Search

We address the question of how to use a machine learned (ML) parameterization in a general circulation model (GCM), and assess its performance both computationally and physically. We take one particular ML parameterization (Guillaumin & Zanna, 2021, https://doi.org/10.1002/essoar.10506419.1) and evaluate the online performance in a different model from which it was previously tested. This parameterization is a deep convolutional network that predicts parameters for a stochastic model of subgrid momentum forcing by mesoscale eddies. We treat the parameterization as we would a conventional parameterization once implemented in the numerical model. This includes trying the parameterization in a different flow regime from that in which it was trained, at different spatial resolutions, and with other differences, all to test generalization. We assess whether tuning is possible, which is a common practice in GCM development. We find the parameterization, without modification or special treatment, to be stable and that the action of the parameterization to be diminishing as spatial resolution is refined. We also find some limitations of the machine learning model in implementation: (a) tuning of the outputs from the parameterization at various depths is necessary; (b) the forcing near boundaries is not predicted as well as in the open ocean; (c) the cost of the parameterization is prohibitively high on central processing units. We discuss these limitations, present some solutions to problems, and conclude that this particular ML parameterization does inject energy, and improve backscatter, as intended but it might need further refinement before we can use it in production mode in contemporary climate models. Plain Language Summary: This paper discusses how machine learning can be used to make climate models more accurate. Specifically, we import an existing machine learning model that predicts how small eddies (in the order of 10–100 km) in the ocean affect larger currents. We test this machine learning model in a different ocean circulation model than the one it was originally designed for, and found that it worked well. However, we also found some limitations: the model works differently at different depths in the ocean, and it does not work as well near the coasts of the ocean. We also found that the model takes a long time to run on normal computers. Overall, we concluded that the model is promising, but more work is needed to make it work well in realistic situations. Key Points: A stochastic‐deep learning model is implemented in an ocean circulation model, MOM6We evaluate the online performance of the stochastic‐deep learning model as a subgrid parameterizationWe identify certain limitations of the machine learned parameterization which otherwise has the potential to improve specific metrics [ABSTRACT FROM AUTHOR]

We develop a doubly periodic version of the Simple Convection‐Permitting E3SM Atmosphere Model (SCREAM) to provide an efficient configuration for this global convection permitting model (GCPM), akin to a single column model often found in conventional general circulation models. The design details are explained, in addition to the extensive case library associated with the doubly periodic SCREAM (DP‐SCREAM) configuration. We demonstrate that doubly periodic cloud resolving models are useful tools to explore the horizontal resolution sensitivity of GCPMs, in addition to replicating biases seen in the global models. Using DP‐SCREAM, we show that SCREAM exhibits behaviors of a scale aware model as it is able to naturally partition between sub‐grid scale (SGS) and resolved vertical transport across the gray zone of turbulence. We show that SCREAM is reasonably scale insensitive when run at resolutions from 1 to 5 km, but can exhibit sensitivity, particularly for the shallow convective regime, when run at resolutions approaching that of large eddy simulations. We conclude that SGS parameterization improvements are likely needed to reduce this scale sensitivity. Plain Language Summary: Advances in computational resources have allowed climate simulations to be performed with very high resolution, which provides higher quality results. However, these simulations require a lot of time and computer resources to perform, which makes these models difficult to use for the common scientist. In this paper we develop a high‐resolution configuration which focuses on a specific point on the globe, enabling it to run fast and to use minimal computational resources. This allows users and developers to gauge how the model may perform before doing a computationally intensive global simulation. We show that this faster configuration is a useful tool to replicate problems that are found in the global model and is a valuable way to assess sensitivities of the model, particularly pertaining to choices made in its resolution. Key Points: Doubly periodic configurations represent an efficient framework to supplement global cloud resolving modelsShallow convection is sensitive to horizontal resolution in Simple Convection‐Permitting E3SM Atmosphere Model (SCREAM) but stratiform clouds are more robustSCREAM exhibits the behaviors of a scale aware model and can represent a range of cloud regimes [ABSTRACT FROM AUTHOR]

