Will it rain Tuesday? Ask a supermodel.

Meteorologists have discovered a way to improve weather forecasts significantly by combining the strengths of many computer models. This approach, called a superensemble, "always performs better than the best model," says study leader T.N. Krishnamurti of Florida State University in Tallahassee.

The National Weather Service and other forecasting agencies around the world routinely run complex computer models to predict how the atmosphere will evolve. The models break the global sky into a three-dimensional grid and solve equations to compute the temperature, humidity, wind speed, and other factors for each grid point. Meteorologists then use the outputs of these models to make forecasts ranging from a few hours ahead to several seasons in the future.

The superensemble works by analyzing the strengths and weaknesses of individual forecasting models. Krishnamurti and his team train their ubermodel by feeding in actual weather measurements and past forecasts made by many separate models.

In the case of weather prediction for a few days ahead, the superensemble reduces the error in forecasts by 20 to 80 percent, compared with individual models. For seasonal climate outlooks, Krishnamurti's model lowered the size of the error by 50 to 70 percent, the researchers report in the Sept. 3 SCIENCE. It even outperforms the best hurricane models, the team says.

The superensemble study, however, will not bowl over practicing forecasters. "There's nothing new about this work," says Robert E. Livezey of the National Weather Service in Silver Spring, Md. Forecasters have known about this approach for a long time but have had difficulty using it because of limitations in computer power. Also, continuous improvements in models complicate the process of training a superensemble.