betaFeedback

WorksPhysics and Astronomy

Guiding future global temperature distributions with simulations and observations

Abstract

Assessing climate risks requires projections of daily temperature distributions, including rare and spatially widespread extremes. Climate simulations are costly and biased, while correcting them toward observations risks suppressing inter-model differences. We train a single machine-learning emulator, a conditional diffusion model, on seventeen CMIP6 models and ERA5 reanalysis to generate spatially coherent daily temperature fields conditioned on model identity, date, and CO2-equivalent concentration. At sampling time, diffusion guidance reduces model biases by steering each emulated climate-model distribution toward observations, with its strength selected through perfect-model experiments using withheld simulated futures. Observational influence decreases with projection horizon and forcing. Across seventeen synthetic targets, guidance more than halves median biases in the 1st and 99th percentiles of daily temperature, improves probabilistic scores and reliability, and reduces median annual global-mean temperature RMSE by 32-57% across three emission scenarios. In a detailed benchmark, guidance also outperforms established bias corrections on daily-field probabilistic scores. Efficient sampling generates large ensembles without further climate simulations. For example, our framework corrects CMIP6's roughly threefold historical overestimate of extreme-heat coverage, yet projects that concurrent extremes will affect nearly half of global land on a typical day in the 2080s under SSP3-7.0.

Climate models provide the foundation for understanding and adapting to a changing climate [1]. Yet they differ in their representation of small-scale processes (e.g., clouds or convection) that strongly affect the response of the planet to greenhouse gases, leading to systematic biases relative to observations [2]. Even models that reproduce historical climate relatively well can project markedly different future warming, making inter-model differences a major source of long-term uncertainty [3, 4, 5]. Correcting model outputs against observations can mitigate these biases, but overcorrection risks erasing the spread that reflects genuine structural uncertainty and muting internal climate variability. Two questions follow: how strongly should observations constrain climate projections, and how can that choice be validated without future observations?

These questions extend beyond mean warming to full distributions of variables such as daily temperature. For example, the probability of local hot days depends on the tails of the distribution, whereas the total area affected is determined by concurrent heat extremes across regions [6]. Accurately estimating these distributions, especially for rare events, requires extensive sampling [7, 8], but climate simulation archives provide few realizations for each model and forcing scenario. Although large ensembles can increase these sample sizes [9, 10], each additional realization requires another computationally expensive simulation [11]. Generative emulators [12] instead learn distributions from existing simulations and efficiently sample new fields without further integrations [13, 14, 15, 16, 17]. Their larger ensembles reduce sampling error but inherit the biases of the training simulations.

Existing methods use historical observations to weight climate models [18] or reduce biases in their output [19, 20, 21, 22, 23, 24]. For a generative emulator, such adjustments operate directly on probability distributions, raising the question of how strongly generated fields should be steered toward observations. Optimizing a distribution to match historical records does not imply accurate projections of unobserved future conditions, because only physics-based models simulate the climate response to future forcing. Steering generated fields toward observations therefore requires explicit control over the relative influence of climate models and observations, and an empirical framework to evaluate this balance against synthetic target futures.

To address these challenges, we develop a generative modeling framework, starting with a single emulator trained on daily temperature data. The data comprise simulations from seventeen Coupled Model Intercomparison Project phase 6 (CMIP6) models [2] under three Shared Socioeconomic Pathway (SSP) scenarios [25] and the ECMWF Reanalysis v5 (ERA5) [26] through 2024.

The emulator, a generative diffusion model [27], learns the distribution of global daily near-surface temperature fields conditioned on model identity, calendar day and year, and carbon dioxide equivalent (CO2_2e) concentration (Materials and Methods). Modeling individual days rather than sequences suffices for the risk measures above, which depend only on the distribution of daily fields.

We denote by pi(x)p_i(\mathbf{x}) the learned distribution of daily fields x\mathbf{x} for climate model ii, and by ptgt(x)p_{\mathrm{tgt}}(\mathbf{x}) that of the ERA5 target. We align each pip_i with ptgtp_{\mathrm{tgt}} by sampling their geometric mean weighted by γ\gamma. Unlike an additive mixture, which would draw each field from one distribution or the other, this product favors fields likely under both:

pγ,i(x)∝pi(x)1−γptgt(x)γ.(1)p_{\gamma,i}(\mathbf{x}) \propto p_i(\mathbf{x})^{1-\gamma} p_{\mathrm{tgt}}(\mathbf{x})^\gamma. \tag*{(1)}

This condition-to-condition guidance builds on guidance methods for diffusion models [28, 29, 30] and samples Eq. (1) by following the weighted sum of the two learned scores (gradients of the log-densities), (1−γ)∇xlog⁡pi+γ∇xlog⁡ptgt(1-\gamma)\nabla_{\mathbf{x}}\log p_i+\gamma\nabla_{\mathbf{x}}\log p_{\mathrm{tgt}} (Materials and Methods). The guidance scale γ∈[0,1]\gamma\in[0,1], set at inference time without further training, moves pγ,ip_{\gamma,i} from the climate model (γ=0\gamma= 0) to the observations (γ=1\gamma= 1; Figure 1). Guiding each of the seventeen models produces an ensemble that we use to evaluate how guidance strength affects both the daily temperature distribution and inter-model spread.

Guidance balances observational fidelity and model-based extrapolation

Figure 1. Guidance balances observational fidelity and model-based extrapolation. Daily temperature fields from a synthetic observation (MPI-ESM1-2-LR, withheld from training after 2024; red) and from the diffusion emulator conditioned on the remaining sixteen CMIP6 models (blue), projected onto the first two principal components. Columns increase the guidance scale γ\gamma from source-model emulation (γ=0\gamma= 0) to target-only emulation (γ=1\gamma= 1). Within the observational period (2015–2024; bottom), target-only emulation best reproduces the target distribution. Far beyond the observed period (2085–2100; top), intermediate guidance better overlaps the target, indicating that extrapolation benefits from retaining information from the physics-based model ensemble. Bold frames indicate the guidance scale closest to the selected value for each period.

We evaluate this framework through perfect-model experiments, or leave-one-out cross-validation (LOOCV), treating each of the seventeen CMIP6 models as a synthetic observational reference whose future, unlike that of ERA5, is known. In each experiment, we train the emulator jointly on the target’s simulation up to 2024, except 2001–2010, and full simulations of the remaining sixteen models. We use the target’s withheld 2025–2100 future to quantify projection errors and determine the optimal guidance strength. In the experiment with MIROC6 as the target, we benchmark against three methods: the bias corrections mean–variance–trend (MVT) [20] and quantile delta mapping (QDM) [19], and classifier guidance [28].

Perfect-model experiments show that the selected observational influence decreases with projection horizon and forcing. Guidance lowers 1st- and 99th-percentile biases, averaged over the three scenarios, for all seventeen synthetic targets, outperforms established bias corrections for MIROC6 on most metrics beyond their calibration period, and reduces the overestimated historical extent of concurrent heat extremes. Our framework shows how generative climate emulation and observations can be combined in a statistically principled manner.

Results

In perfect-model experiments, where a held-out CMIP6 model serves as a synthetic observation with a known future, we fit how strongly guidance should weight the observations, evaluate the resulting schedule on all seventeen targets, and benchmark it against established bias corrections. We then turn to ERA5: we verify the emulator on a withheld observed decade, project global warming with the schedule, and analyze local and regional heat extremes.

We first determine how strongly guidance should weight the observations. No observations of the future exist to tune this weight, so we tune it in perfect-model experiments, whose withheld futures are known. In Figure 1, target-only emulation (γ=1\gamma= 1) best replicates the synthetic target in 2015–2024, while intermediate guidance better captures its withheld 2085–2100 distribution. The same pattern emerges in all seventeen perfect-model experiments when we select the guidance scale that minimizes 1st- and 99th-percentile daily temperature biases. As projections extend farther beyond the historical record, the fitted schedule shifts weight away from the target, decreasing from γ=0.94\gamma= 0.94 in 2025 to 0.84, 0.80, and 0.77 in 2100 under SSP2-4.5, SSP3-7.0, and SSP5-8.5 (Materials and Methods; Supporting Information, Figure S3). The guided projections stay anchored to observations near the present and rely increasingly on simulated dynamics as forcing and lead times grow.

The fitted guidance schedule for three SSP scenarios, shown continuously and rounded to the nearest scale

Figure S3. The fitted guidance schedule, 2025–2100. (A) The guidance scale γ(y,c)\gamma(y,c) of Eq. (12) for each scenario, the continuous values supplied directly to the sampler in the scheduled run and the projections toward ERA5, which sample every day of the year. (B) The same schedule rounded to the nearest scale of the nine-scale sweep (Sampling Designs): γ=1\gamma= 1 through 2026, then 0.875, then 0.75 from 2097 under SSP3-7.0 and from 2085 under SSP5-8.5; the three scenarios coincide before these dates.

We next ask whether the fitted schedule, selected on daily temperature tails, also improves the projected warming trajectory of each withheld target. For each target and scenario, we calculate the root mean squared error (RMSE) between the target’s annual global-mean trajectory and the ensemble mean over 2025–2100 (Supporting Information mation, Evaluation Metrics). Across scenarios, median RMSE drops from 0.47–0.56 °C unguided to 0.24–0.38 °C guided. Averaged over the three scenarios, guidance also reduces area-weighted 1st- and 99th-percentile daily temperature biases for every target model and improves broader distributional metrics (Supporting Information, Tables S1 and S2 and Figs. S5–S9). These scores are in-sample, but the schedule has only three parameters. Statistical bias corrections are the established alternative, but they are calibrated on the historical period and applied unchanged to the future. The MIROC6 experiment (Figure 2B) tests both approaches beyond this calibration period. Guidance reduces the error of the raw CMIP6 trajectory by 41–62%, whereas MVT [20] and QDM [19] reduce it by 8–16% (Supporting Information, Table S3). Classifier guidance (run on a separate emulator trained without MIROC6) underperforms our approach in trajectory error and all daily-field metrics. Over 2081–2100, guided fields have the most uniform rank histograms and lower mean absolute bias than either correction, while MVT better captures local standard deviation and skewness and, applied to the emulator output, has a lower 99th-percentile bias (Supporting Information, Table S3 and Figs. S10–S17).

Figure 2. Guided projections constrain future warming and recover a known synthetic future. Annual global-mean warming above 1850–1900 under SSP2-4.5, SSP3-7.0, and SSP5-8.5; shaded bands show the 5–95% range of each ensemble. (A) Projections guided toward ERA5: the emulator reproduces the observational record through 2024 and yields a narrower range of future warming than raw CMIP6 while retaining inter-model spread. (B) Corresponding perfect-model experiment with MIROC6 as a synthetic observational target, whose future, withheld after 2024, serves as ground truth: guided projections follow the MIROC6 trajectory more closely than raw CMIP6 and the MVT and QDM bias corrections. Gray shading marks the training period.

Projections must also convey their uncertainty, and here our ensembles remain somewhat under-dispersed: the guided 5–95% warming interval covers the withheld MIROC6 trajectory in 62–72% of years (nominal 90%; Supporting Information, Table S3). For a second target, MPI-ESM1-2-LR, the guided ensemble underestimates warming after mid-century (Supporting Information, Held-Out MPI-ESM1-2-LR). In these experiments, guidance refines forced responses and distributions but leaves residual under-dispersion and target-specific biases in long-term warming.

Before projecting from ERA5, we check the emulator against real observations on the decade withheld from training, where the corrections can be scored as well. Over 2001–2010, our emulator replicating ERA5 (γ=1\gamma= 1) leaves a spatially smoother residual bias than MVT and QDM and better preserves the observed temperature differences between neighboring cells, which these per-cell corrections distort, although they better match marginal distribution tails and day-to-day variability (Supporting Information, Tables S4 and S5 and Figs. S18–S22). Beyond 2024, we apply the fitted schedule. Guidance narrows the range of projected global warming while preserving structural differences among the seventeen CMIP6 models (Figure 2A). Over 2081–2100, the guided ensemble retains 20.5–22.9% of the raw CMIP6 5–95% inter-model range, across scenarios (Supporting Information, Table S6). The guided ensemble mean first exceeds 1.5°C in 2028–2029 and 2°C in 2041–2047 across SSPs (Supporting Information, Table S7).

Our emulator’s computational efficiency overcomes the sample-size limitations of standard climate archives: ten daily samples for each guided model produce 312,800 summer days over 2081–2100, against 31,280 in the CMIP6 archive analyzed here. This ten-fold increase resolves rare tail events that raw archives undersample. Observational constraints also reshape the local distributions behind extreme-heat probabilities. At Paris, the emulator replicating ERA5 over 1995–2014 (half in the training set) reduces most of the raw CMIP6 historical bias, particularly in the warm tail (Figure 3A). Guidance and raw CMIP6 identify broadly similar land regions of strongest growth in summer extremes, but guided projections show greater relative increases in exceedance frequency of local historical 99th-percentile thresholds across most land areas, a divergence that grows with forcing (Figure 3B). However, because guidance corrects the historical warm bias, late-century guided extremes remain lower in absolute temperature than raw CMIP6 at equal return periods (Figure 3C). Risk estimates that depend on historical thresholds therefore change once the historical bias is removed.

Figure 3. Spatially resolved analysis of extreme heat. Daily summer temperature statistics for 1995–2014 and 2081–2100 under SSP2-4.5, SSP3-7.0, and SSP5-8.5. (A) Distribution of June–August daily mean temperature at the Paris grid cell for the guided diffusion emulator and raw CMIP6, with ERA5 as the historical reference. (B) Ratio of the probability of exceeding each ensemble’s local 1995–2014 99th-percentile summer temperature in 2081–2100 versus 1995–2014. Values greater than one indicate more frequent future exceedance of historical extremes. (C) Paris temperature exceeded on average once in NN summer days; dots mark the hottest day of each curve, whose extent reflects the number of available samples.

Finally, concurrent heat extremes across regions determine the total area affected, and measuring them requires spatially complete daily fields, which our emulator provides in large numbers. Against local day-of-year 95th-percentile ERA5 thresholds, raw CMIP6 models overestimate the extreme-heat land area in the withheld 2001–2010 decade, with a median extent roughly three times the observed value. The emulator replicating ERA5 instead aligns closely with observations, with a median coverage of 8.1 million km2^2 (5.5% of global land; Figure 4B). Correcting this historical bias does not suppress

Figure 4. Extreme heat affects an increasing share of global land. Total land area contained in contiguous regions of extreme heat under SSP3-7.0. A land grid cell is extreme when its daily mean temperature exceeds the ERA5 1979–2024 95th percentile for that day of year; the same thresholds are applied to all datasets, and only regions of at least two adjacent cells are counted. (A) Distribution of daily affected area for ERA5, the guided diffusion ensemble, and raw CMIP6; circles mark the medians. In the 2001–2010 validation decade, the emulator, which replicates ERA5 before 2025, substantially reduces the CMIP6 overestimate of observed extreme-heat area; the projected distribution then shifts toward much larger affected areas through the century. (B) Representative guided fields for each decade, chosen to have total affected area closest to the median in panel A, each labeled with its date and total area.

the projected response to warming. Under SSP3-7.0, the median affected area grows steadily each decade to 68.6 million km2\mathrm{km}^2, or 46.9% of global land, in the 2080s, converging near the raw CMIP6 baseline (Figure 4A and B).

Discussion

Our experiments show that a single conditional diffusion emulator can expand the scope of statistical questions addressed using a fixed archive of climate simulations and observational data. Its additional, spatially coherent daily fields enable the analysis of quantities often difficult to estimate from standard ensembles, particularly simultaneous regional extremes and sparsely sampled distribution tails. Initial-condition large ensembles offer comparable sample sizes but remain computationally prohibitive for most climate models [9, 10]. Our emulator bypasses this bottleneck by sampling from the joint distribution learned across existing climate archives, so custom analyses need no new simulations, provided the training data span the target climate variability and the emulator learns those distributions.

Condition-to-condition guidance empirically extrapolates historical climate distributions into the future. Reanalyses capture historical variability but contain no signals of future forcing. CMIP6 simulations cover the projection period but carry systematic biases relative to observations. Our perfect-model experiments test this trade-off by replacing the unknown future of ERA5 with held-out model projections. As lead time and forcing grow, the selected scale shifts weight toward the climate models. These scales (γ<1\gamma< 1) and the reduced trajectory and distributional errors across held-out models confirm that combining observations with model dynamics outperforms either alone. Nonetheless, tuning guidance strength is distinct from fully calibrating the predictive distribution. Daily diagnostics and rank histograms indicate improved distributional agreement, and annual uncertainty bands contain the withheld MIROC6 trajectory in most years, but coverage remains below the nominal 90%. We view guidance as a practical framework for adjusting projections to observational reference data. Fully correcting the remaining distributional errors would require test-time calibration algorithms, which trade computational cost for statistical accuracy [31].

Our extreme-heat analyses illustrate how guidance separates historical bias from projected warming trends. In Paris, guiding toward ERA5 removes the historical warm-tail bias of raw CMIP6, shifting the local extreme-heat threshold. The guided ensemble therefore projects a larger relative increase in exceedance frequency, yet lower absolute late-century temperatures than raw CMIP6. This distinction matters whenever future changes are evaluated against model-specific historical thresholds. The extreme-heat area analysis illustrates the same mechanism spatially: guidance reduces the historical CMIP6 bias without suppressing projected warming. Beyond these applications, our guided ensembles can support downstream evaluations, including population exposure under changing heat thresholds, spatially resolved return levels, or regional climate responses at fixed global warming levels.

We acknowledge several limitations. Additional samples reduce Monte Carlo sampling error within the learned distribution but cannot correct errors in the emulator, guidance scheme, or simulations used for training. Because our perfect-model experiments evaluate out-of-model generalization only within the CMIP6 archive, they cannot reveal biases shared across models or confirm calibration against observations. Because our emulator draws days independently, it cannot directly estimate multi-day heat-wave duration, event sequencing, or temporally compound risks. Our analysis is limited to near-surface temperature at moderate spatial resolution; within this scope, guided generative emulation resolves long-term forced warming, daily extreme quantities, and spatial heat coverage across multi-model archives. Extending this framework to temporally continuous, multivariate, and higher-resolution fields would enable broader impact applications, though with higher computational and validation demands.

These limitations suggest future directions. Additional computing power would enable calibration-aware guidance, refined guidance schedules, and more evaluations of tail statistics and spatial dependencies. Because the guidance scale is set at inference time, a single pre-trained emulator can adapt to diverse user requirements, allowing region- and time-specific scales that can be tested in our perfect-model framework. We introduced this methodology to complement physics-based climate modeling: dynamical simulations provide forced responses, reanalyses capture historical variability, and our emulator combines both into a probabilistic ensemble. Climate model archives can now function as continuous, observationally based probability distributions, enabling direct queries for spatial patterns, extreme quantiles, and joint risk statistics.

Materials and Methods

Data. We use daily near-surface temperature fields from seventeen CMIP6 models (Supporting Information, Table S8) over 1979–2100 under SSPs 2-4.5, 3-7.0, and 5-8.5, and the ERA5 reanalysis [26] over 1979–2024 as observational reference. We conservatively remap all fields [32] to a 64×12864 \times128 grid (∼2.8∘\sim2.8^\circ) and pair each with the annual global-mean CO2_2e concentration [33] of its year and scenario (SSP2-4.5 for ERA5). We train on ERA5 and withhold the decade 2001–2010 for validation (Supporting Information, Data Split).

