MindSpore Earth

MindSpore Earth Introduction

Weather phenomena are closely related to human production and life, socioeconomic conditions, military activities, and more. Accurate weather forecasts can mitigate the impact of severe weather events, avoid economic losses, and create ongoing financial revenue in areas such as energy, agriculture, transportation, and entertainment. Currently, weather forecasts mainly use numerical weather prediction models to solve atmospheric dynamic equations that describe weather evolution by processing observational data collected from meteorological satellites, observation stations, radars, etc., thereby providing weather and climate prediction information. The prediction process of numerical models involves massive calculations that require considerable time and computational resources. Compared to numerical models, data-driven deep learning models can effectively reduce computational costs by several orders of magnitude.

MindSpore Earth is an Earth science toolkit developed based on MindSpore. It supports AI meteorological predictions for multiple spatiotemporal scales such as nowcasting, medium-term, and long-term forecasts, as well as disaster weather predictions such as precipitation and typhoons. It aims to provide efficient and easy-to-use AI meteorological prediction software for industrial researchers and engineers, university teachers, and students.

Application Cases

Ocean

Case	Description	Dataset	Model Architecture	NPU
LeadFormer	High-resolution intelligent Arctic sea ice forecasting	Not yet open source	Transformer	✔️

DEM

Case	Description	Dataset	Model Architecture	NPU
[DEM-SRNet][dem-super-resolution-URL]	Global 3-arc-second (90m) high-resolution land-sea digital elevation model	nasadem	EDSR	✔️

El Niño

Case	Description	Dataset	Model Architecture	NPU
[CTEFNet][ensoforecast-URL]	CNN and Transfer Learning-based El Niño prediction model	CMIP5, SODA	CNN	✔️

Nowcasting Precipitation

Case	Description	Dataset	Model Architecture	NPU
[DGMs][dgmr-URL]	Radar data meteorological nowcasting based on deep generative models	Radar data	GAN, ConvGRU	✔️
[NowcastNet][Nowcastnet-URL]	Generative nowcasting precipitation model incorporating physical mechanisms	USA-MRMS dataset	GAN, two-path U-Net	✔️
[PreDiff][PreDiff-URL]	Short-term precipitation forecasting based on latent diffusion models	SEVIR_LR dataset	LDM, Earthformer	✔️

Medium-range Weather Forecast

Case	Description	Dataset	Model Architecture	NPU
[FourCastNet][fourcastnet-URL]	Data-driven global weather prediction model	ERA5 reanalysis dataset	AFNO	✔️
[ViT-KNO][koopman_vit-URL]	Learning Koopman Operator for predicting nonlinear system dynamics	ERA5 reanalysis dataset	ViT	✔️
[GraphCast][graphcast-URL]	Global medium-range weather forecast based on graph neural networks	ERA5 reanalysis dataset	GNN	✔️
[FuXi][fuxi-URL]	Global medium-range weather forecast based on cascaded architecture	ERA5 reanalysis dataset	CNN, Swin Transformer V2	✔️
[SKNO][skno-URL]	Integration of KNO model and SHT operator	ERA5 reanalysis dataset	SKNO	✔️

Earthquake Early Warning

Case	Description	Dataset	Model Architecture	NPU
[G-TEAM][G-TEAM-URL]	Nationwide earthquake early warning system based on data-driven approach	Diting 2.0 dataset	CNN, Transformer	✔️

Core Contributors

Thanks to the following developers for their contributions to MindSpore Earth:

yufan, wangzidong, liuhongsheng, zhouhongye, liulei, libokai, chengqiang, dongyonghan, zhouchuansai, liuruoyan, funfunplus