Google DeepMind published WeatherNext 3 on 3 September, and the number that matters is not the one in the headline. The headline number is resolution: a native 5-kilometre grid, refreshed hourly, roughly five times sharper than WeatherNext 2’s 25-kilometre, six-hour output. The number that matters is the error bar — the Continuous Ranked Probability Score, or CRPS, which the team reports improving by up to 60% against NASA’s IMERG satellite precipitation product. CRPS is a probabilistic score: it grades not a single prediction but the shape of the whole distribution the model emits. Read that carefully. WeatherNext 3 is not promising rain at 14:00. It is telling you how confident it is — and it has measurably gotten better at being honest about its own uncertainty.

The lineage is worth naming. Earlier AI forecasters, WeatherNext 2 among them, learned from numerical weather prediction: supercomputer physics simulations that carry a six-hour data lag. WeatherNext 3 does something structurally different — it trains directly on raw geostationary satellite mosaics and sparse ground-station observations, feeding a single Functional Generative Network mesh transformer. Learning from the observation rather than the simulation is the shift. It is why the model refreshes every hour instead of every six, and why it resolves the thermal fingerprint of a valley that a 25-kilometre grid smears into one flat pixel. Brightband’s independent live leaderboards — not Google’s own benchmarks — put it at the top, which is exactly the replication this desk asks for before believing a launch.

←TODAY: Sept 2026 — an AI model resolves Swiss valley weather hourly at 5 km, free to query in Earth Engine.
→3012: microclimate becomes a queryable design input, every fäcade tuned to its own gravity-well of terrain and sun.
Fulcrum: the forecast is trustworthy only because it now reports how unsure it is — resolution without the error bar is decoration.

Now draw the line to the building site. For a Swiss office the interesting variables are the ones the announcement buries near the end: 100-metre wind speed at roughly turbine height, plus high-resolution cloud cover and surface solar radiation. That is a microclimate layer a fäcade engineer, a Minergie modeller, or a renewables developer has never had at 5-kilometre resolution for free. Coastlines, valleys and mountain ranges — where temperature and humidity swing over a few kilometres — are precisely where the old grids failed, and precisely where most of Switzerland’s built environment sits.

Atelier: For a Büro that models building energy or specifies renewables, the shift is from regional-average weather files to site-specific probabilistic forecasts you can query rather than merely download. The trap is the one PAZ’s own concept panel on nature-inspired optimisers names: a good forecast, like an evolved structural section, is a hypothesis, not a certificate — it returns a likely answer, never a proven one. Monday move: pull one hour of WeatherNext 3 for a live project’s exact coordinates out of Earth Engine and diff it against the Meteonorm or DesignBuilder file you currently trust, then decide — with the CRPS in hand — how much of the gap is real signal.

Hack: Turn a turbine-height wind forecast into the number that actually sizes a grid connection — power. Wind power scales with the cube of velocity, which is why a 5% error in forecast wind speed becomes a ~16% error in predicted output, and why the forecast’s accuracy is worth doing arithmetic over.

rho, area, v = 1.225, 12000, 9.2   # air density, rotor swept area m2, 100m wind m/s
p_watts = 0.5 * rho * area * v**3   # kinetic power in the wind
cp = 0.42                           # realistic turbine capture coefficient
print(round(p_watts * cp / 1e6, 2), "MW")   # ~2.4 MW

Nudge v by half a metre per second and watch the megawatts move — that cubic sensitivity is the whole reason resolving wind at 100 metres instead of 10 kilometres changes a project’s numbers.

The atmosphere keeps a floor of genuine unpredictability, and the model’s honesty about that floor is its best feature, not a flaw. So treat WeatherNext 3 the way you would any wonder-material fresh from the bench: specify it into a workflow only after you have read its error bar and checked it against a source you already trust. Pull the data, diff it, and keep one number in your pocket for every claim it makes.

Source: blog.google