ModelReference
AWI-CM-1-1-MR[24, 25]
BCC-CSM2-MR[26, 27]
CanESM5[28, 29]
CESM2-WACCM[30, 31]
CMCC-CM2-SR5[32, 33]
CMCC-ESM2[34, 35]
EC-Earth3[36, 37]
EC-Earth3-Veg[38, 39]
EC-Earth3-Veg-LR[40, 41]
FGOALS-g3[42, 43]
IPSL-CM6A-LR[44, 45]
MIROC6[46, 47]
MPI-ESM1-2-HR[48, 49]
MPI-ESM1-2-LR[50, 51]
MRI-ESM2-0[52, 53]
NorESM2-LM[54, 55]
NorESM2-MM[56, 57]

Table S8. CMIP6 models used in this study. The column Reference gives the dataset references of the historical and SSP simulations of each model.

Conditional Diffusion Emulator. We emulate the seventeen CMIP6 models and ERA5 with a single conditional diffusion model. This network learns the distribution pθ(x∣m,d,c,y)p_\theta(x \mid m, d, c, y) of a daily temperature field xx, conditioned jointly on the model identity m∈{CMIP61,…,CMIP617,ERA5}m \in\{\mathrm{CMIP6}_1,\ldots,\mathrm{CMIP6}_{17},\mathrm{ERA5}\}, the day of year dd, the annual global-mean CO2_2e concentration cc, and the calendar year yy. We train on the three SSPs jointly and without a scenario index, so the emulator sees the forcing pathway only through cc and yy. Because the scenarios reach a given CO2_2e concentration in different years, cc alone does not identify this pathway, and the climate response depends on the pathway through the lagged ocean response, not only on the current concentration. Conditioning also on yy resolves this ambiguity.

We train a UNet denoiser with elucidated diffusion model (EDM) preconditioning and loss [27] on every daily field of the training data and generate fields with its deterministic Heun sampler (Supporting Information, Machine Learning Setup Details). Since a single network represents all eighteen conditional distributions, the score of any source model and that of ERA5 are available from the same denoiser at every sampling step.

Condition-to-Condition Guidance. We introduce condition-to-condition guidance, a central contribution of this work, which merges the emulated distributions of a climate model and of the observations at sampling time, without retraining. For a given day, CO2_2e concentration and year, let pi(x)=pθ(x∣CMIP6i,d,c,y)p_i(x) = p_\theta(x \mid\mathrm{CMIP6}_i,d,c,y) be the emulated distribution of climate model ii, and ptgt(x)=pθ(x∣ERA5,d,c,y)p_{\mathrm{tgt}}(x) = p_\theta(x \mid\mathrm{ERA5},d,c,y) that of the observations, the target. Guidance from source ii toward the ERA5 target samples their weighted geometric mean pγ,ip_{\gamma,i} of Eq. (1), where the guidance scale γ∈[0,1]\gamma\in[0,1] sets the weight of the target. Eq. (1) assigns high probability only to fields that are likely under both the climate model and the observations. The mixture (1−γ)pi+γptgt(1-\gamma)p_i+\gamma p_{\mathrm{tgt}}, in contrast, would return fields from one or the other (Supporting Information, Interpolating between Two Distributions). Finally, we pool the guided distributions with equal weight,

pγ(x)=1∣I∣∑i∈Ipγ,i(x),(2)p_\gamma(x)=\frac{1}{|\mathcal{I}|}\sum_{i\in\mathcal{I}}p_{\gamma,i}(x), \tag*{(2)}

where I\mathcal{I} contains the seventeen CMIP6 models or, in a perfect-model test, the sixteen source models guided toward the model held-out one. At γ=0\gamma=0, Eq. (2) is the emulated climate-model ensemble; at γ=1\gamma=1, every component becomes ptgtp_{\mathrm{tgt}}. In between, each component moves toward the target and keeps part of what distinguishes its source model. For Gaussian pip_i and ptgtp_{\mathrm{tgt}} with means μi\mu_i and μtgt\mu_{\mathrm{tgt}} and a common covariance, pγ,ip_{\gamma,i} is Gaussian with that covariance and mean (1−γ)μi+γμtgt(1-\gamma)\mu_i+\gamma\mu_{\mathrm{tgt}}, and the spread of the model means around the target shrinks by a factor 1−γ1-\gamma.

To draw a single sample from Eq. (2), we fix a source model i∈Ii\in\mathcal{I} and perform diffusion sampling by integrating the probability flow ordinary differential equation [34] from noise to data with the interpolated score function

∇xtlog⁡pγ,i(xt)≈(1−γ)∇xtlog⁡pθ(xt∣CMIP6i,d,c,y)+γ∇xtlog⁡pθ(xt∣ERA5,d,c,y),(3)\nabla_{x_t}\log p_{\gamma,i}(x_t)\approx(1-\gamma)\nabla_{x_t}\log p_\theta(x_t\mid\mathrm{CMIP6}_i,d,c,y)+\gamma\nabla_{x_t}\log p_\theta(x_t\mid\mathrm{ERA5},d,c,y), \tag*{(3)}

where xtx_t is the partially denoised field at noise level tt.

Sampling starts from random noise and removes it step by step, moving the field along the weighted sum of the two scores (i.e., the directions in which the field becomes more likely under the source model and under ERA5). Eq. (3) approximates the score of Eq. (1) [31], since the geometric mean of two noised distributions is not the noised geometric mean; the approximation is exact in the Gaussian example above and also underlies classifier-free guidance [29]. Unlike classifier guidance [28], which requires a separate classifier, condition-to-condition guidance evaluates both scores with the same parameters θ\theta. Domain Guidance [30] uses the same construction for a single source domain; we generalize it to a pool of sources sharing one target, and the ensemble of Eq. (2) keeps their differences.

Selecting the Guidance Strength. We select γ\gamma values with the perfect-model experiments. We divide 2025–2100 into 32 consecutive blocks and, for each fold, SSP and block, select among nine scales the one whose guided ensemble has the smallest absolute 1st- and 99th-percentile daily temperature bias against the held-out model, averaged over grid cells and both tails (Supporting Information, The Criterion Used to Select the Guidance Scale). Because the selections vary between folds and blocks, we fit one guidance schedule, linear in calendar year and CO2_2e concentration, to all selections pooled over folds and SSPs, and clip it to [0, 1] (Supporting Information, Fitting the Guidance Schedule). We apply this schedule to projections toward ERA5, assuming it transfers because these tests replicate the real-world situation of a target known up to 2024 and projected beyond. Before 2025, where the target is known, we do not extrapolate. Our diffusion model is conditioned on the target only (γ=1\gamma= 1), so it replicates the known record with many realizations. For simplicity, we also refer to this ensemble as guided.

Annual-Variance Rescaling. Independent daily sampling makes emulated annual means average away almost all the year-to-year variability that climate models and observations keep through correlated daily anomalies. We therefore fit a cubic polynomial in time to every emulated annual global-mean series (one per sample of each source model) and multiply the residuals by a factor k≥1k \ge1, the ratio of the year-to-year standard deviation of the target’s annual global mean over 1979–2024 (ERA5, or the held-out model of a perfect-model experiment) to that of the emulator (Supporting Information, Estimating the Rescaling Factor). We apply one kk to all series, years and SSPs of an ensemble, assuming stationary interannual variability and a sampler under-dispersion independent of γ\gamma. We rescale the annual global-mean series of every emulated ensemble, guided or not, including classifier guidance.

Data and Code Availability. The datasets used in this study are freely accessible, under their respective licenses, from the following public repositories: LEAP Pangeo for CMIP6 simulations; Climate Data Store for ERA5 reanalysis [26, 35]. The code will be made available at https://github.com/mandt-lab/gdso. The Supporting Information figures for the other scenarios and for 2030–2050 will be made available at https://github.com/mandt-lab/gdso-figures.

Acknowledgments. We acknowledge funding from National Science Foundation (NSF) through Learning the Earth with Artificial Intelligence and Physics (LEAP) Science and Technology Center (STC) (Award #2019625). Stephan Mandt acknowledges funding from NSF through an NSF CAREER Award IIS-2047418, IIS-2611549, and funding from the Hasso Plattner Research Center at UCI. Pierre Gentine acknowledges funding from the European Research Council Synergy Grant “Understanding and modeling the Earth System with Machine Learning” under the Horizon 2020 research and innovation programme (Grant Agreement No. 855187). Duncan Watson-Parris acknowledges funding from NSF award 2441832. Padhraic Smyth acknowledges funding from NSF award CAIAG 2425932.

Author contributions: F.I., E.T., F.D., and S.M. designed research; F.I., E.T., and J.N. performed research; F.I., E.T., F.D., J.N., V.L., and S.M. analyzed data; F.I., E.T., and S.M. wrote the paper; and F.D., J.N., V.L., P.G., D.W.-P., and P.S. reviewed and edited the paper.

Competing Interests. The authors declare no competing interest.

Supporting Information

Supporting Information Text

Data

This section details the data description in Materials and Methods, Data.

Temperature Fields. We use Coupled Model Intercomparison Project phase 6 (CMIP6) and ECMWF Reanalysis v5 (ERA5) [1] records from 1979 onwards, when the transition to the satellite era improved the quality and spatial completeness of global observations [2]. For each CMIP6 model we use one initial-condition member, r1i1p1f1r1i1p1f1, the primary member used in the Intergovernmental Panel on Climate Change (IPCC) Sixth Assessment Report of the Working Group I (AR6 WGI) [3]. We drop leap days, so that every year has 365 days. The common grid is the 64×12864 \times128 Gaussian grid of CanESM5, the coarsest of the seventeen models used in this work.

CO2_2e Forcing. We obtain the annual global-mean carbon dioxide equivalent (CO2_2e) concentration for 1979–2100 under the three Shared Socioeconomic Pathways (SSPs) from the Minimal CMIP Emulator v1.2 [4], a reduced-complexity model whose effective radiative forcing of greenhouse gases and aerosols we convert to CO2_2e by inverting the CO2_2 forcing expression of Meinshausen et al. [5]. This variable encodes the combined greenhouse-gas and aerosol forcing that drives long-term climate response. Annual resolution is sufficient for our purposes because well-mixed greenhouse gases have atmospheric lifetimes longer than several years [3], and the day-of-year condition carries the seasonal cycle of temperature.

Machine Learning Setup Details

This section details the emulator of Materials and Methods, Conditional Diffusion Emulator: training data, the scaling applied, conditioning, denoiser, noise schedule, training, and sampling. We train eighteen such emulators: one for each of the seventeen perfect-model experiments, or leave-one-out cross-validation (LOOCV) folds, and one (the ERA5 projection model) trained on all seventeen CMIP6 models and ERA5 used as the actual observation.

Data Split. We jointly train the ERA5 projection model on the full CMIP6 simulation record (1979–2100, three SSPs), so that it is exposed to the whole range of forced responses across scenarios, and on ERA5 (1979–2000 and 2011–2024, reserving 2001–2010 for validation). The years 2025–2100 are dedicated to projection, where the ERA5 endpoint of guidance requires extrapolation. Because the SSPs diverge after 2014, we pair the ERA5 temperature fields with CO2_2e values from the SSP2-4.5 trajectory, following the CMIP6 convention of extending the historical period with SSP2-4.5 [6, 7]. Our emulator identifies the scenario only through the combination of CO2_2e and year, and the CO2_2e trajectories of the three SSPs barely diverge over 2015–2024.

In each of the seventeen LOOCV folds, we hold out a single CMIP6 model from the pool and use it as a synthetic observational target for which we also have ground truth in the future. This target model plays the role of ERA5. We train the fold-specific emulator on the daily temperature fields simulated by the remaining sixteen CMIP6 models over 1979–2100. Only the periods 1979–2000 and 2011–2024 of the held-out model enter training. We hold out 2001–2010 to evaluate the emulation quality closer to the present day while keeping the start of each SSP (2015–2024) in the training data. This choice also lets the emulator learn how different SSPs begin to diverge, which is useful for future projections. We withhold consecutive years rather than random days because daily temperature is autocorrelated and random days would leak information from training into validation. During data loading, we divide the daily fields into the historical period (1979–2014) and SSPs 2-4.5, 3-7.0, and 5-8.5, because the three scenarios share the historical period up to 2014. In this way, the emulator sees each daily field once per epoch, in random order.

Data Scaling. We compute a scalar mean and standard deviation over all grid cells and days of the training data, separately for each fold, to apply z-score normalization [8]. We standardize CO2_2e in a similar manner. Because CO2_2e is a function of year and SSP in our setup, its statistics include each possible value once across the full historical and SSP periods, independent of any fold.

Conditioning. Of the four conditioning inputs of Materials and Methods, Conditional Diffusion Emulator, two are categorical: the model identity mm (from a fixed list of CMIP6 models and ERA5) and the day of year dd. We encode each by its integer index and pass it through its own sinusoidal positional embedding eme_m and ede_d, respectively. A learned feed-forward network embeds the z-scored annual global-mean CO2_2e concentration cc as ece_c. We rescale the year yy linearly to [−1,1][-1,1] and embed it with its own learned feed-forward network as eye_y. We concatenate these embeddings and project them to the timestep-embedding dimension (128) to obtain the joint conditioning vector ccomb\mathbf{c}_{\mathrm{comb}}:

ccomb=FFN⁡([em;ec;ed;ey])(4)\mathbf{c}_{\mathrm{comb}} = \operatorname{FFN}([\mathbf{e}_m;\mathbf{e}_c;\mathbf{e}_d;\mathbf{e}_y]) \tag*{(4)}

The vector is added to the timestep embedding throughout the network.

Denoiser. The UNet backbone expands its channels progressively (32→64→128 channels, multipliers [1, 2, 4]), with ResNet blocks at each resolution level and self-attention at the coarsest scale. It also uses elucidated diffusion model (EDM) preconditioning [9] with σdata=1\sigma_{\mathrm{data}} = 1.

Noise Schedule. We use a log-uniform noise schedule over the noise level σ\sigma (the diffusion time tt of Eq. (3) of the main text; σ(t)=t\sigma(t) = t in EDM) with σmin⁡=0.002\sigma_{\min} = 0.002 and σmax⁡=200\sigma_{\max} = 200, where the large σmax⁡\sigma_{\max} ensures the forward process fully destroys the climate signal to prevent low-frequency structure from leaking across noise levels [10].

Training. We train for 150 epochs with AdamW [11] (one-cycle learning rate schedule peaking at 10−310^{-3}, batch size 192, 16-bit bfloat mixed-precision training) and maintain an exponential moving average of the network parameters, which we use for all evaluation. We average the EDM loss [9] over grid cells with weights proportional to their true area, so that each region contributes in proportion to its size, following standard practice in climate emulation [12, 13, 14].

Sampling. We sample with the deterministic Heun sampler of Karras et al. [9], which integrates the probability flow ordinary differential equation (ODE) and enables log-likelihood computation.

Shared Initial Noise Sampling. Since the sampler integrates the probability flow ODE deterministically, a sample is a deterministic function of its initial noise, its conditioning inputs and the guidance scale. As in our earlier work [15], we pass the same initial noise tensors to every guidance scale for a given conditioning and sample index ss, which separates the effect of the guidance scale from sampling variability (Figure S1). Each noise tensor is still sampled independently for each sample index, source model identity, CO2_2e value, day, year, and LOOCV fold.

Figure S1. Sharing the initial noise across guidance scales makes the effect of guidance cheaper to estimate. We sample one condition of the MPI-ESM1-2-LR fold (source model NorESM2-LM, SSP5-8.5, day 195 of 2090) at γ=0\gamma= 0, 0.250.25, 0.50.5, 0.750.75 and 11. For a pair of scales the effect of guidance is the difference between the mean fields at the two scales at each grid cell. In paired estimates both scales start from the same initial noise tensors while in unpaired estimates they start from different ones. A reference effect is computed from 37,498 samples per scale and the paired and unpaired estimates are drawn from separate samples so that no initial noise is used twice. Accuracy is the area-weighted root-mean-square error (RMSE) of an estimated effect map against the reference. (A) For each of the ten scale pairs ((52)=10\binom{5}{2} = 10), the number of additional samples per scale an unpaired estimate needs to be as accurate as a paired estimate from nn samples per scale is plotted. Each curve ends where the available unpaired samples are no longer enough. Colors mark the gap between the two scales, Δγ\Delta\gamma: four pairs share Δγ=0.25\Delta\gamma= 0.25, three share 0.50.5, two share 0.750.75 and one has Δγ=1\Delta\gamma= 1. (B) Estimates from only n=10n = 10 samples per scale against the reference, one point per grid cell, for the four pairs that compare γ=0\gamma= 0 with a guided scale. Red points are unpaired estimates. Blue points are paired estimates shaded by Δγ\Delta\gamma as in A. Points on the dashed line indicate agreement with the reference.

Interpolating between Two Distributions

The two endpoints of a path are the two fixed distributions that it connects: that of a climate model and that of the observations. Both are learned by our conditional diffusion model, pθp_\theta, which returns a different distribution of daily temperature fields for each condition (Materials and Methods, Conditional Diffusion Emulator). The identity condition selects whose distribution it returns. As in the main text, we write the two endpoints as pi=pθ(⋅∣CMIP6i,d,c,y)p_i = p_\theta(\cdot\mid\mathrm{CMIP6}_i, d, c, y) for the source (CMIP6 model ii, i∈{1,…,17}i \in\{1,\ldots,17\} in our case) and ptgt=pθ(⋅∣ERA5,d,c,y)p_{\mathrm{tgt}} = p_\theta(\cdot\mid\mathrm{ERA5}, d, c, y) for the target, so pip_i is the emulated climate model ii and ptgtp_{\mathrm{tgt}} the emulated ERA5. In a perfect-model test, the held-out model plays the role of ERA5 and the sixteen remaining models are the sources. A diffusion sampler does not need to evaluate pθp_\theta itself; at each noise level, it follows the score ∇xlog⁡pθ\nabla_x \log p_\theta, the direction in which a field xx becomes more likely, which the denoiser estimates. Eq. (3) of the main text averages the scores of pip_i and ptgtp_{\mathrm{tgt}} with weights 1−γ1-\gamma and γ\gamma; the paths below show which distribution this average is the score of.

Optimal Interpolating Paths. We call any distribution qq of daily temperature fields a candidate for the interpolating distribution, and let D(p,q)D(p,q) be a measure of distance between two distributions. We look for the interpolating distribution, the candidate that is simultaneously close to both endpoints. For each γ∈[0,1]\gamma\in[0,1], it minimizes the weighted objective:

Jγ(q)=(1−γ)D(pi,q)+γD(ptgt,q),(5)J_\gamma(q) = (1-\gamma)D(p_i,q) + \gamma D(p_{\mathrm{tgt}},q), \tag*{(5)}

where the weight γ\gamma sets how much closeness to the target counts relative to closeness to the source. When DD is convex in qq, the minimizers span, as γ\gamma varies, the Pareto frontier of (D(pi,q),D(ptgt,q))(D(p_i,q), D(p_{\mathrm{tgt}},q)) (which is the set of best compromises between the two distances): a minimizer cannot be moved closer to one endpoint without moving farther from the other. The path traced by varying γ\gamma depends on DD, and we consider three choices of DD below.

