When computer models fail meteorologists

A powdered-sugar coating of snow already had fallen, and the brunt of the storm was imminent.

Yet those computer models that meteorologists rely upon in making snow forecasts once again were at odds yesterday - with each other, and even with nature to a degree or two.

Last night the weather forecasters were calling for up to three inches of snow in and around Philadelphia, with perhaps four in the Pennsylvania suburbs, followed by ice and then rain today.

This might well be the one they get right, but the way things have played out this winter, it would not qualify as shocking if conditions today mocked the forecasts, and every school opened on time. Nor would it be shocking to see the sequence repeated during the next several weeks.

It has not been a banner season for the models or their users. Recall that the forecast for Jan. 10 called for up to three to six inches, and Philadelphia ended up with almost nothing. It hardly even rained. Nine days later, forecasts of an inch or less of snow were buried under three to five inches in some suburbs.

Models have limits, and will for a while, says Jay Searles, a meteorologist at Pennsylvania State University.

True, over the last half-century, computers have revolutionized weather forecasting. As matters of routine, meteorologists look at several computer outputs several times a day, with the United States, Canada and the European Union running major models.

Using "initial" observations of the state of the atmosphere, the computers solve unimaginably complex equations to predict how the atmosphere will change over time in six-hour increments. In subsequent runs, the models are corrected to capture the behavior of the atmosphere.

But some meteorologists complain that models actually have made forecasting more difficult. "There's too much data out there," said Searles.

"There's more to look at, and it's more difficult to keep track of what model is doing what," said Tony Gigi, a meteorologist at the National Weather Service office in Mount Holly, who was on duty yesterday.

In the morning, he was concerned that the models were foundering on a critical detail. The forecasts for the upper atmosphere were about 2 degrees too warm and about 2 degrees too cool closer to the surface. With temperatures close to freezing, 2 degrees could make all the difference in whether the precipitation was liquid or frozen.

In the afternoon, Gigi was trying to figure out what to do about the disparity in snow forecasts between the two U.S. models - the Global Forecast System and the North American Mesoscale models.

The NAM suggested three inches for Philadelphia; the GFS one inch. In the end, the weather service downgraded its accumulation estimate. The weather service ended up bumping down its forecast from three to five inches to two to four.

Forecasters can't blindly follow models, said Searles, because all of them have serious flaws - with a big one right at the get-go.

Models depend on accurate readings of all the atmosphere's conditions, including temperature, pressure and humidity at various levels. Nothing in the Earth's weather system occurs in isolation.

However, in some parts of the world, observations are wanting, especially over the oceans and even some land masses. "There are holes all over the place," said Searles. "Mexico is awful."

He said the technology is probably available to measure the atmosphere comprehensively via satellite; however, such a project would be prohibitively expensive.

If the observations were perfect, the calculations still wouldn't be, he added.

To work out the equations precisely, computers would have to be able to accommodate an almost infinite number of decimal places, and right now they don't have that capacity.

Searles cautioned that forecasters should keep paying attention to their computer models, however flawed; humans, including storm-craving weathermen, can be at least as flawed.