Weather forecasting as we know it was born in the trenches of World War I, when a prophetic young British meteorologist named Lewis Fry Richardson had a stroke of genius. Richardson's idea was to lay a grid over the landscape and calculate the behavior of the atmosphere in each cell. The math would be done by an array of 64,000 people gathered in a tremendous amphitheater. Each person, armed with a mechanical calculator, would compute the weather in a given cell based on physical equations, observations streaming in from the field, and the results passed along from neighboring cells. In one fell swoop, Richardson had anticipated both modern forecasting and parallel computation.
Today, the computers at the National Center for Environmental Prediction (NCEP) of the National Oceanic and Atmospheric Administration (NOAA) and elsewhere, fed by satel-
lites and a global network of observatories, are a realization in silicon of Richardson's dream. These computers do a pretty good job of predicting the weather by laying a coarse grid across the United States. The cells of the grid are currently 32 kilometers (20 miles) on a side, which makes it impossible to predict precisely where a storm is likely to strike. They can say if it will rain, but not if it will rain on your parade.
That was not good enough for the organizers of the 1996 Summer Olympics in Atlanta. They needed to know how clouds would affect equestrian events, if dew would endanger racers at the Velodrome, if the wind would be adequate for sailors or too strong for platform divers. And perhaps most important, they needed to know if a thunderstorm would threaten the elaborate closing ceremonies.
To satisfy the needs of the Olympic organizers, a team of IBM researchers led by Zaphiris Christidis worked closely with goverment meteorologists to develop a system to deliver more precise weather forecasts. Christidis adapted a well-known mathematical model of the atmosphere, the Regional Atmospheric Modeling System (RAMS), developed at Colorado State University, to run on an IBM RS/6000 SP parallel computer.
For the Olympics, the researchers divided the Atlanta area on a grid that could resolve weather events on a scale of between 1.5 and 5 miles and track their evolution in 12-second increments. To make this model run fast enough required careful tuning and optimization, an art that Christidis excels at. "The target was to produce a 24-hour forecast," Christidis explains, "but the constraint was to produce that forecast in just a few hours. After all, if you needed 24 hours to do a 24-hour forecast, you might as well just open a window and stick out your head." In the end, by optimizing the RAMS code and choosing an SP that consisted of 28 processor nodes, Christidis was able to generate a 24-hour forecast in less than three hours.
Another challenge was to make sense of the torrent of meteorological data generated by the detailed model. Lloyd Treinish adapted some of the advanced visualization techniques used in IBM's Visualization Data Explorer® software package to the problem of presenting a detailed three-dimensional map that allows users to instantly interpret the model's predictions. Users can easily customize the output to highlight the factors -- wind speed, temperature, rainfall or dew formation -- most important to their planning decisions.
What brings the project under the umbrella of Deep Computing is the act of combining Christidis's optimized weather code with Treinish's visualization tools to create a system that can be used to make real-world decisions. "It's putting the pieces together," says Treinish. "All the pieces were already out there. We used a model that originated at a university, visualization software developed here at Research and available commercially, and the SP, which is also available commercially. But you can't just buy all these pieces and say, 'OK, now we have an interactive forecaster.'"
The most dramatic test of the forecaster came on August 4, 1996, the day of the closing ceremonies. The National Weather Service's coarse-grained models predicted that the Atlanta area would have thunderstorms, which could have been disastrous if they struck the crowded stadium. Ordinarily, the organizers would have delayed the ceremonies, which would have meant an expensive anticlimax to the games. The IBM system produced an animated 3-D visualization that showed storm clouds sprouting like mushrooms over the Atlanta area. Beneath the clouds were puddles of blue corresponding to predicted rainfall, which swiftly moved across the landscape, missing the stadium by 10 miles. The organizers decided to proceed with the ceremonies. The storm followed the precise track predicted by the IBM system, and the ceremonies went off without a hitch.
The forecaster has since been demonstrated around the United States. At a computer show in San Jose, the IBM system -- running on an SP cluster that was also demonstrating Deep Blue -- correctly predicted the pattern of the next day's rainfall. This feat so impressed reporters that the local press dubbed the system Deep Thunder.
While Deep Thunder is of obvious value to weather bureaus, it has sparked interest in other quarters as well, including insurance companies, airlines, utilities and agriculture. Visualization will be the key to such varied applications. "You have to be able to disseminate the forecast to a user who may not be a meteorologist," Treinish says. "In some cases, you don't even show the weather directly." For example, an insurance company is more interested in the damage a storm will create than in the meteorological details. So a display for insurance executives might combine information about the value and durability of an area's homes with predicted wind speed, showing waves of liability instead of storm dynamics. Christidis is currently working with researchers from Florida State University to adapt their detailed models of hurricane storm tracks to his hardware. A consortium of insurance companies in the Caribbean hopes to use the resulting system to hedge its bets and to offer spot insurance while storms are
bearing down.
In the three years since Atlanta, Deep Thunder has become faster and cheaper. "The equivalent system today is half the price but has more than six times the computing power," Christidis explains. "Now every weather office can
have its own computer to do a local forecast," he says.
Bruce Schechter, who has just completed a Knight Science Journalism Fellowship at MIT, is the author of My Brain Is Open: The Mathematical Journeys of Paul Erdös.
