Weather forecast computer modeling is known in the trade as numerical weather prediction (NWP). It's interesting because it sits at the intersection of three areas that are themselves individually interesting -- first, the prediction of events that seem unpredictable, in this case, the weather; second, following from the first, non-linear systems and chaos theory; and third, computer modeling and great big, impressive machines.

Much of the first part of this inFAQ is loosely adapted from The Handy Weather Answer Book, by Walter A. Lyons (Visible Ink Press, 1997); the part on deterministic chaos is cobbled together from multiple sources; and the information on the models in use is taken partly from USA Today at usatoday.com/weather/wmodlist.htm, partly from Wikipedia at en.wikipedia.org, and partly from the various agencies themselves. This page was re-edited and the links updated in January 2007.

This overview of NWP tries to combine the story of its short history with its science and application. It takes the form of a FAQ, but that's really only a convenience -- it's intended to be read straight through. The couple of pictures on the left are just for decoration. The ones on the right, next to the scrollbar, are clickable.

---

Who formulated the ideas of fronts and air masses?

During World War I, the meteorologists of neutral Norway were largely cut off from weather information from outside of their country because of the restrictions imposed by the warring nations of Europe. In response, Norway established its own dense network of weather stations. Led by the father and son team of Vilhem and Jacob Bjerknes, a group of scientists now known as the the Bergen School went to work analyzing the resulting data.

>From this work they developed the theory of air masses and the weather fronts between air masses. They studied instabilities on the polar front (the boundary between polar and tropical air) and from this developed the basic theory of mid-latitude storms. These theories were to become the foundation of modern meteorology.

So what is numerical weather prediction?

Because of their academic backgrounds in the study of fluid dynamics, the Bergen School scientists understood that air is a heat-conducting fluid and obeys the fundamental physical laws. These are the ones that we learned in school -- the laws of conservation of energy and mass, conservation of momentum (Newton's 2nd and 3rd laws), the laws of thermodynamics, the hydrostatic equation, and the ideal gas law.

We can construct a three-dimensional grid of the atmosphere and use these equations to create a mathematical model. If we plug data that we've gathered on its current state into our grid, we can then solve the equations to predict a future state -- numerical weather prediction. But of course it's not quite that simple.

In our atmospheric model, many of the equations are expressed as nonlinear partial differential equations. Their solution is not usually possible by precise mathematical methods, but is done by approximation using numerical analysis. For a model with thousands of data-points to forecast over any significant period of time requires a huge amount of computation.

Because of the nonlinear nature of the equations, tiny differences in the data that are plugged in to these equations -- that define the initial state of the system -- will yield huge differences in the results. This "sensitive dependence on initial conditions" is the hallmark of any chaotic system. We'll come back to that idea again.

Okay. So how's NWP work?

The actual mathematics involved are beyond the scope of this amateur essay. (But see attractor.html to get just a taste.) However, assuming you've already developed the mathematics for your model, the process goes something like this:

1. First, settle on the area to be looked at and define a three-dimensional grid with an appropriate resolution (20 to 200 kilometers on a side, say, and going maybe 10 kilometers up).
2. Then, gather weather readings (temperature, humidity, barometric pressure, wind speed and direction, precipitation, etc.) for each grid point.
3. Run your assimilation scheme so the data you've gathered actually fits the model you've designed.
4. Now, run your model by stepping it forward in time a few minutes, say.
5. Go back and do step 2 through step 4 again.
6. When you've finally stepped the model forward as far as your forecast outlook (from a day to at most a week), publish your prediction to the world.
7. And finally, analyze and verify how accurately your model predicted the actual weather. Revise it accordingly.

All that produces a numerical prediction. An actual forecast takes much more work -- more about that later.

Who first tried NWP?

While the fundamental notions of numerical weather prediction were first stated by Vilhelm Bjerknes as early as 1904, in 1922 Lewis F. Richardson formally proposed that weather could be predicted by solving the "equations of atmospheric motion."

Called by Mandelbrot "a great scientist whose originality mixed with eccentricity," Richardson soon realized that the amount of calculation would be formidable. Quixoticly, he proposed a weather prediction center in which a giant circular amphitheater would contain some 26,000 accountants equipped with calculators who would make their additions and subtractions as commanded by a sort of conductor.

