2023年9月21日下午，在以“加速行业智能化”为主题的华为全联接大会2023(Huawei Connect2023)昇思MindSpore专题论坛上，昇思MindSpore开源社区发布了MindSpore Earth 0.1地球科学套件。
This suite integrates the AI weather forecast SOTA model at multiple spatial and temporal scales, provides data pre-processing, forecast visualization and other tools, and integrates ERA5 reanalysis, radar echo, and high-resolution DEM data sets, and is committed to efficiently enabling AI+ Integrated research on meteorological and oceanographic forecasts.
气象预报与人们的工作生活息息相关，也是科学智能(AI4Science)领域受到最广泛关注的应用场景之一。昇思MindSpore作为全场景AI融合框架，具有原生支持大模型与AI4Science两大引领创新的能力。
The architecture plan of MindSpore Earth is shown in Figure 1. It covers industry SOTA models for multiple scenarios such as weather forecasting, short-term precipitation, medium-term forecasting, and super-resolution, including GraphCast, ViT-KNO, FourCastNet, DGMR, etc., and the model coverage is industry-leading. The forecast accuracy surpasses traditional numerical models, and the forecast speed is more than a thousand times faster than traditional numerical models.
Figure 1 MindSpore Earth suite architecture planning
**
**
1. Medium-term weather forecast
**
**
Medium-term global weather forecast refers to predicting the weather for about 3 to 10 days in the future on a global scale. Such forecasts are typically based on numerical models simulating changes in atmospheric conditions such as temperature, humidity, air pressure, wind speed and direction, and precipitation. MindSpore Earth provides multiple AI medium-term forecast models:
FourCastNet
MindSpore Earth provides the FourCastNet model, which uses the adaptive Fourier neural operator AFNO. This neural network architecture is an improvement on the Vision Transformer (ViT) model. It constructs the mixed operation steps into continuous global convolutions, in the Fourier Efficient implementation via FFT in the leaf domain reduces spatial mixing complexity to O(Nlog N), which allows flexible and scalable modeling of dependencies across spatial and channel dimensions. This model is the first AI forecast model whose forecast accuracy can be compared with the high-resolution Integrated Forecast System (IFS) model of the European Center for Medium-Range Weather Forecasts (ECMWF).
ViT-KNO
MindSpore Earth provides a lightweight, grid-independent Koopman Neural Operator model designed based on Koopman's global linearization theory and combined with the idea of neural operators. The model architecture is shown in Figure 2. This model was launched by Huawei Advanced Computing and Storage Laboratory in cooperation with Tsinghua University. The model is able to capture complex nonlinear behavior while keeping the model lightweight and computationally efficient. Compared with FourCastNet, ViT-KNO has more efficient training performance and better prediction accuracy.
Figure 2 ViT-KNO model architecture
GraphCast
GraphCast comes from Google's DeepMind, a model that uses GNN to automatically and regressively generate predictions in an "encode-process-decode" architecture. The encoder maps the latitude-longitude input grid of meteorological features at historical moments to a multi-scale internal grid representation; the processor performs multiple rounds of message passing on the multi-grid representation; the decoder maps the multi-grid representation back to latitude-longitude Grid, and output prediction results at the same time. MindSpore Earth has open sourced the icosahedral mesh generation module to realize automatic construction of multi-scale meshes. In addition, to address the attenuation of multi-step prediction accuracy, MindSpore Earth implements rollout multi-step iterative training to reduce model error accumulation.
**
**
2. Short-term precipitation forecast
**
**
MindSpore Earth提供了DGMR降水模型，该模型的主体是一个生成器，配合时间和空间判别器损失以及额外的正则化项进行对抗训练。模型从前四帧雷达序列学习上下文表示，用作采样器的输入，采样器是一个由卷积门控循环单元（GRU）构成的递归网络，它将上下文表示和从高斯分布中取样的潜向量作输入，对未来18个雷达场进行预测。基于MindSpore Earth+昇腾可以进行对降水强度与空间分布进行高效训练与推理。
**
**
3. Digital elevation model over-resolution
**
**
昇思MindSpore团队、华为AI4Sci LAB与清华大学黄小猛团队联合推出适用于全球区域的DEM超分模型，同时发布了全球3弧秒（90 m）海陆DEM数据产品（图3），该成果已发表于《科学通报》（Science Bulletin）上。该模型在RMSE指标、清晰度、细节等方面均优于目前广泛采用的超分模型。该成果是首个分辨率在百米以内的全球海陆DEM数据集，可以满足不同领域和不同层次对海洋测深数据的需求，为不同地形复杂度下全球海陆重力场与地形的关系、探索不同海陆构造单元的均衡机制以及海陆地形对海洋潮流运动的影响等方面的研究提供重要支撑。
Figure 3 High-resolution global land and sea DEM data set
In addition, MindSpore Earth also provides forecast visualization modules, such as wind field visualization (Figure 4); built-in ERA5 reanalysis data set, radar echo data set, high-resolution DEM data, and supports short-term forecast, medium-term forecast and other model training and Evaluate. In the future, MindSpore Earth will continue to provide cutting-edge and efficient AI meteorological and oceanographic models and tools, including Pangu meteorological large model inference functions, long-term climate prediction, downscaling, etc., to enable integrated research of AI + meteorological and oceanic data.
Figure 4 Wind speed visualization effect
For more details, welcome to join the MindSpore Flow & Earth joint SIG group.
MindSpore Earth code warehouse address: https://gitee.com/mindspore/mindscience/tree/master/MindEarth
A programmer born in the 1990s developed a video porting software and made over 7 million in less than a year. The ending was very punishing! Google confirmed layoffs, involving the "35-year-old curse" of Chinese coders in the Flutter, Dart and Python teams . Daily | Microsoft is running against Chrome; a lucky toy for impotent middle-aged people; the mysterious AI capability is too strong and is suspected of GPT-4.5; Tongyi Qianwen open source 8 models Arc Browser for Windows 1.0 in 3 months officially GA Windows 10 market share reaches 70%, Windows 11 GitHub continues to decline. GitHub releases AI native development tool GitHub Copilot Workspace JAVA is the only strong type query that can handle OLTP+OLAP. This is the best ORM. We meet each other too late.