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

Yi Xiao, Lei Bai+4 more|arXiv (Cornell University)|Dec 16, 2023

Meteorological Phenomena and Simulations12 citations

TL;DR

The paper couples the data-driven FengWu weather forecasting model with 4DVar data assimilation using auto-differentiation, enabling AI-based cyclic forecasting without manual adjoint models and showing stable, accurate analyses with simulated 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.

Motivation & Objective

• Motivate eliminating dependence on physical forecast models for initialization in data-driven forecasting systems.
• Propose a cyclic AI forecasting framework FengWu-4DVar by integrating FengWu with 4DVar assimilation.
• Demonstrate that auto-differentiation can replace manually coded adjoint models in the 4DVar pipeline.
• Show that temporal aggregation and model-error considerations improve assimilation accuracy and efficiency.

Proposed method

• Formulate the 4DVar objective J(x0) with background x^b and observation sequence {yτ} using the AI forward model M.
• Use auto-differentiation to compute ∂J/∂x0 without explicit adjoint models and optimize with L-BFGS.
• Introduce a temporal aggregation strategy that uses M1, M3 (and M6) to compute xτ for the assimilation window.
• Adopt a 6-hour assimilation window and 1-hour observation interval to align with operational practices.
• Augment the objective with a model error covariance Qτ to account for AI model errors in the assimilation.
• Implement cyclic forecasting where background fields are updated via M t→t+T and re-initialized through 4DVar.

Experimental results

Research questions

• RQ1 Can an AI-based global weather forecasting model be effectively coupled with 4DVar assimilation to produce accurate, self-contained cyclic forecasts?
• RQ2 Does auto-differentiation obviate the need for manually coded adjoint models in 4DVar when using AI surrogates?
• RQ3 Does temporal aggregation (1-, 3-, 6-hour steps) improve assimilation accuracy and stability compared to single-step integrations?
• RQ4 What are the computational costs and stability characteristics of FengWu-4DVar on realistic simulation data?

Key findings

• FengWu-4DVar yields stable cyclic forecasts over one year using simulated observations with 15% observation coverage.
• Analysis-field RMSE and bias are reduced relative to background fields across reported variables; e.g., t500 RMSE drops from ~0.48 K (background) to ~0.41 K (analysis).
• Analysis fields for z500, t500, u500, and v500 converge to RMSEs of approximately 30 m^2/s^2, 0.4 K, 1.41 m/s, and 1.39 m/s, respectively.
• The end-to-end assimilation (including auto-differentiation on a single Nvidia A100 GPU) runs in about 29.3 seconds per 6-hour window.
• Auto-differentiation equivalence to traditional adjoint-based gradient computation is established and validated in appendices.
• Incorporating model-error variance Qτ into the objective improves assimilation by accounting for AI surrogate errors.