ClimateSuite: ML-Ready Climate Emulation Dataset

Updated 5 July 2026

ClimateSuite is a comprehensive, ML-ready dataset of Earth system simulations covering over 33,000 simulation-years from ten climate models, including intervention scenarios like SAI.
It harmonizes heterogeneous model outputs onto a common 192×288 grid using standardized NetCDF processing and bilinear regridding to preserve large-scale spatial patterns.
The dataset underpins advanced spatiotemporal emulation methods, such as Spatiotemporal Pyramid Flow Matching, demonstrating improved predictive performance and generalization across climate scenarios.

ClimateSuite is a large-scale, ML-ready Earth system simulation dataset introduced alongside Spatiotemporal Pyramid Flow Matching for climate emulation. It is described as “the largest collection of Earth system simulations to date” and “the largest ML-ready climate-scale dataset,” comprising over 33,000 simulation-years across ten climate models, monthly-resolution outputs on a common $192 \times 288$ grid, and—by the authors’ claim—the first ML dataset to include simulations of climate interventions, specifically stratospheric aerosol injection (SAI) (Irvin et al., 1 Dec 2025).

1. Definition and position within climate emulation

ClimateSuite was created to address a specific limitation in climate-emulation benchmarks: prior resources such as ClimateBench were useful but small, derived from a single model, and limited to standard emissions scenarios. ClimateSuite expands that regime by covering ten Earth system models (ESMs), substantially more simulation-years, monthly-resolution outputs, and intervention experiments involving SAI. In the authors’ framing, this broader scope is what enables “superemulators, or surrogates that can emulate multiple ESMs at once,” while also reducing the overfitting risk associated with training on smaller single-model corpora (Irvin et al., 1 Dec 2025).

The dataset is therefore not merely a repository of climate-model output. Its design is tied to probabilistic emulation under heterogeneous forcings, model families, and timescales. This suggests that ClimateSuite occupies the data substrate layer of climate-emulation research: it supplies the breadth needed for models that must generalize across scenarios, across ESMs, and across standard mitigation and intervention settings.

2. Composition, coverage, and harmonization

The paper’s appendix provides the clearest structured accounting of ClimateSuite’s scale. The main text summarizes the dataset as “over 33,000 simulation-years,” while Appendix Table S1 gives the exact total as 33,739 simulation-years.

Aspect	Value
Climate models	10
Simulation-years	33,739
Model-scenario combinations	66
Model-scenario-member combinations	345
Intervention scenario-member combinations	69
Temporal resolution	Monthly
Common grid	$192 \times 288$

The ten included ESMs are BCC-CSM2-MR, CESM2, CESM2-WACCM, CMCC-CM2-SR5, CMCC-ESM2, GFDL-ESM4, IPSL-CM6A-LR, MRI-ESM2-0, NorESM2-LM, and UKESM1-0-LL. To harmonize outputs from these heterogeneous models, all maps are regridded to the common $192 \times 288$ latitude-longitude grid used in CESM2-WACCM. The processing pipeline converts all data to standardized NetCDFs, regrids with bilinear interpolation, and enforces longitudinal periodicity, with the stated aim of preserving large-scale spatial patterns and ensuring physically smooth transitions between adjacent grid cells. For model balance, the authors select three members per climate model because ESMs have highly variable numbers of ensemble members (Irvin et al., 1 Dec 2025).

The paper also records an internal counting discrepancy. One contribution list states that ClimateSuite comprises “276 state-of-the-art simulations from 10 ESMs and 39 SAI simulations,” whereas the appendix tables provide the structured totals of 345 model-scenario-member combinations, 69 intervention scenario-member combinations, and 33,739 years. The appendix tables are presented as the clearest source for the exact counts (Irvin et al., 1 Dec 2025).

3. Scenario families, variables, and forcings

ClimateSuite spans historical simulations from 1850 to near-present and future CMIP6 scenarios to 2100, with SAI runs covering varying date ranges; the paper notes that most SAI experiments are simulated between 2035 and 2070. The standard scenario set comprises historical, SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5. The intervention family includes MA-HISTORICAL and MA-BASELINE controls, several SINGLE-POINT-INJ variants, SSP245-MA-GAUSS variants, two-point injection configurations, SSP245-MA-GAUSS0N-LOWER-0.5, and SAI-1.5. The appendix further summarizes these interventions as spanning equatorial to high-latitude injections, fixed-mass versus controller-based deployments, and temperature targets from $0.5^\circ\mathrm{C}$ to $1.5^\circ\mathrm{C}$ above pre-industrial (Irvin et al., 1 Dec 2025).

The output variables match ClimateBench: surface air temperature and precipitation maps. The forcing inputs begin with the four ClimateBench forcings—cumulative CO $_2$ , cumulative CH $_4$ , spatial SO $_2$ , and spatial black carbon—and add a fifth forcing, stratospheric aerosol optical depth (AOD). The rationale is explicit: the original four forcings do not represent the radiative forcing from SAI, whereas AOD serves as “a direct proxy for the strength and spatial distribution of the forcing produced by SAI.” With that addition, intervention information enters the emulator as a physically aligned forcing field rather than merely as a scenario label (Irvin et al., 1 Dec 2025).

