The European Union has spent four years quietly building the most ambitious simulation in the history of geosciences. Destination Earth — the digital twin of the planet co-led by ECMWF, ESA and EUMETSAT — is targeting one-kilometre global resolution by the end of the decade. ETH Zurich’s Torsten Hoefler and CSCS director Thomas Schulthess are the computer scientists in the room. The strategy paper they co-authored in Nature Computational Science is the blueprint. It is also, read carefully, a confession: the codes underneath today’s weather models extract only about 5 per cent of peak performance from modern CPUs. Climate science coasted on Moore’s Law for forty years, and the bill has come due.

←TODAY: Destination Earth is mid-build; ETH Zurich and CSCS are rewriting the simulation stack from generic grid kernels up.
→3012: Every Wettbewerb brief opens with a twin query — site climate, twenty-five-year stress envelope, biosphere coupling — before the first sketch.
Fulcrum: The same co-design discipline that lets a planetary twin run at 1 km is the discipline that lets a building twin actually keep up with its sensors.

What the dependency graph looks like

Strip Destination Earth down to its real topology and the picture is brutally simple. Three inputs: observational streams (satellites, ground stations, ocean buoys), human-activity data (energy, agriculture, water), and the priors that come from sixty years of weather modelling. One bottleneck: the gap between a one-kilometre grid and the 5%-of-peak code that has to run on it. Two outputs: a continuously-updated state of the Earth, and a scenario engine — Peter Bauer’s two-metre-dike-in-the-Netherlands case, scaled to every coastal nation that has one.

The interesting failure mode is not that the model will get the weather wrong. It is the dependency you forgot to draw: the single supercomputing centre, the single satellite constellation, the single small team that remembers how the assimilation loop actually closes.

Co-design, or the only way out

Hoefler and Schulthess’s prescription is the part of the paper that translates straight into AEC practice. They argue for co-design — hardware and algorithms developed together — with three specific moves: generic data structures, optimised spatial discretisation of the grid, and adaptive timestep lengths. Crucially, they want the scientific code separated from the architecture-specific code, so the simulation can hop processor generations without a rewrite. Antil and colleagues make the parallel point at structural scale in their MATH-DT report (arXiv:2402.10326): a twin starts from this bridge, not a bridge. Destination Earth is the same insight at planetary scale — and the same insight PAZ has documented in its Digital-Twin-en-Obra and Digital-Twin-en-Ingenieria concept panels: a twin earns its keep by staying synchronised with the asset, not by being clever in isolation.

Why a Zurich morning desk cares

Three reasons. First: NEST in Dübendorf, Virtual Singapore and Helsinki 3D+ already gave us the playbook for an instrumented twin — Destination Earth is what happens when you do it for the atmosphere itself. Second: a 1 km global twin makes any city-scale twin computationally honest about its boundary conditions; you stop guessing at the wind. Third: every infrastructure project after roughly 2028 will be tested against scenarios pulled from this twin. The dike, the wind farm, the new tram corridor — all sit downstream of someone’s ensemble run on a Swiss machine.

Atelier: For PAZ this means one practical shift in how we brief Wettbewerb entries. The climate envelope is no longer a 30-year averaged appendix from MeteoSchweiz — it becomes a queryable scenario surface. The PAZ Grasshopper↔Archicad Library is already designed so that climate-loading parameters arrive as a vector, not a fixed number. When ECMWF opens the Destination Earth ensemble outputs publicly, our parametric flow ingests them and the LOIN doesn’t need to change. Brief that capability into your next BEP.

Hack: This Hack teaches you to feel why 1 km global resolution is the actual hard part — by computing the scale-up yourself. The domain is Math: the cube-law that makes planetary compute exponentially expensive.

EARTH_KM2 = 510_000_000           # planet surface in km²
def cells(res_km, layers=137):    # 137 = ECMWF IFS vertical levels
    return int(EARTH_KM2 / res_km**2 * layers)
print(cells(9), "→", cells(1))    # ~863M → ~70B grid points (81×)

Run that. Then remember the Courant condition: halving the grid spacing forces you to halve the timestep, doubling the run again. The ~160× compute bill that drops out is exactly what the co-design paper is funded to absorb. Read every Destination Earth press release differently now.

Move

This week, pull one ECMWF open-data NetCDF file into your usual geometry workflow (xarray plus Rhino.Inside is a clean route). Even a single day of 0.25° forecast data is enough to teach your fingers what boundary-conditions-from-a-twin actually feel like. Then, more importantly: draw your project’s real dependency graph — not the architecture diagram, the dependency graph. Find the third single point of failure you didn’t know you had. That is the exercise.

Sources & Further Reading

• Primary: ETH Zurich News — Scientists begin building highly accurate digital twin of our planet