For decades, global atmospheric models resembled artists forced to paint planetary storms with coarse brushstrokes: a resolution of 25–100 km forced them to 'guess' convection through approximate schemes, introducing systematic errors in rainfall and climate forecasts. Even John von Neumann, while laying the foundations of numerical weather modeling, asserted that the key to accuracy lay in computational power. Yet even today, global storm-resolving models that explicitly reproduce kilometer-scale cells devour around a megawatt-hour of energy per simulated day, like a heavy steamroller flattening the planet at the cost of mountains of carbon dioxide.

STRATA flips this strategy. Instead of hauling the entire atmosphere's burden at once, it acts like a mosaic master: it learns from tiny tiles — local spatial fragments of 128×128 grid cells (~576 km) with a 10-minute time step. Here, physics is an ally: in 10 minutes, acoustic waves in the atmosphere travel only about 200 km, so a local fragment contains almost everything needed to predict its own evolution. But there’s a deeper synchronization too: in that very interval, a cloud droplet grows to the size at which it begins to fall — the model unconsciously catches the natural rhythm of precipitation formation. The architecture absorbs multidimensional dynamics: a 3D patch embedder captures vertical fluxes, while a novel stereographic positional encoding method, StereoRoPE, accounts for Earth’s sphericity and the influence of gravity and solar radiation as external forcings. Convective fields, with their high entropy and tendency for chaotic bursts, are smoothed by a special pixel de-aliasing decoder, like a mosaicist carefully polishing the edges of each smalti tile to prevent noise in the overall picture.

The results exceed expectations: the model generates stable 24-hour forecasts with realistic kilometer-scale dynamics — from the spiraling arms of tropical cyclones to island convection and orographic rainfall pouring on windward slopes. Precipitation power spectra are nearly indistinguishable from the reference SCREAM model, and the key metric Fractions Skill Score confidently beats persistence. But the most staggering breakthrough is energy efficiency: 48 simulated days per megawatt-hour, about 50 times better than classical approaches. On 512 H100 accelerators, STRATA delivers 741 simulated days per real day, while consuming tens of times less electricity and, consequently, cutting emissions of that very CO₂ that drives climate warming.

Such savings move global storm-resolving modeling from a niche computational feat into an accessible tool. It opens the door to ensemble calculations on centennial scales, enabling quantitative assessment of the changing frequency of extreme downpours and the contribution of cloud feedbacks — the main uncertainty in climate projections. The next step is hybridization with coarse-grained global models and advancement toward 100–200 meter resolution, where low clouds become visible, whose balance between reflecting sunlight and trapping heat has so far eluded computation. If the numerical models of the past were heavy hand-woven tapestries, STRATA offers a lightweight mosaic that breathes and lives in real time.