Rescene-ESM: an Earth system model built around a frozen AI weather model
Authors: Minjong Cheon, Kaoru Kakinuma
Organizations: Department of Computer Science, Korea National Open University, Seoul, Republic of Korea · Frontier Research Institute for Interdisciplinary Sciences, Tohoku University, Sendai, Japan
Earth system models underpin climate projections, but their cost limits how many scenarios, ensemble members and impact pathways can be explored. Machine-learning weather models forecast as skillfully as physics-based models at a fraction of the cost, yet coupled to an ocean they have produced weak El Nino variability or have learned it from climate-model output, and none carries the climate through rivers to the people exposed to it. Here we present Rescene-ESM, an Earth system model whose coupler, YAHO, yokes a frozen 20.8-million-parameter AI weather model to physically formulated ocean, sea-ice, land, river, carbon and energy components constrained by observations. Adding the couplings one at a time shows what each contributes: over a climatological ocean the AI atmosphere varies from year to year by a third of the observed amount over the oceans, and an interactive mixed layer, an equatorial wave ocean, the residual of the observed sea surface temperature (SST) equation as stochastic forcing and the land's moisture flux raise this step by step and bring El Nino and its teleconnections, leaving the mean climate almost unchanged. The full model simulates about 1,000 years per day on one graphics processing unit, holds the observed climate for a thousand years without drift, reproduces the observed 1979-2022 global warming, carries forced change with a physically based structure around the frozen network (transient climate response 1.60 K), routes runoff to 212 river gauges with a median seasonal-cycle correlation of 0.79 and, driven by emissions, separates the heatwave exposure of an SSP1 and an SSP3 world by a factor of 4.7 in 2081-2100, three-quarters of it through the warmer climate.
Figures & tables
Fig. 1: Rescene-ESM and its coupler, YAHO. The frozen AI atmosphere (blue): Sonny-ENS-S steps 82 variables at 1.5 ∘ every 12 h from the state and the coupled SST, under a polar filter, a stratospheric sponge and a 120-day relaxation toward the calendar-month climatology plus the forced change. Around the network (orange): a precipitation network, a low-pass observer of the 10-m wind and the land’s soil-controlled moisture flux. The physical surface (green), fitted mostly to ERA5 over 2000–2016 (Methods): radiation diagnostics, the anomaly mixed-layer ocean with Ekman heat transport, the equatorial wave ocean, the wind stress for the ocean (the network’s weather-scale stress, an SST-forced slow part and westerly wind bursts), the stochastic residual of the SST equation, sea ice, land and rivers, whose routing has no fitted constant; every 6 h YAHO passes it fluxes, stress, rain and evaporation and returns the SST and the land’s moisture, with a ledger for each exchange. Forcing, carbon and energy (gray): emissions scaled by population, a carbon cycle fitted to the 1959–2022 CO 2 record, a two-layer energy balance and a forced-response layer that gives the forced change its pattern, carried around each network step by a gauge transform. People in the loop (pink): daily heat, humid heat, soil water and discharge at 4,925 populated cells. Thumbnails are fields of the model and its inputs.
Fig. 2: What each coupling adds. The couplings are added cumulatively to the atmosphere-only model (SST and sea ice held at their calendar climatology): the anomaly mixed layer and sea ice, the equatorial wave ocean with the SST-forced slow stress and wind bursts, the residual of the SST equation as stochastic forcing and the land’s moisture flux (the full model). Each rung is a 30-year, two-member control scored over model years 3–30; the full model uses the same years of its 100-year control. a , Standard deviation of detrended annual-mean T2m in ERA5 1979–2018. b–f , log2 ratio of the same standard deviation in each rung to that of ERA5; titles give the median ratio. g , Year-to-year variability relative to ERA5: global-mean T2m, the median over ocean and over land cells, and sea-ice concentration. h , Niño-3.4 standard deviation (dotted, ERA5) and the pattern correlation with ERA5 of the December–February Z500 and T2m regressions on each run’s own Niño-3.4. i , Annual-mean T2m and Z500 RMSE. Open symbols are references, not rungs: Rescene (left; its 100-year run, no ocean and hence no El Niño) and the same atmosphere driven by observed SST over 1979–2022 (right; AMIP, eight members).
