently feasible horizontal resolutions limited O(10 km) because higher would impede creation ensembles calibration uncertainty quantification, sampling atmospheric oceanic internal variability, broadly exploring quantifying risks. By synergizing decades scientific development with advanced techniques, our aims significantly boost accuracy, interpretability, trustworthiness predictions.

Язык: Английский

---

Generative Data‐Driven Approaches for Stochastic Subgrid Parameterizations in an Idealized Ocean Model

Journal of Advances in Modeling Earth Systems, Год журнала: 2023, Номер 15(10)

Опубликована: Окт. 1, 2023

Abstract Subgrid parameterizations of mesoscale eddies continue to be in demand for climate simulations. These subgrid can powerfully designed using physics and/or data‐driven methods, with uncertainty quantification. For example, Guillaumin and Zanna (2021, https://doi.org/10.1029/2021ms002534 ) proposed a Machine Learning (ML) model that predicts forcing its local uncertainty. The major assumption potential drawback this is the statistical independence stochastic residuals between grid points. Here, we aim improve simulation generative models ML, such as Generative adversarial network (GAN) Variational autoencoder (VAE). learn distribution conditioned on resolved flow directly from data they produce new samples distribution. potentially capture not only spatial correlation but any statistically significant property forcing. We test offline online an idealized ocean model. show are able predict spatially correlated Online simulations range resolutions demonstrated superior baseline ML at coarsest resolution.

Язык: Английский

---

Turbulence Closure With Small, Local Neural Networks: Forced Two‐Dimensional and β‐Plane Flows

Journal of Advances in Modeling Earth Systems, Год журнала: 2024, Номер 16(4)

Опубликована: Апрель 1, 2024

Abstract We parameterize sub‐grid scale (SGS) fluxes in sinusoidally forced two‐dimensional turbulence on the β ‐plane at high Reynolds numbers (Re ∼25,000) 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 16× downscaling factors. Our methodology significantly improves training efficiency speed large eddy simulations runs, while offering insights into physics closure such turbulent flows. approach benefits from extensive hyperparameter searching learning rate weight decay coefficient space, as well use cyclical annealing, which leads to more robust accurate compared fixed rates. either coarse velocity vorticity strain fields inputs, output components deviatoric stress tensor, S d . minimize loss between SGS flux divergence (computed high‐resolution solver) that obtained CNN‐modeled , without requiring energy enstrophy preserving constraints. The success shallow accurately parameterizing this class flows implies stresses have weak non‐local dependence fields; it also aligns our physical conception small‐scales are locally controlled by larger scales vortices their strained filaments. Furthermore, CNN‐parameterizations likely be interpretable.

Язык: Английский

---

Cross‐Attractor Transforms: Improving Forecasts by Learning Optimal Maps Between Dynamical Systems and Imperfect Models

Geophysical Research Letters, Год журнала: 2025, Номер 52(4)

Опубликована: Фев. 17, 2025

Abstract Biased, incomplete numerical models are often used for forecasting states of complex dynamical systems by mapping an estimate a “true” initial state into model phase space, making forecast, and then back to the space. While advances have been made reduce errors associated with initialization forecasts, we lack general framework discovering optimal mappings between spaces. Here, propose using data‐driven approach infer these maps. Our consistently reduces in Lorenz‐96 system imperfect constructed produce significant compared reference configuration. Optimal pre‐ post‐processing transforms leverage “shocks” “drifts” make more skillful forecasts system. The implemented machine learning architecture neural networks custom analog‐adjoint layer makes generalizable across applications.

Язык: Английский

---

Machine Learning for Online Sea Ice Bias Correction Within Global Ice‐Ocean Simulations

Geophysical Research Letters, Год журнала: 2024, Номер 51(3)

Опубликована: Янв. 30, 2024

Abstract In this study, we perform online sea ice bias correction within a Geophysical Fluid Dynamics Laboratory global ice‐ocean model. For this, use convolutional neural network (CNN) which was developed in previous study (Gregory et al., 2023, https://doi.org/10.1029/2023ms003757 ) for the purpose of predicting concentration (SIC) data assimilation (DA) increments. An initial implementation CNN shows systematic improvements SIC biases relative to free‐running model, however large summertime errors remain. We show that these residual can be significantly improved with novel augmentation approach. This approach applies sequential and DA corrections new simulation over training period, then provides set refine weights network. propose machine‐learned scheme could utilized generating conditions, also real‐time seasonal‐to‐subseasonal forecasts.