Forward Kullback–Leibler Divergence Gives the Mixture Path. We first use the forward Kullback–Leibler (KL) divergence, which compares an endpoint pp (pip_i or ptgtp_{\mathrm{tgt}}) with the candidate qq:

KL⁡(p∥q)=Ep[log⁡p(x)q(x)]\operatorname{KL}(p \mathbin{\|} q) = \mathbb{E}_p\left[\log\frac{p(x)}{q(x)}\right]

where the expectation EpE_p is computed under the endpoint (average over fields x\mathbf{x} that the endpoint produces, the emulated climate model for pip_i or the emulated ERA5 for ptgtp_{\mathrm{tgt}}). The divergence is large when the candidate gives low probability to fields that the endpoint produces often. To be close to both endpoints, a candidate must therefore cover the fields of both. Dropping terms that do not involve qq,

Jγ(q)=−∫[(1−γ)pi(x)+γptgt(x)]log⁡q(x) dx+const=KL⁡(pγ,imix∥q)+const,(6)\begin{aligned} J_\gamma(q) &= -\int\left[(1-\gamma)p_i(\mathbf{x})+\gamma p_{\mathrm{tgt}}(\mathbf{x})\right]\log q(\mathbf{x})\,\mathrm{d}\mathbf{x}+\mathrm{const}\\ &=\operatorname{KL}\left(p_{\gamma,i}^{\mathrm{mix}}\mathbin{\|}q\right)+\mathrm{const}, \tag*{(6)} \end{aligned}

with

pγ,imix=(1−γ)pi+γptgt,(7)p_{\gamma,i}^{\mathrm{mix}}=(1-\gamma)p_i+\gamma p_{\mathrm{tgt}}, \tag*{(7)}

where the second line of Eq. (6) adds and subtracts ∫pγ,imix(x)log⁡pγ,imix(x) dx\int p_{\gamma,i}^{\mathrm{mix}}(\mathbf{x})\log p_{\gamma,i}^{\mathrm{mix}}(\mathbf{x})\,\mathrm{d}\mathbf{x}, a term that does not depend on qq and joins the constant. Because the KL divergence is zero when its two arguments coincide and positive otherwise, the minimizer is q=pγ,imixq=p_{\gamma,i}^{\mathrm{mix}} of Eq. (7). This is the arithmetic mixture, which covers the fields of both endpoints, known as the mixture path or m-geodesic. The optimal value is the skewed Jensen–Shannon divergence (1−γ)KL⁡(pi∥pγ,imix)+γKL⁡(ptgt∥pγ,imix)(1-\gamma)\operatorname{KL}(p_i\mathbin{\|}p_{\gamma,i}^{\mathrm{mix}})+\gamma\operatorname{KL}(p_{\mathrm{tgt}}\mathbin{\|}p_{\gamma,i}^{\mathrm{mix}}).

Reverse Kullback–Leibler Divergence Gives the Geometric Path. The reverse KL divergence swaps the roles of the endpoint and the candidate:

KL⁡(q∥p)=Eq[log⁡q(x)p(x)],\operatorname{KL}(q\mathbin{\|}p)=E_q\left[\log\frac{q(\mathbf{x})}{p(\mathbf{x})}\right],

where the expectation EqE_q is now computed under the candidate (average over fields that the candidate qq itself produces). The divergence is large when the candidate produces fields that the endpoint finds unlikely. To be close to both endpoints, a candidate must therefore keep only fields that are plausible under both. Combining the two logarithms,

Jγ(q)=(1−γ)∫q(x)log⁡q(x)pi(x) dx+γ∫q(x)log⁡q(x)ptgt(x) dx=∫q(x)log⁡q(x)pi(x)1−γptgt(x)γ dx=:KL⁡(q∥pγ,i)−log⁡Zγ,(8)\begin{aligned} J_\gamma(q)&=(1-\gamma)\int q(\mathbf{x})\log\frac{q(\mathbf{x})}{p_i(\mathbf{x})}\,\mathrm{d}\mathbf{x}+\gamma\int q(\mathbf{x})\log\frac{q(\mathbf{x})}{p_{\mathrm{tgt}}(\mathbf{x})}\,\mathrm{d}\mathbf{x}\\ &=\int q(\mathbf{x})\log\frac{q(\mathbf{x})}{p_i(\mathbf{x})^{1-\gamma}p_{\mathrm{tgt}}(\mathbf{x})^\gamma}\,\mathrm{d}\mathbf{x}\\ &=:\operatorname{KL}(q\mathbin{\|}p_{\gamma,i})-\log Z_\gamma, \tag*{(8)} \end{aligned}

with

pγ,i=pi(x)1−γptgt(x)γZγ,Zγ=∫pi(x)1−γptgt(x)γ dx.(9)p_{\gamma,i}=\frac{p_i(\mathbf{x})^{1-\gamma}p_{\mathrm{tgt}}(\mathbf{x})^\gamma}{Z_\gamma},\qquad Z_\gamma=\int p_i(\mathbf{x})^{1-\gamma}p_{\mathrm{tgt}}(\mathbf{x})^\gamma\,\mathrm{d}\mathbf{x}. \tag*{(9)}

The normalization constant ZγZ_\gamma does not depend on qq and so does not affect the minimization: minimizing JγJ_\gamma corresponds to minimizing KL⁡(q∥pγ,i)\operatorname{KL}(q\mathbin{\|}p_{\gamma,i}), which, as for the forward divergence, is zero only when q=pγ,iq=p_{\gamma,i}. The minimizer is therefore q=pγ,iq=p_{\gamma,i} of Eq. (9), the weighted geometric mean of Eq. (1) of the main text, known as the geometric path or e-geodesic.

The names of the two paths refer to the coordinate in which each is linear: the mixture path interpolates linearly in the densities, the geometric path in the log-densities [16]. Condition-to-condition guidance follows the geometric path, because at each noise level Eq. (3) of the main text is the score of the weighted geometric mean of the two endpoints pip_i and ptgtp_{\mathrm{tgt}}:

(1−γ)∇log⁡pi+γ∇log⁡ptgt=∇[(1−γ)log⁡pi+γlog⁡ptgt]=∇log⁡[pi1−γptgtγ]=∇log⁡pγ,i,(10)\begin{aligned} (1-\gamma)\nabla\log p_i+\gamma\nabla\log p_{\mathrm{tgt}}&=\nabla\left[(1-\gamma)\log p_i+\gamma\log p_{\mathrm{tgt}}\right]\\ &=\nabla\log\left[p_i^{1-\gamma}p_{\mathrm{tgt}}^\gamma\right]\\ &=\nabla\log p_{\gamma,i}, \tag*{(10)} \end{aligned}

with the last equality holding because ZγZ_\gamma is constant in x\mathbf{x} and vanishes under the gradient. The score used in Eq. (3) of the main text is therefore the score of

pγ,i(xt)∝pθ(xt∣CMIP6i,d,c,y)1−γpθ(xt∣ERA5,d,c,y)γ.(11)p_{\gamma,i}(\mathbf{x}_{t})\propto p_\theta(\mathbf{x}_{t}\mid\mathrm{CMIP6}_i,d,c,y)^{1-\gamma}p_\theta(\mathbf{x}_{t}\mid\mathrm{ERA5},d,c,y)^\gamma. \tag*{(11)}

Eq. (10) involves no approximation: at a given noise level, averaging scores and computing a geometric mean of densities correspond to the same operation. Across noise levels, the reverse diffusion driven by these scores samples the geometric mean of the noise-free conditionals only approximately, because adding noise to a geometric mean of two distributions does not give the geometric mean of the noised distributions (Materials and Methods, Condition-to-Condition Guidance).

Under Total Variation, the Mixture Path Has a Linear Tradeoff. A third choice of DD, total variation (TV), has a simple meaning: it measures how well the best possible classifier can tell the candidate qq from an endpoint pp. For fields drawn from x∼12(q(x)+p(x))x \sim\frac{1}{2}(q(x)+p(x)), that is, from qq or pp with equal probability, the accuracy of this best possible (Bayes-optimal) classifier between qq and pp is

Acc⁡(p,q)=1−12∫min⁡(p(x),q(x)) dx=12(1+TV⁡(p,q)).\operatorname{Acc}(p,q) = 1 - \frac{1}{2} \int\min(p(x),q(x))\,\mathrm{d}x = \frac{1}{2}\left(1+\operatorname{TV}(p,q)\right).

where TV⁡(p,q)=12∫∣p(x)−q(x)∣ dx\operatorname{TV}(p,q) = \frac{1}{2}\int|p(x)-q(x)|\,\mathrm{d}x is the total variation and the second equality uses min⁡(a,b)=(a+b−∣a−b∣)/2\min(a,b)=(a+b-|a-b|)/2. The larger the total variation, the better the classifier tells the two distributions apart.

For the mixture of Eq. (7) as candidate, the differences from the two endpoints are proportional to one another,

pγ,imix−pi=γ(ptgt−pi),pγ,imix−ptgt=(γ−1)(ptgt−pi).p_{\gamma,i}^{\mathrm{mix}}-p_i=\gamma(p_{\mathrm{tgt}}-p_i), \qquad p_{\gamma,i}^{\mathrm{mix}}-p_{\mathrm{tgt}}=(\gamma-1)(p_{\mathrm{tgt}}-p_i).

and therefore

TV⁡(pi,pγ,imix)=γTV⁡(pi,ptgt),\operatorname{TV}(p_i,p_{\gamma,i}^{\mathrm{mix}})=\gamma\operatorname{TV}(p_i,p_{\mathrm{tgt}}),
TV⁡(ptgt,pγ,imix)=(1−γ)TV⁡(pi,ptgt).\operatorname{TV}(p_{\mathrm{tgt}},p_{\gamma,i}^{\mathrm{mix}})=(1-\gamma)\operatorname{TV}(p_i,p_{\mathrm{tgt}}).

Total variation is a metric, so it satisfies the triangle inequality:

TV⁡(pi,ptgt)≤TV⁡(pi,q)+TV⁡(ptgt,q).\operatorname{TV}(p_i,p_{\mathrm{tgt}})\leq\operatorname{TV}(p_i,q)+\operatorname{TV}(p_{\mathrm{tgt}},q).

For q=pγ,imixq=p_{\gamma,i}^{\mathrm{mix}} the equality holds, meaning that the sum of the two TV terms reaches its minimum:

TV⁡(pi,pγ,imix)+TV⁡(ptgt,pγ,imix)=[γ+1−γ]TV⁡(pi,ptgt)=TV⁡(pi,ptgt).\operatorname{TV}(p_i,p_{\gamma,i}^{\mathrm{mix}})+\operatorname{TV}(p_{\mathrm{tgt}},p_{\gamma,i}^{\mathrm{mix}})=[\gamma+1-\gamma]\operatorname{TV}(p_i,p_{\mathrm{tgt}}) =\operatorname{TV}(p_i,p_{\mathrm{tgt}}).

The mixture therefore lies on the linear path between pip_i and ptgtp_{\mathrm{tgt}} in total variation and in accuracy: its distance to pip_i grows, and its distance to ptgtp_{\mathrm{tgt}} shrinks, linearly in γ\gamma.

Which Path Bias Correction Requires. The mixture of Eq. (7) is sampled by drawing from ptgtp_{\mathrm{tgt}} with probability γ\gamma and from pip_i otherwise. Every draw comes from one endpoint or the other, so for well-separated endpoints the mixture is bimodal, with two separate peaks, and no sample lies between them. The geometric mean of Eq. (11) instead assigns high density only where both factors are large, since a field that is unlikely under either endpoint brings the product close to zero. The mixture is a disjunction over the two endpoints: a field is likely if it is likely under one endpoint or the other. The geometric mean is a conjunction: a field is likely only if it is likely under both. Bias correction requires the conjunction: a corrected field should be one that the climate model and the observations could both have produced, in line with the geometric mean in Eq. (11). Figure S2 shows the mixture path and the geometric path side by side for a one-dimensional example built from CMIP6 and ERA5 data.

Figure S2. Mixture and geometric paths between CanESM5 and ERA5 at one Antarctic grid cell (85.1∘85.1^\circS, 140.6∘140.6^\circW), March to May 2001–2010. (A) the mixture path (1−γ)pi+γptgt(1-\gamma)p_i+\gamma p_{\mathrm{tgt}} and (B) the geometric path pi1−γptgtγ/Zγp_i^{1-\gamma}p_{\mathrm{tgt}}^\gamma/Z_\gamma between pip_i, the density of daily near-surface temperature of CanESM5 (blue), and ptgtp_{\mathrm{tgt}}, that of ERA5 (orange), each from 920 daily values. Dashed curves are the paths at γ=0.25\gamma=0.25, 0.50.5 and 0.750.75, darker for larger γ\gamma. Blue and orange ticks below zero mark the daily values of CanESM5 (upper row) and ERA5 (lower row). The paths are defined in Interpretation between Two Distributions.

Fitting the Guidance Schedule

The selections described in Materials and Methods, Selecting the Guidance Strength, give one γ\gamma for each of the seventeen folds, three SSPs and 32 blocks, 17×3×32=1,63217 \times3 \times32 = 1{,}632 values in total (the criterion is defined in Evaluation Metrics, The Criterion Used to Select the Guidance Scale, and the samples in Sampling Designs, The Scale Sweep). Each selection is based on only the twelve days per year of its block, one sample per source model and nine discrete scales, and the selected values vary between folds and blocks. For the projections toward ERA5, no selection is possible, because the future of the target is unknown. The schedule must therefore depend only on quantities known for any future year: the calendar year and the CO2CO_2e concentration of the scenario. To obtain a schedule for all years, we fit one linear function of the calendar year and the CO2CO_2e concentration to all values pooled over folds and SSPs:

γ(y,c)=3.0516−1.0138×10−3y−1.3605×10−4c,(12)\gamma(y,c) = 3.0516 - 1.0138 \times10^{-3} y - 1.3605 \times10^{-4} c, \tag*{(12)}

where yy is the midpoint year of a block and cc its mean CO2CO_2e concentration in ppm, and we clip the result to [0,1][0,1]. Evaluated along the CO2CO_2e pathway of each SSP, the schedule declines from γ=0.940\gamma= 0.940 in 2025 to 0.841, 0.805 and 0.768 in 2100 under SSP2-4.5, SSP3-7.0 and SSP5-8.5, that is −0.0135-0.0135, −0.0180-0.0180 and −0.0231-0.0231 per decade (Figure S3A). We pool over folds so that the schedule is not tuned to any one held-out model, in line with the assumption that it transfers to ERA5 (Materials and Methods, Selecting the Guidance Strength). We pool over SSPs and include the CO2CO_2e term so that the dependence on the scenario enters through the forcing and one function serves all three scenarios. A linear function is the simplest form that lets the scale change gradually with year and CO2CO_2e. Figure S3B shows the same schedule rounded to the nearest of the nine scales of the scale sweep. We use this rounded schedule only to read scheduled-guidance results from the sweep, which covers all seventeen folds (Sampling Designs, The Scale Sweep); the scheduled run and the projections toward ERA5, which sample every day, use the continuous schedule of Figure S3A.

Estimating the Rescaling Factor

Because our emulator samples each day independently, its annual global means keep almost none of the year-to-year variability that climate models and observations show, and we correct this by rescaling (Materials and Methods, Annual-Variance Rescaling). The rescaling acts on each emulated annual global-mean series: the annual global means gyg_y of one sample of one source model, one value for each year yy. It replaces each series gyg_y by

gy′=g~y+k(gy−g~y),k=max⁡(σtgtσemulator,1),(13)g'_y = \tilde{g}_y + k\left(g_y - \tilde{g}_y\right), \qquad k = \max\left(\frac{\sigma_{\mathrm{tgt}}}{\sigma_{\mathrm{emulator}}},1\right), \tag*{(13)}

where g~y\tilde{g}_y is the cubic polynomial in time fitted to gyg_y by least squares. The cubic keeps the forced trajectory, the departure of the source model from the others and the constant offset of the sample, so the residual gy−g~yg_y - \tilde{g}_y is the year-to-year fluctuation. Eq. (13) thus keeps the trend g~y\tilde{g}_y unchanged and stretches the fluctuations around it by the factor kk, which compares how much the annual mean varies from year to year in the target and in the emulator and is never smaller than 1.

For the target, we use its 66 annual global means GyG_y of 1979–2024 (for a held-out CMIP6 model, its historical simulation followed by SSP2-4.5 from 2015). Both the expected temperature of a year and the spread of the annual mean around it can change over time, so we estimate both as functions of the year. Following Guillaumin and Zanna [17], a small neural network learns from these values the expected temperature of each year, μy\mu_y, and then, with μy\mu_y held fixed, the standard deviation around it, sys_y, the year-to-year variability of the target in year yy. The typical deviation over the whole period, σtgt\sigma_{\mathrm{tgt}}, is:

σtgt=(146∑y=19792024sy2)1/2.\sigma_{\mathrm{tgt}} = \left(\frac{1}{46}\sum_{y=1979}^{2024}s_y^2\right)^{1/2}.

For the emulator, we use its samples replicating the target (γ=1\gamma= 1) over the same years, 1979–2024. We use them on the same days of each year as the series we rescale: every day for the designs that sample every day, and the twelve days per year of the scale sweep for the sweep (Sampling Designs). We fit the cubic trend g~y\tilde{g}_y of each of these series, compute the standard deviation of its residuals gy−g~yg_y - \tilde{g}_y, and use the median of these standard deviations over the series as σemulator\sigma_{\mathrm{emulator}}. Two cases need special treatment. The sweep samples only from 2025 and its pattern of days repeats every 31 years, so over 1979–2024 we continue the same pattern backward from 2025. The emulator behind classifier guidance never learned the held-out model and so cannot replicate it; for classifier guidance, we use the series we rescale, over 2025–2100.

Additional Details for Figures

Figure 1. Figure 1 of the main text shows the LOOCV fold in which MPI-ESM1-2-LR is held out as the synthetic observation, under SSP5-8.5, over two periods: 2015–2024 (in-distribution emulation) and 2085–2100 (long-term extrapolation). We build it in four steps: sampling, removal of the seasonal cycle, projection with a principal component analysis (PCA) and density estimation. Sampling: for each period, we draw ten samples each day, for γ∈{0,0.25,0.5,0.75,1}\gamma\in\{0, 0.25, 0.5, 0.75, 1\}, using the same initial noise for every guidance scale (Machine Learning Setup Details, Shared Initial Noise Sampling). Removal of the seasonal cycle: before fitting the PCA and projecting the data, we remove the seasonal cycle to make the impact of guidance clearly visible. We remove it separately for each group of data: the synthetic observation, each source in the simulation data, and each source at each γ\gamma in the generated data. For each group and grid cell, we fit the mean and first three harmonics of the annual cycle (sinusoids with periods of one year, six months and four months) to all days of the period by least squares and subtract the harmonics, similar to Wheeler and Hendon [18]. Projection: for each period, we fit a PCA only to the simulation data of that period, from the sixteen source models and from the held-out model serving as our synthetic observation. We then project onto its first two principal components the samples drawn above from the sixteen sources and, for the synthetic observation, its ground-truth simulation data.

Density estimation: these steps give point clouds in the projected PCA space, one for our method for each period and γ\gamma and one for the synthetic observation for each period. To visualize the distributions and their densities, we compute a 2D kernel density estimation (KDE) for each cloud, a smooth estimate of its density. We discretize the densities and show the highest-density regions, the smallest regions that contain 39.35%, 86.47% and 98.89% of the distribution, shaded darker the smaller the region. These are the fractions inside the 1, 2 and 3 standard-deviation contours of a two-dimensional Gaussian distribution, 1−e−n2/21 - e^{-n^2/2} for n=1,2,3n = 1, 2, 3.

