betaFeedback

WorksComputer Science

A hybrid machine learning-analog framework for downscaling unprecedented climate extremes

Abstract

Extreme weather events are critical for climate impact assessments. While physical models can simulate unprecedented extremes, for example, through ensemble boosting, this approach is computationally demanding, often restricting output to coarse spatial resolutions. Probabilistic generative downscaling offers a solution by enabling efficient sampling from the entire climate distribution, including its extreme tails. Here, we first demonstrate that machine learning generated regional climate model heatwaves over Paris are physically realistic. We then leverage an emulated machine learning dataset to perform analog-based downscaling of coarse global climate model ensemble boosting events. Our hybrid framework effectively bridges high-resolution regional detail with rare-event sampling, providing an efficient tool for extreme event downscaling and capturing multi-model uncertainty.

Introduction

Despite the computational effort dedicated to multi-model initiatives like the Coupled Model Intercomparison Project CMIP [7], their generally coarse spatial scale limits their utility for impact assessments at local to regional scale, in urbanized areas, over complex topography, and along coastlines. Evaluating more localized risks related to heatwaves and compound events benefits from fine-scale detail [4], yet practical analyses remain limited by the inherent scarcity of extreme events and the spatial resolution of standard global climate model outputs. To overcome the sample-size constraints, rare event algorithms such as ensemble boosting [10] and the GKTL algorithm [11] have emerged as tools for generating physically coherent extremes. Ensemble boosting addresses the sample-size constraints by perturbing the initial atmospheric conditions of targeted events days or weeks in advance. It efficiently samples low-probability, high-impact extremes while maintaining physical coherence within climate models. Yet, despite its conceptual efficiency, generating these large ensembles is computationally intensive, particularly in higher-resolution climate models [1].

Dynamical downscaling of these boosted global climate samples presents an even greater technical and computational challenge. Running dynamical regional climate models over hundreds of boosted ensemble members to obtain high-resolution fields is computationally prohibitive. Consequently, extreme event storylines studies frequently rely directly on global climate model outputs [8]. While statistical downscaling is possible, statistical approaches struggle with unprecedented extremes, where observational analogs based on historical archives do not exist for future record-shattering extremes that exceed the historic range [Novelle et al., 2025].

To bridge the gap between low-resolution ensemble-boosted simulations and fine-scale detail, we leverage machine learning emulators to sample climate extremes. Our approach combines physics-based GCM ensemble boosting with a general-purpose, temporally-consistent machine learning (ML) downscaling emulator to construct a large, physically consistent high-resolution climate dataset. By combining high-resolution ML-generated data with low-resolution physical extremes, coarse GCM extremes can be mapped to high-resolution regional states, and climate model uncertainty can be used to construct alternative storylines of extreme events.

Preprint. Under review.

We demonstrate the utility of this framework using Western European heatwaves studied in Fischer et al. [2023], centered on Paris, as a primary proof-of-concept. More specifically, by training the temporally consistent ML emulator EnScale [20] on physics-based GCM-RCM simulations, we create a library of physically consistent, fine-scale climate emulations. The emulator’s computational efficiency allows us to sample 8,000 summers, including extremes, across 8 different GCM-RCM combinations. We show that these ML-generated samples span the range of the most extreme boosted events, allowing us to perform analog matching to instantaneously downscale coarse boosted GCM events. Ultimately, our hybrid framework preserves spatio-temporal consistency while capturing multi-model variability at a fraction of the computational cost of traditional dynamical downscaling.

Methods

Framework overview

As illustrated in Figure 1, our hybrid framework couples low-resolution ensemble boosting with high-resolution ML-generated data to efficiently downscale extreme weather events. In this work, we first address whether the in-distribution extremes generated by the ML emulator (Figure 1, right part) reliably match target RCM dynamics. We use a machine learning emulator pre-trained on physics-based simulations (§ 2.2) to generate an 8,000-summer dataset, including extreme realizations. Concurrently, ensemble boosting (Figure 1, left part) takes an extreme event and generates an ensemble of similar extreme events with a GCM (§ 2.3). Finally, the two parts of the framework are coupled via spatial average pooling and analog matching (§ 2.4). For downscaling, each low-resolution boosted extreme is matched to a corresponding high-resolution field within the emulated dataset. In addition, we also show how analog matching can be used to identify alternative storylines and sample model variability by selecting extreme samples from other models without spatial field constraints.