Fig. 3: The coupled mean climate. Annual-mean departure of the 100-year control (model years 3–100, two members) from ERA5 1979–2018: a , 2-m temperature; b , SST; c , MSLP; d , 500-hPa geopotential height; e , 200-hPa zonal wind; f , land precipitation. Titles give the area-weighted RMSE and the pattern correlation with ERA5.
Fig. 4: Historical warming, 1979–2022. a–c , Annual means over populated land of the daily maximum ( a , c ) and minimum ( b ) temperature, area-weighted ( a , b ) and population-weighted ( c ), as anomalies from 1979–1988, for ERA5 and the two members of the first historical run; legends give linear trends. d , Global T2m of ERA5 and of the energy balance with λ=1.96 W m -2 K -1 ; 2011–2022 (shaded) was withheld from the temperature constraint. e , f , Trend of the annual daily-maximum temperature over populated land in ERA5 and in the first member.
Fig. 5: Structure of the CO 2 response. Abrupt doubling of CO 2 against a control at fixed CO 2 , last five of 20 years. The vertical structure and the humidity change are set by the forced-response layer (moist adiabat in the tropics, fixed relative humidity) and carried around the network by the gauge transform; panels b and c show that the coupled model keeps them. a , Latitude–pressure section of zonal-mean warming per kelvin of global T2m warming. b , Tropical (20 ∘ S–20 ∘ N) warming profile per kelvin of tropical surface warming, with the moist adiabat. c , Tropical specific-humidity change per kelvin of global T2m warming, with the change at constant relative humidity. d , Zonal-mean T2m warming per kelvin of global warming. e , Annual global warming of the model and of the energy balance.
Fig. 6: Emission-driven futures and the people exposed. The CMIP7 medium emissions scenario with emissions scaled by the SSP1, SSP2 and SSP3 population pathways. a , Atmospheric CO 2 from the carbon cycle, with Mauna Loa and the scenario’s MAGICC concentration. b , Warming above 1850–1900. c , Heatwave days per person, ERA5 and the historical runs, then the three pathways. d , Heatwave exposure and the split of the SSP3–SSP1 difference into the part from the number of people and the part from the warmer climate. e , Change in heatwave days per year, 2081–2100 against 1986–2005, under SSP2. f , Where the additional exposure of SSP3 against SSP1 lives.
Extended Data Fig. 1: Stability of the 100-year control. Annual means of a , global T2m; b , open-ocean SST; c , Arctic and d , Antarctic sea-ice extent, with the ERA5 mean and ±2 standard deviations (gray); titles give the linear drift per century. e , Global top-of-atmosphere imbalance. f , Zonal-mean Z500 anomaly.
Extended Data Fig. 2: Taylor diagrams [ 40 ] . a , Time-mean climate as departures from the zonal mean. b , Variability patterns: the leading empirical orthogonal functions of extratropical MSLP (NAM, SAM), the DJF regressions of T2m and Z500 on Niño-3.4 (ENSO) and the interannual standard deviation of T2m and MSLP (IAV). Points: Rescene-ESM and the no-noise ablation (100-year controls) and the same atmosphere driven by observed SST (AMIP, 1979–2022); ERA5 is the reference at correlation 1 and normalized standard deviation 1.
Extended Data Fig. 3: The ocean alone under prescribed winds , the experiments of KIST-Ocean [ 41 ] . a , Equatorial (2 ∘ S–2 ∘ N) thermocline anomaly after a 5-day westerly wind burst over 130–170 ∘ E, 10 ∘ S–10 ∘ N; dashed, the Kelvin-wave speed of 2.7 m s -1 . b , Niño-3.4 SST response to the westerly and an equal easterly burst. c , Thermocline at the center of a persistent cyclonic or anticyclonic wind-stress curl at 15.5 ∘ N, 135.5 ∘ E. d–g , Niño-3.4 SST of the ocean driven by ERA5 wind stress and surface fluxes and by climatological forcing, from the ERA5 state at the start, for the 2014/15 and 2015/16 El Niño (inside the 2000–2016 fit period) and the 1997/98 and 2018/19 El Niño (outside it); titles give the correlation with ERA5.
