Deep Thunder increases the resolution by using a mini-supercomputer to include additional information about the local area that can affect weather conditions. Geographical information about the local area (such as the topographic layout of the metropolitan area) and coastal information (such as the city’s exposure) is fed into the supercomputer. “The differences are subtle, but at a local level they may be important,” says Treinish. “The way thunderstorms occur in New York is different from the way they occur in Miami.”
This information is then fed into one of IBM’s pSeries Cluster 1600 computers–effectively a mini-supercomputer–to calculate forecasts every 30 minutes. Trials are currently being run in several cities, including Miami, FL; New York, NY; Kansas City, KS; and Baltimore, MD. “We are running a four-dimensional physics model for each of these metropolitan areas,” says Treinish.
Having one-kilometer resolution does not necessarily mean you can resolve detail at that level, says Clive Wilson, manager of the mesoscale modeling group at the United Kingdom’s Met Office, in Exeter. The Met Office currently provides the United Kingdom with up to four-kilometer resolution but is conducting trials of models that work at higher resolutions. The advantage of going down to these sorts of levels is that it has the effect of improving the overall accuracy of weather forecasts for a larger area.
IBM’s Anthony Praino, who also works on Deep Thunder, agrees. There is a trade-off, he says, to increasing the resolution versus the accuracy. In much the same way that increasing the pixilation of a display will give an overall more accurate picture, the actual variation of the individual pixels can alter more dramatically. So while the accuracy for these one-kilometer regions might be poorer than other forecasts, the accuracy for a larger region becomes more detailed and more accurate.
Nevertheless, there is a fundamental problem with making claims about accuracy when using such high resolutions, says Lord: “How do you verify it?” The only way would be to create a network of ground-based sensors covering one-kilometer grids.