Integrating “data driven” prediction models into DA for NWP

There has been a recent surge in development of accurate, data-driven global atmospheric models (e.g. FourCastNet, PanguWeather, GraphCast, GenCast). Evaluation (and training) of these models has generally been focused on forecast skill metrics, mostly with 1-10 day lead times. However, forecasts rely on accurate initial conditions that, in the context of operational weather prediction, generally come from a cycled data assimilation (DA) system. These emulators are generally much faster to run than traditional physics-based models, which could allow for efficient testing of new DA algorithms or running existing DA algorithms with very large ensemble sizes. While there have been a few DA studies that utilize emulators, these studies have focused on synthetic observation experiments. At NOAA/PSL we have examined the performance of currently available “off-the-shelf” data-driven prediction models in cycled DA, by replacing the Global Forecast System (GFS) model with its data-driven counterpart and assimilating a reduced observation network. These experiments have identified some deficiencies in the current generation of data-driven prediction models that limit their usefulness in cycled DA workflows. I will present results illustrating some of these deficiencies and discuss potential strategies for changing the way the models are trained to address them.