Ensemble Kalman Filters Bring Weather Models Up to Date

‘Twas the morning after Christmas, and in Santa’s wake an unforeseen visitor descended upon Europe. On December 26, 1999, a storm called Lothar barrelled out of the North Atlantic with hurricane-force winds. In Versailles, at the French royal palace, gusts of more than 100 miles per hour toppled ten thousand trees, including a Corsican pine that had been planted by Napoleon. Across Europe, three and a half million people lost electricity, some for as long as 20 days. More than 100 people and 400 million trees died as a result of the storm, which finally dissipated over Poland. The storm caught most European weather services with their pants down. Britain, Germany, and Switzerland predicted Lothar only 18 hours before landfall. The French weather service, Météo-France, did better, issuing storm warnings 30 hours in advance—but even they reduced the forecasted wind speeds from 90 mph to 70 mph because their computer models were so far out of line with the British ones. Why did European weather forecasters, for the most part, strike out on the storm of the century? Would better satellites or more powerful computers have helped? Surprisingly, the answer seems to be no. The most critical weakness of the weather models came in a little-known area called “data assimilation.” As meteorologist Per Unden, of the Swedish Meteorological and Hydrological Institute, wrote a month later, “The forecast problems [were] most likely due to data assimilation difficulties only.” Data assimilation is the glue that binds raw data with the physics-based equations that go into computer weather models. These equations, like all differential equations, require “initial values” to be fed into them. If you tell them the temperature, velocity, and pressure of the air in every cubic inch of the Earth’s atmosphere, the equations can predict how that state will evolve. The problem is that nobody knows the correct initial conditions. Observations from balloons, buoys, and satellites provide some information—but only for specific places and times. And instruments can always break or malfunction. The weather service’s own previous forecasts also offer an abundance of information, but some of it will be outdated or incorrect. “If either the forecast or the measurement is terrible, we should ignore it,” says Dennis McLaughlin of the Massachusetts Institute of Technology. “But in most cases, each contains some information.” The trick is to blend the two sources, integrating new data into the model without tossing out the still-valuable information embodied in the old predictions. Atmospheric and oceanic scientists are constantly looking for better ways to do this. Some of the most promising schemes are actually new variations on an old idea: Kalman filters, a data assimilation method that has been used for years in inertial guidance systems for airplanes and spacecraft. Perhaps their most famous application to date was guiding the Apollo spacecraft to the Moon.