There are different strategies for training neural networks (NNs) as subgrid‐scale parameterizations. Here, we use a 1D model of the quasi‐biennial oscillation (QBO) and gravity wave (GW) parameterizations as testbeds. A 12‐layer convolutional NN that predicts GW forcings for given wind profiles, when trained offline in a
Atmospheric predictability from subseasonal to seasonal time scales and climate variability are both influenced critically by gravity waves (GW). The quality of regional and global numerical models relies on thorough understanding of GW dynamics and its interplay with chemistry, precipitation, clouds, and climate across many scales. For the foreseeable future, GWs and many other relevant processes will remain partly unresolved, and models will continue to rely on parameterizations. Recent model intercomparisons and studies show that present-day GW parameterizations do not accurately represent GW processes. These shortcomings introduce uncertainties, among others, in predicting the effects of climate change on important modes of variability. However, the last decade has produced new data and advances in theoretical and numerical developments that promise to improve the situation. This review gives a survey of these developments, discusses the present status of GW parameterizations, and formulates recommendations on how to proceed from there.
more » « less- PAR ID:
- 10487231
- Publisher / Repository:
- American Meteorological Society
- Date Published:
- Journal Name:
- Journal of the Atmospheric Sciences
- Volume:
- 81
- Issue:
- 2
- ISSN:
- 0022-4928
- Format(s):
- Medium: X Size: p. 237-262
- Size(s):
- p. 237-262
- Sponsoring Org:
- National Science Foundation
More Like this
-
more » « less
Abstract big ‐data regime (100‐year), produces realistic QBOs once coupled to the 1D model. In contrast, offline training of this NN in asmall ‐data regime (18‐month) yields unrealistic QBOs. However, online re‐training of just two layers of this NN using ensemble Kalman inversion and only time‐averaged QBO statistics leads to parameterizations that yield realistic QBOs. Fourier analysis of these three NNs' kernels suggests why/how re‐training works and reveals that these NNs primarily learn low‐pass, high‐pass, and a combination of band‐pass filters, potentially related to the local and non‐local dynamics in GW propagation and dissipation. These findings/strategies generally apply to data‐driven parameterizations of other climate processes. -
more » « less
Abstract Aquaplanet experiments are important tools for understanding and improving physical processes simulated by global models; yet, previous aquaplanet experiments largely differ in their representation of subseasonal tropical rainfall variability. This study presents results from aquaplanet experiments produced with the Model for Prediction Across Scales‐Atmosphere (MPAS‐A)—a community model specifically designed to study weather and climate in a common framework. The mean climate and tropical rainfall variability simulated by MPAS‐A with varying horizontal resolution were compared against results from a recent suite of aquaplanet experiments. This comparison shows that, regardless of horizontal resolution, MPAS‐A produces the expected mean climate of an aquaplanet framework with zonally symmetric but meridionally varying sea‐surface temperature. MPAS‐A, however, has a stronger signal of tropical rainfall variability driven by convectively coupled equatorial waves. Sensitivity experiments with different cumulus parameterizations, physics packages, and vertical grids consistently show the presence of those waves, especially equatorial Kelvin waves, in phase with lower‐tropospheric convergence. Other models do not capture such rainfall‐kinematics phasing. These results suggest that simulated tropical rainfall variability depends not only on the cumulus parameterization (as suggested by previous studies) but also on the coupling between physics and dynamics of climate and weather prediction models.
-
more » « less
Abstract Stochastic parameterizations account for uncertainty in the representation of unresolved subgrid processes by sampling from the distribution of possible subgrid forcings. Some existing stochastic parameterizations utilize data‐driven approaches to characterize uncertainty, but these approaches require significant structural assumptions that can limit their scalability. Machine learning models, including neural networks, are able to represent a wide range of distributions and build optimized mappings between a large number of inputs and subgrid forcings. Recent research on machine learning parameterizations has focused only on deterministic parameterizations. In this study, we develop a stochastic parameterization using the generative adversarial network (GAN) machine learning framework. The GAN stochastic parameterization is trained and evaluated on output from the Lorenz '96 model, which is a common baseline model for evaluating both parameterization and data assimilation techniques. We evaluate different ways of characterizing the input noise for the model and perform model runs with the GAN parameterization at weather and climate time scales. Some of the GAN configurations perform better than a baseline bespoke parameterization at both time scales, and the networks closely reproduce the spatiotemporal correlations and regimes of the Lorenz '96 system. We also find that, in general, those models which produce skillful forecasts are also associated with the best climate simulations.