Extended Data Fig. 4: El Niño in the coupled control. a , Monthly Niño-3.4 SST anomaly of the control (first 30 years shown) and of ERA5 1979–2008. b , Normalized spectra with the 95% range of a fitted first-order autoregressive process. c , Autocorrelation. d–i , Regression on Niño-3.4 of SST ( d , e ), zonal wind stress ( f , g ) and thermocline depth ( h , i ; 20 ∘ C isotherm in ORAS5, the wave ocean’s h in the model) per standard deviation of the index. j , k , Lag regression of equatorial thermocline depth on Niño-3.4; negative lags: h leads. l , Standard deviation of Niño-3.4 by calendar month. m–o , The links of the Bjerknes loop: Niño-4 zonal stress per kelvin of Niño-3.4 SST, Niño-3 thermocline depth per unit of Niño-4 stress and Niño-3 SST per 10 m of Niño-3 thermocline depth.
Extended Data Fig. 5: ENSO teleconnections. December–February T2m ( a , c ) and Z500 ( b , d ) regressed on the Niño-3.4 index in ERA5 and in the control. The model’s patterns follow its own ENSO; no statistical teleconnection is imposed.
Extended Data Fig. 6: ENSO hindcasts. Eight-member hindcasts started on 1 May of each year 1997–2021 (25 starts), scored raw against ERA5. a , Anomaly correlation of the ensemble-mean Niño-3.4 index and of persistence. b , RMSE of the ensemble mean, of persistence and of climatology, and the ensemble spread. c–f , Members (thin) and ensemble mean (thick) for the 1997/98 and 2015/16 El Niño, the 2010/11 La Niña and the 2020/21 La Niña, with ERA5 (black). Starts in 2017–2021 are outside every regression of the surface components; the configuration was chosen partly on these hindcasts.
Extended Data Fig. 7: Daily precipitation and the Indian monsoon , 40-year control against ERA5 1979–2018. a–c , Probability density of daily precipitation over the globe, the tropical ocean (10 ∘ S–10 ∘ N) and the contiguous United States; dashed and dotted lines mark the 99th and 99.99th percentiles; titles give the 99th percentiles of ERA5 and the model. d , e , Climatological daily precipitation over 70–100 ∘ E by latitude and day of year. f , Indian monsoon precipitation over 10–25 ∘ N (15-day mean).
Extended Data Fig. 8: Patterns of the CO 2 response. Years 10–19 of abrupt 2 × , 3 × and 4 × CO 2 against the fixed-CO 2 control. a–c , T2m change per kelvin of global warming (titles give the global change). d–f , Precipitation change (titles give the global change and the hydrological sensitivity).
Extended Data Fig. 9: Rivers. Discharge of the historical run at 212 GRDC gauges against observations (climatology from 1990). a , Mean discharge; shading, a factor of 1.5. b , Correlation of the seasonal cycle at each gauge. c–f , Seasonal cycles of the four largest basins in the set at their main-stem gauges. No constant is fitted to discharge.
Extended Data Fig. 10: Reservoirs. The historical runoff routed without and with the 7,302 ISIMIP3a dams (Hanasaki release rule; no constant fitted to discharge), against GRDC climatologies from 1990. a , Seasonal-cycle correlation without and with reservoirs, colored by the upstream capacity relative to the annual flow. b , Distribution of the seasonal range relative to the mean flow at the 129 gauges with upstream capacity above 10% of the annual flow. c–f , Seasonal cycles of the most regulated large rivers; the purple lines add water consumption (sensitivity run, Methods).
