In 1831, Charles Darwin received a letter from one of his former professors. It brought news of a voyage that, as Darwin later admitted, determined his whole career.
“Capt F. wants a man (I understand) more as a companion than a mere collector & would not take any one however good a Naturalist who was not recommended to him likewise as a gentleman,” the letter from John Stevens Henslow said. “The voyage is to last 2 years & if you take plenty of books with you, any thing you please may be done.”
The invitation was, of course, to join the HMS Beagle as its naturalist. Observations made over the five-year journey would allow Darwin to form his theory of evolution — gaining him generational fame. But his wasn’t the only scientifically slick mind on board.
Robert FitzRoy, the Beagle’s captain, would know celebrity status of his own, albeit not for successfully captaining a hydrological survey of South America.
Years after their excursion, FitzRoy would be appointed to lead the Meteorological Department of the Board of Trade, which would later become Britain’s Met Office. His job was to collect marine observations and analyze wind direction data in Royal Naval ship logs, using his experience to draw complex charts for ocean currents and rainfall — experience he would soon employ to pioneer weather predictions.
These “forecasts,” as he would call them, began as a means to prevent potential shipwrecks. With a rapidly expanding telegraph network, FitzRoy began collecting real-time weather data from the coasts at his London office in “a race to warn the outpost before the gale reaches them.”
Further developing his methods, he began giving out general forecasts in the 1860s, stating the probable weather two days ahead.
In “The Weather Book: A Manual of Practical Meteorology,” FitzRoy speaks of weather forecasting as a rather elementary science. “With a barometer, two or three thermometers, some brief instructions, and an attentive observation, not of instruments only, but the sky and atmosphere, Meteorology may be utilized,” he wrote. This was, however, a gross understatement.
While his work saved many lives at sea, FitzRoy’s attempts at weather forecasting were often inaccurate, laying him open to routine mockery — ridicule that might have been one of the probable causes behind his suicide.
FitzRoy was, in truth, one of the first to make a science out of weather prediction; a science that now has over a century of hindsight to benefit from. Today, the accuracy of three-day forecasts is around 97 percent, almost as accurate as a one-day forecast three decades ago.
How is it, then, that our systems still manage to miss the occasional shower? If it’s an error as minuscule as three percent, is it my weather app that needs changing?
It makes you wonder, what else is being left out of the equation?
When a world has been as suffused with weather forecasts as ours, we’re committed to noticing only when they’re wrong. Miss out on predicting a few minutes of rain, and all forecasts are to be dismissed.
The limitations of weather communication make these predictions seem far less accurate than they are, and our bone-encased satchels of preconceived notions often choose to believe slanderous word-of-mouth over hard science.
A litany of weather apps has only served to deplete our already flawed understanding of forecasts — oversimplifying everything by representing the ever-dynamic weather through a reductive set of icons.
In reality, the scale of meteorology is spatiotemporal. Events range from short-lived phenomena affecting small geographic areas to patterns that determine temperature across the entire Earth.
To predict these events, factors like the wind, temperature, humidity, and the state of the barometer are to be observed; factors stupendously intertwined. Temperature affects pressure and humidity, pressure influences the distribution of heat and humidity, and humidity influences temperature. Predicting all these perturbations involves mathematical equations far too complicated.
Weather apps, by nature, are competing in a game of probabilities, and how each interprets the initial state of the atmosphere could result in varying levels of accuracy and sophistication. But in terms of overall quality,
single-day forecasts are almost as accurate as they’re ever going to be.
In his 1922 book, “Weather Prediction by Numerical Process,” British mathematician Lewis Fry Richardson described in detail his fantastical vision of building a giant “forecast factory.”
He described a huge building with a large, spherical central chamber. Painted upon the wall of this chamber is a map of the world, divided into equal-sized red and white checkers; each box is occupied by a human who’s to “work upon the weather of the part of the map where each sits.”
But this proved a task too complex to be picked by hand. It took Richardson six weeks to calculate a forecast six days ahead — hardly of any use retrospectively. As he explained in his book, he would need a staff of sixty-four thousand human “computers” if he were to produce anything meaningful, and even that was assuming that with practice, each could work some ten times faster than he’d managed to do in the first go.