-
more » « less
Abstract Neural networks (NNs) are increasingly used for data‐driven subgrid‐scale parameterizations in weather and climate models. While NNs are powerful tools for learning complex non‐linear relationships from data, there are several challenges in using them for parameterizations. Three of these challenges are (a) data imbalance related to learning rare, often large‐amplitude, samples; (b) uncertainty quantification (UQ) of the predictions to provide an accuracy indicator; and (c) generalization to other climates, for example, those with different radiative forcings. Here, we examine the performance of methods for addressing these challenges using NN‐based emulators of the Whole Atmosphere Community Climate Model (WACCM) physics‐based gravity wave (GW) parameterizations as a test case. WACCM has complex, state‐of‐the‐art parameterizations for orography‐, convection‐, and front‐driven GWs. Convection‐ and orography‐driven GWs have significant data imbalance due to the absence of convection or orography in most grid points. We address data imbalance using resampling and/or weighted loss functions, enabling the successful emulation of parameterizations for all three sources. We demonstrate that three UQ methods (Bayesian NNs, variational auto‐encoders, and dropouts) provide ensemble spreads that correspond to accuracy during testing, offering criteria for identifying when an NN gives inaccurate predictions. Finally, we show that the accuracy of these NNs decreases for a warmer climate (4 × CO2). However, their performance is significantly improved by applying transfer learning, for example, re‐training only one layer using ∼1% new data from the warmer climate. The findings of this study offer insights for developing reliable and generalizable data‐driven parameterizations for various processes, including (but not limited to) GWs.
-
more » « less
Abstract Four state-of-the-science numerical weather prediction (NWP) models were used to perform mountain wave (MW)-resolving hindcasts over the Drake Passage of a 10-day period in 2010 with numerous observed MW cases. The Integrated Forecast System (IFS) and the Icosahedral Nonhydrostatic (ICON) model were run at Δ
x ≈ 9 and 13 km globally. The Weather Research and Forecasting (WRF) Model and the Met Office Unified Model (UM) were both configured with a Δx = 3-km regional domain. All domains had tops near 1 Pa (z ≈ 80 km). These deep domains allowedquantitative validation against Atmospheric Infrared Sounder (AIRS) observations, accounting for observation time, viewing geometry, and radiative transfer. All models reproduced observed middle-atmosphere MWs with remarkable skill. Increased horizontal resolution improved validations. Still, all models underrepresented observed MW amplitudes, even after accounting for model effective resolution and instrument noise, suggesting even at Δx ≈ 3-km resolution, small-scale MWs are underresolved and/or overdiffused. MW drag parameterizations are still necessary in NWP models at current operational resolutions of Δx ≈ 10 km. Upper GW sponge layers in the operationally configured models significantly, artificially reduced MW amplitudes in the upper stratosphere and mesosphere. In the IFS, parameterized GW drags partly compensated this deficiency, but still, total drags were ≈6 times smaller than that resolved at Δx ≈ 3 km. Meridionally propagating MWs significantly enhance zonal drag over the Drake Passage. Interestingly, drag associated with meridional fluxes of zonal momentum (i.e.,) were important; not accounting for these terms results in a drag in the wrong direction at and below the polar night jet. Significance Statement This study had three purposes: to quantitatively evaluate how well four state-of-the-science weather models could reproduce observed mountain waves (MWs) in the middle atmosphere, to compare the simulated MWs within the models, and to quantitatively evaluate two MW parameterizations in a widely used climate model. These models reproduced observed MWs with remarkable skill. Still, MW parameterizations are necessary in current Δ
x ≈ 10-km resolution global weather models. Even Δx ≈ 3-km resolution does not appear to be high enough to represent all momentum-fluxing MW scales. Meridionally propagating MWs can significantly influence zonal winds over the Drake Passage. Parameterizations that handle horizontal propagation may need to consider horizontal fluxes of horizontal momentum in order to get the direction of their forcing correct.