The bold frames mark, for each period, the guidance scale closest to the selected value. For 2085–2100, the selected value is that of the continuous γ\gamma schedule fitted over 2025–2100, which we round to the nearest scale shown; for 2015–2024, where the target is known, it is γ=1\gamma= 1 (Materials and Methods, Selecting the Guidance Strength). We show fixed scales, rather than the schedule, to display how γ\gamma changes the sampled distribution. As γ\gamma increases toward 1, the sampled distribution changes from multimodal to the single target distribution. It is multimodal at small γ\gamma because it pools the sixteen sources, each of which forms its own compact cloud (Figure S4).

Figure S4. Figure 1 of the main text decomposed by source model. Same fold (MPI-ESM1-2-LR as synthetic observation), scenario (SSP5-8.5), windows, guidance scales, projection and density estimate as Figure 1. Each of the sixteen source models is drawn as its own kernel density with the same three highest-density bands (39.35, 86.47 and 98.89%), filled at 20% opacity in the color given in the key; the gray outlines are the bands of the sixteen sources pooled, the blue distribution of Figure 1, and the coral bands are the synthetic observation. The key shows each source’s color as it appears on its own; where sources overlap their fills stack and the color deepens. The pooled distribution is a mixture of individually compact clouds, and every cloud contracts toward the emulator’s version of the synthetic observation (γ=1\gamma= 1) as γ\gamma increases.

Sampling Designs

Every evaluation and projection in the main text and in this Supporting Information is based on one of the five sampling designs below. Figure 1 of the main text and Figure S4, which illustrate how guidance acts, use their own samples, described in Additional Details for Figures.

The Emulator Replicating ERA5. The emulator conditioned on ERA5 (γ=1\gamma= 1) replicates the observed record with seventeen samples on every day of 1979–2024. Table S4 and Table S5 and Figure S18–Figure S22 use these samples over 2001–2010, the decade withheld from training and used for validation. Figure 2A, Figure 3 and Figure 4 of the main text use them before 2025.

RawGuidedMVTQDM
Mean absolute bias [°C]0.7380.2070.2290.226
RMSE [°C]0.9520.3050.3450.341
1st-percentile bias [°C]1.5430.5490.5420.439
99th-percentile bias [°C]2.2540.4430.4380.357
Standard-deviation ratio deviation0.2800.0550.0430.035
TX90p bias [percentage points]13.202.612.642.37
Wasserstein distance [°C]0.9550.2860.2920.278
Laplacian roughness [°C]0.6460.2450.3600.355
RI, per cell0.02510.00260.00300.0028

Table S4. Distributional error against ERA5, 2001–2010. Raw CMIP6 output, guided samples at γ=1\gamma= 1, and the raw output after MVT and QDM bias correction, each with seventeen members pooled, on every day of the decade. Every metric is computed at each grid cell against ERA5 and reduced to one scalar by area weighting. Lower is better, and the smallest value of each row is in bold. The RI floor is 0.0003. The metrics are defined in Evaluation Metrics, Distributional Metrics.

regionquantityraw CMIP6guidedMVTQDM
High Mountain Asiadisplacement [°C]0.8560.3342.5162.480
roughness ratio (ERA5 1.00)1.0971.0260.9000.896
Europedisplacement [°C]0.3980.0751.5971.579
roughness ratio (ERA5 1.00)1.3101.0320.9790.937

Table S5. Spatial consistency at two adjacent grid-cell pairs. The cell pairs (Evaluation Metrics, Spatial-Consistency Metrics) are 37.7°N and 40.5°N at 87.2°E over High Mountain Asia, whose observed mean temperatures differ by 14.2°C, and the same two latitudes at 39.4°E, in eastern Anatolia inside the Europe box, where they differ by 6.6°C. Displacement is given as an absolute value in °C, and smaller is better. Roughness ratio, over each pair’s regional box, is relative to ERA5, where 1 is a perfect match. Figure S21 and Figure S22 show the joint distributions of the two pairs. Guided samples at a fixed γ=1\gamma= 1 over 2001–2010, every day of the year, seventeen members, as in Table S4. The metrics are defined in Evaluation Metrics, Spatial-Consistency Metrics.

Figure S18. Performance comparison of condition-to-condition guidance, MVT and QDM (applied to the raw CMIP6 output), evaluated against ERA5 over 2001–2010. Nine distributional error diagnostics. Columns, left to right: raw CMIP6 output; guided diffusion samples (γ=1\gamma= 1); and the same raw output after mean–variance–trend (MVT) and quantile delta mapping (QDM) bias correction. Rows, top to bottom: mean bias; ratio of modeled to ERA5 daily standard deviation; 1st- and 99th-percentile bias; bias in the exceedance frequency of each cell’s ERA5 90th-percentile warm-day (TX90p) and 10th-percentile cold-day (TN10p) thresholds, in percentage points; 1-Wasserstein distance between the daily distributions; skewness bias; and excess-kurtosis bias.

Figure S22. Performance comparison of condition-to-condition guidance, MVT and QDM (applied to the raw CMIP6 output), evaluated against ERA5 over 2001–2010. Mean bias and the joint distribution of two adjacent grid cells over Europe, 2001–2010. As Figure S21, for the box 35°–60°N, 0°–40°E and its cell pair: 37.7°N and 40.5°N, both at 39.4°E, whose ERA5 mean temperatures differ by 6.6°C. The dotted reference line is the principal axis of the ERA5 joint distribution for this pair.

The Projections toward ERA5. We guide each of the seventeen CMIP6 models toward ERA5 with the continuous γ\gamma schedule, with ten samples per source CMIP6 model on every day of 2025–2100. Figure 2A, the 2081–2100 half of Figure 3 and Table S6 and Table S7 use all ten samples. Figure 4 uses one sample per source model and day, so that the guided and raw CMIP6 distributions are based on the same number of maps.

SSP2-4.5SSP3-7.0SSP5-8.5
Raw CMIP6 inter-model range [°C]1.1491.6182.103
Guided inter-model range [°C]0.2590.3320.482
Retained inter-model spread22.5%20.5%22.9%

Table S6. Inter-model spread retained by the guided ensemble, 2081–2100. The metrics are defined in Evaluation Metrics, Projection and Extreme-Heat Metrics.

LevelScenarioFirst 20-year periodCrossing year5–95% range0.98 ∘C^\circ\mathrm{C}0.67 ∘C^\circ\mathrm{C}IPCC AR6
1.5∘C1.5^\circ\mathrm{C}SSP2-4.52020–203920292028–2032202420362021–2040 [2012–2031 to 2037–2056]
SSP3-7.02020–203920292028–2031202520352021–2040 [2013–2032 to 2033–2052]
SSP5-8.52019–203820282027–2030202420332018–2037 [2011–2030 to 2029–2048]
2∘C2^\circ\mathrm{C}SSP2-4.52038–205720472045–2051204320542043–2062 [2028–2047 to 2075–2094]
SSP3-7.02036–205520452042–2048204020512037–2056 [2026–2045 to 2053–2072]
SSP5-8.52032–205120412039–2043203820452032–2051 [2023–2042 to 2044–2063]

Table S7. Years in which the guided ensemble mean first exceeds 1.5∘C1.5^\circ\mathrm{C} and 2∘C2^\circ\mathrm{C} above 1850–1900. The first 20-year period above each level, the crossing year and its 5–95% range across the 170 guided series, with the assessed warming of 0.85∘C0.85^\circ\mathrm{C} from 1850–1900 to 1995–2014. The next two columns give the crossing year with that warming at its very likely bounds, 0.980.98 and 0.67∘C0.67^\circ\mathrm{C}. The last column gives the first 20-year period above each level assessed by IPCC AR6 WGI (Table 4.5 of the report), with its very likely (5–95%) range in brackets [3]. The crossing year is defined in Evaluation Metrics, Warming-Level Crossing Year.

The Scheduled Run. Sampling every day with ten samples per source model is computationally expensive, so we use this design only for two LOOCV folds, as case studies: those with MIROC6 and MPI-ESM1-2-LR as the synthetic observation. In each, we guide the sixteen remaining models toward the held-out one with the continuous γ\gamma schedule, with ten samples per source model on every day of 2025–2100. For MIROC6, we also sample, at the same density, the diffusion emulator without guidance (γ=0\gamma= 0) and classifier guidance. Figure 2B of the main text and Figure S23 use all ten samples. The comparisons of methods (Table S3 and Table S9) and the maps and distributions of the daily fields (Figure S10–Figure S17 and Figure S24) use one of the ten samples per source model and day, so that every ensemble has sixteen members, one per source model, as in the raw CMIP6 archive used in this work. The guided panels of Figure S25–Figure S27 use one sample per source model on only the twelve days per year of the scale sweep (the same days as their unguided panels; see The Scale Sweep).

Table S3. Held-out MIROC6: condition-to-condition guidance against MVT, QDM, and classifier guidance. Upper table: annual global-mean warming over 2025–2100 (complement to Figure 2B of the main text), with trajectory RMSE and band coverage. Lower table: daily fields over 2081–2100, each metric computed at each grid cell against MIROC6’s own daily fields and reduced to one scalar by area weighting. Raw CMIP6 is the ensemble of the sixteen source simulations. Emulator γ=0\gamma= 0 is the emulator conditioned on the same sixteen source models without guidance. Guided is the emulator guided from the same sixteen models toward MIROC6 with the γ\gamma schedule. MVT and QDM are applied either to the raw CMIP6 data (on raw CMIP6) or to the output of the emulator without guidance (on emulator output). Classifier guidance is the second emulator described in the text. In both tables each column has sixteen members, one per source model, with one sample per member and day, and the lower table pools the members. The bands of the emulator without guidance, of the guided emulator and of classifier guidance are rescaled (Materials and Methods), and those of raw CMIP6, MVT and QDM are not. The RI floor is 0.0001. The smallest value of each metric except the band coverage is in bold. The metrics are defined in Evaluation Metrics, Trajectory Metrics and Distributional Metrics.

MetricScenarioRaw CMIP6Emulator γ=0\gamma= 0GuidedMVTQDMMVTQDMClassifier guidance
on raw CMIP6on raw CMIP6on emulator outputon emulator output
Annual global-mean warming, 2025–2100
Trajectory RMSE [°C]SSP2-4.50.3630.3680.2140.3110.3050.3030.3130.514
SSP3-7.00.4670.4460.2000.4310.4080.4000.3930.459
SSP5-8.50.5320.5110.2010.4900.4660.4650.4560.360
Band coverageSSP2-4.50.820.780.620.470.870.050.710.24
SSP3-7.00.780.790.720.410.830.030.720.25
SSP5-8.50.830.760.670.290.890.050.720.39
Daily fields, 2081–2100
Mean absolute bias [°C]SSP2-4.51.3881.3940.3730.6190.5990.5940.5600.824
SSP3-7.01.3631.3720.4120.8380.7700.7540.7030.824
SSP5-8.51.4041.4260.5150.9470.8670.8660.7970.900
RMSE [°C]SSP2-4.52.0752.0710.5930.8100.7940.7910.7571.225
SSP3-7.01.9561.9730.5861.0010.9330.9300.8711.203
SSP5-8.51.9782.0100.7301.1571.0821.0971.0241.281
Standard-deviation ratio deviationSSP2-4.50.3310.3390.0750.0520.0590.0530.0570.177
SSP3-7.00.3410.3430.0920.0720.0870.0650.0830.168
SSP5-8.50.3490.3550.1080.0810.1230.0710.1140.168
Skewness biasSSP2-4.50.2200.2140.1070.0950.1030.0910.0900.196
SSP3-7.00.2100.2020.1100.1030.1140.1010.1060.191
SSP5-8.50.2130.2120.1210.1030.1430.1030.1360.187
Excess-kurtosis biasSSP2-4.50.5810.5700.2030.2440.2180.2040.1660.414
SSP3-7.00.5860.5630.2100.3010.2800.2260.2240.410
SSP5-8.50.5660.5720.2350.3100.3330.2420.2860.426
TX90p bias [percentage points]SSP2-4.512.4112.544.248.428.088.047.569.28
SSP3-7.013.7413.645.3112.5111.3211.2110.279.90
SSP5-8.514.4414.266.0614.0712.7612.8611.7111.31
TN10p bias [percentage points]SSP2-4.59.789.892.734.594.384.674.286.42
SSP3-7.08.999.242.775.214.665.204.425.66
SSP5-8.58.929.113.485.684.805.744.665.32
1st-percentile bias [°C]SSP2-4.52.6262.5820.6720.7020.7400.7580.7701.333
SSP3-7.02.4482.4420.6700.7910.7770.8350.7791.175
SSP5-8.52.4482.4710.8020.9770.8951.0530.9281.264
99th-percentile bias [°C]SSP2-4.52.2492.2520.7260.7310.8690.6290.7371.558
SSP3-7.02.4672.4190.8871.0531.2590.8321.1071.605
SSP5-8.52.6932.6541.0701.2481.6340.9871.4261.706
CRPS [°C]SSP2-4.51.4861.4851.0191.1141.0821.1061.0691.238
SSP3-7.01.4641.4661.0221.2021.1331.1741.1071.230
SSP5-8.51.4671.4721.0391.2491.1531.2251.1321.261
RI, per cellSSP2-4.50.05390.05280.00590.03620.02220.03700.01960.0329
SSP3-7.00.05550.05450.00720.06210.03290.05900.02680.0351
SSP5-8.50.06020.05860.01060.08190.03630.07950.03230.0457
Laplacian roughness [°C]SSP2-4.51.2531.2850.4610.4600.4550.4630.4571.159
SSP3-7.01.2481.2810.5170.4850.4710.4890.4701.170
SSP5-8.51.2701.3000.5900.5300.5120.5350.5131.193
Wasserstein distance [°C]SSP2-4.51.5591.5710.4660.6680.6500.6470.6150.961
SSP3-7.01.5441.5540.5000.8810.8180.8020.7540.947
SSP5-8.51.5951.6110.6050.9960.9280.9210.8601.030

Table S3.

MetricScenarioRaw CMIP6Guided
Annual global-mean warming, 2025–2100
Trajectory RMSE [∘C^\circ\mathrm{C}]SSP2-4.50.3920.257
SSP3-7.00.4240.397
SSP5-8.50.6210.394
Band coverageSSP2-4.50.840.34
SSP3-7.00.790.17
SSP5-8.50.630.38
Daily fields, 2081–2100
Mean absolute bias [∘C^\circ\mathrm{C}]SSP2-4.51.1190.391
SSP3-7.01.1350.613
SSP5-8.51.3610.651
RMSE [∘C^\circ\mathrm{C}]SSP2-4.51.3960.583
SSP3-7.01.4220.839
SSP5-8.51.6590.865
Standard-deviation ratio deviationSSP2-4.50.2170.062
SSP3-7.00.2400.075
SSP5-8.50.2450.076
Skewness biasSSP2-4.50.2160.111
SSP3-7.00.2310.124
SSP5-8.50.2190.120
Excess-kurtosis biasSSP2-4.50.5430.261
SSP3-7.00.5760.292
SSP5-8.50.5420.294
TX90p bias [percentage points]SSP2-4.513.972.42
SSP3-7.013.973.70
SSP5-8.516.734.04
TN10p bias [percentage points]SSP2-4.55.023.73
SSP3-7.04.915.87
SSP5-8.55.056.45
1st-percentile bias [∘C^\circ\mathrm{C}]SSP2-4.51.7370.906
SSP3-7.01.6541.146
SSP5-8.51.5471.154
99th-percentile bias [∘C^\circ\mathrm{C}]SSP2-4.52.6050.438
SSP3-7.02.7900.566
SSP5-8.53.2100.644
CRPS [∘C^\circ\mathrm{C}]SSP2-4.51.2971.016
SSP3-7.01.2911.060
SSP5-8.51.3681.061
RI, per cellSSP2-4.50.03920.0049
SSP3-7.00.03680.0109
SSP5-8.50.04890.0128
Laplacian roughness [∘C^\circ\mathrm{C}]SSP2-4.51.1780.356
SSP3-7.01.1730.423
SSP5-8.51.1840.475
Wasserstein distance [∘C^\circ\mathrm{C}]SSP2-4.51.2630.440
SSP3-7.01.2680.650
SSP5-8.51.4780.690

Table S9. Table S9. Held-out MPI-ESM1-2-LR: condition-to-condition guidance against raw CMIP6. The metrics, samples and conventions of Table S3, for raw CMIP6 and the emulator guided from the sixteen remaining models. The RI floor is 0.0001. The metrics are defined in Evaluation Metrics, Trajectory Metrics and Distributional Metrics.

Figure S10. LOOCV performance comparison of condition-to-condition guidance, MVT, QDM (applied to raw CMIP6 data), and classifier guidance for the MIROC6 fold. Nine distributional error diagnostics against the held-out MIROC6, 2081–2100, SSP3-7.0. Columns, left to right: the raw output of the sixteen CMIP6 source models; the emulator guided from the same sixteen models with the γ\gamma schedule; the raw output after MVT and QDM bias correction; and classifier guidance. Rows, top to bottom: mean bias; ratio of modeled to MIROC6 daily standard deviation; 1st- and 99th-percentile bias; bias in the exceedance frequency of each cell’s MIROC6 90th-percentile warm-day (TX90p) and 10th-percentile cold-day (TN10p) thresholds, in percentage points; 1-Wasserstein distance between the daily distributions; skewness bias; and excess-kurtosis bias. Computed with one sample per source model and day of 2081–2100 for the emulator and for classifier guidance.

Figure S17. LOOCV performance comparison of condition-to-condition guidance, MVT, and QDM (applied to the unguided emulator output) for the MIROC6 fold. Seasonal cycle of the global-mean temperature, held-out MIROC6, 2081–2100. As Figure S13, with the methods of Figure S15.

Figure S23. LOOCV performance of condition-to-condition guidance for the MPI-ESM1-2-LR fold. Global-mean warming of the held-out MPI-ESM1-2-LR and of the guided ensemble, 1979–2100. Annual global-mean warming above 1850–1900 under SSP2-4.5, SSP3-7.0 and SSP5-8.5 (top to bottom) for MPI-ESM1-2-LR (black), the sixteen remaining CMIP6 models (blue, 5–95% band) and the diffusion emulator (green, ensemble mean and 5–95% band). The emulator replicates MPI-ESM1-2-LR (γ=1\gamma= 1) over 1979–2024 with ten samples per day, and is guided from the sixteen source models with the γ\gamma schedule over 2025–2100 with ten samples per source model and day. Yellow shading marks the years of MPI-ESM1-2-LR simulation in the emulator’s training data. Warming and bands are constructed as in Figure 2 of the main text.

