prediction solutions since 2018, utilizing the NCAR Model for Prediction Across Scales
(MPAS). The initial implementation replaced the legacy global Weather Research and
Forecasting model and its downstream applications, followed by a collaboration with UCAR to
develop a next-generation high performance computing global model based on IBM’s POWER9
processor and NVIDIA Tesla V100 GPU acceleration. In 2019, this yielded the world’s first
high-resolution, convective-allowing, and hourly-updating global weather forecast model – the
TWC Global High-Resolution Atmospheric Forecasting (GRAF) System – using a variable
resolution 15/3km mesh consisting of 24 million cells. GRAF provided hourly updating
precipitation forecasts for TWC’s short-term weather forecast platforms and applications.
A 72-hour version of GRAF has also been providing 4x-daily forecasts on a variable 15/4km
mesh, with the 4-km refinement across CONUS and Europe. This extended GRAF solution has
become a foundational asset and tool across media applications, including widespread use
throughout the broadcast television weather industry.
More recently, significant HPC expansion and unification of the multiple GRAF applications
have been underway. With over 200 compute nodes, a more advanced TWC weather prediction
system is now possible and will consist of high-resolution, convective-allowing, and
hourly-updating global 72-hour forecasts. Data assimilation (DA) will play an integral role for
next-generation GRAF, with a transition from a partially-cycled GSI (Gridpoint Statistical
Interpolation) approach to a fully-cycled, rapidly-updating JEDI (Joint Effort for Data
assimilation Integration) implementation. The TWC JEDI development efforts will progress
from 3DVar to Hybrid 3D- and 4D-EnVar, using a wide range of observation datasets such as
radar, satellite, conventional, and TWC proprietary sources. The ensemble component of the
hybrid DA approach will be based on an in-house MPAS configuration. Current work has been
related to (i) generating a 12-km background error covariance matrix based on 5 months of
retrospective GRAF forecasts and (ii) automating a JEDI-based workflow using GPU MPAS.
The presentation will provide an overview of the GRAF configuration and expansion, model
updates, ongoing JEDI DA, and a roadmap summary including the future of AI for GRAF.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner