Over the warm waters of the Indian Ocean, every six to eight weeks, a vast wave of moist air gathers. It moves eastward at a pedestrian pace—about five meters per second—and takes forty to ninety days to circle half the planet. Meteorologists call it the Madden–Julian oscillation, or MJO, and this wave influences when monsoon rains begin in India and when cold snaps hit mid-latitudes. But the wave itself is capricious: it strengthens, then vanishes, and predicting its next move has been painfully difficult.
A new study turns this wave into a puppet theater. Low-dimensional indices—two numbers specifying the wave's phase and amplitude, plus the season and ocean temperature—serve as the strings. The puppeteer is a video diffusion model: a neural network that learns to turn random noise into a sequence of frames, much like a photographer developing a picture from fog. It was trained on sixty years of weather records from ERA5, a consistent set of atmospheric data gathered from satellites, probes, and weather stations. Now, just pull the strings, and the model will sketch sixty years of plausible tropical waves in half an hour.
This puppet theater can stage plays that never occur in nature. You can set eternal summer and see how the wave changes. You can force the ocean to swing between El Niño and La Niña every two years and observe that after the warm phase the wave becomes stronger and extends farther east, exactly as observations show. Or you can remove all strings but one and find out what that single string does to the orchestra.
This approach for the first time links the low-dimensional theory of the MJO to the full complexity of the tropical atmosphere. Previously, scientists either solved equations for a few numbers and lost cloud details, or ran giant models and drowned in minutiae. Now the neural network serves as a bridge: it takes simple commands and returns a rich picture in which both the wave itself and its associated secondary oscillations—Kelvin and Rossby waves, familiar to anyone who has watched ripples on water—are visible. The puppet theater does not replace the real atmosphere, but it lets us peek behind the scenes and understand which strings truly matter.
Generating a 60-year MJO record takes about 30 minutes, comparable in computational cost to models of intermediate complexity. Moreover, the model can create counterfactual scenarios, such as a perpetual MJO with a constant period, which are impossible to observe in nature.
The idea of 'prompting' a climate model resembles weather control in stories where scientists set parameters and obtain predictable yet realistic responses. However, here there is no manipulation of the real atmosphere, only exploration of its virtual twins.