    
|
|
This article is the sixth in a series on NASA High-Performance Computing and Communications (HPCC) research teams. Around 7 p.m. on a mild spring evening, a weather balloon rises 100,000 feet above the National Weather Service station in Sterling, Va. A purse-sized instrument called a radiosonde records temperature, humidity, atmospheric pressure and wind speed and direction. Twice daily this ritual is repeated here and at 7,000 other locations worldwide. NASA uses these and many other observations to better understand the prolonged weather patterns of Earth's climate system. In a technique known as data assimilation, agency scientists combine the observations with forecast models that mimic the atmosphere's physics. "We have to put the observations together to make a coherent, time-evolving, quantitative moving picture of the current climate and eventually say more about the future climate," said Steve Cohn of NASA's Data Assimilation Office, based at Goddard Space Flight Center, Greenbelt, Md. Producing a trustworthy climate documentary is no easy task, as balloon readings are indirect. "To get temperatures, radiosondes measure resistance and then convert to degree readings," said Peter Lyster, an Office scientist affiliated with the University of Maryland. Similarly, NASA relies on satellite instruments for global coverage detecting various types of radiant energy. Indirect means imperfect, so scientists have to worry about errors when combining observations with forecast models. NASA HPCC is helping the Office find better ways of dealing with errors as the number of observations increase and models get more complex. The three-year Earth and Space Sciences agreement "allows us to do some prototyping of new computational approaches," said Office head Ricky Rood. Lessons learned serve their mission to provide climate data sets to the international research community. From the moon to the atmosphere Every instrument has a percentage error that is estimated from observation comparisons. "Data assimilation uses the errors to determine the appropriate weights to apply to observations," explained Lyster, the team's principal investigator. "If you measure a room's temperature with two thermometers, and one says 30 degrees and the other 29 degrees, how do you combine those temperatures to get the best estimate? You give more weight to the more accurate instrument." NASA HPCC allows prototyping of new computational approaches. In weighing errors on a global scale, the Office's guide is engineer Rudolf Kalman. During the 1960s, he developed a way to "estimate the current state of a [changing] system, which you can only imperfectly model given noisy, imperfect observations," Cohn said. NASA first used the Kalman filter to help steer Apollo missions to the moon; today, it is commonly employed in aircraft navigation systems and numerous other applications. While difficult enough, navigation involves pinning down only position and velocity. Balloons, satellites, surface instruments, ships and aircraft make over 400,000 observations a day. Kalman filtering of that full data set is not possible, even with current supercomputers, so "we test in mock-up problems," Cohn said. In a first-of-a-kind study, the team applied the Kalman filter to the dispersal of methane gas through the stratosphere (10 to 50 kilometers above Earth's surface). Joining one carbon and four hydrogen atoms, methane is one of the greenhouse gases that contribute to global warming. Among its sources is vegetation interacting with standing flood water. Using several scalable parallel supercomputers over 18 months, researchers assimilated eight days of 1992 methane data from NASA's Upper Atmosphere Research Satellite. "Methane comes from the surface and is pumped into the stratosphere through the tropics," said Richard Ménard, Office scientist from the University of Maryland. As it is blown about, methane traces how air and other gases like ozone circulate in the stratosphere.
"What the Kalman filter does for this problem is unique," Ménard said. "For every structure, we also have an uncertainty being calculated. That uncertainty evolves." Working with the forecast model, the filter moves the methane information and evolving uncertainties along the wind patterns. "As a result, observational information is carried over very long distances," he said. Add the filter's ability to reduce errors over time, and the study showed a much sharper than expected contrast between high and low values of methane, indicating less general mixing in the stratosphere. "The Kalman filter maximizes the amount of information you can extract from the observations," Ménard said. Evolving uncertainties and reducing errors are what make the Kalman filter so computationally intensive, a problem for petaflops (one million billion floating-point operations per second) computers when applied to all the observations, Lyster said. (See Mix of technologies spurs future supercomputer.) The Office cannot wait 10 years for petaflops capabilities when they sometimes must generate forecasts within 12 hours for NASA satellite or aircraft missions, Rood said. Accordingly, last year the Office began implementing the parallel Physical Space Analysis System, which preserves the Kalman filter feature that connects the impact of forecast and observational errors in one location with all other locations. Improving on NASA's previous approach that analyzed data regionally, "the system solves for the whole globe at one time so it is internally consistent," explained Jing Guo, the Office and General Sciences Corp.
An earlier NASA HPCC team tooled the system for parallel computing, a collaboration with NASA's Jet Propulsion Laboratory, building what Lyster calls "the Ferrari." "It is hand-made and very fast," he said. "What we are trying to do for the Office is build a Corvette, which is high-performance but mass-produced. We want the system to be reliable and fast and still give correct answers in a diverse environment where other programmers don't necessarily know anything about HPCC." Lyster's team is applying standard software engineering practices to ready the system for daily usage. That effort includes checking results against climate averages to secure its role as "the scientific infrastructure to accommodate new types of observations," Rood said. Among the new observations are those from the U.S.-Japan Tropical Rainfall Measuring Mission and the upcoming NASA Earth Observing System satellites. "Tropical rainfall is a signature of vertically integrated heating, which drives atmospheric circulation," said Mission deputy project scientist Arthur Hou. With this satellite's data added, the system will be able to probe a variety of impacts. For instance, during El Niño "atmospheric heating follows the maximum sea surface temperature to the Eastern and Central Pacific, changing the circulation pattern and regional climate throughout the world," Hou said. More information about this research is available at Lyster's World Wide Web site.
|
