ArchesClimate-SSP, trained with an energy-score loss, reaches 0.9739 K land-temperature RMSE on unseen SSP5-3.4, beating MESMER-M's 1.0963 K

AC-SSP generates long-term climate statistics consistent with IPSL-CM6A-LR on scenarios unseen during training, including the SSP5-8.5 extrapolation where CO2 concentrations exceed the training range. Most prior machine-learning emulators rely on prescribed sea surface temperatures or are limited to fixed warming levels and cannot internally predict SSP forcings; this work conditions on forcing concentrations and rolls out autoregressively to generate full spatiotemporal fields on unseen scenarios. On the held-out SSP5-3.4 test scenario, the 5-member ensemble's land surface temperature RMSE is 0.9739 K versus MESMER-M's 1.0963 K; under SSP5-8.5 the model tracks the IPSL warming trend and grid-cell temperature distribution, with only the Arctic mode showing a larger density shortfall.

Forcing dropout and multi-step pushforward fine-tuning are key training designs that let the model learn the forcing-climate relationship. The authors introduce forcing dropout grouped by physical coupling (six aerosol species, ten ozone bands) and replace dropped channels with a learned null token rather than zeros, while a probability pall controls how often the model sees the full forcing set; this differs from prior practice of simply concatenating forcing channels. Ablations show pall=0.8 and pall=1.0 have similar RMSE (0.9941 K and 0.9800 K), but pall=0.8 better matches the ERF×TCR reference in methane response sign, ozone response error, and N2O over-attribution (×9.27 versus ×12.41); pushforward length pf=4 gives interannual variability of 0.6961, closer to IPSL's 0.6337 than pf=2's 0.9394.

The model recovers the correct response sign for most forcings but severely over-attributes warming to N2O and also over-attributes to CO2. The work tests forcing attribution per gas using piControl-held ablations against an ERF×TCR reference, and identifies that CO2 and N2O are highly collinear in the training corpus, leaving N2O without an independent constraint. Under SSP5-3.4, holding N2O at pre-industrial levels produces about 2.0 K cooling by 2100 versus a TCR reference of only 0.2 K, roughly a tenfold gap; the CO2 ablation produces 2.00 K cooling, exceeding the 1.5 K TCR reference; the methane response closely tracks its TCR reference.

The framework transfers to another Earth system model, CanESM5, with single-forcing attribution behavior matching IPSL's. The authors adapt the preprocessing pipeline (grid, vertical levels, variable set) and apply AC-SSP to CanESM5, showing that the pattern of CO2/N2O over-attribution and close methane agreement is not specific to IPSL. The CanESM5 model tracks target trends across SSP4-3.4, SSP5-3.4, and SSP5-8.5, with only a marginal underestimate of the SSP5-3.4 temperature peak; its single-forcing ablations use CanESM5's own TCR=2.74 K as reference.

The result is aimed at researchers using IPSL-CM6A-LR and CanESM5 output and providing forcing concentrations (not emissions) as input, applicable to generating monthly 2015-2100 scenario rollouts with three-dimensional atmosphere and upper-ocean variables. It enables rapid generation of unseen SSP scenarios, including overshoot pathways like SSP5-3.4, and can produce multi-member ensembles from a single initialization for probabilistic analysis. The method assumes forcings are provided as concentrations, leaving direct emission-to-climate-state mapping to future work; the ocean covers only down to 147.4 m depth.

Several points remain for a careful reader: under abrupt-4xCO2 the energy-score objective makes the model jump almost immediately to a near-final state and plateau, failing to reproduce the deterministic baseline's slow ocean-mediated transient warming, and the authors propose three untested fixes; the N2O over-attribution is attributed to CO2-N2O collinearity in the training corpus, and resolving it would require adding single-forcing N2O ablation data; the Arctic mode density shortfall is attributed to extrapolation, high polar interannual variability, and latitude-weighted loss acting together; and the energy-score model shows about 22% RMSE spread across random seeds, so single-run performance differences should be interpreted cautiously.