Extended Data Fig. 11: Emission scenarios. CMIP7 very low, medium and high emissions from 2022, each with its own non-CO 2 forcing and aerosol pattern and, for the exposure, its own population pathway (SSP1, SSP2, SSP3). a , Atmospheric CO 2 from the carbon cycle, with Mauna Loa and each scenario’s MAGICC concentration (dashed). b , Global warming above 1850–1900 of the energy balance (thick) and the coupled model’s global-mean T2m (thin). c , Heatwave days per person (9-year running mean). d , e , T2m and heatwave days per year in 2081–2100, high minus very low emissions. f , Heatwave person-days per year in 2081–2100, SSP3 world minus SSP1 world, split into the part from the warmer climate and the part from the number of people (Methods).
Rescene
Rescene-ESM
Atmosphere network
Sonny (frozen), deterministic 6-h step, 69 variables, with a 0.4-million-parameter wrapper
Sonny-ENS-S (frozen), 20.8 million parameters, 12-h step with Monte Carlo dropout, 82 variables, SST as an input
Free-run stability
Slow clock: the large scales blended toward the fixed 1979–2018 calendar climatology every 6 h (weight 0.85 after 10 days)
Polar filter, stratospheric sponge and a 120-day relaxation toward the calendar climatology plus the forced change
Day-to-day weather
Band-limited stochastic perturbation
The network’s own dropout ensemble
Ocean and sea ice
Prescribed calendar climatology
Anomaly mixed layer, equatorial wave ocean, stochastic SST residual, sea ice with thermodynamics and drift
Land and rivers
None
Soil-water bucket with groundwater and stomata, moisture flux into the boundary layer, rivers and reservoirs
Forced change
None
Carbon cycle, two-layer energy balance, forced-response layer, gauge transform
Supplementary Table 1: From Rescene to Rescene-ESM. Rescene is the predecessor that made a frozen weather model stable for a century without interactive components. Metrics: 100-year controls against ERA5 1979–2018 (Rescene-ESM: model years 3–100).
Experiment
Length, members
Forcing and use
Control
100 yr, 2
No forced change; mean climate, variability, ENSO
1000-year control
1000 yr, 1
No forced change; annual-mean fields and monthly Niño-3.4; long-term stability and ENSO modulation (Supplementary Fig. 20)
Daily control
40 yr, 1
As the control, with a daily record; precipitation, blocking, cyclones
Supplementary Table 2: Experiments. All runs use the same frozen network, coefficients and flux adjustment, except the no-noise ablation, which has its own flux adjustment.
T2m / Z500
T2m variability
Sea-ice
Niño-3.4
ENSO
RMSE
Ocean
Land
Global
variability
(K)
Z500 r
Rescene
1.69 K / 33.3 m
1.19
0.63
0.48
–
–
–
Atmosphere only
0.60 K / 20.5 m
0.34
0.65
0.25
0.00
0.00
–
+ mixed layer and sea ice
0.65 K / 21.0 m
0.53
0.67
0.28
0.47
0.01
–
+ wave ocean, slow stress, bursts
0.65 K / 21.1 m
0.56
0.72
0.25
0.48
0.40
0.03
+ SST noise
0.65 K / 21.0 m
0.68
0.73
0.45
0.50
0.85
0.65
Supplementary Table 3: Coupling ladder. 30-year, two-member controls with the couplings added cumulatively (model years 3–30; the full model: the same years of the 100-year control; Rescene: its 100-year run). Variability: standard deviation of detrended annual means relative to ERA5 1979–2018 (median over ocean or land cells; global mean; sea-ice concentration). ENSO teleconnection: pattern correlation with ERA5 of the DJF Z500 regression on each run’s own Niño-3.4. AMIP: the same atmosphere driven by observed SST, 1979–2022. Rescene’s value for single ocean cells reflects its stochastic weather perturbation, which varies each cell while its global mean varies by half the observed amount.
Metric
ERA5
Rescene-ESM
No-noise ablation
AMIP / Rescene
T2m RMSE (K)
–
0.64
0.66
0.57 / 1.69
Z500 RMSE (m)
–
21.1
21.1
20.7 / 33.3
MSLP RMSE (hPa)
–
4.16
4.12
4.25 / 0.91
U200 RMSE (m s -1 )
–
1.73
1.80
–
Land precipitation / evaporation RMSE (mm day -1 )
–
0.67 / 0.34
0.66 / 0.37
–
Global T2m drift (K per century)
0
+0.00
+0.02
–
Supplementary Table 4: Coupled climate against ERA5 , 100-year control (model years 3–100). The same atmosphere driven by observed SST over 1979–2022 (AMIP, eight members) and Rescene are shown for reference; AMIP teleconnections are by the measure of Fig. 2, on which Rescene-ESM has 0.84. River rows: historical runs.