This forcing design is central to the dataset’s policy relevance. ClimateSuite encodes both standard emissions trajectories and intervention perturbations in a form that can condition generative climate models directly.

4. Role in spatiotemporal climate emulation

ClimateSuite is tightly coupled to the Spatiotemporal Pyramid Flow (SPF) model introduced in the same paper. SPF uses a three-stage pyramid at decadal, yearly, and monthly timescales, and ClimateSuite’s monthly-native resolution is what makes that multiscale factorization practical at scale. The paper states that each stage of the flow-matching model is conditioned on temporally aligned forcings from the corresponding timescale, so monthly stages receive monthly-aligned forcings, yearly stages yearly-aligned forcings, and so on (Irvin et al., 1 Dec 2025).

The underlying flow-matching formalism is written as

$\frac{d\boldsymbol{x}_t}{dt}=\boldsymbol{v}_t(\boldsymbol{x}_t),$

with SPF extending standard conditional flow matching by partitioning the generative trajectory across spatial and temporal scales. ClimateSuite matters here because it supplies the broad monthly corpus required to train the same network under decadal, annual, and monthly refinement paths. The paper’s temporal hierarchy uses a $\times 10$ factor from decadal to yearly and a $192 \times 288$ 0 factor from yearly to monthly, so the dataset’s native monthly resolution matches the finest stage of the pyramid directly (Irvin et al., 1 Dec 2025).

A practical implication is that ClimateSuite is not simply “large.” Its structure is aligned to conditional, multiscale emulation with prescribed forcings, including nonstationary intervention scenarios.

5. Experimental use and reported performance

The paper uses ClimateSuite in three distinct ways: pretraining SPF before fine-tuning on ClimateBench, direct multi-model held-out scenario evaluation, and held-out climate-model intervention generalization. The pretraining result is one of the clearest demonstrations of the dataset’s value. A 600M-parameter SPF without ClimateSuite pretraining achieves yearly CRPS/RMSE of $192 \times 288$ 1 and monthly $192 \times 288$ 2, whereas the ClimateSuite-pretrained 600M SPF improves to yearly $192 \times 288$ 3 and monthly $192 \times 288$ 4. The paper emphasizes that this gain is not simply a scale effect, because the non-pretrained 600M model underperforms the 200M model at yearly scale, while the pretrained model outperforms both (Irvin et al., 1 Dec 2025).

Direct evaluation on ClimateSuite is reported against a 600M UNet baseline for yearly emulation. On held-out SSP2-4.5 across ten climate models, average RMSE improves from $192 \times 288$ 5 to $192 \times 288$ 6 and average CRPS from $192 \times 288$ 7 to $192 \times 288$ 8. On the held-out UKESM1-0-LL SAI intervention scenario, RMSE improves from $192 \times 288$ 9 to $192 \times 288$ 0 and CRPS from $192 \times 288$ 1 to $192 \times 288$ 2. The paper summarizes this as “good generalization to held-out scenarios across climate models,” while also noting that for the held-out SAI case the global means are close to the simulation for most of the century but poorer at the start (Irvin et al., 1 Dec 2025).

These results position ClimateSuite as an evaluation benchmark as well as a training corpus. It supports both cross-scenario generalization and the more demanding setting in which an intervention experiment is held out for an entire climate model.

6. Availability, limitations, and broader significance

ClimateSuite is stated to be publicly available through the SPF repository at https://github.com/stanfordmlgroup/spf, and the processed data are described as standardized NetCDFs. The manuscript also identifies the raw upstream sources: forcings from input4MIPs via ESGF, standard CMIP6 historical and SSP simulations via ESMValCore and acccmip6, SAI experiments via Globus, and UKESM1-0-LL SAI from the Met Office ARISE portal (Irvin et al., 1 Dec 2025).

The paper also identifies several limitations. First, because ClimateSuite is built from existing ESM ensembles, it “likely limit[s] generalization to unseen parameterizations or extreme forcings.” Second, harmonization via regridding may smooth some fine-scale inter-model differences. Third, although the dataset expands policy relevance by including SAI, the intervention archive is concentrated in CESM2-WACCM and UKESM1-0-LL, with most SAI coverage between 2035 and 2070. The excerpt further notes that the paper does not specify a dataset DOI, license, or detailed user instructions beyond the public GitHub release statement (Irvin et al., 1 Dec 2025).

In the broader climate-software landscape, ClimateSuite occupies a distinct niche. AIRCC-Clim is a user-facing regional probabilistic scenario generator for temperature and precipitation under custom emissions pathways (Estrada et al., 2021), Climate2Energy converts climate-model outputs into energy-system inputs across wind, solar, hydropower, heating, and cooling (Wohland et al., 13 Aug 2025), and cubble provides an analysis-facing data abstraction for multivariate spatio-temporal workflows (Zhang et al., 2022). This suggests that ClimateSuite’s primary identity is not scenario emulation for policy users, sectoral conversion, or data wrangling, but large-scale supervised substrate for probabilistic multi-model climate emulation.