Twice in the past 100 years, a typical winter weather disturbance moving up the Atlantic coast stalled suddenly just south of New England - right in the face of a huge mass of bitterly cold air streaming down from Canada.

The fallout from those two huge collisions in the sky were the "great blizzards" of 1888 and of 1978.The four-day storm that began March 11, 1888, struck without warning from forecasters, who had called for overnight rain and then predicted "fair weather throughout the Atlantic states" for the very hours when astounding amounts of snow - up to 50 inches in some part of New England - fell.

In 1978, conversely, the National Weather Service's just-activated computerized prediction program warned New England two days in advance that a major winter disturbance was on the way.

But most local forecasters, skeptical of the technology, waited almost until the snow started before issuing their alerts. The computer rarely has been that accurate since.

The blizzard that struck 10 years ago caused 99 deaths and nearly $1 billion in losses in the Northeast. The storm dumped 27 inches of snow in Boston.

Today, despite a decade of continued progress in meteorology, scientists admit that a blizzard or other storm could catch them - and the public - unaware.

The problem, they say, is that the weather involves so many things going on at once that at times it almost defies prediction.

Think of the atmosphere above Earth as a multi-layer cake. But the layers in this cake keep shifting so that little pockets of ingredients sometimes get scrambled together when they bump into each other. Predicting what will happen when these pockets bump, which is the key to it all, depends on understanding the complexities of fluid mechanics.

"You have to predict what's going to happen everywhere to know what will happen in any specific place," says Roland B. Smith, a professor of geology and geophysics at Yale University. That's because the weather directly overhead today depends on what happened yesterday and days before at the surface and high in the atmosphere, over land and the oceans.

"Just how the upper-level winds steer ground-level disturbances is still unclear. And ground-level interactions also depend on local topography - hills and mountains and coast lines or flat plains.

"It's all very difficult physics, complex fluid dynamics. Some of it seems to be random, unpredictable," says the 43-year-old Smith, a former aerospace engineer turned atmospheric scientist.

Reading the temperature, wind direction, wind speed, atmospheric pressure and air density at each atmospheric layer has improved steadily thanks to hundreds of instrument-carrying balloons, orbiting satellites and airplanes, ships, buoys and ground stations that report the environment around them.

But there remain huge blind spots in the data grid - both horizontally across the entire surface of the globe and vertically from ground level to the top of the atmosphere. For instance:

- Ground-level data-collecting stations, which formerly were spaced roughly 125 miles apart over much of the Earth are now 50 to 60 miles apart in many places, but are absent in many others.

- Although hundreds of balloons rise twice a day over North America, Europe and Asia, few are sent aloft over the oceans or the southern hemisphere. That means no vertical sampling is being done above much of the Earth.

- Satellites are getting better in gathering information from areas not sampled directly, but it is difficult for one of those eyes in the sky to read winds, pressures, temperatures and other information down through the layer after layer of atmosphere below. And clouds blind satellites.

All the information is pulled together twice a day and fed into some of the world's fastest computers, which are programmed with six mathematical equations. Some come straight from Sir Isaac Newton - such as his laws of motion and of friction - while more recently formulated ones deal with the thermodynamics of liquids and gases and the conservation of energy.

The computers take the answers from the equations and plug them into programs that predict what the numbers will be 12 hours later. Those projections then are manipulated to produce weather forecasts that are sent to governmental and private weather services around the world.

Dramatic improvement in these forecasts is unlikely in the near future.

"Even if we could double the data and do the calculations four times a day, the improvement would be negligible," Smith says. "We seem to have hit a wall."

One problem is that the precise effects of heat as it strikes the different layers of the atmosphere are still little understood, he says. Thus, the equations dealing with heat are weak.

Another is that the atmosphere itself is riddled with small pockets of instability, just as the ocean at a beach is full of small eddies, side currents and whirlpools that come and go almost instantly.

Even more crippling is what some contemporary mathematicians are calling the theory of chaos.

"Sometimes when you take accurate data and put it into a precise equation, you get chaos, not the predicted result - chaos instead of determinism. It's kind of amazing and neat scientifically but not useful in trying to improve forecasting," according to Smith.

As pessimistic as he is about the future, Smith and others acknowledge the progress made.

Consider the two great blizzards.

The 1888 storm formed in the Gulf of Mexico and crossed Alabama, Georgia and the Carolinas. Then it went out to sea and disappeared.

Collection of weather data at the time was largely the job of the U.S. War Department's Signal Service. Based on weather information telegraphed to Washington from Army forts across the nation, the service issued a nationwide forecast each morning.

The Signal Service forecasters were unaware that the 1888 storm had turned north over the Atlantic Ocean and, several hundred miles offshore, was racing toward the Northeast.

Worst hit were New York's Hudson Valley and central New England. New York City and Trenton, N. J., received 21 inches of snow. White Plains, a few miles to the north of Manhattan, got 32 inches, and further upstate, Saratoga Springs received 50.

Ninety years later, the National Weather Service had highly sophisticated equipment, including the new computerized prediction program it had just put on line.

The 1978 storm was tracked all the way as it, too, moved up the coast and hit land slightly east of the 1888 blizzard's path. A furious combination of wind, cold and snow moved across Connecticut early Monday morning, Feb. 6, 1978, and didn't leave until the night of Feb. 7.

The drifting snow was so overwhelming that Gov. Ella Grasso closed state roads Monday afternoon, leaving thousands of office workers stranded at home. Parts of Rhode Island were buried under more than 48 inches.

From a scientific point of view, both blizzards were structurally the same: coastal frontal cyclones. These huge storms, always rotating counterclockwise, are standard weather fare along the Atlantic coast from September to April.

By the time they reach New England, their winds often are coming off the ocean, blowing in from somewhere between north and east - hence their name: nor'easter. Dozens of nor'easters parade up the coast each year.

The unusual thing about the 1888 and 1978 blizzards was that each nor'easter ran into a huge Canadian cold front and then was stopped south of Long Island by the upper-level winds of the jet stream, which had looped down from Canada well into the United States.

The west-to-east jet stream wrapped around the rotating surface storms and held them stationary for a day or two. Ordinarily, the jet stream quickly steers such storms away from New England and out into the ocean.

The then-new National Weather Service computer model had seen the developing pattern Feb. 4 and predicted a major snowstorm.

"It was amazing. The computer forecast was 100 percent accurate," says Mel Goldstein, director of the Weather Center at Western Connecticut State University in Danbury and one of the few forecasters who took the computer at face value.

But the computer models "are inconsistent because there are just too many unknowns and variables."

That's why local forecasters thrive. All receive the same printouts of the computer's predictions and satellite photos and weather maps. Then they try to take account of all the local variables and draw on their experiences with similar situations as they make their own predictions.

"The model is correct 85 percent of the time. So anyone can be 85 percent correct by just reading the Weather Service forecast," Goldstein says.