Trained model checkpoints and validation data for "GCMagicc v1: a fast generative emulator for multivariate climate-impact ensembles"
Year: 2026
Type: dataset
Abstract: This record archives the large model and evaluation objects accompanying GCMagicc v1.0.1 and the manuscript GCMagicc v1: a fast generative emulator for multivariate climate-impact ensembles (N. Meinshausen, M. Meinshausen, J. Lewis, Z. R. J. Nicholls, S. Schöngart, A. Self, X. Shen, K. Spiller and E. Vogel), submitted to Geoscientific Model Development (EGUsphere preprint egusphere-2026-5176). The companion repository GCMagicc_v1_manuscript holds the code, compact prepared data, selected figures and detailed provenance. The projected ensembles themselves are distributed at gcmagicc.org. You do not need every file here to read the paper or reproduce the compact-data figures. DATASET_GUIDE.md in this record is the full guide to what each file contains, which result uses it, and when you would want it. Summary: Which files do I need? Run the released GCMagicc modelgcmagicc-checkpoints.tar.gz Run the released GCMagicc-CE modelgcmagicc-ce-checkpoints.tar.gz Aerosol interventionNIC-AER.zip + gcmagicc-checkpoints.tar.gz Resolution experimentNIC-RES.zip + gcmagicc-checkpoints.tar.gz GCMagicc-PM emergent-constraint figureNIC-PM.zip GCMagicc-XS extrapolation workflowNIC-XS.zip (optionally the regenerated CSV and NPZ) Audit the main validation-diagnostics figureboth validation-diagnostics-*-v20260821.sqlite Observational-alignment figure, older validation panelsvalidation-metrics-snapshot-20260420.sqlite Verify downloadsSHA256SUMS.txt Contents Trained model archives (Apache-2.0) gcmagicc-checkpoints.tar.gz (7.92 GB) — deterministic archive of the 37 released checkpoint and normalisation files for the default GCMagicc model (production A5 lineage), including the multiresolution cascade and dependence/lag models. gcmagicc-ce-checkpoints.tar.gz (7.82 GB) — the 24 released checkpoint and normalisation files for the verified GCMagicc-CE configuration (T1 lineage). Reproduction bundles — sanitized research bundles with code, configuration, manifests, a pinned environment and prepared data; each has its own README (Apache-2.0 code, CC-BY-4.0 data/figures/docs) NIC-AER.zip (15.4 MB) — aerosol-pattern intervention: factual forcing versus an identical run with aerosol effective radiative forcing set to zero; released diagnostic is the factual-minus-aerosol-off response of tasmax - tasmin at HEALPix nside=64. NIC-PM.zip (340 MB) — GCMagicc-PM perfect-model and emergent-constraint workflow: per-model checkpoints, prepared trend tables, and leave-one-family-out residual sampling for scenario warming intervals. The figure/quantile step is CPU-only and verified byte-for-byte (seed 0, 2000 draws). NIC-RES.zip (402 MB) — two-part resolution experiment: spatial outputs across the HEALPix cascade (nside 1–256) and an end-to-end inference time/memory benchmark by output resolution. NIC-XS.zip (798 MB) — GCMagicc-XS extrapolation-skill workflow, omitting the two smoothed global predictors and testing on high-warming and idealized experiments held out from training; includes the two small XS checkpoints, evaluation and plotting code, manifests and frozen outputs. Validation database snapshots (CC-BY-4.0) — three immutable snapshots, not interchangeable versions of one input. The April snapshot belongs to the historical audit and observational-alignment products; the two August files are a matched pair for the current validation-diagnostics set. validation-metrics-snapshot-20260420.sqlite (11.92 GB) — 20 April 2026 historical validation-metrics database; gofnc records grouped by public version, metric domain and experiment (4,539,079 rows, of which 1,265,222 SSP2-4.5 hold-out). validation-diagnostics-metrics-v20260821.sqlite (12.06 GB) — 21 August 2026 metrics database behind the main validation-diagnostics figure and its ten alternate-draw supplements. validation-diagnostics-edist-v20260821.sqlite (2.70 GB) — promoted 21 August 2026 energy-distance database, the matched pair to the file above. GCMagicc-XS regenerated products (CC-BY-4.0) — regenerated together on 13 July 2026; useful research outputs, but not the checksum-locked historical CSV from which the selected manuscript XS plots were compacted. gcmagicc-xs-regenerated-full_monthly_results-20260713-shac172dc10.csv (467 MB) — regenerated monthly summary by driving model, scenario, test/training status, variable and month. gcmagicc-xs-member-draws-20260713-shaed7d10cf.npz (454 MB) — raw held-out-scenario global-mean stochastic member draws, keyed by <scenario>|<model>|<variable>, shape (n_draws, time). Integrity and what is not included SHA256SUMS.txt gives the SHA-256 checksum of every other file here, including the guide; verify with sha256sum -c SHA256SUMS.txt. The authoritative machine-readable mapping of object IDs, filenames, sizes, checksums and model names is data/external_data_manifest.json in the GitHub companion. Deliberately not redistributed: the 1,528,623,568-byte historical GCMagicc-XS figure-source CSV (the companion's compact plotted points reproduce the selected figures), and the approximately 4.6 TB normalized data_DAT_*.h5 input collection, needed only for selected raw GPU regeneration steps. Abstract of the accompanying paper Projecting the impacts of climate change requires large ensembles of climate variables that match historical observations, align with the warming ranges assessed by the IPCC, and can efficiently run new future emissions scenarios, including the newest generation of climate model scenarios (CMIP7) and pathways consistent with countries’ Paris Agreement pledges. Generating such ensembles at the scale needed for impact studies is normally computationally prohibitive. We close this gap with GCMagicc, a hybrid model that pairs a simple physical climate model with machine learning to generate ensembles of 10 climate variables at the resolution of full-scale Earth system models, without relying on GPU resources or retraining for new scenarios. Being trained on 32 CMIP6 Earth system models and observational/reanalysis data, GCMagicc complements rather than replaces Earth system models. We apply it to a range of future pathways: the canonical set of SSP scenarios in the latest Intergovernmental Panel on Climate Change (IPCC) report (1.2–6.1 °C warming, min-max range across scenarios of 5-95 percentile ranges), current policies (2.3–4.0 °C); national pledges under the Paris Agreement (1.5–3.3 °C) and the CMIP7 range from the so-called ‘VL’ to ‘H’ scenarios (1.2–4.2 °C), releasing a large public dataset. As an illustration of the usefulness of GCMagicc, we perform an attribution analysis on the severe 2025 Iranian drought using GCMagicc ensembles (and three CMIP6 ESM large ensembles as comparisons) with and without anthropogenic forcings. Our GCMagicc results suggest a strong anthropogenic signal, with a median probability of drought conditions at least as severe as the observed ones of 29% (median) with anthropogenic forcing and zero probability (median) under natural-forcing only simulations. In the future, drought conditions are projected to materially worsen, amplifying the potential for severe agricultural and food security impacts and geopolitical conflicts that use water scarcity as a weapon.
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7204589306