Figure 1. Overview of the hybrid extreme-event framework. Low-resolution ensemble boosting leverages a single extreme event as an initial condition to generate a large sample of physically consistent extremes with a global climate model (darker shades indicate higher temperatures). In parallel, the ML emulator generates 8,000 years of RCM-resolution data that can then be used to find high resolution analogs for the low resolution extremes from ensemble boosting.

EnScale setup

The EnScale machine learning model was developed to emulate regional climate models (RCMs) given global climate model (GCM) inputs [20]. It is a generative model that learns the conditional distribution of high-resolution RCM targets given GCM predictors, enabling the sampling of multiple realizations of high-resolution fields for a given GCM field. In its multivariate setup, it predicts temperature, precipitation, surface wind, and solar radiation jointly, spatially consistent across Central Europe. EnScale splits the full GCM-RCM map into several physically-motivated steps: It first performs a “coarse correction”, predicting a coarsened (average-pooled) version of the RCM data from GCM data. This step accounts for large-scale mismatches between GCM and RCM. The coarsened RCM fields are then iteratively refined through a sequence of super-resolution steps, each enhancing the spatial resolution further until the target resolution is recovered. Each stage is trained by minimizing the energy score, a strictly proper scoring rule that ensures EnScale’s learned conditional distribution matches the target distribution.

In this paper, we use the version EnScale-t, which explicitly models the temporal dependence of the RCM data through an autoregressive framework. Both EnScale and EnScale-t have been shown to accurately reproduce the multivariate conditional distribution. Compared to other state-of-the-art generative models, EnScale(-t) performs competitively while requiring substantially lower computational costs [15]. Since we use only the temporal variant throughout, we refer to EnScale-t simply as EnScale in the following.

Here, we use eight RCM-GCM combinations spanning the years 1971-2099 (with training data between 1971-2030 and 2040-2089) at a daily resolution; an overview of the models is given in Table 1. The GCMs are from CMIP5 [22], and the RCMs are from CORDEX [13], derived from historical runs and RCP8.5. Although this emulator allows sampling from a specific target RCM, the shared representations learned during joint training mean that the generated samples from a specific RCM-GCM combination are not expected to perfectly reproduce the isolated characteristics of the original physical model. We introduce an adaptation to the standard EnScale-t model. In its original configuration, temporal dependence is modeled at every spatial resolution. Here, for simplicity, we use a slightly adjusted version and restrict temporal autocorrelation modeling exclusively to the lowest spatial resolution (as suggested in the preprint [19] of Schillinger et al. [ [20]). Specifically, while the coarsened RCM data is modeled autoregressively, the super-resolution module is applied independently to each daily timestep without temporal conditioning. Across our evaluations, this modification yields no qualitative performance differences compared to the full baseline (Figure A.2).

RCM / GCMCNRM-CM5MIROC5MPI-ESM1-2-LR
ALADIN63model 1model 2
CCLM4-8-17model 3model 4model 5
REMO2015model 6
RegCM4-6model 7model 8

Table 1. RCM–GCM combinations in the emulator dataset. We enumerate all models according to the numbering displayed here. Numbers indicate model combinations that exist and were used both for training the machine learning model and in the data analysis presented here.

ML sampling and sample sizes Using EnScale, we sample 8,000 emulator summers and compare them against a baseline of 8 ground-truth RCM-GCM combinations, which span 10 years of data each from 2030 to 2039 (totaling 80 summers). This means that, for each original RCM summer, we have 100 EnScale samples. While the emulator is therefore expected to produce longer and more extreme events than are present in the original RCM data, the aim is to validate whether these generated extremes are physically plausible.

Ensemble boosting setup

The ensemble boosting method [10] is a rare event algorithm that allows for a computationally efficient simulation of extreme weather events. Starting from a set of extreme events simulated in a large ensemble (“parent events”), it consists of adding round-off-sized perturbations to the atmospheric humidity field a few days or weeks before the peak of the parent event. In this way, a new ensemble of simulations is created, which quickly deviates substantially from their parent event.

Ensemble boosting simulations for one extreme parent event

Figure 2. Ensemble boosting simulations for one extreme parent events. Shown are the temperature anomalies in Paris for the GCM’s parent simulation (black) and the boosting ensemble (1200 simulations, gray).

