Predicting rain in the next ten to ninety minutes is its own problem, separate from forecasting tomorrow's weather. Conventional high-resolution numerical models can do it, but they need repeated data assimilation, initialisation and spin-up before they produce anything, and that lead time is exactly what you do not have when you are deciding whether a road is about to flood.

Machine learning changes the economics. A model learns how storms evolve directly from frequent observations, and once trained it produces a forecast quickly. The heavy computation moves to training time, where waiting is acceptable, instead of prediction time, where it is not.

The authors build a compact radar-only system and aim it at Mumbai, which is a genuinely hard case rather than a convenient one. Monsoon convection, the interaction between land and sea, and rainfall that turns intense over small areas all make short-term prediction there difficult. A method that holds up in that setting is being tested against the conditions that matter.