There are countless examples of disasters exacerbated by inaccurate predictions. In one case, in October 2003, the U.S. Global Forecast System (GFS) Model, on which most U.S. forecasts are based, predicted that in 48 hours about 2.5 centimeters of rain would fall off the coast of the state of Washington, as well as an insignificant amount further inland. A storm did indeed appear, on 20 October, but it dumped more than 30 cm of water inland, making 21 October the wettest day in Seattle history and causing devastating floods in seven counties [see photo, “Whoops!”]. The model’s limited resolution was the likely culprit in this forecast bust.

Basically, the model had missed conditions off the coast of Asia, including several tropical storms and a developing typhoon. In their early stages, these storms were just too small to register on the model’s 35-km resolution. They kicked off a wavelike disturbance in the atmosphere that carried tropical moisture toward the North American coast. In effect, it was an airborne tsunami. When this wave hit the already developing storm, the residents of Washington had a major flood.

Clearly, spacing the data points more closely together on the grid would improve forecast accuracy. But halving the distance between grid points in all three dimensions increases the number of points by a factor of eight—two in each spatial dimension. That increase in spatial density also requires halving the time step, because the ratio between the time step and the grid spacing has to stay within certain parameters or the model produces an excessive number of errors and the resulting forecast is useless. Today the time step is typically about 15 seconds.

Therefore, doubling the resolution of the model would increase the computing requirements by a factor of 16. If you tried to look a week ahead with a global model and one of today’s computers, it would take most of a day to execute the whole program.

That’s too much time, because the computers used today to do weather simulations have other duties. For example, in the United States, the same computers used to run the global models also tackle data assimilation and compute the regional models, as well as long-range climate and ocean models, which are growing in importance. And the computer runs the entire set of models again from new data every 12 hours for global models and as often as every three hours for regional predictions.

So today’s models are a compromise between resolution and realistic run time. For example, the U.S. GFS Model has a 35‑km horizontal resolution and a 60-level vertical resolution. To look ahead 10 days, it takes a little more than an hour and a half on a 308‑processor IBM Power4 supercomputer, a system with an average processing capacity of about 2 teraflops.

Of course, Moore’s Law, the anticipated periodic doubling of transistor density and therefore processing power of ICs, suggests that in about 10 years meteorologists will be able to increase by a factor of eight the resolution of numerical weather models and still expect to churn through them in a reasonable amount of time.

Researchers are already testing such higher-resolution models as simulations. These simulations are currently impractical for operational forecasting, as they take longer than real time to produce results. In some cases, the simulations cover the entire globe, but most of them concentrate on worrisome events that could take shape in specific regions—for example, hurricanes or severe weather during the spring and early summer in the continental United States. The simulations have demonstrated that higher-resolution models can improve predictions of such difficult weather features as hurricane intensity and heavy rainfall.

Today, forecasters can’t predict storm intensity more than a few hours in advance with any confidence. With high-resolution modeling, researchers conjecture that forecasters will be able to accurately estimate hurricane intensity several days in advance.

In the absence of sufficiently powerful computers, forecasters have to rely on clever techniques to minimize the potential of errors being magnified by the chaotic nonlinearity of weather. One of their key strategies is to run groups of models, called ensembles, rather than single models.

To produce a forecast ensemble, a numerical weather prediction system repeatedly runs the same weather model but changes the initial state slightly each time by adding small increments to the original atmospheric measurements. A temperature reading of 21 oC at a certain grid point, for instance, might be changed by a 10th of a degree. A relative humidity at a different grid point might be increased by half a percent, and so on.

A range of reasonable forecasts emerges. As forecasters execute the programs repeatedly, one member of this forecast ensemble appears most frequently, and meteorologists regard that as the most likely forecast. They don’t discard the other members but rather use them to calculate the probability that a particular forecast will come to pass—for example, a 60 percent chance of rain or a 70 percent chance of a hurricane’s striking a particular section of coastline.

Ensembles were key to the confidence researchers had in the numerical forecasts of Hurricane Katrina’s path. On 25 August, as the storm passed over Florida, the predictions generated by different members were all over the place—atmospheric conditions magnified small changes dramatically. For instance, the winds in the area as Katrina approached Florida were so divergent that a minor difference in the predicted location at that point led to major differences in predictions of the storm’s path further out.

Forecasters therefore were unable to say with any real confidence where Katrina was heading. But by 26 August, the ensemble predictions began to coalesce into a smaller range of potential forecasts [see maps, “Aiming for New Orleans”]. By 27 August, one path clearly had a very high degree of probability, and forecasters could confidently predict the hit near New Orleans, which occurred on 29 August. Accurately communicating this changing degree of uncertainty to decision makers and the public is a critical challenge for forecasters.

Ensembles, of course, add to the processing burden. But the advantage of the ensemble approach is that the burden can be easily shared among many computers—even ones separated by continents or oceans. And thanks to international cooperation in ensemble modeling, we may be able to use larger ensembles long before 10- or 20-teraflop computers become available.

Modeling organizations in different countries are already sharing their research as well as results of complete model runs. A forecaster in the United States is likely to check his results against predictions about the same weather event computed by his colleagues in the United Kingdom, for example.

But collaboration by several countries in simultaneously producing a single ensemble forecast is not yet possible. That kind of collaboration would be a major step forward: it would allow individual facilities to each run fewer versions of higher-resolution models, resulting in greater accuracy.

In hopes of making collaboration on individual ensembles routine, the world’s major forecast centers are starting to work together. Already the United States and Canada are sharing their ensemble members and combining them to produce a single forecast. Such collaboration requires that forecasters standardize independently developed models by putting the predictions from the different models onto a common grid and then calibrating the system to remove the biases inherent in each model. Some models, for example, predict consistently warmer or colder temperatures than others.

Work is under way to combine the ensemble models of the world’s major global forecast centers. In January 2005, the 187 nations of the World Meteorological Organization (WMO), based in Geneva, under the auspices of the United Nations, launched a 10-year effort to accelerate the current rate of improvement in forecasts and reduce worldwide fatalities from weather events by half. The project is called The Observing System Research and Predictability Experiment (Thorpex). Its aims include researching the benefits of producing a single international forecast by combining the ensembles of all the various forecast systems.