Figure S24. LOOCV performance of condition-to-condition guidance for the MPI-ESM1-2-LR fold. Nine distributional error diagnostics against the held-out MPI-ESM1-2-LR, SSP3-7.0, 2080–2100. Columns, left to right: the raw output of the sixteen CMIP6 source models; the emulator guided from the same sixteen models with the γ\gamma schedule. Rows, top to bottom: mean bias; ratio of modeled to MPI-ESM1-2-LR daily standard deviation; 1st- and 99th-percentile bias; bias in the exceedance frequency of each cell’s MPI-ESM1-2-LR 90th-percentile warm-day (TX90p) and 10th-percentile cold-day (TN10p) thresholds, in percentage points; 1-Wasserstein distance between the daily distributions; skewness bias; and excess-kurtosis bias. Native 64×12864 \times128 grid, Robinson projection, every day of 2080–2100, with one sample per source model and day for the emulator. The sixteen members are pooled into a single sample of 122,640 days per cell and each diagnostic is computed once on that sample; MPI-ESM1-2-LR contributes its own 7,665 days. Color scales are shared along each row. They are centered on zero, except for the standard-deviation ratio, centered on 1 between 0.75 and 1.25, and the Wasserstein distance, which starts at zero.

Figure S25. LOOCV performance of condition-to-condition guidance for the MPI-ESM1-2-LR fold. Distribution of the daily global-mean temperature, held-out MPI-ESM1-2-LR. Kernel density of the daily area-weighted global-mean temperature under SSP2-4.5 (A, B), SSP3-7.0 (C, D) and SSP5-8.5 (E, F), over 2030–2050 (A, C, E) and 2080–2100 (B, D, F). In each panel, left: the emulator conditioned on the sixteen source models without guidance (γ=0\gamma= 0); right: the emulator guided from the same models with the γ\gamma schedule. Both halves use one sample per source model on the same twelve days per year, and MPI-ESM1-2-LR (dotted) is shown on those days (Sampling Designs). The emulator curve pools the sixteen source models.

Figure S27. LOOCV performance of condition-to-condition guidance for the MPI-ESM1-2-LR fold. Ensemble calibration against the held-out MPI-ESM1-2-LR. Area-weighted rank histogram of MPI-ESM1-2-LR among the sixteen ensemble members at each grid cell and day under SSP2-4.5 (A, B), SSP3-7.0 (C, D) and SSP5-8.5 (E, F), over 2030–2050 (A, C, E) and 2080–2100 (B, D, F). In each panel, left: the emulator conditioned on the sixteen source models without guidance (γ=0\gamma= 0); right: the emulator guided from the same models with the γ\gamma schedule. Both halves use one sample per source model on the same twelve days per year (Sampling Designs). The dashed line marks the uniform frequency 1/171/17 expected of a calibrated sixteen-member ensemble. The RI floor is 0.0037.

The Scale Sweep. The design used to select the guidance strength covers nine discrete scales γ∈{0,0.125,…,0.875,1}\gamma\in\{0, 0.125, \ldots, 0.875, 1\}, all seventeen folds of the LOOCV and all three SSPs over 2025–2100. We draw one sample per source CMIP6 model, day, and scale on twelve days per year, about a month apart and shifted from year to year so that every calendar day is used over the period. Results with the scheduled guidance strength that we read from the sweep require no further sampling: for each year, we use the samples of the sweep at the scale nearest the value of the γ\gamma schedule of Eq. (12) for that year. This rounding gives γ=1\gamma= 1 in 2025 and 2026, 0.875 afterward, and 0.75 from 2097 under SSP3-7.0 and from 2085 under SSP5-8.5; under SSP2-4.5, the scale stays at 0.875 through 2100 (Figure S3B). The two designs that guide with the schedule on every day, The Scheduled Run and The Projections toward ERA5, use the continuous schedule instead. At γ=1\gamma= 1 the sweep holds ten samples of the emulator conditioned on the held-out model and no sixteen-member ensemble, so we score 2025 and 2026 with the γ=0.875\gamma= 0.875 samples. Sparse sampling makes all seventeen folds and every scale affordable, at the cost of depth: twelve days a year over a twenty-one-year window give 252 values per grid cell and member, so only about 2.5 of them lie below the 1st percentile of a single cell, and as many above the 99th. The selection criterion averages over grid cells and members, which makes this depth sufficient. Every seventeen-fold statistic in the main text comes from this design: the trajectory root mean squared error (RMSE), the percentile biases and the distributional metrics. Because the γ\gamma schedule is fitted to selections made on these folds, these statistics are in-sample for the choice of γ\gamma. Tables S1 and S2, Figure S5–Figure S9 and the unguided panels (γ=0\gamma= 0) of Figure S25–Figure S27 use this design.

Figure S5. LOOCV performance across seventeen folds. Mean bias of the emulated ensemble against each held-out model, without and with guidance, SSP3-7.0, 2080–2100. (A–Q) One LOOCV fold each, named by the held-out model. Left of each pair: the emulator conditioned on the sixteen source models without guidance (γ=0\gamma= 0). Right: the emulator guided toward the held-out model with the γ\gamma schedule rounded to the nearest sampled scale. Each map is the mean at every grid cell minus that of the held-out model. Samples from the scale sweep, with one sample per source model on twelve days per year, the sixteen source models pooled and the held-out model’s run on the same days (Sampling Designs).

Figure S9. LOOCV performance across seventeen folds. Ensemble calibration against each held-out model, without and with guidance, SSP3-7.0, 2080–2100. Folds, pairs, γ\gamma and sampling as in Figure S5. Area-weighted rank histogram of the held-out model among the sixteen ensemble members at each grid cell and day. The dashed line marks the uniform frequency 1/171/17 expected of a calibrated sixteen-member ensemble. The RI floor is 0.0037.

The Emulator Conditioned on Each Model Identity. To evaluate the two endpoints between which guidance interpolates, we sample the emulator conditioned on each of the seventeen CMIP6 models without guidance (γ=0\gamma= 0), with ten samples on every day of 2080–2100 (and of 2030–2050 for the figures at https://github.com/mandt-lab/gdso-figures), and conditioned on ERA5 (γ=1\gamma= 1), with ten samples on every day of 2001–2010. Figure S28–Figure S33 use this design.

Figure S28. Performance of the diffusion emulator. Mean bias of the diffusion emulator conditioned on each of eighteen model identities, SSP3-7.0, 2080–2100. The emulator is trained jointly on seventeen CMIP6 models over 1979–2100 and on ERA5 over 1979–2000 and 2011–2024. In (A–Q) it is conditioned on one CMIP6 model without guidance, the γ=0\gamma= 0 endpoint of condition-to-condition guidance, sampled ten times for every day of 2080–2100 (years of CMIP6 record included in the training set) under SSP3-7.0 and compared with that model’s own daily temperature on the same days. In (R) it is conditioned on ERA5 without guidance, the γ=1\gamma= 1 endpoint, sampled ten times for every day of 2001–2010 (decade of ERA5 record withheld from training set), and compared with ERA5 daily temperature. Each panel shows, at each grid cell, the emulator’s mean daily temperature over the window minus the one of the target.

Figure S33. Performance of the diffusion emulator. Seasonal cycle of the global-mean temperature of the diffusion emulator conditioned on each of eighteen model identities, SSP3-7.0, 2080–2100. Setup, panels, sampling and windows as in Figure S28. Area-weighted global-mean temperature by calendar month, averaged over the years of the window, for the emulator (blue), with the ten samples pooled, and for the target (black, dotted).

Baseline Methods

This section describes the three methods against which the main text benchmarks condition-to-condition guidance: two statistical bias corrections, mean–variance–trend (MVT) and quantile delta mapping (QDM), and classifier guidance. We evaluate all three in the held-out MIROC6 fold, and the two bias corrections also against ERA5 over 2001–2010.

Mean–Variance–Trend Transform (MVT). We implement the MVT correction of Xu et al. [19] independently at each grid cell. Let xsrc(τ)x_{\mathrm{src}}(\tau) be the daily temperature of a source CMIP6 model on date τ\tau, Tsrc(τ)T_{\mathrm{src}}(\tau) its nonlinear monthly trend, and Csrc(d)C_{\mathrm{src}}(d) its climatology after removing that trend, for calendar day dd of τ\tau. The source weather anomaly is

asrc(τ)=xsrc(τ)−Tsrc(τ)−Csrc(d).(14)a_{\mathrm{src}}(\tau) = x_{\mathrm{src}}(\tau) - T_{\mathrm{src}}(\tau) - C_{\mathrm{src}}(d). \tag*{(14)}

We estimate the nonlinear trends from monthly temperature with ensemble empirical mode decomposition (EEMD). The corrected temperature is

x^MVT(τ)=Ctgt(d)+asrc(τ)σtgt(d)σsrc(d)+TMME(τ)−Bm,(15)\widehat{x}_{\mathrm{MVT}}(\tau) = C_{\mathrm{tgt}}(d) + a_{\mathrm{src}}(\tau) \frac{\sigma_{\mathrm{tgt}}(d)}{\sigma_{\mathrm{src}}(d)} + T_{\mathrm{MME}}(\tau) - B_m, \tag*{(15)}

where TtgtT_{\mathrm{tgt}} and CtgtC_{\mathrm{tgt}} are the target’s nonlinear monthly trend and its climatology after removing that trend, σtgt(d)\sigma_{\mathrm{tgt}}(d) and σsrc(d)\sigma_{\mathrm{src}}(d), the standard deviations of the target and source anomalies on calendar day dd, rescale the daily variability, and TMMET_{\mathrm{MME}} is the nonlinear trend of the raw-CMIP6 multi-model ensemble mean. The monthly trend correction is

Bm=T‾MME,mH−T‾tgt,mH,(16)B_m = \overline{T}_{\mathrm{MME},m}^{H} - \overline{T}_{\mathrm{tgt},m}^{H}, \tag*{(16)}

where the overline denotes the average over the calibration years HH in calendar month mm, the month of τ\tau. Thus, MVT combines the target climatology and variability with the projected nonlinear trend of the source-model ensemble. We use five EEMD trials with a noise width of 0.2 and recompute the daily standard deviations after excluding anomalies larger than three times their initial standard deviation. Applied to the unguided emulator (γ=0\gamma= 0), MVT estimates the monthly source trend and the calibration statistics jointly from all samples.

ERA5 as the Target. We compute TMMET_{\mathrm{MME}} from all seventeen raw CMIP6 models. Historical calibration uses 1979–2000 and 2011–2014, and each SSP adds 2015–2024, so the calibration years are 1979–2000 and 2011–2024. The decade 2001–2010 is excluded from all calibration statistics and retained for validation.

MIROC6 as the Target. MIROC6 replaces ERA5 as the target and is not corrected, and we recompute TMMET_{\mathrm{MME}} from the remaining sixteen raw CMIP6 models. We use the same calibration years and withhold the MIROC6 simulation over 2001–2010 and 2025–2100 for validation. Because EEMD requires a continuous monthly series, we fill the missing target values over 2001–2010 by linear interpolation, separately for each calendar month, only to estimate TtgtT_{\mathrm{tgt}}. The target climatology, variability and trend-bias statistics use only the calibration years.

Quantile Delta Mapping (QDM). We implement additive QDM following Cannon et al. [20], independently at each grid cell. For a source temperature xsrc(τ)x_{\mathrm{src}}(\tau), its percentile in the local projected source distribution is

p(τ)=Fsrc,proj(τ)[xsrc(τ)].p(\tau) = F_{\mathrm{src,proj}(\tau)}[x_{\mathrm{src}}(\tau)].

The corrected temperature is

x^QDM(τ)=Ftgt,H−1[p(τ)]+xsrc(τ)−Fsrc,H−1[p(τ)],(17)\widehat{x}_{\mathrm{QDM}}(\tau) = F_{\mathrm{tgt},H}^{-1}[p(\tau)] + x_{\mathrm{src}}(\tau) - F_{\mathrm{src},H}^{-1}[p(\tau)], \tag*{(17)}

where Ftgt,HF_{\mathrm{tgt},H} and Fsrc,HF_{\mathrm{src},H} are the empirical target and source distributions over the calibration years. The additive delta term preserves the source’s modeled temperature change at percentile p(τ)p(\tau). We estimate the distributions separately for each grid cell and calendar month, pooling each month with its two adjacent months, and estimate Fsrc,proj(τ)F_{\mathrm{src,proj}(\tau)} from a centered 30-year window, shifted at the beginning and end of the record to keep 30 complete years. Applied to the unguided emulator, QDM pools all samples to estimate the source distributions.

ERA5 as the Target. We estimate the target distributions from ERA5 and correct each of the seventeen CMIP6 sources independently. Historical calibration uses 1979–2000 and 2011–2014, and each SSP adds 2015–2024. The decade 2001–2010 is excluded from the calibration distributions and retained for validation.

MIROC6 as the Target. MIROC6 supplies the target distributions over the calibration years and the ground truth over 2001–2010 and 2025–2100, and is not a source. The remaining sixteen models are the sources, and we correct both their raw simulations and the corresponding unguided emulator samples.

Classifier Guidance. We compare condition-to-condition guidance with classifier guidance [21] in the MIROC6 fold of our LOOCV framework. In this fold, MIROC6 over 2001–2010 and 2025–2100 is the withheld synthetic observation, and the other sixteen CMIP6 models are the sources. With classifier guidance, the correction toward MIROC6 must come from the classifier, not from information already contained in the emulator. We therefore train a second diffusion emulator on the sixteen source models over 1979–2100, with all MIROC6 simulations (historical and future) held out and all other hyperparameters unchanged. The classifier keeps the encoder and conditioning embeddings of this second emulator and adds an attention head. We train it with a cross-entropy loss to identify which of the seventeen CMIP6 models a noised field comes from, over the same noise levels as the diffusion model. Its training data contain MIROC6 only over 1979–2000 and 2011–2024.

We find that classifier guidance underperforms condition-to-condition guidance. This is likely because the classifier is trained to classify rather than to provide accurate gradients, so its corrections are less precise. In contrast, condition-to-condition guidance adds scores that the emulator has learned directly.

Classifier guidance has two scales, η\eta for the pull toward the target and κ\kappa for the push away from the source model:

∇xtlog⁡p~i(xt)=∇xtlog⁡pθ(xt∣CMIP6i,d,c,y)+∇xt[ηlog⁡pϕ(MIROC6∣xt,d,c,y)−κlog⁡pϕ(CMIP6i∣xt,d,c,y)],(18)\nabla_{x_t}\log\tilde{p}_i(x_t) = \nabla_{x_t}\log p_\theta(x_t \mid\mathrm{CMIP6}_i,d,c,y) + \nabla_{x_t}\left[\eta\log p_\phi(\mathrm{MIROC6}\mid x_t,d,c,y)-\kappa\log p_\phi(\mathrm{CMIP6}_i\mid x_t,d,c,y)\right], \tag*{(18)}

where p~i\tilde{p}_i is the distribution sampled by classifier guidance from source ii, pθp_\theta is the second diffusion emulator and pϕp_\phi is the classifier.

We apply guidance only at noise levels σ<10\sigma< 10 with η=κ=2.5\eta= \kappa= 2.5. We select this threshold and tune both scales together (η=κ\eta= \kappa) by sampling within the held-out 2001–2010 period and comparing to the MIROC6 ground truth.

Evaluation Metrics

This section defines every metric used to evaluate the emulator, both in the perfect-model experiments and against ERA5, and every quantity used to describe its projections.

Notation. A source model is a CMIP6 model on which the emulator is conditioned and from which guidance starts (i.e., the γ=0\gamma= 0 endpoint of Eq. (3) of the main text). In the perfect-model setting, one CMIP6 model is held out and treated as the truth, and the remaining Nm=16N_m = 16 models are the source models. In the projections toward ERA5, all Nm=17N_m = 17 models are source models. For a given setting (no guidance, γ=0\gamma= 0; guidance at scale γ\gamma; or a bias correction), let

x^y,d,s,m,j,k,y=1,…,Ny, d=1,…,Nd, s=1,…,Ns, m=1,…,Nm,\hat{x}_{y,d,s,m,j,k}, \qquad y = 1,\ldots,N_y,\ d = 1,\ldots,N_d,\ s = 1,\ldots,N_s,\ m = 1,\ldots,N_m,

denote the ss-th sample of daily near-surface temperature generated from source model mm at latitude index jj, longitude index kk, on day dd of year yy. The truth is the held-out model’s own simulation, a single realization,

xy,d,j,ktruth,x^{\mathrm{truth}}_{y,d,j,k},

and is replaced by xy,d,j,kERA5x^{\mathrm{ERA5}}_{y,d,j,k} wherever the evaluation is against observations. The grid is 64 × 128, global, and all temperatures are in °C. The sampling designs of Sampling Designs differ in NdN_d and NsN_s: the scale sweep has Nd=12N_d = 12 and Ns=1N_s = 1, and the other designs sample every day (Nd=365N_d = 365) and each states its NsN_s. Raw CMIP6 and the MVT and QDM bias corrections have Ns=1N_s = 1. At γ=1\gamma= 1 the source model plays no role: the ensemble then consists of NsN_s independent samples of the emulator conditioned on the target (ERA5 or the held-out model), and Nm=1N_m = 1. Each table states the window of its metrics, and NyN_y is the number of years in that window.

Throughout, an ensemble member is one source model together with its samples. Unless a table states otherwise, we use the pooled reduction, in which the index pair (s,m)(s,m) is flattened into a single sample axis of length NsNmN_sN_m before any statistic is computed.

Area-Weighted Average. Every spatial average in this work is area-weighted, and the 64 × 128 grid is Gaussian in latitude. Let Aj,kA_{j,k} be the true area of cell (j,k)(j,k) supplied with the grid; summed over the grid, these areas recover the surface area of the Earth. The normalized weight is

wj,k=Aj,k∑j′,k′Aj′,k′,∑j,kwj,k=1,w_{j,k} = \frac{A_{j,k}}{\sum_{j',k'} A_{j',k'}}, \qquad\sum_{j,k} w_{j,k} = 1,

so that the area-weighted spatial mean of any field fj,kf_{j,k} is ⟨f⟩=∑j,kwj,kfj,k\langle f\rangle= \sum_{j,k} w_{j,k}f_{j,k}. Most metrics below reduce a bias map to a scalar through ⟨⋅⟩\langle\cdot\rangle applied to its absolute value, so that errors of opposite sign in different regions cannot cancel.

Trajectory Metrics. These two metrics score the projected annual global-mean temperature, one value per year, against that of the ground truth. The trajectory RMSE measures how closely the ensemble mean follows the true trajectory, and the band coverage how often the ensemble’s spread contains it.

Trajectory RMSE. Let gy,m=⟨1NdNs∑d,sx^y,d,s,m,⋅,⋅⟩g_{y,m}=\left\langle\frac{1}{N_dN_s}\sum_{d,s}\hat{x}_{y,d,s,m,\cdot,\cdot}\right\rangle be the annual global-mean temperature of member mm in year yy, rescaled where this applies (Materials and Methods, Annual-Variance Rescaling), and gytruthg_y^{\mathrm{truth}} the same quantity for the held-out model. With gˉy=1Nm∑mgy,m\bar{g}_y=\frac{1}{N_m}\sum_m g_{y,m} the ensemble mean, the trajectory RMSE over the NyN_y years of a period is

RMSEtraj=1Ny∑y=1Ny(gˉy−gytruth)2.\mathrm{RMSE}_{\mathrm{traj}}=\sqrt{\frac{1}{N_y}\sum_{y=1}^{N_y}\left(\bar{g}_y-g_y^{\mathrm{truth}}\right)^2}.

Tables S1, S3 and S9 report it.

