A weather forecast is a physics problem. The atmosphere is a thin, turbulent fluid on a rotating sphere, heated unevenly by a star 150 million kilometres away. For fifty years we solved that problem by brute force: discretise the fluid, integrate the equations and buy a bigger supercomputer. According to Google’s own announcement, its new WeatherNext 3 model now delivers 50% more accurate precipitation forecasts a day or more ahead. It combines real-time satellite observations with a learned model to produce high-resolution hourly forecasts without immense supercomputing power.

That line appears in a long roundup Google published about AI in science. Among the other items: AlphaGenome Atlas maps all 9 billion possible single-letter changes in the human genome, Flood Hub forecasts now cover 2 billion people in more than 150 countries, and the breast-cancer study with Imperial College London and the NHS found AI could flag 25% of interval cancers missed in 175,000 women’s mammograms. Most of it belongs to other desks. The rain number lands on a building site.

What was actually measured

Here is the physicist’s question. “50% more accurate” compared with what baseline, on which metric, and at what error bar? The blog post does not say, and a press roundup is not a verification paper. Treat the figure as a claim waiting for its methods section. The trade-off, stated plainly: a learned model is only as good as the past it learned from, and a warming climate keeps producing weather outside that past.

The lineage deserves its credit. Earlier DeepMind weather models such as GraphCast learned from ECMWF’s ERA5 reanalysis. That is decades of global observations rebuilt, cell by cell, by forecasters and data engineers whose names never reach a headline. Every learned forecast runs on their archive. PAZ’s gold reference on Google DeepMind’s science programme tracks the same move: prediction learned from the physical record rather than integrated from scratch.

←TODAY: In 2026 a learned global model claims a 50% precipitation gain at 24 h+, and it already runs inside consumer products.
→3012: Earth’s axis will have precessed about 14° around its 25,772-year cone, and forecasts will still be spherical physics projected onto flat sheets.
Fulcrum: A forecast only helps a site if you know the size of the cell it describes on the ground.

Why the grid matters in Zurich

Sun and Earth set the terms. A forecast lives on a latitude-longitude grid wrapped around a sphere. Global learned models have commonly worked at 0.25°, and a quarter-degree is not the same distance everywhere. At Zurich’s 47.4° N, lines of longitude sit closer together than at the equator, so each cell is about 19 km wide and 28 km tall. One cell can hold both the Limmat valley and the foot of the Uetliberg. Swiss topography does the rest: MeteoSwiss runs its own ICON-based ensemble at kilometre scale precisely because the Alps break coarse grids.

Atelier: For a Bauleitung team scheduling concrete pours, facade glazing or roof waterproofing, a better one-to-two-day rain forecast lowers the cost of every weather-exposed step. The risk is that the office trusts a single number before knowing how it performs in its own valley. Monday move: add two columns to the site diary, forecast rain versus measured rain for each planned pour, and log MeteoSwiss alongside one AI-based forecast for a month before letting either one decide a pour.

Hack: Convert a forecast grid cell into metres on your site before you trust it. A degree of latitude spans about 111 km almost everywhere. A degree of longitude shrinks with the cosine of your latitude, which is the 3D-to-2D projection problem in a single line. Run this for your site’s latitude:

import math
lat, cell = 47.37, 0.25          # Zurich, grid spacing in degrees
dy = cell * 111.32               # km north-south
dx = dy * math.cos(math.radians(lat))  # km east-west
print(f"cell ≈ {dx:.1f} km × {dy:.1f} km")  # ≈ 18.8 × 27.8

If the cell is larger than the valley your crane stands in, read the forecast as a regional signal, not a site measurement.

The late-2070s view is simple. The forecasts that did damage were the ones people specified into schedules before they had checked them against local rain. Put a rain gauge on the hoarding, keep the log and let your own measurements decide how far to trust the model.

Source: blog.google