Язык: Английский

---

The K‐Profile Parameterization Augmented by Deep Neural Networks (KPP_DNN) in the General Ocean Turbulence Model (GOTM)

Journal of Advances in Modeling Earth Systems, Год журнала: 2024, Номер 16(9)

Опубликована: Сен. 1, 2024

Abstract This study utilizes Deep Neural Networks (DNN) to improve the K‐Profile Parameterization (KPP) for vertical mixing effects in ocean's surface boundary layer turbulence. The deep neural networks were trained using 11‐year turbulence‐resolving solutions, obtained by running a large eddy simulation model Ocean Station Papa, predict turbulence velocity scale coefficient and unresolved shear KPP. DNN‐augmented KPP schemes (KPP_DNN) have been implemented General Turbulence Model (GOTM). KPP_DNN is stable long‐term integration more efficient than existing variants of with wave effects. Three different schemes, each differing their input output variables, developed trained. performance models utilizing compared those employing traditional deterministic first‐order second‐moment closure turbulent parameterizations. Solution comparisons indicate that simulated mixed becomes cooler deeper when are included parameterizations, aligning closer observations. In framework, shear, which used calculate ocean depth, has greater impact on magnitude diffusivity does. KPP_DNN, depends not only forcing, but also depth buoyancy forcing.

Язык: Английский

---

Impact of Improved Surface Flux Parameterization on Simulation of Radiation Fog Formation in the Yangtze River Delta, China

Journal of Geophysical Research Atmospheres, Год журнала: 2025, Номер 130(9)

Опубликована: Апрель 30, 2025

Abstract Meteorological conditions within the boundary layer play significant roles in radiation fog formation, which typically occur under stable conditions. The stratification surface are represented by stability parameter ( ζ ), calculated as ratio of reference height z to Monin‐Obukhov length L (i.e., = / ). Current schemes exhibit uncertainties strong > 1). Grachev2007 scheme for 1 and Li2014 Li2015 calculating implemented into Weather Research Forecasting model coupled with Chemistry (WRF‐Chem). Two successive events Yangtze River Delta simulated compare improved default scheme. Both high‐pressure characterized clear sky light wind during nighttime. results indicate that dominate before improves threat scores formation. Regarding flux, due reduced thermal resistance parameterization, increased heat exchange enhances cooling from sensible flux 1, is conducive turbulent mixing, dynamic drag reduces speed 1. This weakens contribution shear kinetic energy, ultimately promoting findings this paper applicable simulations other regions, such plain areas covered grassland, cropland, or vegetation, providing support improving simulation.

Язык: Английский

---

A Machine Learning‐Based Model Infers the Sea Surface Velocity of Surface Water and Ocean Topography (SWOT)

Geophysical Research Letters, Год журнала: 2025, Номер 52(9)

Опубликована: Май 5, 2025

Abstract High‐resolution sea surface velocity (SSV) is crucial for advancing our understanding of ocean sub‐mesoscale processes, energy cascades, etc. The recently launched Surface Water and Ocean Topography (SWOT) satellite measures height with a resolved resolution. Based on geostrophic balance, the so‐called in SWOT estimated. Although SWOT‐derived not true as it does consider separation balanced unbalanced motions, offers valuable insights into both ageostrophic velocities. Here we propose machine learning‐based model to infer SSV using drifter data. result demonstrates error between velocities from total are reduced by about 50%. Furthermore, kinetic inferred aligns more closely reanalysis data, particularly at low latitudes. This study thus presents promising approach inferring global

Язык: Английский

---

Addressing Out‐of‐Sample Issues in Multi‐Layer Convolutional Neural‐Network Parameterization of Mesoscale Eddies Applied Near Coastlines

Journal of Advances in Modeling Earth Systems, Год журнала: 2025, Номер 17(5)

Опубликована: Май 1, 2025

Abstract This study addresses the boundary artifacts in machine‐learned (ML) parameterizations for ocean subgrid mesoscale momentum forcing, as identified online ML implementation from a previous (Zhang et al., 2023, https://doi.org/10.1029/2023ms003697 ). We focus on condition (BC) treatment within existing convolutional neural network (CNN) models and aim to mitigate “out‐of‐sample” errors observed near complex coastal regions without developing new, architectures. Our approach leverages two established strategies placing BCs CNN models, namely zero replicate padding. Offline evaluations revealed that these padding significantly reduce root mean squared error (RMSE) by limiting dependence random initialization of weights restricting range out‐of‐sample predictions. Further suggest consistently reduces across various retrained models. In contrast, sometimes intensifies certain despite both performing similarly offline evaluations. underscores need BC treatments trained open water data when predicting near‐coastal forces parameterizations. The application padding, particular, offers robust strategy minimize propagation extreme values can contaminate computational or cause simulations fail. findings provide insights enhancing accuracy stability circulation with coastlines.

Язык: Английский

---