As an example, we use existing ensemble boosting simulations of heatwave storylines in the Greater Paris region, described in Fischer et al. [ [8]. For these simulations, three parent heatwaves are selected based on the 5-day daily maximum temperature anomaly in Greater Paris in a CESM2 initial condition large ensemble for 2005–2035. We use simulations initialized 4–15 days before the peak anomaly in the respective parent event, with 100 ensemble members per lead time. This yields 3 parent events × 12 initialization days × 100 ensemble members = 3600 simulated heatwaves, each of which is run for 50 days. The heatwaves occur in the model years 2016, 2028, and 2030 [8].

Downscaling via analogs

In § 4.1, we employ an analog approach to downscale the GCM ensemble-boosted extreme events to RCM resolution. We choose the analog framework over direct ML-based downscaling to leverage the pre-trained emulator without the need or expense of retraining. Directly applying the trained emulator to boosted data would be challenging for several reasons: CESM2 was not part of the original training set, and the emulator’s required predictor variables and domain do not align with the saved output variables of the boosted dataset.

For analog-based downscaling, we use a two-step approach that first filters candidates by local Paris heatwave intensity using 2m temperature data and subsequently selects the most structurally similar event based on spatial pattern alignment. To construct the analog pool, we coarsen the 8,000 summers of EnScale data to the GCM grid using conservative remapping via CDO [21]. Then, for a given Paris heatwave generated by the GCM, we identify all candidate 5-day heatwaves in the coarsened EnScale dataset whose intensity falls within ±0.1°C of the GCM event; we describe the details in § 4.1. From this pre-filtered subset, we select the event with the highest spatial similarity across the entire spatial field as the best analog. To assess spatial pattern alignment between the coarsened EnScale field xx and the boosted GCM field yy, we employ the Structural Similarity Index (SSIM; [24]) on 5-day temperature averages. Conceptually, SSIM aligns with object-based meteorological evaluation metrics such as the SAL (Structure, Amplitude, Location) index [25]. SSIM evaluates spatial similarity by simultaneously comparing localized mean, contrast, and structural information:

SSIM⁡(x,y)=(2μxμy+C1)(2σxy+C2)(μx2+μy2+C1)(σx2+σy2+C2)(1)\operatorname{SSIM}(x,y) = \frac{(2\mu_x\mu_y + C_1)(2\sigma_{xy} + C_2)}{(\mu_x^2 + \mu_y^2 + C_1)(\sigma_x^2 + \sigma_y^2 + C_2)} \tag*{(1)}

where μx\mu_x and μy\mu_y denote the spatial means, σx2\sigma_x^2 and σy2\sigma_y^2 represent the spatial variances, and σxy\sigma_{xy} is the spatial covariance between the two fields. The terms C1=(K1L)2C_1 = (K_1L)^2 and C2=(K2L)2C_2 = (K_2L)^2 are stabilizing constants, where LL is the range of the temperature data and K1=0.01K_1 = 0.01, K2=0.03K_2 = 0.03. The index ranges from −1 to 1, where 1 indicates perfectly identical spatial structures. Once the analog with the maximum SSIM is selected, we map it back to its original high-resolution EnScale counterpart, yielding a physically plausible high-resolution analog for the GCM extreme event.

Standardized Anomalies

To compare heatwaves across different models—especially those that fall outside the training RCM distributions—we compute standardized anomalies. Because EnScale is trained jointly on all RCM-GCM combinations, individual model samples are influenced by the broader ensemble. Therefore, we standardize the data using the mean and variance computed across all models. For a given temperature Tt,yT_{t,y} on day tt of year yy (pre-smoothed with a 5-day running mean), the standardized anomaly Zt,yZ_{t,y} is computed as:

Zt,y=Tt,y−μt,yσt,y(2)Z_{t,y} = \frac{T_{t,y} - \mu_{t,y}}{\sigma_{t,y}} \tag*{(2)}

where the seasonal mean μt,y\mu_{t,y} and standard deviation σt,y\sigma_{t,y} are calculated over a rolling window of ±2 days and a reference climatology of ±15 years centered on the target date (t,y)(t,y):

μt,y=1N∑j=−1515∑i=−22Tt+i,y+j(3)\mu_{t,y} = \frac{1}{N} \sum_{j=-15}^{15} \sum_{i=-2}^{2} T_{t+i,y+j} \tag*{(3)}
σt,y=1N∑j=−1515∑i=−22(Tt+i,y+j−μt,y)2(4)\sigma_{t,y} = \sqrt{\frac{1}{N} \sum_{j=-15}^{15} \sum_{i=-2}^{2} (T_{t+i,y+j} - \mu_{t,y})^2} \tag*{(4)}

The total number of RCM days used to compute the climatology is 31years × 5days = 155 samples.