Richardson's first attempts failed because the method predicted pressure changes far larger than any that had ever been observed. This unexpected result was years later found to be caused by the way he had approximated the solutions to the equations. His idea -- which, it should be emphasized, was basically sound -- was dismissed and forgotten.

Numerical weather prediction would have to wait for the proper tool.

When were the first practical attempts at NWP?

The electronic computer was conceived in the 1940's, when mathematician John von Neumann developed the prototype of the stored program electronic machine, the forerunner of today's modern computers. He soon turned his interest to NWP and formed a meteorology project in 1946 at Princeton's Institute for Advanced Study. There meteorologist Jule Charney began working on the problem of numerical weather prediction.

After figuring out why Richardson's first attempts 25 years earlier had failed, Charney was able to formulate equations that could be solved on a modern digital computer. The first successful numerical prediction of weather was made in April 1950, using the ENIAC computer at Maryland's Aberdeen Proving Ground. Within several years research groups worldwide were experimenting with "weather by the numbers." The first operational weather predictions, using an IBM 701 computer, were begun in May 1955 in a joint Air Force, Navy, and Weather Bureau project.

What role do computers play now?

Because of the amount of computation, meteorologists have required the fastest computers to do their numerical modeling. NWP has advanced greatly in six decades, in large part due to the spectacular growth in speed and capacity of digital computers. To give you a feel for it, one of the first commercial computers used for atmospheric research was the IBM 1620 which could perform about a thousand (103) operations per second. Today's massively parallel supercomputers can clip along in the low hundreds of teraflops -- trillions (1012) of floating point operations per second -- over a billion (109) times faster.

Atmospheric scientists and climate researchers (not just NWP people) use and generate huge amounts of data. The National Center for Atmospheric Research (NCAR) estimated that in 1997 they maintained computer files totaling 30 terabytes -- 30 trillion (1012) bytes -- of data. In late 2000 total data on their Mass Storage System had grown to over 200 terabytes, by early 2003 that number continued growing exponentially to over a petabyte (a petabyte is 1024 terabytes, a mega-gigabyte, or 1015 bytes), and by July 2006 NCAR's Mass Storage System surpassed 3 petabytes of data storage.

(At the end of 2006, about three quarters of all supercomputers were designed as computer clusters using thousands of processors (usually off-the-shelf), networked and programmed to work together. Today, since actual testing of nuclear bombs is no longer allowed, about half of the fastest of the world's supercomputers are used for simulations for atomic research. Maybe that's a good thing.

As for storage, starting later in 2007 CERN's Large Hadron Collider is expected to generate 15 petabytes of data each year in particle physics experiments. And in the middle of 2006, Google's largest computer cluster was estimated to have 4 petabytes of RAM!. It's mass storage is far larger.)

But there are inherent limits to numerical weather prediction that even the fastest computers can't overcome.

Is the weather even predictable or is the atmosphere chaotic?

The answer is both. We know that weather forecasters are right only part of the time, and that they often give their predictions as percentages of possibilities. So can forecasters actually predict the weather or are they not doing much more than just playing the odds?

Part of the answer seems easy. If the sun is shining, a pleasing breeze is from the north, and the only clouds in the sky are nice little puffy ones, then even we can predict that the weather for the afternoon will stay nice -- probably. So the weathermen are actually doing their jobs.

But in spite of the predictability of the weather -- at least in the short-term -- the atmosphere is in fact chaotic, not in the usual sense of "random, disordered, and unpredictable," but rather, with the technical meaning of a deterministic chaotic system, that is, a system that is ordered and predictable, but in such a complex way that its patterns of order are revealed only with new mathematical tools.

Who discovered deterministic chaos?

Well, not too new. The French mathematical genius Poincaré studied the problem of determined but apparently unsolvable dynamic systems a hundred years ago working with the three-body problem. And the Americans Birkhoff and Smale, and many others world-wide contributed greatly to the study of dynamic systems.