Lead (months)
3
6
8
10
12
Anomaly correlation, 1997–2021
0.83
0.77
0.72
0.69
0.53
RMSE (K), 1997–2021
0.44
0.94
1.04
0.78
0.62
Ensemble spread (K)
0.24
0.61
0.79
0.66
0.50
Persistence RMSE (K)
0.84
1.16
1.34
1.01
0.70
Climatology RMSE (K)
0.68
1.09
1.30
0.99
0.57
Anomaly correlation, 2017–2021
0.48
0.72
0.79
0.84
0.94
Supplementary Table 5: Niño-3.4 hindcast skill (ensemble mean of eight members, starts on 1 May). Values for 2017–2021 (five starts) are outside every regression of the surface components.
Diagnostic
ERA5
Rescene-ESM
Figure
El Niño / La Niña events per 40 years
10.0 / 10.9
8.2 / 11.0
SI 9
El Niño mean / maximum peak (K)
1.76 / 2.84
1.99 / 4.37
SI 9
Pacific tripole index s.d. / 13-year low-pass s.d. (K)
Supplementary Table 6: Diagnostics in the format of AI coupled-model studies (ACE2-SOM, SamudrACE, DLESyM, ACE2-NEMO). Control runs; ERA5 1979–2018 (daily) or 1979–2022 (ENSO). ED, Extended Data Fig.; SI, Supplementary Fig.
2 × CO 2
3 × CO 2
4 × CO 2
Global T2m change (K)
1.60
2.51
3.13
Global precipitation change (%)
+2.4
+2.8
+2.6
Hydrological sensitivity (% K -1 )
1.52
1.11
0.85
Land / ocean warming ratio
1.30
1.24
1.21
Arctic (north of 60 ∘ N) amplification
2.40
2.14
2.11
SSP1 population
SSP2 population
SSP3 population
Supplementary Table 7: Forced response and futures. Abrupt experiments: years 10–19 against the fixed-CO 2 control. Futures: 2081–2100, CMIP7 medium emissions scaled by population, and each scenario’s own emissions with its own population.
ERA5
Rescene-ESM
Noise ×0.5
Noise ×0.7
Gulf Stream / Kuroshio SST s.d., cells (K)
0.89 / 0.80
0.32 / 0.35
0.45 / 0.52
0.56 / 0.57
Gulf Stream / Kuroshio SST s.d., box means (K)
0.50 / 0.52
0.23 / 0.23
0.29 / 0.37
0.37 / 0.37
Extratropical (30–60 ∘ ) monthly SST s.d. ratio
1
0.52
0.65
0.72
Interannual T2m s.d. ratio, ocean / land
1 / 1
0.68 / 0.76
0.72 / 0.78
0.70 / 0.75
Niño-3.4 s.d. (K)
0.90
1.04
0.99
0.86
T2m / SST RMSE (K)
–
0.64 / 0.37
0.63 / 0.37
0.64 / 0.37
Supplementary Table 8: Extratropical SST noise (sensitivity). 30-year controls, first member, model years 3–30. The 80 leading modes of the SST-equation residual over 25–65 ∘ latitude, fitted as the 20 tropical ones, are added to them at 0.5 and 0.7 of their fitted size. Boundary currents: s.d. of monthly SST anomalies averaged over the cells of the Gulf Stream (35–45 ∘ N, 75–45 ∘ W) and Kuroshio (30–40 ∘ N, 140–170 ∘ E) Extension boxes, and of the box means. Ratios: median over cells of the model’s s.d. over ERA5’s (1979–2022); mean errors against ERA5 1979–2018.