“Perhaps some day in the dim future it will be possible to advance the computations faster than the weather advances and at a cost less than the saving to mankind due to the information gained. But that is a dream,” Richardson concluded.
The scale of his idea seems whimsical, but it serves to represent the complexity of predicting how molecules in the atmosphere would interact.
As a computational problem, prediction of weather globally is comparable to a simulation of the human brain, or of the evolution of the early Universe, according to one analysis. This explains why it would take decades for Richardson’s ideas to be realized.
Our current most accurate forecast system, the numerical weather prediction (NWP) method, involves solving a set of seven equations with seven unknowns that govern the evolution of the atmosphere, called partial differential equations (PDEs).
Weather stations, satellites, ships, radars, balloons, and buoys are used to collect data on current atmospheric conditions. This data is then used to integrate these partial differential equations forward — solving them over time to predict the future state of the system.
At present, the most accurate systems rely on supercomputers that can perform about 12 quadrillion calculations each second. But even several quadrillions a second of calculations can’t predict the weather with a hundred percent accuracy.
According to the National Oceanic and Atmospheric Administration, a seven-day forecast can accurately predict the weather about 80 percent of the time, while a five-day forecast can accurately predict the weather approximately 90 percent of the time. However, anything beyond 10 days is accurate only about half the time.
Is this a limit manifesting from human incompetence? Unlikely. AI tools like GraphCast draw upon decades of historical weather data to make reasonable assessments about the weather on a certain day, but even those aren’t accurate all of the time.
Due to a lack of training data, these statistical models are unlikely to predict rare events. Their forecasts also tend to be of the deterministic kind — giving a single forecast for what it thinks is the most likely weather, in stark contrast to the probabilistic forecasts we’re used to.
This intrinsic limit is instead established due to the vastness of the initial determinants: With each cubic meter of air on Earth containing about 10 trillion trillion molecules, minor deviancy of even a small fraction of those could compound into wildly divergent effects. The atmosphere is chaotic, and
the farther you look into the future, the more unpredictable the weather gets.
During the 139th meeting of the AAAS, meteorologist Edward Lorenz posed a provocative question: “Does the flap of a butterfly’s wings in Brazil set off a tornado in Texas?”
“Since we do not know exactly how many butterflies there are, nor where they are located, let alone which ones are flapping their wings at any instant, we cannot, if the answer to our question is affirmative, accurately predict the occurrence of tornadoes at a sufficiently distant future time,” said Lorenz.
This idea, embraced as the “butterfly effect,” was meant to illustrate the unpredictability of complex dynamic systems, even if it mistakenly imparts a destructive capability to butterflies.
It anticipates the maximal potential of weather forecasting, something that has been up for debate since the 1960s, when Lorenz determined a two-week predictability limit of the atmosphere. Future accounts mostly support this hypothesis, even if there’s considerable room for improvement.
Improvements that, in many cases, involve accounting for the unpredictability brought forth by climate change. Seasonal temperature anomalies, unusual precipitation, and changing natural climate patterns are making extreme weather events more common. Since our experience with weather forecasting relies on conditions that are no longer the norm, our models need adjusting.
Equally important is broader access to forecasting capabilities utilized by the likes of the U.S. and Europe. The north Indian Ocean, for example, accounts for 6% of the global tropical cyclones annually, but causes more than 80% of the global fatalities due to cyclones.
This results from the poor accessibility of early warning systems in the poorest countries. High-income countries can invest significantly more on weather forecasting and communication, 15 to 20 times greater than lower-income ones.
These countries are more likely to benefit from AI-based weather forecasting models, which can generate forecasts in minutes as opposed to the hours it would take conventional physics-based models — all at a fraction of the current cost.
But everything ultimately hinges on people relying on these forecasts rather than dismissing them as a roll-of-a-dice prediction. Tailoring complex global forecasts to smaller regions is complicated enough, but communicating this information to resident populations is where the real challenge lies.