Numerical Weather Prediction at The Weather Company: Overview of a Global Rapidly Updating Forecast System

The Weather Company (TWC), an IBM Business, is releasing a global weather prediction system that is run at convective-allowing resolution over large parts of the world. The system, referred to as IBM GRAF (Global high-Resolution Atmospheric Forecast System), provides rapidly-updating global forecasts for the day ahead time period.

The Model for Prediction across Scales (MPAS), developed by the National Center for Atmospheric Research (NCAR), serves as the core global numerical weather prediction system, while the Community Gridpoint Statistical Interpolation (GSI) system serves as the core data assimilation (DA) system. IBM GRAF is configured to run at 3 km horizontal resolution over ~30% of the world (mostly land areas) and smoothly transition, over a 1200 km wide zone, to 15 km over the rest of the world. In order to run this complex and computationally expensive configuration, several enhancements to MPAS, GSI, and the surrounding software suites have been made, including a) porting MPAS to a GPU-accelerated supercomputer, b) optimizing the voronoi mesh generation algorithms, c) adapting a convective parameterization (nTiedtke) to support variable resolution, d) interfacing the MPAS native mesh to the GSI native grid and e) diagnosing and exporting datasets from the model output to enable downstream applications to utilize the forecast data.

In order to run MPAS on GPU-accelerated systems, TWC, IBM Research, and NCAR have partnered to develop a GPU-accelerated software port of MPAS and to integrate this port into a future MPAS release. TWC is running MPAS operationally on a cluster of IBM Power9 systems containing a total of 296 nVidia GPUs. As of August 2019, MPAS is running several times faster on one nVidia Volta v100 GPU as compared to 1 Intel Broadwell CPU.

The GSI data assimilation system is currently running as a cycled 3DVAR system and will later transition to run as a cycled hybrid ensemble-variational system. Conventional observations (METARs, SYNOPs, Rawinsondes) are being assimilated and additional observations will be added soon, including radar, satellite (GOES-R, Himawari, POES, etc.), and high-frequency commercial aircraft observations.

Initially, the forecast output from IBM GRAF is contributing to weather conditions and 0-8 hour forecasts that are provided to consumers through TWC apps and through default weather from several search engines and device manufacturers. IBM GRAF will also be distributed throughout TWC business lines to provide numerous product offerings including graphical forecasts provided by television stations, aviation-based forecasts of convective hazards and turbulence, precipitation accumulations forecasts for agriculture applications and many more.