Very low, SSP1
Medium, SSP2
High, SSP3
Warm-spell index
Days per year, 2023–2050
9.4
9.9
10.2
Ratio to 1991–2020, area-weighted
2.20
2.33
2.38
Ratio to 1991–2020, population-weighted
2.02
2.01
2.10
Heatwave exposure, 2023–2050, against 57.7 in 1991–2020
Exposure
139
144
150
Supplementary Table 9: Heat exposure on the metrics of multi-model studies. Warm-spell index as in Dash et al. (npj Nat. Hazards, 2025; factor of 2–3 over 2021–2050 in 19 CMIP6 models), over the populated cells; in 1991–2020 the model has 4.27 and ERA5 5.00 days per year. Exposure is the product of the period-mean heatwave days and population, in billion person-days per year (hence 486 against the 487 of Supplementary Table 7). Each rise is given with its split into climate / population / their product, in percent. India, China, Pakistan and Bangladesh hold 73% of the people exposed at 3 K of warming in Kong et al. (Lancet Planet. Health, 2026); on the 1.5 ∘ grid, ERA5 exceeds the wet-bulb limit of 30.6 ∘ C quoted there on at most 0.3 days per year in any populated cell.
Supplementary Fig. 1: What sets the coupled ENSO. a , Standard deviation of the 30-day residual of the mixed-layer SST equation on ERA5 2000–2016, the stochastic tendency of the model. b , Variance of the equatorial Pacific zonal stress at synoptic and intraseasonal scales relative to ERA5, in 120-day free runs of the network without a 10-m wind closure, with the closure on the full departure (ablation) and with it acting on the departure’s running mean over 10 days and over 5 days (Rescene-ESM). c , Niño-3 feedback rates of the SST equation (thermocline, damping and net growth) and the Niño-3 standard deviation with observed inputs, in the no-noise ablation and in Rescene-ESM. d , Seasonal phase locking, e , spectrum and f , hindcast error and spread of the ablation and of Rescene-ESM.
Supplementary Fig. 2: Energy and water budgets of the 100-year control. a , Global top-of-atmosphere imbalance (annual means) and ERA5 for the same calendar years. b , Ocean heat content change. c , Land water budget over 100 years; every term is a ledger of the run. d , River budget.
Supplementary Fig. 3: Sea ice. a , d , Seasonal cycle of Arctic and Antarctic sea-ice extent (concentration above 0.15). b , c , e , f , Sea-ice concentration at the seasonal maxima, March in the Arctic and September in the Antarctic.
Supplementary Fig. 4: Seasonal cycle by latitude. Zonal-mean calendar-month SST ( a ) and T2m ( c ) of ERA5 and the departure of the control ( b , d ).
Supplementary Fig. 5: Annular modes and interannual variability. a , b , d , e , Leading empirical orthogonal function of monthly MSLP anomalies over the extratropical Northern (NAM) and Southern (SAM) Hemisphere, regressed on the principal component. c , f , Ratio of the interannual standard deviation of T2m and MSLP to ERA5, mapped as log2 ; titles give the median ratio.
Supplementary Fig. 6: ENSO and heat over the full control. a , Monthly Niño-3.4 anomaly over the 100 years, with ERA5 1979–2022. b , Spectra (40-year Hann windows, half overlapping). c , Standard deviation by calendar month. d , Annual heat content of the ocean mixed layer; the title gives its linear drift.
Supplementary Fig. 7: ENSO feedbacks. a , c , Regression of the 10-m zonal wind on the zonal SST gradient (the Bjerknes wind feedback; titles give the Niño-4 mean). b , d , Regression of the upward turbulent heat flux on SST (the heat-flux damping; titles give the Niño-3 mean).
Supplementary Fig. 8: Mean precipitation and its ENSO response. a , ERA5 annual-mean precipitation, 1979–2018. b , Departure of the 40-year control. c , d , December–February precipitation regressed on Niño-3.4 in ERA5 and in the control; the title gives the pattern correlation over 30 ∘ S–30 ∘ N.
Supplementary Fig. 9: ENSO events. a , Niño-3.4 spectra of 40-year blocks (thin) and their mean (thick). b , Distribution of monthly Niño-3.4 anomalies. c , El Niño and La Niña events per 40 years (five-month running mean beyond ±0.5 K for at least five months); numbers give the mean peak.