Band Coverage. Each sample of each source model gives one annual series, gy,s,m=⟨1Nd∑dx^y,d,s,m,⋅,⋅⟩g_{y,s,m}=\left\langle\frac{1}{N_d}\sum_d\hat{x}_{y,d,s,m,\cdot,\cdot}\right\rangle, rescaled where this applies (Materials and Methods, Annual-Variance Rescaling). In each year yy, the 5–95% band is the interval between the empirical 5th and 95th percentiles of the NsNmN_sN_m values gy,s,mg_{y,s,m}. The band coverage is the fraction of the NyN_y years in which gytruthg_y^{\mathrm{truth}} falls inside the band. The band is nominally 90%, so 0.90 is the target, and a lower coverage means that the ensemble is under-dispersive. Tables S1, S3 and S9 report it.

Tables S3 and S9 compute both trajectory metrics on the warming above 1850–1900 (Warming-Level Crossing Year), as Figure 2B of the main text shows it, instead of on the annual global-mean temperature.

Distributional Metrics. These metrics compare, at every grid cell, the daily temperatures of an ensemble with those of the truth on the same days. Together they test the mean, the spread, the shape and the tails of the daily distribution, the calibration of the ensemble and the spatial texture of its mean bias.

Write xˉ^j,k=1NyNdNsNm∑y,d,s,mx^y,d,s,m,j,k\hat{\bar{x}}_{j,k}=\frac{1}{N_yN_dN_sN_m}\sum_{y,d,s,m}\hat{x}_{y,d,s,m,j,k} for the climatological-mean field of the ensemble and xˉj,ktruth\bar{x}^{\mathrm{truth}}_{j,k} for the corresponding truth climatology.

Mean Absolute Bias and RMSE. The pointwise mean-bias map and its two scalar summaries, the mean absolute bias and the RMSE, are

Bj,kμ=xˉ^j,k−xˉj,ktruth,⟨∣Bμ∣⟩,RMSEμ=∑j,kwj,k(Bj,kμ)2.B^\mu_{j,k}=\hat{\bar{x}}_{j,k}-\bar{x}^{\mathrm{truth}}_{j,k},\qquad\left\langle\lvert B^\mu\rvert\right\rangle,\qquad\mathrm{RMSE}_\mu=\sqrt{\sum_{j,k}w_{j,k}\left(B^\mu_{j,k}\right)^2}.

Tables S2, S3, S4 and S9 report the RMSE, and Tables S3, S4 and S9 the mean absolute bias.

Standard-Deviation Ratio Deviation. With σj,kpred\sigma^{\mathrm{pred}}_{j,k} and σj,ktruth\sigma^{\mathrm{truth}}_{j,k} the standard deviations at cell (j,k)(j,k) across (y,d,s,m)(y,d,s,m) and (y,d)(y,d) respectively, the reported scalar is ⟨∣σj,kpred/σj,ktruth−1∣⟩\left\langle\left\lvert\sigma^{\mathrm{pred}}_{j,k}/\sigma^{\mathrm{truth}}_{j,k}-1\right\rvert\right\rangle. It is zero when the daily spread matches exactly and penalizes over- and under-dispersion equally. Tables S2, S3, S4 and S9 report it.

Skewness and Excess-Kurtosis Bias. The pointwise standardized third and fourth central moments are

Skewj,kpred=E[(x^⋅,j,k−xˉ^j,kσj,kpred)3],Kurtj,kpred=E[(x^⋅,j,k−xˉ^j,kσj,kpred)4]−3,\mathrm{Skew}^{\mathrm{pred}}_{j,k}=\mathbb{E}\left[\left(\frac{\hat{x}_{\cdot,j,k}-\hat{\bar{x}}_{j,k}}{\sigma^{\mathrm{pred}}_{j,k}}\right)^3\right],\qquad\mathrm{Kurt}^{\mathrm{pred}}_{j,k}=\mathbb{E}\left[\left(\frac{\hat{x}_{\cdot,j,k}-\hat{\bar{x}}_{j,k}}{\sigma^{\mathrm{pred}}_{j,k}}\right)^4\right]-3,

the expectation computed empirically over (y,d,s,m)(y,d,s,m), and analogously for the truth. The reported scalars are ⟨∣Skewpred−Skewtruth∣⟩\left\langle\lvert\mathrm{Skew}^{\mathrm{pred}}-\mathrm{Skew}^{\mathrm{truth}}\rvert\right\rangle and the same for kurtosis: the skewness bias and the excess-kurtosis bias. They diagnose shape discrepancies that the mean and variance miss. Tables S2, S3 and S9 report both.

Percentile Biases. Let qα,j,kpredq^{\mathrm{pred}}_{\alpha,j,k} and qα,j,ktruthq^{\mathrm{truth}}_{\alpha,j,k} denote the empirical α\alpha-quantiles at cell (j,k)(j,k), computed over (y,d,s,m)(y,d,s,m) and (y,d)(y,d) respectively. The quantile-bias map is Bj,kQα=qα,j,kpred−qα,j,ktruthB^{Q_\alpha}_{j,k}=q^{\mathrm{pred}}_{\alpha,j,k}-q^{\mathrm{truth}}_{\alpha,j,k} and the reported scalar is ⟨∣BQα∣⟩\left\langle\lvert B^{Q_\alpha}\rvert\right\rangle, the 1st-percentile bias for α=0.01\alpha=0.01 and the 99th-percentile bias for α=0.99\alpha=0.99. Tables S2, S3, S4 and S9 report both.

TX90p and TN10p Biases. The names follow two indices of the Expert Team on Climate Change Detection and Indices (ETCCDI) [22]: TX90p, the percentage of days on which the daily maximum temperature exceeds its 90th percentile, and TN10p, the percentage of days on which the daily minimum temperature falls below its 10th percentile. We apply them to daily mean temperature, with one threshold per grid cell: the 90th or 10th percentile of the truth’s daily values in the window (Percentile Biases). The warm-day exceedance frequency of the ensemble is

Fj,kTX90p=1NyNdNsNm∑y,d,s,m1[x^y,d,s,m,j,k>q0.9,j,ktruth].F^{\mathrm{TX90p}}_{j,k}=\frac{1}{N_yN_dN_sN_m}\sum_{y,d,s,m}\mathbf{1}\left[\hat{x}_{y,d,s,m,j,k}>q^{\mathrm{truth}}_{0.9,j,k}\right].

The bias map is Fj,kTX90p−Fj,kTX90p,truthF^{\mathrm{TX90p}}_{j,k}-F^{\mathrm{TX90p},\mathrm{truth}}_{j,k}, where the truth’s own exceedance frequency is 0.1 up to the discreteness of the empirical percentile, so that an unbiased ensemble scores zero. It is reported in percentage points, and the scalar is its area-weighted absolute mean. TN10p is defined analogously with 1[x^<q0.1,j,ktruth]\mathbf{1}[\hat{x}<q^{\mathrm{truth}}_{0.1,j,k}]. Tables S2, S3 and S9 report both, and Table S4 the TX90p bias.

Wasserstein Distance. At each cell the 1-Wasserstein distance between the daily distributions of the ensemble and the truth is the mean absolute difference of their quantile functions,

W1,j,k=∫01∣qu,j,kpred−qu,j,ktruth∣ du,W_{1,j,k}=\int_0^1\left\lvert q^{\mathrm{pred}}_{u,j,k}-q^{\mathrm{truth}}_{u,j,k}\right\rvert\,du,

and the scalar is ⟨W1⟩\left\langle W_1\right\rangle. Tables S2, S3, S4 and S9 report it.

Continuous Ranked Probability Score. Continuous ranked probability score (CRPS) is a proper scoring rule that penalizes jointly the distance from the ensemble to the truth and the miscalibration of its spread. At each grid cell and day, the S=NsNmS = N_sN_m pooled values x^1,…,x^S\hat{x}_1,\ldots,\hat{x}_S of the ensemble are scored against the truth xx of that cell and day with the fair estimator

CRPS=1S∑s∣x^s−x∣−12S(S−1)∑s,s′∣x^s−x^s′∣,\mathrm{CRPS} = \frac{1}{S}\sum_s \lvert\hat{x}_s-x\rvert- \frac{1}{2S(S-1)}\sum_{s,s'}\lvert\hat{x}_s-\hat{x}_{s'}\rvert,

averaged over the NyNdN_yN_d days at each cell and then over cells with area weights. Table S2, Table S3 and Table S9 report it.

Table S2. Full distributional evaluation for each of the seventeen LOOCV iterations. Every metric is computed at each grid cell against the held-out model’s own daily fields over 2080–2100 and reduced to one scalar by area weighting. The row of each held-out model gives the mean over the three scenarios. The median rows give the median over the seventeen held-out models for each scenario. Bold marks a guided value smaller than the unguided one. The RI floor is 0.0037. Unguided is the emulator conditioned on the sixteen source models with no guidance. Guided uses the fitted γ\gamma schedule rounded to the nearest of the nine sampled scales (Sampling Designs). The metrics are defined in Evaluation Metrics, Distributional Metrics.

Held-out modelRMSE [°C]Standard-deviation ratio deviationSkewness biasExcess-kurtosis biasTX90p bias [percentage points]TN10p bias [percentage points]1st-percentile bias [°C]99th-percentile bias [°C]CRPS per cell [°C]Laplacian roughness [°C]RI per cellWasserstein distance [°C]
Unguided (γ=0\gamma= 0)
AWI-CM-1-1-MR1.1280.1960.2090.5048.606.161.7492.0321.2230.02661.1591.018
BCC-CSM2-MR1.6570.4080.2860.7577.6120.753.1892.0751.4870.09131.0081.578
CanESM52.4660.2420.2030.5557.6524.993.5971.6961.8210.15751.2442.096
CESM2-WACCM1.1460.2270.2310.5024.7712.062.2521.6081.2340.03470.6141.090
CMCC-CM2-SR51.6500.1960.2230.5355.189.312.4511.3911.2940.03010.7481.221
CMCC-ESM21.4570.2590.2270.5204.899.882.5351.5641.2260.02630.7531.132
EC-Earth31.6560.1580.1980.5998.348.812.1851.5041.3720.03530.9801.311
EC-Earth3-Veg1.7060.1600.1960.5948.059.222.2351.4941.3770.03780.9701.326
EC-Earth3-Veg-LR1.6540.1540.1980.60512.368.611.8311.8431.4540.05020.6161.448
FGOALS-g32.9970.3330.2930.52518.606.452.1203.6651.8050.06841.0162.077
IPSL-CM6A-LR1.5880.1990.2590.5626.406.351.6741.9391.3280.02610.8551.170
MIROC62.0310.3500.2220.59113.399.492.6382.4621.4760.05771.2931.595
MPI-ESM1-2-HR1.1630.2040.2170.5418.426.501.5602.3501.2190.02781.1291.066
MPI-ESM1-2-LR1.4910.2380.2240.57714.755.181.7092.9411.3160.04371.2141.348
MRI-ESM2-01.1510.2290.2210.5087.387.791.5782.1131.2640.02830.7741.057
NorESM2-LM1.3470.2180.2330.4758.376.061.5322.2381.2750.02491.1241.076
NorESM2-MM1.5280.2330.2400.45612.684.681.1482.6911.3320.03090.6221.202
Median, SSP2-4.51.5900.2210.2220.5498.028.222.0251.9661.3440.03460.9981.284
Median, SSP3-7.01.5070.2250.2230.5317.837.852.0691.9381.3110.03200.9751.295
Median, SSP5-8.51.6090.2310.2260.5427.718.632.1142.1511.3390.03590.9661.295
Guided
AWI-CM-1-1-MR0.7680.0700.1330.3167.774.020.7761.0141.1330.02390.4700.705
BCC-CSM2-MR0.6940.0990.1730.3704.525.651.1000.7741.0110.01980.4550.589
CanESM50.8230.1160.1530.3475.155.161.2440.9021.1010.01500.5880.634
CESM2-WACCM0.7620.0670.1270.2584.844.420.8710.8171.0910.01680.3190.527
CMCC-CM2-SR50.6670.0550.1180.2554.433.390.8460.5941.0400.01220.2830.533
CMCC-ESM20.9610.0720.1160.2284.358.191.5010.8271.1110.02680.8980.837
EC-Earth30.9270.0520.1110.2295.744.110.9090.5921.1290.01800.3070.739
EC-Earth3-Veg0.5870.0460.1170.2973.213.170.6580.7561.0490.00930.2070.490
EC-Earth3-Veg-LR0.8100.0860.1550.2844.125.281.3271.6091.0910.01120.3390.514
FGOALS-g31.8110.1120.1590.31312.204.761.5461.6001.3880.05030.4991.287
IPSL-CM6A-LR0.8610.0860.1460.2534.403.380.8260.7861.1420.01690.9020.696
MPI-ESM1-2-HR0.6440.0050.1290.2375.183.160.7521.1241.0280.01100.5220.530
MPI-ESM1-2-LR0.6880.0820.1340.2933.585.150.5330.9621.0840.01510.4700.618
MRI-ESM2-00.7670.0780.1380.2873.855.941.0150.7331.1340.01730.3740.617
NorESM2-LM1.0010.0760.1420.2689.074.430.9391.0171.1830.02120.9430.793
NorESM2-MM0.9700.0760.1470.2944.494.410.8801.0871.2050.02680.3470.824
Median, SSP2-4.50.6880.0660.1320.2824.403.590.8790.7381.0990.01330.3400.567
Median, SSP3-7.00.7810.0760.1320.2774.514.460.9270.8181.0920.01630.3750.628
Median, SSP5-8.50.8780.0830.1360.2904.765.161.0070.8311.1060.01850.4060.696

Table S2.

Rank Histogram and Reliability Index. For each (y,d,j,k)(y,d,j,k) let rr be the number of the SS pooled ensemble values strictly below the truth, so that r∈{0,…,S}r \in\{0,\ldots,S\} and a calibrated ensemble gives ranks uniform over the K=S+1K=S+1 bins. The rank histograms of the figures show the area-weighted frequency of each rank over all cells and days. The reliability index (RI) is computed at each grid cell from the frequencies fℓ,j,kf_{\ell,j,k} of the ranks over the T=NyNdT=N_yN_d days at that cell,

RIj,k=∑ℓ=0S(fℓ,j,k−1K)2,\mathrm{RI}_{j,k}=\sum_{\ell=0}^{S}\left(f_{\ell,j,k}-\frac{1}{K}\right)^2,

zero at perfect calibration, and the reported scalar is ⟨RI⟩\langle\mathrm{RI}\rangle. Computing it per cell prevents miscalibrations of opposite sign in different regions from canceling, as they would in a histogram pooled over cells. A finite sample of TT ranks per cell gives even a perfect ensemble a positive floor RI0=(K−1)/(TK)\mathrm{RI}_0=(K-1)/(TK), which we report alongside the value. Table S2, Table S3, Table S4 and Table S9 report it as RI, per cell.

Laplacian Roughness. Small-scale checkerboard structure in a bias field indicates that the spatial coherence of the underlying maps has not been preserved. For the mean-bias field the discrete Laplacian at an interior cell is

(∇2Bμ)j,k=Bj+1,kμ+Bj−1,kμ+Bj,k+1μ+Bj,k−1μ−4Bj,kμ,(\nabla^2 B^\mu)_{j,k}=B^\mu_{j+1,k}+B^\mu_{j-1,k}+B^\mu_{j,k+1}+B^\mu_{j,k-1}-4B^\mu_{j,k},

with the longitude index wrapping around the globe. The reported scalar is the area-weighted mean of its absolute value over the cells C\mathcal{C} of every latitude row except the northernmost and the southernmost, where the stencil is incomplete:

LapRough=∑(j,k)∈Cwj,k∣(∇2Bμ)j,k∣∑(j,k)∈Cwj,k.\mathrm{LapRough}=\frac{\sum_{(j,k)\in\mathcal{C}}w_{j,k}\lvert(\nabla^2B^\mu)_{j,k}\rvert}{\sum_{(j,k)\in\mathcal{C}}w_{j,k}}.

The bias field is averaged over (y,d,s,m)(y,d,s,m) before the Laplacian is applied, so this measures systematic small-scale structure in the climatological bias rather than in any single sample. Table S2, Table S3, Table S4 and Table S9 report it.

Spatial-Consistency Metrics. A per-cell correction can repair every cell’s marginal distribution and still leave the relation between neighboring cells wrong. The two quantities of Table S5 test that relation inside a regional box, the first at the pair of north–south adjacent cells whose MVT mean biases differ most. Write xAx_A and xBx_B for the daily temperatures at the pair’s southern and northern cell, over all sampled days and members.

Displacement. For ERA5 we compute the principal axis of the joint distribution of (xAERA5,xBERA5)(x_A^{\mathrm{ERA5}},x_B^{\mathrm{ERA5}}) over all sampled days: the line through its mean along the leading eigenvector of the 2×22\times2 covariance matrix. This line is the reference, rather than the diagonal xB=xAx_B=x_A, because the two cells have different climates and a perfect emulator’s points would lie along it. For each method we then compute the signed perpendicular distance from that line to the mean of its own joint distribution,

D=(x‾−x‾ERA5)⋅n,n∝(−m^,1),D=(\overline{x}-\overline{x}^{\mathrm{ERA5}})\cdot\mathbf{n},\qquad\mathbf{n}\propto(-\hat{m},1),

with x‾=(x‾A,x‾B)\overline{x}=(\overline{x}_A,\overline{x}_B) and x‾ERA5\overline{x}^{\mathrm{ERA5}} the means of the joint distributions of the method and of ERA5, m^\hat{m} the slope of the principal axis and n\mathbf{n} the unit normal to it. D=0D=0 means the method reproduces the observed relation between the two cells; a negative DD means its points sit below that relation, i.e. the contrast between the cells is too weak. Table S5 reports ∣D∣\lvert D\rvert as the displacement.

Roughness Ratio. The mean absolute five-point Laplacian of a daily field, over the interior cells of the regional box,

ρ=1∣C∣∑(j,k)∈C∣(∇2x)j,k∣,\rho=\frac{1}{\lvert\mathcal{C}\rvert}\sum_{(j,k)\in\mathcal{C}}\lvert(\nabla^2x)_{j,k}\rvert,

averaged over days and members and divided by the same quantity for ERA5. A value of 1 matches the observed spatial texture; below 1 the fields are too smooth. This differs from the Laplacian roughness defined earlier, which is applied to the climatological bias field rather than to individual daily fields. Table S5 reports it.

Projection and Extreme-Heat Metrics. These quantities describe the projections of the main text. Except where stated, they are not scored against a held-out truth.

Retained Inter-Model Spread. This metric measures how much of the disagreement between the climate models about late-century warming is preserved by guidance toward ERA5. For each scenario, write wy,s,mw_{y,s,m} for the warming above 1850–1900 (Warming-Level Crossing Year) in year yy of sample ss of the projection guided from source model mm, after the annual-variance rescaling (Materials and Methods, Annual-Variance Rescaling; Ns=10N_s = 10, Nm=17N_m = 17), and wy,mraww^{\mathrm{raw}}_{y,m} for that of raw CMIP6 model mm, both as drawn in Figure 2A of the main text. Each model is reduced to its mean warming over 2081–2100, averaged over its samples,

Wm=120Ns∑y=20812100∑s=1Nswy,s,m,Wmraw=120∑y=20812100wy,mraw.W_m = \frac{1}{20N_s}\sum_{y=2081}^{2100}\sum_{s=1}^{N_s} w_{y,s,m}, \qquad W_m^{\mathrm{raw}} = \frac{1}{20}\sum_{y=2081}^{2100} w^{\mathrm{raw}}_{y,m}.

