Why do forecasters run the same forecast many times with slightly different starts?

Forecasters run the same forecast many times because weather is chaotic: a tiny difference in the starting conditions can grow into a completely different outcome, so one forecast cannot show the whole range of what might happen. Instead of making a single forecast of the most likely weather, they produce a set, an ensemble, which gives an indication of the range of possible future states of the atmosphere.

The story begins with a shortcut. In 1961 Edward Lorenz restarted a weather simulation from the middle of an earlier run. He entered the initial condition 0.506 from the printout instead of the full 0.506127, and the result was a completely different weather scenario. The differences between the two runs more or less steadily doubled every four days or so, until all resemblance disappeared somewhere in the second month. A rounding error had grown into different weather.

Lorenz first described the effect with a seagull causing a storm, and was persuaded to use a butterfly and a tornado by 1972. For a talk that year, Philip Merilees concocted the title Does the flap of a butterfly's wings in Brazil set off a tornado in Texas? The butterfly does not power or directly create the tornado: its flap is part of the initial conditions, and one set of conditions leads to a tornado while the other does not.

Ensembles apply the lesson. Forecasters account for two sources of uncertainty: imperfect initial conditions amplified by chaos, and imperfections in the model itself. They run many simulations with perturbed starting conditions and varied model parameters. Ideally the real future state falls within the ensemble spread, and the amount of spread is related to the uncertainty of the forecast. Ensemble forecasts began to be prepared by two big centres, ECMWF and NCEP, only in 1992.