A blindfolded sapper searches for a mine by tossing pebbles at random. He cannot see the map or calculate where the charge is buried. But if he throws enough pebbles and listens for which one lands with a duller thud, he can pinpoint the danger. That is roughly how a new method for predicting El Niño works: the warm current that every few years alters weather across the entire planet.

Previously, the adjoint method was used. It resembles an interpreter who listens to a speech while simultaneously translating it back into the original language to check every word. Such translation requires twice the effort and a separate model that can 'rewind' the calculation. For complex ocean and atmosphere models, building a reverse model is nearly impossible.

The new approach does away with the interpreter. Instead of an exact calculation, it runs random samples: the computer repeatedly tweaks the initial conditions and watches how fast the warm patch grows. Each sample is like a tossed pebble. From the sound of hundreds of pebbles, you can locate the mine without ever seeing it.

A test on the classic Zebiak–Cane model showed that the pattern of warm and cold patches matches that from the adjoint method. A warm patch of about 0.2 degrees in the eastern Pacific, a cold one in the center, and the thermocline dips by a hundred meters. The Niño 3.4 index dynamics over a year differ by less than a hundredth of a degree. Meanwhile, the calculation takes 3 seconds instead of 15: five times faster.

The blind sapper has learned to find the mine faster than a sighted engineer with blueprints. And this skill is not limited to El Niño: the method can be applied to any forecast model where a reverse calculation is too complex. Pebbles tossed at random sometimes work better than exact formulas.