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

Yi Xiao, Lei Bai, Wei Xue, Kang Chen, Tao Han, Wanli Ouyang2024 · DOI

TL;DR

This paper introduces FengWu-4DVar, a novel AI-driven weather forecasting system that couples a deep learning weather model (FengWu) with a traditional data assimilation technique (4DVar). By leveraging auto-differentiation, it bypasses the need for complex adjoint models, enabling more efficient and accurate cyclic weather predictions without relying on physical models for initial conditions.

What Problem Does This Paper Address?

Traditional weather forecasting relies on complex physical models and data assimilation to generate initial conditions for predictions. While AI models like FengWu have shown impressive forecasting accuracy, they still depend on these traditional systems for their initial states. This reliance limits their computational efficiency and accuracy ceiling. The paper aims to break this dependency by creating a self-contained AI forecasting system that can perform both forecasting and data assimilation, thus eliminating the need for traditional physical models in the operational forecasting loop.

Methods

The core innovation is the coupling of the FengWu AI forecasting model with the Four-Dimensional Variational (4DVar) assimilation algorithm. Key techniques include:

  • Auto-differentiation: Instead of manually coding complex adjoint models required by traditional 4DVar, the authors use the auto-differentiation capabilities of deep learning frameworks (like PyTorch) to implicitly derive the adjoint model. This significantly reduces engineering effort and computational cost.
  • Temporal Aggregation Strategy: To combat the rapid error accumulation in AI models over longer forecast horizons, they train multiple FengWu models with different forecast lengths (1-hour, 3-hour, 6-hour). These are then strategically combined within the 4DVar objective function to better assimilate observations that are further away in time from the initial state.
  • Differentiable Background Error Covariance Matrix: They construct a differentiable background error covariance matrix (B) that incorporates horizontal correlations using spherical convolutions, allowing it to be integrated into the auto-differentiation process.

Datasets

  • Training Data: ERA5 dataset (1979-2015) was used to train the FengWu forecasting models (M₁, M₃, M₆).
  • Evaluation Data: Simulated observations derived from the ERA5 dataset were used for evaluating the FengWu-4DVar system.

Key Insights and Ideas

  • AI-driven Data Assimilation is Feasible: The paper demonstrates that AI models can be effectively integrated into data assimilation frameworks, paving the way for fully AI-driven weather forecasting systems.
  • Auto-differentiation is a Game-Changer: Eliminating the need for manual adjoint model development is a significant breakthrough, making advanced assimilation techniques more accessible for AI models.
  • Temporal Aggregation Mitigates AI Error Accumulation: The strategy of using multiple AI models with different forecast horizons is crucial for maintaining assimilation accuracy over longer periods, addressing a key weakness of autoregressive AI models.
  • Self-Contained AI Forecasting is Achievable: The system can perform stable, cyclic forecasting and assimilation for extended periods without external physical models, offering a glimpse into the future of weather prediction.

Key Results

  • Stable Cyclic Predictions: FengWu-4DVar demonstrated stable and accurate cyclic predictions for at least one year on simulated observational data.
  • Reduced RMSE: Analysis fields consistently showed smaller RMSEs compared to background fields, indicating successful assimilation. For example, t500 RMSE decreased by approximately 15%.
  • Computational Efficiency: The system can assimilate observations within a 6-hour window in less than 30 seconds on a single Nvidia A100 GPU.
  • Robustness to Initial Errors: The system showed robustness to initial state errors, correcting them within about 20 days of assimilation.
  • Impact of Temporal Aggregation: Using temporal aggregation significantly improved assimilation results, reducing RMSE values.
  • Impact of Window Size: Larger assimilation windows led to decreased RMSE, as expected due to more available observations.

Limitations and Future Work

  • Accuracy Improvement Potential: While stable, the accuracy of the analysis field still has room for improvement.
  • Background Error Covariance: Currently, only horizontal correlations are considered in the B matrix. Future work will incorporate vertical layer and cross-variable correlations to enhance accuracy.
  • Simulated Observations: The current study uses simulated observations. Validation with real-world observational data is a crucial next step.
  • Generalization: While the authors suggest the potential for generalization to other AI forecasting models, this needs to be demonstrated.

Summary updated

Abstract

In recent years, the development of artificial intelligence has led to rapid advances in data-driven weather forecasting models, some of which rival or even surpass traditional methods like Integrated Forecasting System (IFS) in terms of forecasting accuracy. However, 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.Four-dimensional variational assimilation (4DVar) is one of the most popular data assimilation algorithms and has been adopted in numerical weather prediction centers worldwide. This research aims at extending the ability of data-driven weather forecasting models by coupling them with the 4DVar algorithm, i.e., alternatively running the AI forecast and 4DVar to realize a long-term self-contained forecasting system.In the 4DVar algorithm, the forecasting model is embedded into the objective function so that the flow dependencies are taken into account. Realizing the 4DVar algorithm in which the flow dependencies are expressed by the AI weather forecasting model is still a new area to be explored. Previous research has demonstrated the feasibility of using AI forecasting models in 4DVar as flow dependencies with the aid of auto-differentiation in simple dynamic systems, but scaling to the more complicated global weather forecasting faces additional challenges. For example, a differentiable background error covariance matrix needs to be constructed so that auto-differentiation can be implemented. Furthermore, the rapid error accumulation of AI forecasting models reduces the accuracy of flow dependencies in 4DVar and hinders the assimilation accuracy.In this research, we address these challenges by leveraging the following techniques. First, we take advantage of the “torch-harmonics” package developed by Nvidia to implement the differentiable spherical convolution for representing horizontal correlations in the background error matrix. Second, we reformulate the 4DVar objective function to take into account the cumulative error of AI weather forecasting model so that the objective function can better represent the error statistics. Third, the temporal aggregation strategy with different time-length AI forecasting models is employed to efficiently build flow dependencies so as to reduce the iterative error of AI forecasting model and improve assimilation accuracy.We conduct this research on the global AI weather forecasting model, FengWu, and couple it with 4DVar to implement the AI weather forecasting system prototype, FengWu-4DVar. Our experiments were conducted with the FengWu forecasting model at 1.4° resolution and the ERA5 simulation observations. With an observation proportion of 15% and the assimilation window of 6 hours, FengWu-4DVar is capable of generating reasonable analysis fields and achieving stable and efficiently cyclic assimilation and forecasting for at least one year, and the root mean square error on the potential height of the analysis field at 500hPa is less than 25m2/s2 on average. Moreover, assimilating observations in a 6-hour window can be realized in less than 30 seconds on one GPU of NVIDIA A100.
BibTeX
@misc{xiao2024coupling,
  author = {Yi Xiao and Lei Bai and Wei Xue and Kang Chen and Tao Han and Wanli Ouyang},
  title = {{Coupling the Data-driven Weather Forecasting Model with 4D Variational Assimilation}},
  year = {2024},
  doi = {10.5194/egusphere-egu24-2857},
}
Figure