Machine Learning Global Simulation of Nonlocal Gravity Wave Propagation

TL;DR

The first-ever global simulation of atmospheric GW fluxes using machine learning (ML) models trained on the WINDSET dataset to emulate global GW emulation in the atmosphere, as an alternative to traditional single-column parameterizations is presented.

Abstract

Global climate models typically operate at a grid resolution of hundreds of kilometers and fail to resolve atmospheric mesoscale processes, e.g., clouds, precipitation, and gravity waves (GWs). Model representation of these processes and their sources is essential to the global circulation and planetary energy budget, but subgrid scale contributions from these processes are often only approximately represented in models using parameterizations. These parameterizations are subject to approximations and idealizations, which limit their capability and accuracy. The most drastic of these approximations is the "single-column approximation" which completely neglects the horizontal evolution of these processes, resulting in key biases in current climate models. With a focus on atmospheric GWs, we present the first-ever global simulation of atmospheric GW fluxes using machine learning (ML) models trained on the WINDSET dataset to emulate global GW emulation in the atmosphere, as an alternative to traditional single-column parameterizations. Using an Attention U-Net-based architecture trained on globally resolved GW momentum fluxes, we illustrate the importance and effectiveness of global nonlocality, when simulating GWs using data-driven schemes.

Figures (7)

• Figure 1: The three architectures used for global GW resolved momentum flux simulation. The three architectures, described in section \ref{['subsec:model_arch']} employ three different degrees of nonlocality. On one end, M1 uses single-column background data to predict the fluxes within that column. A timeslice is therefore a single vector of length 366. Intermediately, M2 uses background information in a 3$\times$3 stencil to predict the fluxes within the single-column at the center of the stencil. A timeslice for M2 has dimensions 3 $\times$ 3 $\times$ 366. On the other end, M3 uses global maps of the background field to predict global maps of fluxes. A timeslice for M3, thus, has dimensions 366 $\times$ 64 $\times$ 128.
• Figure 2: Mean predicted fluxes from globally nonlocal model, M3, for May 2015 at 200 hPa height. (a) and (b) respectively show the true mean and the predicted mean zonal flux ($u'\omega'$) for May 2015. (c) and (d) show the true mean and the predicted mean meridional flux ($v'\omega'$). The WINDSET dataset contains input variables and momentum fluxes which were normalized using a constant mean and standard deviation. Mean predicted fluxes for Models M1 and M3 are shown in Figures \ref{['fig:comp_zonal']} and \ref{['fig:comp_meridional']}.
• Figure 3: R$^2$ value for M3 for (a) zonal flux and (b) meridional flux predictions for May 2015. R$^2$ denotes the percent variance captured by the predictor. A higher R$^2$ value indicates better prediction.
• Figure 4: Mean predicted fluxes compared with the (top left) true ERA5 flux from the (top right) M1: single-column ANN, (bottom left) M2: 3x3 nonlocal columns ANN, and (bottom right) M3: globally nonlocal Attnetion U-Net CNN, for May 2015 at 200 hPa height. The figure compares the true mean and the predicted mean vertical flux of zonal momentum ($u'\omega'$) for the 3 models trained for the same number of epochs. The 1x1 and 3x3 ANNs had identical hyperparameters and the 3x3 input was processed and propagated into a single 1x1 column input by applying a 3x3 2D convolution layer. Even though the 1x1 ANN roughly captures the gross structure of the fluxes, and identifies the stationary hotspots in the midlatitudes to a certain degree, the predictions have a clear strong bias. Moreover, M1 incorrectly predicts the sign of the zonal flux over most of the Northern polar region and the sign of the meridional flux over most of the Southern polar region. Introducing nonlocal leads to a drastic improvement in performance, reduced model overfitting, and produces better prediction. The globally nonlocal UNet provides the best prediction.
• Figure 5: Same comparison as in Figure \ref{['fig:comp_meridional']}, but for the meridional flux of vertical momentum ($v'\omega'$) at 200 hPa.