But its principles were rediscovered in the early 1960s by the meteorologist Edward Lorenz of MIT. While working with a simplified model in fluid dynamics, he solved the same equations twice with seemingly identical data. But on the second computer run, trying to save a little time on his very slow machine (this was nearly fifty years ago), he truncated his data from six to three decimal places, thinking it would make no difference to the outcome. He was surprised to get totally different solutions. He had serendipitously rediscovered "sensitive dependence on initial conditions."

Lorenz went on to elaborate the principles of chaotic systems, and is considered to be the father of this area of study. He is often credited with having coined the term "butterfly effect" -- can the flap of a butterfly's wings in Brazil spawn a tornado in Texas? (But see the note.)

(James Yorke of the University of Maryland is usually credited with having spawned this (somewhat misleading) new use of the word "chaos.")

What are the characteristics of a chaotic system?

Deterministic chaotic behavior is found throughout the natural world -- in dripping faucets and orbiting moons; in reacting chemicals and beating hearts; in spreading epidemics and predator-prey relationships; in financial markets and class grading patterns; and, of course, in the dynamics of the earth's atmosphere.

All these phenomena share common characteristics, some of which are:

Sensitive dependence on initial conditions:
Starting from extremely similar but slightly different initial conditions, any complex system will rapidly move to different states. This behavior has two important implications:

1. Exponential amplification of errors --
   Any mistakes in describing the initial state of a model of a complex system will guarantee completely erroneous results; and
2. Unpredictability of long-term behavior --
   Even extremely accurate starting data will not allow long-term results from a model. The model has to be stepped forward in time just a bit, the resulting data measured and plugged back in, and the model restarted from there.

Aperiodic:
A complex system never, ever repeats itself exactly (although it may be seen to come close).

Non-random:
Although a complex system may at some level contain random elements (Brownian motion, say, or quantum effects), the system's behavior is not random. It is chaotic, in our sense of the word.

Local instability, but global stability:
In the smallest scale the behavior of a complex system is completely unpredictable, while in the large scale the behavior of the system "falls back into itself," that is, restabilizes.

One can see that there are theoretical constraints on what numerical modeling can do.

Some real-world problems NWP must overcome

There are many hundreds of world-wide weather stations, weather buoys, observations from ships and aircraft, weather balloons and radiosondes, information from doppler radar, plus satellites. In spite of that, some of the data needed to plug into the grid for initializing a computer run of a model is either missing or doesn't fit the model grid either in time or space. The methods developed to deal with this are referred to as data assimilation.

Further, atmospheric processes that happen on scales smaller than that of the model's grid scale but that signicantly affect the atmosphere (such as the large amount of convection that can occur in thunderstorms, cloud formation and the release of latent heat, etc.) must be accounted for. The procedure to do this that is incorporated into the models is called parameterization. As the Meteorological Service of Canada has stated, "Parameterizations can be (and usually are) complex models in their own right."

How does resolution affect the model?

Now, obviously using a finer resolution for the model grid will more accurately reflect the actual atmosphere, and all else being equal, the prediction will more accurately forecast the weather. But the finer the resolution, the more numbers that have to be crunched on the same computers.

So, in practice, models that cover large areas (like the whole Northern hemisphere) have coarser resolution than those that cover smaller areas (like just the USA, say) and so are not going to be as accurate in the small scale. Further, it's worth noting that models that work with smaller areas can predict only for shorter time periods, since as time passes it's inevitable that weather from outside the model area (and therefore not accounted for in the model) will have influenced the weather inside the model area.

To overcome this limitation, a finer grid for a smaller area of interest is nested inside a larger, coarser grid. This method is very widely used, but adds its own complications which must be accounted for in the models.

How can NWP models be categorized?

Many models, or more specifically, the forecast products derived from them, don't fit into a neat scheme. Increasingly, the same models are being used to output many different products. For just two examples, the UKMET Unified Model is used to output predictions from global scale down to just the UK, and NOAA's Global Forecast System also outputs predictions for different scales and time-frames.

