IBM Introducing the World’s Highest-Resolution Global Weather Forecasting Model

IBM announced today that they will be introducing the world’s highest-resolution global weather forecasting model later in 2019--the IBM Global High-Resolution Atmospheric Forecasting System (GRAF). The model will be the first hourly-updating weather model that is able to predict something as small as a thunderstorm virtually anywhere on the planet.

The current top global forecasting models, the European (ECMWF) model and the U.S. GFS model, subdivide the global atmosphere into grid boxes that average 9 kilometers and 13 kilometers on a side, respectively, then solve the mathematical equations of atmospheric flow to generate a forecast for each of those grid cells. The new GRAF model employs a variable-resolution grid, resembling a honeycomb, that can be configured with higher resolution over areas of particular interest (Figure 1). Over land areas, the GRAF grid elements will have average resolutions of 3 km, which is about 3 - 4 times greater than the European and GFS models.

Figure 1. The framework on which GRAF is based enables forecasters to combine a global view of the atmosphere with a higher-resolution view of a particular region, such as North America, with lower resolution over the oceans. This image is from a previous model, MPAS, on which GRAF is based (see below). Image credit: UCAR.

GRAF achieves what has been something of a “holy grail” for global numerical weather forecasting--the ability to run at a resolution so fine that no approximations are needed to simulate how individual thunderstorms behave. Previous global models have always been forced to use what are called “convective parameterization” schemes to approximate what is going on with thunderstorms within each individual grid box, since numerous thunderstorms could be present within the box. These schemes can be a major source of error in weather forecasts. The new GRAF model does not need a convective parameterization scheme when running at 3 km resolution, since the model will permit the existence of individual thunderstorms. In many cases, this will allow the model to make more accurate forecasts in situations where thunderstorms are present.

In addition, GRAF runs hourly, compared to the 6-hourly runs available for the GFS model and the 12-hourly runs of the European model (though 6-hourly data from the model is available to some users for forecasts out to 90 hours in the future). NOAA’s High-Resolution Rapid Refresh (HRRR) model does run hourly at 3-kilometer resolution, but only covers the U.S. The new GRAF model will be like the HRRR model, but global in coverage.

Initially, the highest-resolution version of the GRAF model will run at least 12 hours into the future; a lower-resolution version now being tested extends out as far as 120 hours. Verification data to show the accuracy of the forecasts is not yet available, but we expect it to be later in 2019 once the model becomes operational. In particular, it will be very interesting to see how well the model does with rainfall forecasts for landfalling hurricanes. This is a critically important quantity to get right, as we saw with Hurricane Florence in 2018 and Hurricane Harvey in 2017.

Terabytes of data

All of this is possible because the GRAF system runs on a hardware platform capable of processing the massive volumes of data required. GRAF includes graphics processing units (GPUs) that accelerate workloads for faster performance. This is the first operational global model that will be run on GPUs in addition to traditional CPUs (central processing units). GPUs, which are commonly used in video games and other graphics-intensive software, are much more efficient than CPUs when many calculations have to be done in parallel, which is the case every time a global model steps forward in time. GPUs are used by the U.S Department of Energy in the Summit and Sierra supercomputers, the two most powerful computers in the world.

More specifically, GRAF will be composed of 84 nodes of the IBM Power Systems AC922 server and will use 3.5 petabytes of capacity from the IBM Elastic Storage Server to ensure data is available to keep the model fed. Elastic storage expands as needed to fulfill model demands, which reduces the total amount of storage needed over time.

Unique data sources

Every weather model requires weather observations to initialize the forecast, and GRAF uses the same system as the GFS to get the initial weather conditions. However, GRAF has the capability to crowdsource additional weather data by using pressure sensor readings sent from barometers found within smartphones, if people opt in to share that information. In addition, hundreds of thousands of pesonal weather stations, many operated by amateur weather enthusiasts, will also be able to contribute data to the model. These additional data sources will require some challenging quality assurance algorithms, and it will be very interesting to see by how much this crowd-sourced data improves forecasts.

A collaboration between IBM and the National Center for Atmospheric Research

This newest weather prediction system is made possible by The Weather Company’s open-source collaboration with the National Center for Atmospheric Research (NCAR). GRAF incorporates the latest-generation global weather model--the Model for Predictions Across Scales, or MPAS--which was developed by NCAR in conjunction with the Los Alamos National Laboratory. “This is a great example of how long-term basic research funded by the federal government has created an industry opportunity that is both good for the bottom line and protects lives and property,” said Antonio Busalacchi, president of the University Corporation for Atmospheric Research, which manages NCAR on behalf of the National Science Foundation.

In summary, the new GRAF model is a promising new addition to the weather forecasting business. What is especially cool is the potential for the model to utilize the power of the Internet to bring in additional high-resolution crowd-sourced data. Anybody with The Weather Channel or Weather Underground app on their smartphone will benefit from the improved forecasts and can help make these forecasts even better by choosing to opt in and share their personal weather station or smartphone weather data.

Bob Henson co-wrote this post.