A frozen AI weather model becomes the atmosphere of an Earth system model: about 1,000 simulated years per day on one GPU, with El Niño near observed amplitude

An AI weather model never trained for climate simulation can serve, frozen, as the atmosphere of an Earth system model and keep the coupled system stable for a thousand years. Earlier AI climate models were mostly driven by prescribed sea surface temperature and simulated only the atmosphere; existing AI atmosphere-ocean couplings either produced weak El Niño or learned it from a physics-based climate model, and none routed the land's water to rivers or ran from emissions through a carbon cycle to exposed populations. In the single-member 1000-year control, global-mean 2-m temperature trends per century have standard errors and 100-year means that stay within 0.04 K of each other, and the 2-m temperature and SST errors remain 0.64–0.65 K and 0.37–0.39 K in every century; the same atmosphere driven by observed SST gives 0.57 K and 20.7 m, against 1.69 K and 33.3 m for the predecessor Rescene.

The couplings are added one at a time, and each rung shows what it contributes. Over a climatological ocean the AI atmosphere keeps the mean climate of its relaxation target but its year-to-year variability collapses: annual-mean 2-m temperature varies by a median 0.34 of ERA5 over the ocean and 0.25 in the global mean. In 30-year, two-member controls on the same frozen network, a mixed layer and sea ice raise the ocean variability ratio to 0.53; an equatorial wave ocean with slow wind stress gives a Niño-3.4 standard deviation of 0.40 K; adding the ERA5 SST-equation residual as stochastic forcing raises it to 0.85 K with a Z500 teleconnection of 0.65; adding the land's moisture flux gives Niño-3.4 of 1.04 K and a Z500 teleconnection of 0.81, matching the 0.81 of the same atmosphere driven by observed SST, while the 2-m temperature and Z500 errors stay within 0.60–0.65 K and 20.5–21.1 m at every rung.

The coupler YAHO reads and writes the frozen network's state by variable name, passing heat, water and momentum between the AI atmosphere and the physical surface components every 6 hours and keeping a ledger of every exchange. Run freely, the frozen network diverges after about six months, and it has never seen the soil and carries no precipitation. A Fourier filter poleward of 45 degrees, a stratospheric sponge and a relaxation with a 120-day time scale keep the network bounded; a 5.9-million-parameter diagnostic network supplies precipitation; the 10-m wind is moved toward a relation with the 850-hPa wind fitted on ERA5; and the land's evaporation anomaly enters the lowest 150 hPa with a cap of 60 W m-2 of latent heat. All components run on the GPU with the network, and a two-member 100-year control takes 4.9 hours on one NVIDIA A40.

Emission-driven scenario runs carry heatwave exposure through to populations and separate the shares from a warmer climate and from the number of people. Between consistent worlds, emissions set three-quarters of the difference in heat exposure; within the medium emissions scenario, the number of people sets four-fifths of a difference less than half as large. Heatwave exposure in 2081–2100 reaches 104, 323 and 487 billion person-days per year in the very low, medium and high worlds; the very low and high worlds differ by 383 billion person-days, of which 290 billion come from the warmer climate and 93 billion from the number of people. Runoff routed along the DDM30 network reaches 212 GRDC gauges with a median seasonal-cycle correlation of 0.79.

The demonstrated scope is a frozen Sonny-ENS-S atmosphere coupled to an anomaly mixed layer, an equatorial wave ocean, sea ice, a soil-water bucket, rivers and reservoirs, a carbon cycle and an energy balance, evaluated in 100-year and 1000-year controls, 1979–2022 historical runs, CO2-doubling experiments and 2022–2100 scenario runs. The next step explicitly proposed in the source is coupling a three-dimensional ocean to the same frozen atmosphere.

The mean climate is held near the ERA5 climatology by the 120-day relaxation and the flux adjustment, so its errors measure how far the coupled model departs from that target; the network's training covers 1979–2018, so the daily statistics of the controls are in-sample with respect to its training; and the final configuration was chosen among candidates on free-run El Niño and on the May hindcasts, so the 2017–2021 starts lie outside the surface-component regressions but not outside that choice. In the 1000-year control, Niño-3.4 standard deviations over 30-year windows range over 0.68–1.18 K, so a record of a few decades constrains the amplitude only within that range.