Supplementary Fig. 10: Decadal variability and the western boundary currents. a , 13-year low-pass Pacific tripole index. b , Its spectrum. c , d , Spectra of SST anomalies averaged over the Gulf Stream and Kuroshio Extensions; legends give the standard deviation of the box mean and of individual cells. e , Distribution of the Gulf Stream box anomaly.
Supplementary Fig. 11: Sea-ice variability. Standard deviation of monthly sea-ice concentration anomalies in ERA5 and in the control, Northern ( a , b ) and Southern ( c , d ) Hemispheres; titles give the area mean and its ratio to ERA5. Contours: climatological ice edge.
Supplementary Fig. 12: Northern Hemisphere blocking , two-dimensional index on daily Z500, 40-year control against ERA5 1979–2018. a , b , Frequency of blocked days in December–February. c , Frequency over 45–70 ∘ N by longitude, winter and the whole year.
Supplementary Fig. 13: Western North Pacific tropical cyclones on the network’s own 12-hourly states (ERA5 at 00 and 12 UTC). a , Seasonal cycle of cyclone centers below 1000 hPa (20-day mean). b , Cyclones per year by minimum MSLP. c , d , Density of centers below 994 hPa.
Supplementary Fig. 14: Historical surface climate, 1979–2022. Annual changes from 1979–1988 of a , global T2m; b , ice-free SST; c , d , upward latent and sensible heat fluxes over the ice-free ocean, for ERA5 and the first member of each of the three historical runs.
Supplementary Fig. 15: Land–atmosphere coupling. Correlation between monthly anomalies of root-zone soil water and of the daily-maximum temperature in the same month ( a , b ) and one month later ( c , d ), local summer, ERA5 against the historical runs. e , f , Their distributions over land cells.
Supplementary Fig. 16: Transient response. a , Global T2m change under 1%-per-year CO 2 (with CO 2 , dashed) and the abrupt experiments; the transient climate response is the mean of years 61–80 (shaded). b , Global precipitation change. c , Precipitation against temperature change, with the regression of the 1%-per-year run.
Supplementary Fig. 17: Return from a warm perturbation. The control restarted with +2 K open-ocean SST. a , Monthly difference in open-ocean SST and global T2m from the control; the title gives the e-folding time of the SST difference. b , c , SST difference in years 1 and 3.
Supplementary Fig. 18: Future rivers and hazards. a , Change in mean discharge and b , in the annual maximum daily discharge at 241 gauges, 2081–2100 against 1995–2014, SSP2 population, with reservoirs. c , River-flood days and d , drought months per person (five-year running means) in the historical runs and the three population pathways.
Supplementary Fig. 19: Component checks against observations. a , Elasticity of evaporation to stomatal resistance from Penman–Monteith inversions at 64 FLUXNET sites by biome; bars, medians. b , The elasticity against the GLEAM transpiration share at each site, with the fit the land uses. c , Change of land evaporation at 560 ppm from stomatal closure alone, with the atmosphere held fixed. d , Atmospheric CO 2 of the carbon cycle driven by historical emissions from its 1750 spin-up, with its one fitted constant, against Mauna Loa, 1959–2022.
Supplementary Fig. 20: The 1000-year control (one member, model years 3–1000). a , Global T2m anomaly, annual and 30-year running mean; the title gives the trend and its standard error, adjusted for lag-1 autocorrelation. b , Niño-3.4 standard deviation in 30-year windows and 100-year blocks, with ERA5 1979–2022. c , Niño-3.4 power spectrum of 40-year blocks (thin) and their mean (thick), with ERA5 1979–2022. d , Annual-mean T2m and SST error against ERA5 1979–2018 in each century. e , Power spectrum of annual global T2m, with its variance by time scale. f , T2m, last 100 years minus the first 100 years.
Climate modeling uses physical simulations and data-driven methods to understand, predict, and project Earth’s climate and its response to external forces.
