10:15 AM - 10:30 AM
[MGI27-06] ClimaX-LETKF: A pure data-driven artificial intelligence-based ensemble weather forecasting system
Keywords:Weather prediction, Ensemble forecasting, Data assimilation, Artificial intelligence
Numerical weather prediction (NWP) model is the core of present weather prediction, which provides us with essential information for decision making. The NWP models solve the equations of the dynamical and physical processes of the atmosphere. Employing NWP simulations is computationally expensive, and parameterizations of sub-grid scale phenomena can be a major source of model uncertainty, requiring lots of effort for manual parameter tuning. Motivated by these issues, artificial intelligence (AI) has been adapted to the field of atmospheric science and has shown its usefulness in many tasks such as downscaling, error correction, parameter tuning, nowcasting and weather prediction. Starting with the GraphCast released by Google, kinds of AI-based weather prediction (AIWP) models have been developed and have shown their performances on weather predictions competitive to or better than the most reliable operational forecasts. However, studies on assimilation of real-world observations with these AI models are still limited, and there are no studies on ensemble weather predictions, which are essential for probabilistic weather predictions and risk assessments.
Here, we present ClimaX-LETKF, the first pure data-driven AI-based ensemble weather forecasting system. This system uses an AI-based foundation model ClimaX with assimilation of conventional observations a.k.a. Prepbufr by local ensemble tranform Kalman filter (LETKF). Namely, ClimaX-LETKF realizes weather forecast only by data: real-world observation data, and data-driven AIWP model rather than process-driven NWP models. The current system assimilates all the types of observations in Prepbufr such as radiosonde, airplane, ship, radar and satellite-derived winds. We have succeeded in its stable operation for multiple years.
We investigated two different inflation methods, relaxation to prior spread (RTPS) and relaxation to prior perturbation (RTPP). In contrast to NWP models, our system was more stable with RTPP than with RTPS, implying AI-based model’s inability to move the atmospheric field back to its attractor when it is out of the attractor due to inconsideration of physical equations of the atmosphere, which is a well-known problem of AI-based models. Since the RTPS inflates analysis perturbations, their physical imbalance in analysis remains in the initial states for subsequent forecasts, making a precise prediction more difficult. In contrast, the RTPP yields the perturbation with the mixture of analysis and background perturbations, reducing the impacts of the imbalance of the analyses. Relationships of RMSE and spread of the forecasts as a function of forecast lead time were also investigated. Both RMSE and spread increased as the lead time gets longer, indicating the chaotic nature of the model. Two inflation methods showed almost equivalent RMSE, but RTPP showed the larger spread, which was closer to RMSE and desirable. Spatial distributions of time-averaged RMSE and the number of assimilated observations correlated with each other to some extent, but there were considerable exceptions such as small errors in temperature in the low latitude region, in spite of sparse observations in substantial areas there or not very high accuracy of temperature, winds or surface pressure in the areas with dense observations.
Our successful operation of ensemble weather prediction assimilating real-world observations provides us with a lot of beneficial information to realize real-time AI-driven weather forecasting. Further investigation of the methods of training of the model, assimilated observations, assimilation and inflation will make better understanding of AIWP and data assimilation and more precise predictions possible.
Here, we present ClimaX-LETKF, the first pure data-driven AI-based ensemble weather forecasting system. This system uses an AI-based foundation model ClimaX with assimilation of conventional observations a.k.a. Prepbufr by local ensemble tranform Kalman filter (LETKF). Namely, ClimaX-LETKF realizes weather forecast only by data: real-world observation data, and data-driven AIWP model rather than process-driven NWP models. The current system assimilates all the types of observations in Prepbufr such as radiosonde, airplane, ship, radar and satellite-derived winds. We have succeeded in its stable operation for multiple years.
We investigated two different inflation methods, relaxation to prior spread (RTPS) and relaxation to prior perturbation (RTPP). In contrast to NWP models, our system was more stable with RTPP than with RTPS, implying AI-based model’s inability to move the atmospheric field back to its attractor when it is out of the attractor due to inconsideration of physical equations of the atmosphere, which is a well-known problem of AI-based models. Since the RTPS inflates analysis perturbations, their physical imbalance in analysis remains in the initial states for subsequent forecasts, making a precise prediction more difficult. In contrast, the RTPP yields the perturbation with the mixture of analysis and background perturbations, reducing the impacts of the imbalance of the analyses. Relationships of RMSE and spread of the forecasts as a function of forecast lead time were also investigated. Both RMSE and spread increased as the lead time gets longer, indicating the chaotic nature of the model. Two inflation methods showed almost equivalent RMSE, but RTPP showed the larger spread, which was closer to RMSE and desirable. Spatial distributions of time-averaged RMSE and the number of assimilated observations correlated with each other to some extent, but there were considerable exceptions such as small errors in temperature in the low latitude region, in spite of sparse observations in substantial areas there or not very high accuracy of temperature, winds or surface pressure in the areas with dense observations.
Our successful operation of ensemble weather prediction assimilating real-world observations provides us with a lot of beneficial information to realize real-time AI-driven weather forecasting. Further investigation of the methods of training of the model, assimilated observations, assimilation and inflation will make better understanding of AIWP and data assimilation and more precise predictions possible.