Towards a Self-contained Data-driven Global Weather Forecasting Framework
Authors: Yi Xiao, Lei Bai, Wei Xue, Hao Chen, Kun Chen, Kang Chen, Tao Han, Wanli Ouyang
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that under the ERA5 simulated observational data with varying proportions and noise levels, Feng Wu4DVar can generate accurate analysis fields; remarkably, it has achieved stable self-contained global weather forecasts for an entire year for the first time, demonstrating its potential for realworld applications. Additionally, our framework is approximately 100 times faster than the traditional 4DVar algorithm under similar experimental conditions, highlighting its significant computational efficiency. |
| Researcher Affiliation | Collaboration | 1Department of Computer Science and Technology, Tsinghua University, Beijing, China 2Shanghai Artificial Intelligence Laboratory, Shanghai, China 3Qinghai University and Intelligent Computing and Application Laboratory of Qinghai Province, China 4School of Information Science and Technology, Fudan University, Shanghai, China. |
| Pseudocode | Yes | Algorithm 1 Cyclic Forecasting with Feng Wu-4DVar |
| Open Source Code | Yes | 1The code of Feng Wu-4DVar is available at https://github.com/Open Earth Lab/Feng Wu-4DVar. |
| Open Datasets | Yes | We have trained two models using ERA5 dataset of year 1979-2015, including a 1-hour forecasting model M1, which is embedded into the 4DVar algorithm for representing flow dependencies, and a 6-hour forecasting model M6, which is employed for making forecasts. ... In this study, all observations are simulated observations generated from the ERA5 reanalysis dataset (Hersbach et al., 2020). |
| Dataset Splits | No | The paper mentions using the ERA5 dataset for training and simulated observations but does not specify explicit train/validation/test splits (e.g., percentages, sample counts, or cross-validation setup) for the datasets used in the experiments. |
| Hardware Specification | Yes | assimilating observations in a 6-hour window can be realized in less than 30 seconds on one NVIDIA A100 GPU, 100 times faster than the traditional 4DVar on 256 processors of the PI-SUGON high-performance computer. |
| Software Dependencies | No | The paper mentions the use of "torch-harmonics package" and "Py Torch" but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | We simulate five atmospheric variables (each with 13 pressure levels) and four surface variables, resulting in a total of 69 predictands. ... The spatial resolution we test is 128 256. ... the observation proportion in our experiments is 15%, indicating that only 15% of the locations have observations. The standard deviation of observation noise is 0.001 times the standard deviation of the variable distribution. ... when the observation proportion is between 5% and 15% and the assimilation window is set to 6 hours, Feng Wu-4DVar is able to generate reasonable analysis fields and achieve stable and efficient cyclic assimilation and forecasting for at least one year. |