Machine-learning weather models now match or exceed operational numerical prediction in the medium range, but when they are run freely for years they drift, blow up or lose their variability, and dedicated climate emulators are expensive to train. Here we present Rescene, a 0.4 M-parameter wrapper that turns a frozen 1.5∘, 6-hourly, 69-variable weather model into a stochastic climate emulator. A deterministic slow clock anchors the frozen model to a calendar climatology, and a band-limited AR(1) perturbation, whose spectral envelope is fixed to the observed anomaly spectrum, restores the weather that anchoring removes. A 100-year, 8-member Rescene run shows no detectable drift (−0.004±0.014 K per 100 years), and on the 2022 test year the 16-member ensemble is calibrated (spread/skill 0.89, 0.97 and 0.93 for Z500, T2m and MSLP) with an extratropical synoptic amplitude of 97% of ERA5 on the tuning year. The free run reproduces the geography of day-to-day variability (pattern correlation 0.83 for Z500) and 71% of the observed blocking frequency, where the deterministic wrapper alone has none. Interventions on the frozen operator show a scale-local, approximately linear response in which Z500 perturbations grow upscale 6--14 times more than downscale per 6 h. All constants are fitted on a training year against a rule declared beforehand, and we show that common evaluation shortcuts would have overstated variability and skill. Rescene makes an existing weather model usable for century-long ensembles at negligible training cost.
Minjong Cheon
Department of Computer Science and Engineering, Sejong University, Seoul, South Korea
Global AI weather forecasting still relies mainly on uniform-resolution models, making it hard to combine regional refinement, two-way regional-global coupling, and affordable training cost. We introduce StretchCast, a global-regional AI forecasting framework built on a variable-resolution stretched cubed-sphere (SCS) mesh that preserves a closed global domain while concentrating resolution over a target region. Within this framework, we develop a one-step predictor, SCS_Base Model, and a rollout-oriented multistep predictor, SCS_FCST4 Model, to test the feasibility of SCS-based forecasting and the benefit of joint multistep training. Experiments use ERA5 with 69 variables over 1998-2022. Because training compute remains limited, this study uses a coarse-resolution proof-of-concept configuration rather than a final high-resolution system. Even with only about 7,776 effective global grid cells and roughly 0.875 degree resolution over the center-refined face, the 23M-parameter SCS_Base Model yields stable multivariate forecasts. With 83M parameters and training cost on the order of hours, SCS_FCST4 Model delivers competitive medium-range anomaly-correlation evolution over the target region after unified reprojection, especially for geopotential height, specific humidity, and part of the lower-tropospheric winds, while maintaining smooth cross-face continuity and realistic multiscale structure in typhoon and spectral analyses. These results support StretchCast as a practical lightweight foundation for global-regional AI weather forecasting.
Jin Feng
Institute of Urban Meteorology, China Meteorological Administration (IUM, CMA), Beijing, China · LeadSee Team, Beijing Meteorological Service
Language-model agents are increasingly asked to carry out open-ended scientific research, yet their results are usually graded against a known answer, a rubric, or a language-model reviewer, none of which can tell whether a new scientific model is valid. The AI Science Exam for El Nino-Southern Oscillation (SciExam for ENSO) is a benchmark in which agents build low-order stochastic models of ENSO, the dominant mode of interannual climate variability, from real observations. Within a six-hour budget, agents process the observations, write their own diagnostics, which are then frozen, and develop a model using only these diagnostics as feedback. Hidden graders then test whether the model reproduces ENSO's statistics, recovers unobserved variables, and forecasts held-out years, and score a published model in the same way. Across twelve agent systems, six produce models that score higher than the published model, mainly through better reconstruction and forecasting. The simplified forms of the stronger models are each compatible with one of the two competing explanations of ENSO's warm-cold asymmetry, an open debate that the task never mentions. Controlled runs of the top system under varied information suggest that its scores do not come from recalling the dated observational record and that the information it receives shapes how it builds its model. SciExam for ENSO can thus evaluate agent research where no answer is known, and the results suggest that agents can already build competitive models whose structures bear on questions that scientists still debate.
Yinling Zhang, Langchen Liu, Dongbin Xiu +4
The Ohio State University · Yale University · University of California, San Diego +2