The 5–95% inter-model range RR is the difference between the empirical 95th and 5th percentiles of the seventeen values WmW_m, and RrawR^{\mathrm{raw}} is the same range of the WmrawW_m^{\mathrm{raw}}. The retained inter-model spread is the fraction R/RrawR/R^{\mathrm{raw}}. Averaging over ten samples and twenty years removes nearly all year-to-year variability, so RR measures the spread between models. Table S6 reports RrawR^{\mathrm{raw}}, RR and the retained inter-model spread.

Exceedance-Probability Ratio. Summer is June–August north of the equator and December–February south of it. At a grid cell, every daily summer temperature of a dataset over 1995–2014 is pooled across members, samples, days and years, and qαq_\alpha is the α\alpha quantile of this pool. With P^w(x>qα)\hat{P}_w(x > q_\alpha) the fraction of pooled values above qαq_\alpha in window ww,

Rα=P^2081−2100(x>qα)P^1995−2014(x>qα).R_\alpha= \frac{\hat{P}_{2081-2100}(x > q_\alpha)}{\hat{P}_{1995-2014}(x > q_\alpha)}.

Each dataset is referred to its own qαq_\alpha, so its 1995–2014 exceedance is 1−α1-\alpha and RαR_\alpha cannot exceed 1/(1−α)1/(1-\alpha), that is 10 at the 90th percentile and 100 at the 99th. A constant offset added to every value of a dataset leaves RαR_\alpha unchanged. Figure 3B of the main text shows it.

Temperature Exceeded Once in NN Days. The nn pooled daily values of a dataset and window are sorted in decreasing order, x(1)≥x(2)≥⋯≥x(n)x_{(1)} \ge x_{(2)} \ge\cdots\ge x_{(n)}. The value of rank ii is exceeded on average once in Ni=(n+1)/iN_i = (n+1)/i days, and intermediate values of NN are interpolated linearly in log⁡N\log N. A curve therefore ends at its hottest day, N=n+1N=n+1, and each level up to N=(n+1)/10N=(n+1)/10 is reached on at least ten sampled days. The emulator draws each day independently, whereas real summer temperatures persist from day to day, so NN counts days, not a return period in years. Figure 3C of the main text shows it.

Area of Contiguous Extreme Heat. A land cell is extreme on a given day when its daily-mean temperature exceeds the 95th percentile of ERA5 at that cell over 1979–2024, computed from all days within five days of the same calendar day. Extreme cells that share an edge form one region, with the grid wrapped across the date line. The daily statistic is the total area of all regions of at least two cells, summed from the true cell areas; the land surface of the model grid is 146.4 million km2\mathrm{km}^2. The statistic is computed on every daily map, one per member and day, and the value of a decade is the median over all of its maps: 3,650 for ERA5, and 62,050 for the guided emulator and for raw CMIP6, each with seventeen members (Sampling Designs). Figure 4 of the main text shows it.

Warming-Level Crossing Year. We express every series in Figure 2 of the main text as warming above 1850–1900: we reference it to its own 1995–2014 mean and add the warming of 0.85 °C from 1850–1900 to 1995–2014 assessed by IPCC AR6 WGI (3). The guided projections start in 2025, so we reference them to the 1995–2014 mean of the emulator that replicates the target (γ=1\gamma= 1) before 2025. The year in which the guided ensemble mean crosses 1.5 or 2 °C is the midpoint of the first 20-year period whose mean warming exceeds the level (3), rounded down to a calendar year. Table S7 reports the crossing year and its 5–95% range.

The Criterion Used to Select the Guidance Scale. We select the guidance scale that minimizes the bias of the tails of the daily distribution. For a given fold, scenario and block (Materials and Methods, Selecting the Guidance Strength), we compute at every grid cell the 1st and 99th percentiles of daily temperature for the held-out model and, separately, for each of the NsNmN_sN_m samples of the ensemble. The ensemble percentile is the mean of the sample percentiles, unlike the pooled percentiles of Percentile Biases. The bias maps Bq0.01(γ)B^{q0.01}(\gamma) and Bq0.99(γ)B^{q0.99}(\gamma) are the ensemble percentiles minus those of the held-out model. We then select

γ∗=arg min⁡γ 12(⟨∣Bq0.01(γ)∣⟩+⟨∣Bq0.99(γ)∣⟩)\gamma^* = \underset{\gamma}{\operatorname{arg\,min}}\ \frac{1}{2}\left(\left\langle\left|B^{q0.01}(\gamma)\right|\right\rangle+ \left\langle\left|B^{q0.99}(\gamma)\right|\right\rangle\right)

over the nine sampled scales (Sampling Designs, The Scale Sweep).

Fidelity of the Unguided Emulator

Before assessing guidance, we check the two endpoints it interpolates between: the emulator without guidance, conditioned on each of the eighteen model identities in turn (Sampling Designs, The Emulator Conditioned on Each Model Identity). The γ=0\gamma= 0 endpoint is the emulator conditioned on one of the seventeen CMIP6 models, evaluated over 2080–2100 under SSP3-7.0, as a representative case for the long-term future. The γ=1\gamma= 1 endpoint is the emulator conditioned on ERA5, evaluated over 2001–2010, the decade withheld from its observational training record. Figure S28–Figure S30 map the biases of both endpoints, and Figure S31–Figure S33 show the rank histograms, density, and seasonal cycle of the daily global-mean temperature.

Figure S30. Performance of the diffusion emulator. 99th-percentile bias of the diffusion emulator conditioned on each of eighteen model identities, SSP3-7.0, 2080–2100. As Figure S29, for the 99th percentile.

Figure S29. Performance of the diffusion emulator. 1st-percentile bias of the diffusion emulator conditioned on each of eighteen model identities, SSP3-7.0, 2080–2100. Setup, panels, sampling and windows as in Figure S28. Each panel shows the 1st percentile of the emulator’s daily temperature distribution at each grid cell, over the window’s days and with the ten samples pooled, minus the same percentile of the target’s.

Figure S31. Fig. S31. Performance of the diffusion emulator. Rank histograms of the diffusion emulator conditioned on each of eighteen model identities, SSP3-7.0, 2080–2100. Setup, panels, sampling and windows as in Figure S28. Each panel shows the area-weighted frequency of the rank of the target’s daily temperature among the ten emulator samples of the same day and grid cell. The dashed line marks the uniform frequency 1/11 expected of a calibrated ten-member ensemble.

Conditioned on a CMIP6 model, the emulator reproduces that model’s daily temperature distribution closely: under SSP3-7.0 the mean, 1st-percentile and 99th-percentile biases are small for all seventeen models (Figure S28–Figure S30, panels A–Q). The observational endpoint (panel R), evaluated on a decade withheld from training, has larger residuals than most model endpoints.

Perfect-Model Validation with a Scheduled Guidance Strength

This section details the perfect-model evaluation of the guidance schedule summarized in the Results of the main text: first in all seventeen folds, then in the folds with MIROC6 and MPI-ESM1-2-LR held out (Sampling Designs names the design of each table and figure).

All Seventeen Folds under SSP3-7.0. Guidance improves the calibration of the emulated ensemble against every held-out model, but for several of them the guided ensemble has a bias of one sign over most of the globe. Figure S5–Figure S9 show each of the seventeen folds under SSP3-7.0 over 2080–2100, without and with guidance, with the samples of the scale sweep. Guidance lowers the per-cell reliability index for every held-out model, most for CanESM5 and BCC-CSM2-MR, the two with the highest index without guidance (Figure S9). Over most of the globe the guided ensemble is warmer than some held-out models, such as FGOALS-g3 and the two NorESM2 models, and colder than others, such as MPI-ESM1-2-LR, CMCC-ESM2 and EC-Earth3-Veg (Figure S5 and Figure S8).

Figure S8. LOOCV performance across seventeen folds. Distribution of the daily global-mean temperature for each held-out model, without and with guidance, SSP3-7.0, 2080–2100. Folds, pairs, γ\gamma and sampling as in Figure S5. Kernel density of the daily area-weighted global-mean temperature for the emulator without guidance (dashed), the guided emulator (solid) and the held-out model on the same days (dotted).

Comparison with Two Bias Corrections and Classifier Guidance beyond the Calibration Period. Guidance reproduces the daily temperature fields of the held-out MIROC6 over 2081–2100 more closely than two statistical bias corrections, MVT [19] and QDM [20] (Table S3 and Figure S10–Figure S17). We calibrate both corrections toward MIROC6 over 1979–2000 and 2011–2024 and apply them unchanged to 2081–2100 (Baseline Methods). We apply them to two inputs. The first, the raw output of the sixteen source CMIP6 models, reproduces the pipeline that guidance would replace. The second, the output of the emulator conditioned on the same sixteen models without guidance (γ=0\gamma= 0), forms an ablation that keeps the emulator fixed and changes only how its output is adjusted toward MIROC6. Every ensemble has sixteen members, one per source model (Sampling Designs).

Under all three SSPs and with either input, guidance has the smallest error on eight metrics: the mean absolute bias, the RMSE, the 1st-percentile bias, the TX90p and TN10p biases, CRPS, the reliability index and the Wasserstein distance. The guided rank histograms are also the closest to uniform (Figure S11 and Figure S15). Both corrections produce fields warmer than MIROC6 in every calendar month (Figure S13 and Figure S17), and in their rank histograms MIROC6 falls below all sixteen members more often than at any other rank.

Figure S11. LOOCV performance comparison of condition-to-condition guidance, MVT, QDM (applied to raw CMIP6 data), and classifier guidance for the MIROC6 fold. Ensemble calibration against the held-out MIROC6, 2081–2100. Area-weighted rank histogram of MIROC6 among the sixteen ensemble members at each grid cell and day, for (A) SSP2-4.5, (B) SSP3-7.0 and (C) SSP5-8.5. Columns, left to right: raw output of the sixteen CMIP6 source models; the emulator guided from the same sixteen models with the γ\gamma schedule; the raw output after MVT and QDM bias correction; and classifier guidance. The dashed line marks the uniform frequency 1/171/17 expected of a calibrated sixteen-member ensemble. Every day of 2081–2100, with one sample per source model and day for the emulator and for classifier guidance.

Figure S13. LOOCV performance comparison of condition-to-condition guidance, MVT, QDM (applied to raw CMIP6 data), and classifier guidance for the MIROC6 fold. Seasonal cycle of the global-mean temperature, held-out MIROC6, 2081–2100. Global-mean temperature by calendar month for MIROC6 (black, dotted) and each method (color), for (A) SSP2-4.5, (B) SSP3-7.0 and (C) SSP5-8.5.

Figure S15. LOOCV performance comparison of condition-to-condition guidance, MVT, and QDM (applied to the unguided emulator output) for the MIROC6 fold. Ensemble calibration against the held-out MIROC6, 2081–2100. As Figure S11, with the output of the emulator conditioned on the same sixteen models without guidance (γ=0\gamma= 0) in the leftmost column and as the input to MVT and QDM, and without classifier guidance.

The other five metrics favor one or both corrections in some of the six combinations of scenario and input. MVT has a smaller standard-deviation ratio deviation and a smaller skewness bias than guidance in all six. MVT applied to the emulator output also has a smaller 99th-percentile bias under all three scenarios, and QDM applied to the emulator output has a smaller excess-kurtosis bias under SSP2-4.5. Both corrections, with either input, have a smaller Laplacian roughness than guidance under SSP3-7.0 and SSP5-8.5 and similar values under SSP2-4.5. This metric favors guidance against ERA5 over 2001–2010 (Table S4) but no longer beyond the calibration period. Applying the corrections to the emulator output instead of to raw CMIP6 lowers most of their errors slightly.

Classifier guidance [21] steers a diffusion model toward a target with the gradient of a separately trained classifier, and we compare it with condition-to-condition guidance in the same setting (Baseline Methods, Classifier Guidance). Like condition-to-condition guidance, it is end-to-end: a diffusion model trained on the raw CMIP6 models generates the corrected fields directly. We train a second dedicated diffusion emulator on the sixteen source models over 1979–2100, without MIROC6. We also train a classifier that recognizes MIROC6 among the CMIP6 models, using MIROC6 data only from 1979–2000 and 2011–2024. During sampling, we condition the trained emulator on each source model and steer it toward MIROC6 with the gradient of this classifier. We use a classifier scale of 2.5, applied only at diffusion noise levels below 10. Classifier guidance therefore uses MIROC6 over the same years as the corrections and our emulator. It differs from condition-to-condition guidance in the training set of the emulator and in the guidance mechanism, so the comparison measures the two methods as a whole, each with its own emulator. It appears in the rightmost column of Figure S10, Figure S11 and Figure S12 and as a separate curve in Figure S13.

Figure S12. LOOCV performance comparison of condition-to-condition guidance, MVT, QDM (applied to raw CMIP6 data), and classifier guidance for the MIROC6 fold. Distribution of the daily global-mean temperature, held-out MIROC6, 2081–2100. Kernel density of the daily area-weighted global-mean temperature for each method (color) and for MIROC6 (dotted), for (A) SSP2-4.5, (B) SSP3-7.0 and (C) SSP5-8.5.

Condition-to-condition guidance has a smaller trajectory RMSE than classifier guidance and a smaller error on all thirteen metrics of the daily fields under all three scenarios (Table S3). QDM, applied to raw CMIP6 or to the emulator output, also has a smaller error than classifier guidance on every metric of the daily fields except the TX90p bias under SSP3-7.0 and SSP5-8.5. Unlike the two corrections, classifier guidance produces fields colder than MIROC6, in every calendar month under SSP2-4.5 and SSP3-7.0 and in most months under SSP5-8.5 (Figure S13).

Held-Out MPI-ESM1-2-LR. Guidance does not reproduce every held-out model equally well. We repeat the scheduled run with MPI-ESM1-2-LR as the synthetic observation, with the same design and the same γ\gamma schedule (Table S9 and Figure S23–Figure S27). Until the middle of the century the guided ensemble follows MPI-ESM1-2-LR closely: over 2030–2050, its distribution and seasonal cycle of the daily global-mean temperature match those of MPI-ESM1-2-LR, and its rank histograms are close to uniform (Figure S25–Figure S27). Later it warms less than MPI-ESM1-2-LR (Figure S23). Over 2080–2100 it is colder than MPI-ESM1-2-LR in every calendar month, and MPI-ESM1-2-LR ranks high among its members, most of all under SSP3-7.0 and SSP5-8.5 (Figure S26 and Figure S27). Under SSP3-7.0 the guided ensemble is colder than MPI-ESM1-2-LR over most of the globe, whereas the raw CMIP6 ensemble is warmer there (Figure S24). Its trajectory RMSE stays below that of the raw ensemble mean under all three scenarios, but its 5–95% band contains MPI-ESM1-2-LR in only 17–38% of the years (Table S9). In the scale sweep, averaged over the three scenarios, guidance also lowers the trajectory RMSE of this fold relative to the unguided emulator, while its band coverage falls (Table S1). Over 2081–2100 the guided ensemble has a smaller error than raw CMIP6 on every metric of the daily fields except the TN10p bias under SSP3-7.0 and SSP5-8.5 (Table S9).

Held-out modelTrajectory RMSE [°C]Band coverage
unguidedguidedunguidedguided
AWI-CM-1-1-MR0.160.351.000.44
BCC-CSM2-MR0.700.190.710.65
CanESM51.180.240.320.56
CESM2-WACCM0.410.190.980.70
CMCC-CM2-SR50.640.270.890.67
CMCC-ESM20.500.650.950.18
EC-Earth30.430.461.000.31
EC-Earth3-Veg0.420.280.990.62
EC-Earth3-Veg-LR0.320.201.000.67
FGOALS-g31.510.800.000.17
IPSL-CM6A-LR0.540.180.890.83
MIROC60.830.240.530.64
MPI-ESM1-2-HR0.390.391.000.24
MPI-ESM1-2-LR0.820.350.840.37
MRI-ESM2-00.280.131.000.87
NorESM2-LM0.290.541.000.23
NorESM2-MM0.780.580.900.26
Median, SSP2-4.50.560.240.920.58
Median, SSP3-7.00.470.281.000.54
Median, SSP5-8.50.560.380.950.50

Table S1. Perfect-model skill for each of the seventeen LOOCV iterations. For the unguided and guided case, the trajectory RMSE and the band coverage over 2025–2100. The row of each held-out model gives the mean over the three scenarios. The final rows give the median over the seventeen held-out models for each scenario. Bold marks a guided trajectory RMSE smaller than the unguided one. Unguided is the emulator conditioned on the sixteen source models with no guidance. Guided uses the fitted γ\gamma schedule rounded to the nearest of the nine sampled scales (Sampling Designs). The metrics are defined in Evaluation Metrics, Trajectory Metrics.

Figure S26. LOOCV performance of condition-to-condition guidance for the MPI-ESM1-2-LR fold. Seasonal cycle of the global-mean temperature, held-out MPI-ESM1-2-LR. Global-mean temperature by calendar month under SSP2-4.5 (A, B), SSP3-7.0 (C, D) and SSP5-8.5 (E, F), over 2030–2050 (A, C, E) and 2080–2100 (B, D, F), for MPI-ESM1-2-LR (black, dotted), the emulator without guidance (dashed) and the guided emulator (solid). Both emulators use one sample per source model on the same twelve days per year, and MPI-ESM1-2-LR is shown on those days (Sampling Designs). The source models are pooled before the monthly means are computed.

Comparison with Established Bias Correction against ERA5

The emulator replicating ERA5 (γ=1\gamma= 1), whose ensemble we also call guided (Materials and Methods, Selecting the Guidance Strength), is competitive with the two statistical corrections that guidance would replace, when all three are scored against

observations. We compare it with MVT and QDM applied to the raw output of the seventeen CMIP6 models (Table S4 and Figure S18–Figure S22). All four columns are scored against ERA5 [2] over 2001–2010, a decade that lies outside both the corrections’ calibration windows (1979–2000 and 2011–2024) and the emulator’s observational training years, on every day of the year and with seventeen members each (Sampling Designs, The Emulator Replicating ERA5). This historical decade is the only period in which ERA5 allows the comparison. The held-out MIROC6 comparison above scores the same corrections beyond their calibration period, where they are applied in practice (Table S3).

On the aggregate diagnostics the three corrections are close, and guidance improves on the mean field and on spatial structure (Table S4). All three reduce the mean absolute bias of raw CMIP6 output, and guidance has the smallest mean absolute bias and RMSE. The clearest separation is in the spatial texture of the residual bias: the Laplacian roughness of the mean-bias field is smallest for guidance, larger and similar for MVT and QDM, and largest for raw CMIP6. QDM is the best of the four on the quantities it is constructed to fix, the 99th-percentile bias and the standard-deviation ratio deviation. The Wasserstein distance marginally favors QDM.

Per-Cell Correction Displaces the Gradient between Neighboring Cells. Aggregate scores are spatial averages of per-cell quantities, so they do not reveal how a correction treats the relation between neighboring cells. MVT and QDM are fitted independently at each grid cell, and neither constrains the difference between adjacent cells. Figure S21 and Figure S22 and Table S5 examine this relation at the cell pair of each of two regional boxes (Evaluation Metrics, Spatial-Consistency Metrics). Over High Mountain Asia the two cells are about 300 km apart, and their ERA5 mean temperatures differ by 14.2∘^\circC. The displacement of each method’s joint daily distribution from the principal axis of the ERA5 distribution is smallest for guidance, larger for raw CMIP6 and largest for MVT and QDM, all in the same direction. Both corrections therefore leave the inter-cell gradient further from ERA5 than raw CMIP6 does. The Europe example behaves the same way. A per-cell correction acts on each cell’s marginal distribution without improving the spatial coherence of the map to which it belongs, a limitation reported independently for precipitation [14, 23]. Our generative emulator instead generates each field as a whole, and we designed the two-region test to detect this difference.

