A mosaic of large colored tiles. Each tile is the average precipitation over an entire month for an area the size of a small country. For a general understanding, the picture is fine, but if you need to know whether a downpour hit a specific city yesterday, it's useless. Global climate models work exactly like this: they give a blurry picture because computing every gust of wind and every cloud for the entire Earth is too expensive.
A new approach, named Longwang, tackles the problem differently. It takes a coarse monthly precipitation map (squares about 100 km on a side) and turns it into a daily one with cells of roughly 10 km. The secret is that the neural network doesn't learn from 'bad-good' pairs but from the very nature of rain. It absorbs physical patterns: how moisture is carried by wind, how precipitation depends on terrain, what typical rain fronts look like. Then, looking at the blurry image, it fills in plausible details — much like an artist who reconstructs a full picture from a sketch.
The most impressive part is that the method works where others falter: it reproduces rare but intense downpours, which are the ones most often causing floods. On historical data and even in future warming scenarios, it preserves the spatial structure of precipitation and its temporal connectivity — rain doesn't appear chaotically but moves across the map meaningfully.
Thus, the coarse mosaic of climate models gains sharpness, allowing us to peer into the daily weather of the distant future, and prepare for it today.
The name Longwang refers to the Chinese dragon king of rain, which perfectly captures the essence: the neural network 'conjures' detailed precipitation from coarse climate data.
In Kim Stanley Robinson's novel 'The Ministry for the Future,' precise local prediction of heat waves and floods changes global politics — Longwang takes a step toward such a reality.