Exploring the Use of Machine Learning Weather Models in Data Assimilation

TL;DR

This study evaluates the tangent linear (TL) and adjoint (AD) models of both GraphCast and NeuralGCM to assess their viability for integration into a DA framework and compares them with those of the Model for Prediction Across Scales - Atmosphere (MPAS-A), a well-established numerical weather prediction (NWP) model.

Abstract

The use of machine learning (ML) models in meteorology has attracted significant attention for their potential to improve weather forecasting efficiency and accuracy. GraphCast and NeuralGCM, two promising ML-based weather models, are at the forefront of this innovation. However, their suitability for data assimilation (DA) systems, particularly for four-dimensional variational (4DVar) DA, remains under-explored. This study evaluates the tangent linear (TL) and adjoint (AD) models of both GraphCast and NeuralGCM to assess their viability for integration into a DA framework. We compare the TL/AD results of GraphCast and NeuralGCM with those of the Model for Prediction Across Scales - Atmosphere (MPAS-A), a well-established numerical weather prediction (NWP) model. The comparison focuses on the physical consistency and reliability of TL/AD responses to perturbations. While the adjoint results of both GraphCast and NeuralGCM show some similarity to those of MPAS-A, they also exhibit unphysical noise at various vertical levels, raising concerns about their robustness for operational DA systems. The implications of this study extend beyond 4DVar applications. Unphysical behavior and noise in ML-derived TL/AD models could lead to inaccurate error covariances and unreliable ensemble forecasts, potentially degrading the overall performance of ensemble-based DA systems, as well. Addressing these challenges is critical to ensuring that ML models, such as GraphCast and NeuralGCM, can be effectively integrated into operational DA systems, paving the way for more accurate and efficient weather predictions.

Figures (5)

• Figure 1: Variations in the function for the correctness check of the GraphCast tangent linear model for a 6-hour forecast length when the initial conditions for variables $\Phi$, $T$, $u$, $v$, $w$, and $q$ are separately perturbed, where $\alpha$ is the scale factor of initial perturbations. (a) Logarithm of the absolute value of $\frac{J(\alpha)}{\alpha} - 1$ for each variable, indicating the accuracy of the linear approximation. (b)-(c) Scatter plots of the nonlinear forward difference vs. tangent linear results for (b) temperature and (c) specific humidity. The slope/intercept of the linear regression line, plotted in red, are (b) 1.0/1.18×10-6 and (c) 0.997/2.31×10-9, respectively.
• Figure 2: (a) Background geopotential heights (contoured) and zonal wind (shaded) at 00 UTC on January 1, 2022. The cross marks the location of the imposed perturbation. (b)-(c) Horizontal distribution of the TL response in zonal wind 6 hours into the forecast to a zonal wind perturbation at the initial time for GraphCast (left) and MPAS-A (right). (d)-(e) Vertical cross-sections of the TL response in zonal wind along the longitude line at 33$^{\circ}$N for GraphCast (left) and MPAS-A (right).
• Figure 3: Horizontal distribution of adjoint sensitivity 6 hours prior to the response function of $T=1$ marked by the cross in temperature (top row), zonal wind (middle row), and specific humidity (bottom row). The left column presents the GraphCast adjoint results, while the right column presents the MPAS-A adjoint results.
• Figure 4: Vertical cross-sections of adjoint sensitivity 6 hours prior to response function of $T_{ad}=1$ at the location marked by the cross along the longitude line at 33 degrees north temperature (top row), zonal wind (middle row), and specific humidity (bottom row). The left column presents the GraphCast adjoint results, while the right column presents the MPAS-A adjoint results.
• Figure 5: The horizontal distribution (top row) and vertical cross sections (bottom row) of NeuralGCM adjoint sensitivity six hours prior to $T=1$ marked by the cross in temperature (left column) and zonal wind (right column).