Accurate medium-range global weather forecasting with 3D neural networks

https://www.nature.com/articles/s41586-023-06185-3

My goal is to show how to write “better” (mainly in terms of grammar in this case).

目的：提高读者的科技英语写作水平(不是跟作者过不去)。

Green highlight is used to show good writing.

绿色: 好好学习。

Yellow means "questionable."

黄色：有问题。

Blue: Pay attention.

蓝色：需要关注。

我一般分析摘要。今天，分析“讨论”部分。

Discussion

In this paper, we present Pangu-Weather, an AI-based system that trains deep networks for fast and accurate numerical weather forecasting. The major technical contributions include the design of the 3DEST architecture and the application of the hierarchical temporal aggregation strategy for medium-range forecasting. By training the models on 39 years of global weather 【reanalysis?】data, Pangu-Weather produces better deterministic forecast results on reanalysis data than the world’s best NWP system, the operational IFS of ECMWF, while also being much faster. In addition, Pangu-Weather is excellent at forecasting extreme weather events and performing ensemble weather forecasts. Pangu-Weather reveals the potential of using large pre-trained models for various downstream applications, showing the same trend as other AI scopes, such as computer vision^26,27, natural language processing^28,29, cross-modal understanding³⁰ and beyond.

评论：删on reanalysis data。It’s ok to use “reveals,” but better to use “demonstrates.”

Despite the promising forecast accuracy on reanalysis data, our algorithm has some limitations. First, throughout this paper, Pangu-Weather was trained and tested on reanalysis data, but real-world forecast systems work on observational data. There are differences between these data sources; thus, Pangu-Weather’s performance across applications needs further investigation. Second, some weather variables, such as precipitation, were not investigated in this paper. Omitting these factors may cause the current model to lack some abilities, for example, making use of precipitation data for the accurate prediction of small-scale extreme weather events, such as tornado outbreaks^31,32. Third, AI-based methods produce smoother forecast results, increasing the risk of underestimating the magnitudes of extreme weather events. We studied a special case, cyclone tracking, but there is much more work to do. Fourth, temporal inconsistency can be introduced by using models with different lead times. This is a challenging topic worth further investigation.

评论：一个常见的“错误”是用paper替代study。因为这一段是讲“不足”，“can be”应该是“need to be”。这里，“worth further investigation”应该删further。（这也是一个常见的“错误”。）

Looking to the future, there is room for improvement for both AI-based methods and NWP methods. On the AI side, further gains can be found by incorporating more vertical levels and/or atmospheric variables, integrating the time dimension and training four-dimensional deep networks^33,34, using deeper and/or wider networks, or simply increasing the number of training epochs. All of these directions call for more powerful GPU clusters with larger memories and higher FLOPS (floating point operations per second), which is the current trend of the AI community. On the NWP side, post-processing methods can be developed to alleviate the predictable biases of NWP models. We expect that AI-based and NWP methods will be combined in the future to bring about even stronger performance.

评论：Too brief.

Follow me, if you want to improve your writing skill:

如果你想提高科技英语写作能力，请跟我来...

http://blog.sciencenet.cn/blog-306792-1146690.html