This paper presents efficient and practical implementations of sequential data assimilation methods for the Simplified Parameterizations, primitive‐Equation DYnamics (SPEEDY) Model, a well‐known numerical model, into the data assimilation community for climate prediction. In the SPEEDY model, the time evolution of dynamics is performed via the second‐order Leapfrog integration scheme; this time integrator relies on two steps: the position and the velocity. The computational implementation of SPEEDY blends the time integrator and the spatial discretization of dynamics to accelerate algebraic computations. Thus, there is no access to the right‐hand side function of the ordinary differential equations governing the time evolution of model dynamics. Consequently, the SPEEDY model is often treated as a black box wherein positions and velocities work as inputs and outputs. Since observations in operational data assimilation only match position states, we can exploit augmented vector states to propagate analysis innovations from positions to velocities. For this purpose, we formulate three variants of ensemble‐based filters and perform numerical experiments to assess their accuracies. We consider two scenarios for the experiments: an ideal case wherein positions and velocities can be observed and a more realistic one wherein measurements are only accessible for position states. Besides, we discuss the effects of the ensemble size on the accuracies of our formulations and, even more, the typical case in which velocities are not updated across assimilation steps. The results reveal that all filter formulations' accuracies remain the same in terms of Root‐Mean‐Square‐Error by neglecting observations from velocities (a realistic scenario) even for cases wherein the number of measurements decreases to 6% of model components. Furthermore, for all discussed filter implementations, the propagation of analysis increments from position to velocities improves up to 100% the performance of filter implementations wherein velocities are not updated, a typical operational scenario. [ABSTRACT FROM AUTHOR]

Data‐driven parameterizations offer considerable potential for improving the fidelity of General Circulation Models. However, ensuring that these remain consistent with the governing equations while still producing stable simulations remains a challenge. In this paper, we propose a combined Variational‐Multiscale (VMS) Artificial Neural Network (ANN) discretization which makes no a priori assumptions on the model form, and is only restricted in its accuracy by the precision of the ANN. Using a simplified problem, we demonstrate that good predictions of the required closure terms can be obtained with relatively compact ANN architectures. We then turn our attention to the stability of the VMS‐ANN discretization in the context of a single implicit time step. It is demonstrated that the ANN parameterization introduces nonphysical solutions to the governing equations that can significantly affect or prevent convergence. We show that enriching the training data with nonphysical states from intra‐time step iterations is an effective remedy. This indicates that the lack of representative ANN‐induced errors in our original, exact training inputs underpin the observed instabilities. In turn, this suggests that data set enrichment might aid in resolving instabilities that develop over several time steps. Plain Language Summary: Computer models of weather and climate have coarse resolutions, which prevent them from accurately predicting the effect of small atmospheric motions, for instance in low‐lying clouds, on the global climate. Recent studies indicate that improvements might be had by using models for the small motions that are informed purely by data, for example Artificial Neural Network (ANN) models. However, capitalizing on this potential is challenging in practice, since ANNs can introduce instabilities to the numerical model for larger‐scale motions. In this study, we zero in on these instabilities by introducing a model framework in which prediction accuracy is only limited by an ANN's ability to predict the net influence of small atmospheric motions. We contextualize this framework by contrasting it to more commonly used approaches. Using a simple test case based on the motions underpinning the development of low clouds, we illustrate that standard training procedures do not prepare ANN models for their interaction with the rest of the numerical model, and that this can contribute to numerical instability. We demonstrate an effective remedy for a specific type of instability, and suggest how this technique might also be used to treat other types of instabilities prompted by ANN models. Key Points: Artificial Neural Networks (ANNs) can parameterize exact closure terms of a Variational Multiscale method without a priori model form assumptionsThe ANN parameterization outperforms state‐of‐the‐art schemes in offline testsInstabilities in the numerical model comprising the ANN parameterization are amendable through data set enrichment [ABSTRACT FROM AUTHOR]