Toward Integration of Artificial Intelligence to Improve Operational Data Assimilation for Numerical Weather Prediction

The Environmental Modeling Center at the National Centers for Environmental Prediction (part of NOAA’s National Weather Service) recently published its long term development plans for future operational data assimilation (DA) for numerical weather prediction (Kleist et al., 2024, doi: https://doi.org/10.25923/kqra-6z66). One aspect of the plan focuses on the need for integration of Artificial Intelligence (AI) and traditional DA to advance science and operational numerical predictions. This includes work around the use of observations (quality control, data selection, bias correction), forward operator emulation (computational efficiency and enhanced complexity when connecting observations to models), background error representation, improved background for assimilation, model error representation and estimation, use of emulators to replace tangent-linear and adjoint operators, generation of large ensembles, and hybridization of traditional assimilation methodologies with machine learning techniques. Several pilot projects are underway in an effort to exploit such techniques to improve future operational assimilation. This presentation will provide an overview of some of the highlights from several projects that use AI to 1) to generate superobservations of atmospheric motion vectors, 2) perform bias correction of satellite radiance observations, 3) evaluation of tangent-linear and adjoint versions of GraphCast for potential use in 4DVar, 4) building multivariate background error covariance operators, and 5) performing soil moisture retrievals from satellites that are suitable for assimilation into a coupled system. Plans for broadening such activities based on our experiences thus far will also be discussed.