16:15 〜 16:30
[AAS02-10] Evaluation of an AI weather forecast model FourCastNet trained on 30-year datasets of ERA5 and on a global cloud-system resolving model NICAM
キーワード：人工知能、数値予報、天気予報、畳み込みニューラルネットワーク、AI、GPU
In recent years, AI weather forecast has rapidly flourished. FourCastNet (Fourier ForeCasting Neural Network, Pathak et al., 2022) is one of them. In this model, the input data is divided into p x p sized patches, where p denotes the number of grid points. Subsequently, these patches and their positions are embedded in a higher dimensional space, and then performed spatial and channel mixing using Fast Fourier Transform.
In the previous study, the ERA5 dataset (Hersbach et al., 2020) was used for FourCastNet training and inference. In this study, trials on the ERA5 dataset and NICAM-AMIP dataset (Kodama et al., 2015) is conducted. The models’ computational cost and performance are evaluated, and potential applications are discussed.
By comparing runtime data at different resolutions, the spatial computational complexity of FourCastNet is estimated to be O((n/p2)log(n/p2)), while the time complexity is estimated to be O(n3T) (n: number of grid points, T: time period of training data, p2: number of grids which consist one patch). The data used must have no missing values globally for all pressure levels or altitudes, regardless of the terrain. This requirement poses a challenge in terms of data imputation when using FourCastNet. In all execution results, temperature shows a consistent trend of high accuracy throughout the entire forecast period. In contrary, relative humidity tends to have lower accuracy. Such difference may be because temperature tends to be stronger dominated by the climatological seasonal cycle compared to relative humidity. It is worth investigating accuracy excluding seasonal variations in the future.
Once trained, FourCastNet can infer with significantly lower computational cost compared to prediction with traditional mechanical and thermodynamic numerical models. We believe that in the near future, AI models could be utilized for short-term precipitation forecasts, and for surrogates of parameterization schemes that handle sub-grid scale components in climate models.
Pathak et al., 2022: https://doi.org/10.48550/arXiv.2202.11214.
Hersbach et al., 2020: https://doi.org/10.1002/qj.3803.
Kodama et al., 2015: https://doi.org/10.2151/jmsj.2015-024.
In the previous study, the ERA5 dataset (Hersbach et al., 2020) was used for FourCastNet training and inference. In this study, trials on the ERA5 dataset and NICAM-AMIP dataset (Kodama et al., 2015) is conducted. The models’ computational cost and performance are evaluated, and potential applications are discussed.
By comparing runtime data at different resolutions, the spatial computational complexity of FourCastNet is estimated to be O((n/p2)log(n/p2)), while the time complexity is estimated to be O(n3T) (n: number of grid points, T: time period of training data, p2: number of grids which consist one patch). The data used must have no missing values globally for all pressure levels or altitudes, regardless of the terrain. This requirement poses a challenge in terms of data imputation when using FourCastNet. In all execution results, temperature shows a consistent trend of high accuracy throughout the entire forecast period. In contrary, relative humidity tends to have lower accuracy. Such difference may be because temperature tends to be stronger dominated by the climatological seasonal cycle compared to relative humidity. It is worth investigating accuracy excluding seasonal variations in the future.
Once trained, FourCastNet can infer with significantly lower computational cost compared to prediction with traditional mechanical and thermodynamic numerical models. We believe that in the near future, AI models could be utilized for short-term precipitation forecasts, and for surrogates of parameterization schemes that handle sub-grid scale components in climate models.
Pathak et al., 2022: https://doi.org/10.48550/arXiv.2202.11214.
Hersbach et al., 2020: https://doi.org/10.1002/qj.3803.
Kodama et al., 2015: https://doi.org/10.2151/jmsj.2015-024.