Figure S21. Performance comparison of condition-to-condition guidance, MVT and QDM (applied to the raw CMIP6 output), evaluated against ERA5 over 2001–2010. Mean bias and the joint distribution of two adjacent grid cells over High Mountain Asia. (A) Mean bias against ERA5 within 20∘20^\circ–45∘45^\circN, 60∘60^\circ–110∘110^\circE, cell edges drawn; columns are raw CMIP6 output, guided diffusion samples (γ=1\gamma= 1), MVT and QDM. Circles mark the two north–south adjacent cells, at 37.7∘37.7^\circN and 40.5∘40.5^\circN and both at 87.2∘87.2^\circE. (B) Every sampled day and ensemble member at those two cells, plotted as temperature at 40.5∘40.5^\circN against temperature at 37.7∘37.7^\circN, with those for ERA5 shown as black dots. The dotted line is the principal axis of the ERA5 joint distribution. Every day of the year, seventeen ensemble members.

Measurements on individual daily fields point the same way. The mean-bias maps of MVT and QDM show a speckled residual (Figure S21A and Figure S22A), which could in principle be an artifact of averaging over time. The daily fields show that their texture itself departs from the observed one. By the roughness ratio, the corrected daily fields of MVT and QDM are 2–10% too smooth in both regions, the raw CMIP6 fields are too rough, and the guided fields are within a few percent of the observed texture, although over Europe the MVT fields are closer to it (Table S5).

The corrections adjust the daily fields of each CMIP6 model, whereas the guided column draws its seventeen members from the emulator conditioned on ERA5, so the guided column is not a day-by-day correction of a particular simulation. We compare the methods this way because we condition the emulator on ERA5 alone (γ=1\gamma= 1) for the past, which observations describe best, and blend simulations with observations only in projections. We also compare them over the long-term future, with MIROC6 as the synthetic observation (Perfect-Model Validation with a Scheduled Guidance Strength, Comparison with Two Bias Corrections and Classifier Guidance beyond the Calibration Period).

LOOCV across Seventeen Folds

The figures of this section show the perfect-model test in each of the seventeen LOOCV folds under SSP3-7.0 over 2080–2100. Each figure compares the emulator conditioned on the sixteen source models without guidance (γ=0\gamma= 0) with the same emulator guided toward the held-out model. Figs. S5–S7 map the mean, 1st-percentile and 99th-percentile biases, Fig. S8 shows the distribution of the daily global-mean temperature, and Fig. S9 the rank histograms. Table S2 gives the metrics of every fold, and All Seventeen Folds under SSP3-7.0 discusses the figures. The same figures for SSP2-4.5 and SSP5-8.5 over 2080–2100, and for the three scenarios over 2030–2050, which we use as an additional mid-term validation period, are available at https://github.com/mandt-lab/gdso-figures for completeness.

Figure S7. LOOCV performance across seventeen folds. 99th-percentile bias of the emulated ensemble against each held-out model, without and with guidance, SSP3-7.0, 2080–2100. As Figure S6, for the 99th percentile.

Figure S6. LOOCV performance across seventeen folds. 1st-percentile bias of the emulated ensemble against each held-out model, without and with guidance, SSP3-7.0, 2080–2100. Folds, pairs, γ\gamma and sampling as in Figure S5. Each map is the 1st percentile of daily temperature at every grid cell minus that of the held-out model. One color scale for all panels.

LOOCV for MIROC6 (Considered as Synthetic Observation) – Comparison of Condition-to-Condition Guidance, MVT, QDM (Applied to Raw CMIP6 Data), and Classifier Guidance

The figures of this section compare condition-to-condition guidance with MVT and QDM applied to the raw output of the sixteen source models, and with classifier guidance, in the LOOCV fold in which MIROC6 is held out, over 2081–2100. Figure S10 maps nine distributional errors under SSP3-7.0. Figure S11–Figure S13 show, under the three scenarios, the rank histograms, the distribution of the daily global-mean temperature and its seasonal cycle. Table S3 gives the metrics, and Comparison with Two Bias Corrections and Classifier Guidance beyond the Calibration Period discusses them. The figures for SSP2-4.5 and SSP5-8.5 over 2081–2100, and all four figures for the three scenarios over 2030–2050, which we use as an additional mid-term validation period, are available at https://github.com/mandt-lab/gdso-figures for completeness.

LOOCV for MIROC6 (Considered as Synthetic Observation) – Comparison of Condition-to-Condition Guidance, MVT, and QDM (Applied to the Unguided Emulator as an Ablation)

The figures of this section show the ablation of the MIROC6 comparison. The output of the emulator conditioned on the sixteen source models without guidance (γ=0\gamma= 0) replaces raw CMIP6, both in the leftmost column and as the input to MVT and QDM. Figure S14 maps the nine distributional errors under SSP3-7.0, and Figure S15–Figure S17 show the rank histograms, the distribution of the daily global-mean temperature and its seasonal cycle. Table S3 gives the metrics. The figures for SSP2-4.5 and SSP5-8.5 over 2081–2100, and all four figures for the three scenarios over 2030–2050, which we use as an additional mid-term validation period, are available at https://github.com/mandt-lab/gdso-figures for completeness.

Figure S14. LOOCV performance comparison of condition-to-condition guidance, MVT, and QDM (applied to the unguided emulator output) for the MIROC6 fold. Nine distributional error diagnostics against the held-out MIROC6, 2081–2100, SSP3-7.0. Columns, left to right: the emulator conditioned on the sixteen CMIP6 source models without guidance (γ=0\gamma= 0); the emulator guided from the same sixteen models with the γ\gamma schedule; and the unguided output after MVT and QDM bias correction. Rows, top to bottom: mean bias; ratio of modeled to MIROC6 daily standard deviation; 1st- and 99th-percentile bias; bias in the exceedance frequency of each cell’s MIROC6 90th-percentile warm-day (TX90p) and 10th-percentile cold-day (TN10p) thresholds, in percentage points; 1-Wasserstein distance between the daily distributions; skewness bias; and excess-kurtosis bias. Computed with one sample per source model and day of 2081–2100 for the emulator.

Figure S16. LOOCV performance comparison of condition-to-condition guidance, MVT, and QDM (applied to the unguided emulator output) for the MIROC6 fold. Distribution of the daily global-mean temperature, held-out MIROC6, 2081–2100. As Figure S12, with the columns of Figure S15.

Comparison of Condition-to-Condition Guidance, MVT, and QDM against ERA5 over 2001–2010

The figures of this section compare guidance at full scale (γ=1\gamma= 1) with MVT and QDM applied to the raw output of the seventeen CMIP6 models, all scored against ERA5 over 2001–2010. Figure S18 maps nine distributional errors. Fig. S19 shows the ensemble calibration and the distribution of the daily global-mean temperature, and Figure S20 the seasonal cycle and the spatial power spectrum. Figure S21 and Figure S22 show the joint distribution of two adjacent grid cells over High Mountain Asia and over Europe. Table S4 and Table S5 give the metrics, and Comparison with Established Bias Correction against ERA5 discusses them.

Figure S20. Performance comparison of condition-to-condition guidance, MVT and QDM (applied to the raw CMIP6 output), evaluated against ERA5 over 2001–2010. Seasonal cycle and spatial power spectrum of the daily temperature fields. (A) Global-mean temperature by calendar month for ERA5 (black, dotted) and for each method (color). (B) Radially averaged spatial power spectrum of the daily fields against wavenumber, on logarithmic axes, truncated at 0.5 cycles per grid interval, the Nyquist wavenumber of each axis.

Figure 19. Performance comparison of condition-to-condition guidance, MVT and QDM (applied to the raw CMIP6 output), evaluated against ERA5 over 2001–2010. Ensemble calibration and the distribution of global-mean daily temperature for each method. (A) Area-weighted rank histogram: the rank of ERA5 among the seventeen ensemble members at each grid cell, weighted by cell area; the dashed line marks the uniform frequency 1/181/18 expected of a calibrated seventeen-member ensemble. (B) Kernel density of the daily global-mean temperature for each method (color) and for ERA5 (dotted).

LOOCV for MPI-ESM1-2-LR (Considered as Synthetic Observation)

The figures of this section show the perfect-model test in the LOOCV fold in which MPI-ESM1-2-LR is held out. Figure S23 shows the annual global-mean warming over 1979–2100. Figure S24 maps nine distributional errors of raw CMIP6 and of the guided ensemble under SSP3-7.0 over 2080–2100. Figure S25–Figure S27 compare the emulator without and with guidance over 2030–2050 and 2080–2100, through the distribution and seasonal cycle of the daily global-mean temperature and the rank histograms. Table S9 gives the metrics, and Held-Out MPI-ESM1-2-LR discusses them. The figures for SSP2-4.5 and SSP5-8.5 over 2080–2100, and for the three scenarios over 2030–2050, the mid-term period of Figure S25–Figure S27, are available at https://github.com/mandt-lab/gdso-figures for completeness.

Evaluation of the Unguided Emulator

The figures of this section evaluate the emulator without guidance, conditioned on each of the seventeen CMIP6 models under SSP3-7.0 over 2080–2100 (γ=0\gamma= 0) and on ERA5 over 2001–2010 (γ=1\gamma= 1). Figure S28–Figure S30 map the mean, 1st-percentile and 99th-percentile biases, and Figure S31 shows the rank histograms. Figure S32 and Figure S33 show the distribution and the seasonal cycle of the global-mean temperature. Fidelity of the Unguided Emulator discusses them. The same figures for SSP2-4.5 and SSP5-8.5 over 2080–2100, and for the three scenarios over 2030–2050, a mid-term period that is also part of the training record, are available at https://github.com/mandt-lab/gdso-figures for completeness; panel RR (ERA5, 2001–2010) shows the same values in all of them, since it depends on neither the scenario nor the projection period.

Figure S32. Performance of the diffusion emulator. Distribution of the daily global-mean temperature of the diffusion emulator conditioned on each of eighteen model identities, SSP3-7.0, 2080–2100. Setup, panels, sampling and windows as in Figure S28. Kernel density of the area-weighted daily global-mean temperature for the emulator (blue), with the ten samples pooled, and for the target (black, dotted).

References

  1. [1]H Hersbach, et al., ERA5 hourly data on single levels from 1940 to present (Copernicus Climate Change Service (C3S) Climate Data Store (CDS)) (2023) https://doi.org/10.24381/cds.adbb2d47.
  2. [2]H Hersbach, et al., The ERA5 global reanalysis. Q. J. Royal Meteorol. Soc. 146, 1999–2049 (2020).
  3. [3]IPCC, Climate Change 2021: The Physical Science Basis eds. V Masson-Delmotte, et al. (Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA), (2021).
  4. [4]J Tsutsui, Minimal CMIP Emulator (MCE v1.2): a new simplified method for probabilistic climate projections. Geosci. Model. Dev. 15, 951–970 (2022).DOI
  5. [5]M Meinshausen, et al., The shared socio-economic pathway (SSP) greenhouse gas concentrations and their extensions to 2500. Geosci. Model. Dev. 13, 3571–3605 (2020).
  6. [6]NP Gillett, et al., The Detection and Attribution Model Intercomparison Project (DAMIP v1.0) contribution to CMIP6. Geosci. Model. Dev. 9, 3685–3697 (2016).DOI
  7. [7]BC O’Neill, et al., The Scenario Model Intercomparison Project (ScenarioMIP) for CMIP6. Geosci. Model. Dev. 9, 3461–3482 (2016).DOI
  8. [8]J Han, M Kamber, J Pei, 3 - data preprocessing in Data Mining: Concepts and Techniques (Third Edition), The Morgan Kaufmann Series in Data Management Systems. (Morgan Kaufmann), pp. 83–124 (2012).
  9. [9]T Karras, M Aittala, T Aila, S Laine, Elucidating the design space of diffusion-based generative models in Advances in Neural Information Processing Systems, eds. AH Oh, A Agarwal, D Belgrave, K Cho. (2022).arxiv.org/abs/2206.00364
  10. [10]ND Brenowitz, et al., Climate in a bottle: Towards a generative foundation model for the kilometer-scale global atmosphere (2025) arXiv:2505.06474.arxiv.org/abs/2505.06474
  11. [11]I Loshchilov, F Hutter, Decoupled weight decay regularization in International Conference on Learning Representations. (2019).
  12. [12]T Nguyen, JK Jewik, H Bansal, P Sharma, A Grover, Climatelearn: Benchmarking machine learning for weather and climate modeling in Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track. (2023).arxiv.org/abs/2307.01909
  13. [13]JD Willard, et al., Analyzing and exploring training recipes for large-scale transformer-based weather prediction. Artif. Intell. for Earth Syst. 4, 240061 (2025).
  14. [14]M Aich, et al., Conditional diffusion models for downscaling and bias correction of Earth system model precipitation. Geosci. Model. Dev. 19, 1791–1808 (2026).arxiv.org/abs/2404.14416
  15. [15]F Immorlano, et al., Technical report: Towards unified diffusion models for multi-model climate emulation at scale (2025) arXiv:2511.22970.arxiv.org/abs/2511.22970
  16. [16]SI Amari, Information Geometry and Its Applications, Applied Mathematical Sciences. (Springer) Vol. 194, (2016).
  17. [17]AP Guillaumin, L Zanna, Stochastic-deep learning parameterization of ocean momentum forcing. J. Adv. Model. Earth Syst. 13, e2021MS002534 (2021).DOI
  18. [18]MC Wheeler, HH Hendon, An all-season real-time multivariate mjo index: Development of an index for monitoring and prediction. Mon. Weather. Rev. 132, 1917–1932 (2004).doi.org/10.1175/1520-0493(2004)132<1917:aarmmi>2.0.co;2
  19. [19]Z Xu, Y Han, CY Tam, ZL Yang, C Fu, Bias-corrected CMIP6 global dataset for dynamical downscaling of the historical and future climate (1979–2100). Sci. Data 8, 293 (2021).DOI
  20. [20]AJ Cannon, SR Sobie, TQ Murdock, Bias correction of GCM precipitation by quantile mapping: How well do methods preserve changes in quantiles and extremes? J. Clim. 28, 6938–6959 (2015).DOI
  21. [21]P Dhariwal, A Nichol, Diffusion models beat GANs on image synthesis in Proceedings of the 35th International Conference on Neural Information Processing Systems, NIPS ’21. (Curran Associates Inc., Red Hook, NY, USA), (2021).arxiv.org/abs/2105.05233
  22. [22]X Zhang, et al., Indices for monitoring changes in extremes based on daily temperature and precipitation data. WIREs Clim. Chang. 2, 851–870 (2011).DOI
  23. [23]P Hess, M Drüke, S Petri, FM Strnad, N Boers, Physically constrained generative adversarial networks for improving precipitation fields from Earth system models. Nat. Mach. Intell. 4, 828–839 (2022).arxiv.org/abs/2209.07568
  24. [24]T Semmler, et al., AWI AWI-CM1.1MR model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2018) https://doi.org/10.22033/ESGF/CMIP6.359.
  25. [25]T Semmler, et al., AWI AWI-CM1.1MR model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.376.
  26. [26]X Xin, et al., BCC BCC-CSM2MR model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2018) https://doi.org/10.22033/ESGF/CMIP6.1725.
  27. [27]X Xin, et al., BCC BCC-CSM2MR model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.1732.
  28. [28]NC Swart, et al., CCCma CanESM5 model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.1303.
  29. [29]NC Swart, et al., CCCma CanESM5 model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.1317.
  30. [30]G Danabasoglu, NCAR CESM2-WACCM model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.10024.
  31. [31]G Danabasoglu, NCAR CESM2-WACCM model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.10026.
  32. [32]T Lovato, D Peano, CMCC CMCC-CM2-SR5 model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2020) https://doi.org/10.22033/ESGF/CMIP6.1362.
  33. [33]T Lovato, D Peano, CMCC CMCC-CM2-SR5 model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2020) https://doi.org/10.22033/ESGF/CMIP6.1365.
  34. [34]T Lovato, D Peano, M Butenschön, CMCC CMCC-ESM2 model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2021) https://doi.org/10.22033/ESGF/CMIP6.13164.
  35. [35]T Lovato, D Peano, M Butenschön, CMCC CMCC-ESM2 model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2021) https://doi.org/10.22033/ESGF/CMIP6.13168.
  36. [36]EC-Earth Consortium (EC-Earth), EC-Earth-Consortium EC-Earth3 model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.181.
  37. [37]EC-Earth Consortium (EC-Earth), EC-Earth-Consortium EC-Earth3 model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.251.
  38. [38]EC-Earth Consortium (EC-Earth), EC-Earth-Consortium EC-Earth3-Veg model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.642.
  39. [39]EC-Earth Consortium (EC-Earth), EC-Earth-Consortium EC-Earth3-Veg model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.727.
  40. [40]EC-Earth Consortium (EC-Earth), EC-Earth-Consortium EC-Earth3-Veg-LR model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2020) https://doi.org/10.22033/ESGF/CMIP6.643.
  41. [41]EC-Earth Consortium (EC-Earth), EC-Earth-Consortium EC-Earth3-Veg-LR model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2020) https://doi.org/10.22033/ESGF/CMIP6.728.
  42. [42]L Li, CAS FGOALS-g3 model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.1783.
  43. [43]L Li, CAS FGOALS-g3 model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.2056.
  44. [44]O Boucher, et al., IPSL IPSL-CM6A-LR model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2018) https://doi.org/10.22033/ESGF/CMIP6.1534.
  45. [45]O Boucher, et al., IPSL IPSL-CM6A-LR model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.1532.
  46. [46]H Tatebe, M Watanabe, MIROC MIROC6 model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2018) https://doi.org/10.22033/ESGF/CMIP6.881.
  47. [47]H Shiogama, M Abe, H Tatebe, MIROC MIROC6 model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.898.
  48. [48]J Jungclaus, et al., MPI-M MPIESM1.2-HR model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.741.
  49. [49]M Schupfner, et al., DKZR MPI-ESM1.2-HR model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.2450.
  50. [50]KH Wieners, et al., MPI-M MPIESM1.2-LR model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.742.
  51. [51]KH Wieners, et al., MPI-M MPIESM1.2-LR model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.793.
  52. [52]S Yukimoto, et al., MRI MRI-ESM2.0 model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.621.
  53. [53]S Yukimoto, et al., MRI MRI-ESM2.0 model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.638.
  54. [54]Ø Seland, et al., NCC NorESM2-LM model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.502.
  55. [55]Ø Seland, et al., NCC NorESM2-LM model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.604.
  56. [56]M Bentsen, et al., NCC NorESM2-MM model output prepared for CMIP6 CMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.506.
  57. [57]M Bentsen, et al., NCC NorESM2-MM model output prepared for CMIP6 ScenarioMIP (Earth System Grid Federation) (2019) https://doi.org/10.22033/ESGF/CMIP6.608.

Paper details

vlnc:W_49tdewtq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents