A forecast calculated, not guessed
Lewis Fry Richardson’s 1922 book, Weather Prediction by Numerical Process, proposed something no one had attempted before: forecasting the weather not by pattern-matching against past conditions but by directly solving the physical equations governing how pressure, temperature and wind actually change over time. To demonstrate the idea, he worked by hand, while serving with a Quaker ambulance unit in France during the First World War, through a retrospective six-hour forecast for a single date, 20 May 1910, using real observed data as a starting point. The calculation took him weeks to complete and produced a wildly wrong answer, predicting a pressure rise of 145 hectopascals over six hours when conditions had actually stayed essentially unchanged, an error large enough on its face to make the entire method look hopeless rather than merely difficult.
Weeks of hand calculation, a wildly wrong answer
Richardson’s method itself, breaking the atmosphere into a grid and stepping the physical equations forward in small time increments, was sound; what defeated his specific attempt was a practical detail rather than a conceptual flaw. Later analysis showed that his input data contained noise that his calculation amplified rather than smoothed out, an error avoidable with the right data-processing technique, and that once modern researchers applied appropriate smoothing to his original 1910 data, his method produced a forecast reasonably close to what actually happened. Richardson himself understood the deeper problem was not accuracy but speed: he calculated that a useful real-time forecast using his method would require an enormous number of human calculators working in coordination, an idea he described vividly as a “forecast factory,” a hall of tens of thousands of people each computing one small piece of the atmosphere, directed by a conductor using coloured lights to keep the calculations synchronised with real weather as it happened.
A flaw in the data, not the method
The gap between Richardson’s method and a usable forecast closed only once a machine could calculate faster than the weather itself changed. In 1950, a team led by Jule Charney and John von Neumann used the ENIAC, one of the first electronic computers, to produce the first computer-generated weather forecasts, working from a deliberately simplified version of the full atmospheric equations that a machine of that era could actually handle in a reasonable time. This crossed the threshold Richardson had identified as the real obstacle: computation time finally fell below the length of the forecast period it covered, meaning a prediction could, for the first time, be produced before the weather it described had already happened. Sweden’s Meteorological and Hydrological Institute followed in 1954 with the first operational forecast used for practical purposes, and the United States established its own Joint Numerical Weather Prediction Unit, combining the Air Force, Navy and Weather Bureau, in 1955.
The forecast factory fantasy
Richardson’s basic method, numerically solving the governing equations of atmospheric motion on a grid, has held up as the entire foundation of modern weather forecasting, extended rather than replaced by every subsequent generation of computing power. Norman Phillips built a working climate model by 1956 capable of representing weather patterns over monthly and seasonal timescales, and by the late 1960s researchers had combined ocean and atmosphere processes into a single general circulation model. Operational forecasting using increasingly complete versions of the underlying physical equations spread through West Germany and the United States by 1966, the United Kingdom by 1972, and Australia by 1977, with each step made possible not by a new theoretical insight but by enough additional computing power to handle a more complete and finer-grained version of the same equations Richardson had tried to solve by hand.
ENIAC crosses the speed threshold
What Richardson’s original attempt could not resolve on its own was the practical limit computing speed placed on any numerical approach, a limit that remained the field’s central constraint for decades after his book appeared. His forecast factory vision, absurd as a literal proposal involving tens of thousands of human calculators, was in effect a correct description of the computational scale the problem actually required, later supplied not by people but by electronic machines whose capacity kept expanding roughly in step with the field’s ambitions. This dependency on raw computing power is why numerical weather prediction’s history reads less as a series of conceptual breakthroughs after 1922 and more as a long, steady translation of an already-correct method into a form fast enough to be useful, a pattern distinct from most of the other discoveries in this collection.
Decades of the same equations, more computing power
This is worth the time for how directly it connects a spectacular, embarrassing failure to a correct underlying idea, and for how clearly it shows an entire scientific field waiting, for roughly three decades, on a piece of technology rather than a piece of theory. Richardson’s own honesty about his forecast’s dramatic error, and his willingness to publish the failed attempt in full detail alongside the reasoning behind it, makes the material a useful example of how a method can be right even when its first demonstration is wrong. Readers should not expect Richardson’s 1922 book to have produced a usable forecast; its value lies entirely in the method it laid out, later vindicated once computers rather than human calculators could run the same equations fast enough to matter. The forecast factory image alone makes this worth reading closely.