A recent feature piece from the Texas Advanced Computing Center (TACC) explores the relationship between the rise of powerful supercomputers and advances in weather forecasting. The most accurate atmospheric models accommodate a host of variables that can affect weather patterns. Heat, radiation, and the rotation of the Earth are just a few of the many factors that must be taken into account. The data is collected and converted into mathematical formulae, which the computers transform into weather forecasts.
For some time, climate forecasts were limited to so-called global weather models, which have a resolution of 100 kilometers (km) per grid-point. Despite being the current standard upon which all official predictions are based, such models lack granularity and may omit significant details. For example, two towns that are nearby each other, one on a hill and the other in a valley, will be shown as having the same weather experience, when in reality there may be subtle, or not so subtle, differences.
Masao Kanamitsu, an expert in atmospheric modeling and a leading researcher at Scripps Institution of Oceanography, is working on creating more precise weather models. Kanamitsu’s experience goes back to the mid-1990s when he ran climate models using Cray supercomputers and Japan’s Earth Simulator. Nowadays, he uses the Ranger supercomputer at the Texas Advanced Computing Center.
To improve regional predictions, Kanamitsu and others working in the field use a process called downscaling. According to the article, the “technique takes output from the global climate model, which is unable to resolve important features like clouds and mountains, and adds information at scales smaller than the grid spacing.”
Kanamitsu is using downscaling to improve microclimate forecasts in California. By integrating additional factors — like topography, vegetation, and river flow — into the subgrid of California, Kanamitsu is achieving a resolution of 10 kilometers (km) with hourly predictions.
The method’s ability to fine-tune local forecasts using global data points seems counterintuative, a sentiment that seems to be shared by Kanamitsu. He states: “We’re finding that downscaling works very well, which is amazing because it doesn’t use any small-scale observation data. You just feed the large-scale information from the boundaries and you get small-scale features that match very closely with observations.”
The work requires the powerful capabilities of systems like Ranger at TACC, which excels at producing long historical downscaling in a short period of time. Kanamitsu’s climate simulations have even outperformed those of the National Weather Service, and were the topic of 10 papers in 2010.