FengWu-4DVar: Coupling the Data-driven Weather Forecasting Model with 4D Variational Assimilation

TL;DR

This work presents FengWu-4DVar, a cyclic weather forecasting framework that couples the AI-based FengWu model with Four-Dimensional Variational (4DVar) assimilation to produce self-contained analysis fields without relying on physical forecast models. By exploiting auto-differentiation, the method bypasses manual adjoint-model development and introduces a temporal aggregation strategy to assimilate observations within a fixed 6-hour window, using 1-hour and 3-hour surrogates (and a 6-hour surrogate for forecasting). Experimental results on simulated ERA5-derived data on a $128\times256$ grid show that analysis RMSE and Bias improve over background fields and that the system can sustain year-long cyclic forecasts with reasonable computational efficiency (about 29.3 seconds per assimilation on an Nvidia A100). The approach contributes to data assimilation by enabling AI-based cyclic forecasting without physical models and highlights future work on expanding background error structures and validating with real observations.

Abstract

Weather forecasting is a crucial yet highly challenging task. With the maturity of Artificial Intelligence (AI), the emergence of data-driven weather forecasting models has opened up a new paradigm for the development of weather forecasting systems. Despite the significant successes that have been achieved (e.g., surpassing advanced traditional physical models for global medium-range forecasting), existing data-driven weather forecasting models still rely on the analysis fields generated by the traditional assimilation and forecasting system, which hampers the significance of data-driven weather forecasting models regarding both computational cost and forecasting accuracy. In this work, we explore the possibility of coupling the data-driven weather forecasting model with data assimilation by integrating the global AI weather forecasting model, FengWu, with one of the most popular assimilation algorithms, Four-Dimensional Variational (4DVar) assimilation, and develop an AI-based cyclic weather forecasting system, FengWu-4DVar. FengWu-4DVar can incorporate observational data into the data-driven weather forecasting model and consider the temporal evolution of atmospheric dynamics to obtain accurate analysis fields for making predictions in a cycling manner without the help of physical models. Owning to the auto-differentiation ability of deep learning models, FengWu-4DVar eliminates the need of developing the cumbersome adjoint model, which is usually required in the traditional implementation of the 4DVar algorithm. Experiments on the simulated observational dataset demonstrate that FengWu-4DVar is capable of generating reasonable analysis fields for making accurate and efficient iterative predictions.

Figures (10)

• Figure 1: Three development phases of cyclic weather forecasting systems.
• Figure 2: Schematic diagram of implementing 4DVar on the FengWu forecasting model (corresponding to one step optimization at time $t=t_0$). The temporal aggregation strategy is represented by $\mathcal{M}_3$ and $\hat{\mathcal{M}}_3$ in the figure, enabling more direct and efficient propagation of gradients. The auto-differentiation strategy is represented by $\hat{\mathcal{M}}_1$ and $\hat{\mathcal{M}}_3$ in the figure and is automatically implemented by PyTorch, requiring no manual code writing.
• Figure 3: 4DVar Assimilation Results on FengWu. The x-axis in each sub-figure represents lead time, at a 6-hour interval over a one-year lead time. The y-axis represents the latitude-weighted RMSE and Bias defined in Equation \ref{['equ:rmse']} and Equation \ref{['equ:bias']}. The blue lines correspond to the analysis fields and the black lines correspond to the background fields.
• Figure 4: Effect of Window Size. The x-axis in each sub-figure represents lead time, at a 6-hour interval over a one-year lead time. The y-axis represents the latitude-weighted RMSE defined in Equation \ref{['equ:rmse']}. All the reults on reported on analysis fields. The blue, orange, green lines correspond to the 6-hour, 3-hour, and 1-hour windows, respectively.
• Figure 5: Effect of Temporal Aggregation Strategy (during 4DVar). The label "w/ TA (4DVar)" corresponds to the experiment in which the temporal aggregation strategy is applied; the label "w/o TA (4DVar)" corresponds to that in which the temporal aggregation strategy is removed.