Accurate Forecasts Tied to Supercomputer Spending
As large swaths of the country are hit by snowstorm after snowstorm and western states battle a decade’s long dry spell with California declaring a drought emergency, weather is on nearly everyone’s mind. For many years now, the populace has relied on powerful supercomputers to forecast the weather, so it may be surprising to learn the practice is still fraught with uncertainty.
In the words of Nashua Telegraph reporter David Brooks, supercomputers are almost as uncertain about the weather as we are. That be overstating things a bit, but the atmosphere is a chaotic system and that chaos can only be tamed so much.
In professional circles, there’s a debate about how to best predict atmospheric conditions and events. It centers on the degree of accuracy and uncertainty associated with the world’s most prominent large-scale weather models.
Weather models all employ the same basic techniques. They collect data from satellites, weather balloons and other sites and plug them into physics equations, for computers to calculate the answers. The process is repeated many times to illustrate, for example, the most likely track a storm will take. This all has to happen very quickly. Forecast accuracy depends on the quality and number of input data, the available computing power and the code design.
“Some models seem to do a better job forecasting certain things out two, three, four days, but when you get closer to the actual event, another model might do a better job of something like snowfall amounts,” notes Mary Stampone, an assistant professor of geography at the University of New Hampshire in Durham and also the official state climatologist. “It takes the expertise of trained meteorologists to sort through all this computer data to see which (model) is more reasonable.”
The issue even extends to the Olympics, where weather concerns have highlighted the performance discrepancy between America’s Global Forecast System, used by the National Weather Service, and the European Centre for Medium-Range Weather Forecasts (ECMWF). The consensus among meteorologists is that the ECMWF provides the more accurate global forecasting model. With two systems on the current TOP500 list at number 51 and 52 – the European model has significantly more computing power than the US model.
The limitations of the US model came to the light when the ECMWF warned that Hurricane Sandy would hit the East Coast days ahead of the American hurricane model, which showed the storm heading off-shore away from land.
“The state of operational U.S. numerical weather prediction is an embarrassment to the nation and it does not have to be this way,” wrote Cliff Maas, a professor of atmospheric sciences at the University of Washington on his weather blog last year. “Taiwan, Germany, England, the European Center, Canada, and other nations have more computer power for their weather prediction services. Our nation has had inferior numerical weather prediction for too long. New computers are an obvious and relatively easy first step, because they make everything possible.”
The two main forecasting supercomputers used by the National Weather Service are underpowered, with only one-tenth the computing power of the European center, according to Maas. The good news is that the systems are currently undergoing a $25 million upgrade as part of the Hurricane Sandy supplemental bill.
“If the U.S. did invest more money and people into making the model better, then the forecast would be better,” said Jeff Masters, meteorology director at the online forecasting service Weather Underground. “The money we spend on weather forecasts and improving them pays for itself.”
Meteorologists agree that leadership-class supercomputers are required to improve forecasts for this computationally-intensive application. That means spending money on petascale supercomputers and preparing for the exascale era, something the ECMWF has already started doing. Working with the Cray-formed CRESTA project, the European center has been refining its Integrated Forecast System (IFS) model, which provides medium-range weather forecasts to its 34 European member states.
The global grid size for simulations is currently based on a 16 km resolution, but researchers are working to get that down to 2.5 km global weather forecast model by 2030. To be exascale-ready, IFS needs to run efficiently on a thousand times more cores. Advances achieved by CRESTA so far have enabled IFS to harness 200,000 CPU cores on Titan, the fastest supercomputer in the United States. This is the most cores ever harnessed by a weather model and it marks the first use of the 5 km resolution model that will be needed in medium range forecasts in 2023.
Even with the best data and the ultimate computing machine, weather forecasting will never be perfect. Weather systems are a stochastic problem, which means they are sufficiently complex that a future state cannot be determined with absolute certainty, only with a certain probability.
Yet, there’s no question that sophisticated weather models and storm tracking tools help protect the country from the effects of severe weather events. Considering that devastating weather events, like tornadoes, hurricanes and floods, can cost the country billions of dollars a year, the boost in computing power and enhancements to other weather-related services are a wise investment.