But it still can be useful to categorize forecast models by their three main characterics:

1. Forecast area, or scale:
   • global, but typically a hemisphere,
   • regional, North America or Europe, for example, and
   • relocatable;
2. Resolution, or the size of the grid: from over a hundred kilometers on a side down to just a few kilometers;
3. Outlook, or time-frame:
   • short-range, meaning one to two days out, and
   • medium-range, going out from three to 10 days.

What models are in common use (operational)?

Worldwide, there are a couple of dozen computer forecast models being used. In the United States, while there are maybe a dozen operational models, about half that number are in common use, each putting out different products. This list is by no means exhaustive, but here are some of them, grouped by scale:

global:

The European Centre for Medium-Range Weather Forecasting itself has half a dozen different models. In the States, ECMWF refers to their 40 km grid, medium-range, global product.

UKMET usually refers to the United Kingdom Meteorological Office 40 km, 6 day, global product of their Unified Model (which goes down to 4 km resolution for just the UK itself).

NOAA's GFS, the Global Forecast System, replaced the AVN in 2002, which itself absorbed the MRF (Medium Range Forecast). The GFS is a 35 to 70 km resolution, medium-range, global model. Because all output from its various products is freely available over the Internet, it is widely used by forecasters.

regional:

The NAM, North American Model, is the former ETA, renamed in 2005. The ETA gets its name from the Eta coordinate system, a mathematical system that takes into account topographical features such as mountains. It uses a nested grid, but with better resolution the ETA has absorbed the otherwise similar NGM (Nested Grid Model). According to NCEP, the ETA has "outperformed all the others in forecasting amounts of precipitation."

The GEM, Global Environmental Multiscale model, developed by the Meteorological Service of Canada, provides as one of its product families output for North America.

NOAA's RUC, Rapid Update Cycle, provides very high frequency updates of current conditions and short-range forecasts, primarily for aviation users. It puts out 12 hour forecasts every hour.

relocatable:

The HIRLAM, HIgh Resolution Limited Area Model, is a highly-configurable, high-resolution (down to 5 km), system of short-range models developed and maintained as a cooperative program by all five Scandanavian nations, plus Ireland, the Netherlands, and Spain. (See the site for Denmark, for example.) In use since 1985, it's the main or only forecast model used by most of the program members.

The MM5, Mesoscale Model version 5, is an experimental model developed at Penn State and NCAR and used at various universities. It is globally relocatable, accounts for terrain, allows for variable scaling from global down to a few kilometers, has multiple-nesting capabilites, and more. For an example at regional scale, see atmos.washington.edu/mm5rt/info.html, where it is producing high-resolution (down to 4 km), 48 hour forecasts. It can be run on as little as a single PC running Linux, although with a very coarse grid.

How do you get from a computer model to a weather forecast?

More work is needed to help overcome the inherent weaknesses of numerical models and turn output from a model into an actual forecast. Statistical post-processing develops relations between model predictions and observed weather. The use of Model Output Statistics is one long-standing technique.

These relations are then used to help translate the model output to specific forecast products. For example, two products of dozens produced each day are plots of forecast surface conditions from the NAM model, and 7-day maximum temperature predictions from the GFS.

No single run of a model can be useful beyond six or seven days -- remember "sensitive dependence on initial conditions"? To produce a long-range forecast, ensemble forecasting is widely used, in which the ouput from many model runs is statistically combined, smoothing out the inevitable errors of a single model.

The various statistical techniques typically result in percentages of probabilities, which explains why we so often hear the phrase, "Chance of rain..."

How good are these models and the predictions based on them?

The short answer is, Not bad, and a lot better than forecasting without them. The longer answer is in three parts:

Some of the models are much better at particular things than others; for example, as a USA Today article pointed out, the AVN "tends to perform better than the others in certain situations, such as strong low pressure near the East Coast," and "the ETA has outperformed all the others in forecasting amounts of precipitation." For more on this subject, here's a slide show from NCEP.

The models are getting better and better as they are validated, updated, and replaced -- the "new" MRF replaced the old, was then absorbed by the AVN, which was itself replaced by the GFS. The NGM was replaced by the ETA, now renamed the NAM. (Who's on third?)

And that's why we'll always need the weatherman -- to interpret and collate the various computer predictions, add his local knowledge, look out the window, and come up with a real forecast.