For the ensemble boosting, the same approach is used to calculate standardized anomalies. As climatology, we use the 100-member CESM2 large ensemble running for the period 1850–2100 under historical and SSP370 scenarios, respectively [5, 16]. This allows for the construction of centered climatologies for all ensemble boosting simulations.

Weighted least squares

To evaluate heatwave persistence and physicality, we use a unified Weighted Least Squares (WLS) regression framework. Intuitively, this method fits a linear trend to each model class simultaneously, and we use it to perform hypothesis testing to determine whether the ML emulator EnScale exhibits statistically significant slope departures from physics-based baselines.

For any target variable YY evaluated against its driver xx, we estimate the single interaction model:

Y=α0+β1x+αRCMIRCM+αGCMIGCM+βint,RCM(x×IRCM)+βint,GCM(x×IGCM)+ε(5)Y = \alpha_0 + \beta_1 x + \alpha_{\mathrm{RCM}} I_{\mathrm{RCM}} + \alpha_{\mathrm{GCM}} I_{\mathrm{GCM}} + \beta_{\mathrm{int,RCM}}(x \times I_{\mathrm{RCM}}) + \beta_{\mathrm{int,GCM}}(x \times I_{\mathrm{GCM}}) + \varepsilon \tag*{(5)}

where IRCMI_{\mathrm{RCM}} and IGCMI_{\mathrm{GCM}} are binary indicators (1 if the data comes from RCM or GCM respectively, 0 otherwise).

Using EnScale as the baseline, β1\beta_1 isolates the emulator’s slope, while the interaction terms βint,RCM\beta_{\mathrm{int,RCM}} and βint,GCM\beta_{\mathrm{int,GCM}} directly measure the slope differences of the RCM and GCM relative to EnScale. Thus, the null hypotheses are that these interaction terms are βint,RCM=0\beta_{\mathrm{int,RCM}} = 0 and βint,GCM=0\beta_{\mathrm{int,GCM}} = 0, respectively. A two-tailed tt-test within a WLS ANCOVA framework ([e.g. 2]) evaluates whether the slope differences are statistically significant. To ensure robustness, each bin is weighted by its sample count NN, giving more influence to densely populated bins and preventing rare events from skewing the trend line.

To evaluate heatwave persistence, we map this framework directly to an exponential decay model P(d)∝exp⁡(βd)P(d) \propto\exp(\beta d) by setting Y=ln⁡P(d)Y = \ln P(d) (the log-probability of an event lasting dd days) and x=dx = d (event duration in days). In this setup, β1\beta_1 measures the emulator’s duration decay rate, from which we derive the daily continuation probability pcont=exp⁡(β1)p_{\mathrm{cont}} = \exp(\beta_1) and the atmospheric memory scale τ=−1/β1\tau= -1/\beta_1.

Validation of EnScale heatwaves

To use EnScale data for downscaling ensemble boosted heatwaves, we first validate heatwaves in EnScale using a twofold approach. First, for in-distribution heatwaves, we verify that RCM-simulated heatwaves in Paris closely match ML-generated heatwaves in magnitude, temporal alignment, and spatial fields. We then extend this analysis to out-of-distribution heatwaves.

In-distribution heatwaves

To ensure that emulated heatwaves do not emerge spuriously, we first verify that the temperature time series of both RCMs and EnScale align with the temporal progression of their driving GCMs, as visualized in Figure 3 (for an overview separated by GCM, see Figure A.1). Overall, the correlation between GCMs and RCMs and emulated RCMs, respectively, is high. The lower individual member correlation of EnScale and the GCMs (0.61–0.74 for the year 2033) paired with a higher ensemble-mean correlation (0.82–0.93 for the year 2033; Table A.1) demonstrates that EnScale successfully separates fluctuations from the large-scale atmospheric variability imposed by the driving GCM. While 2033 is selected here as a representative year, this behavior is consistent and robust across the other years as well.

Average daily temperature time series in the MPI model and corresponding physical RCMs and EnScale for the summer 2033

Figure 3. Average daily temperature time series in the MPI model and corresponding physical RCMs and EnScale for the summer 2033. Thick lines show mean values for the different model classes (one per RCM), while thin lines display individual realizations.

Building on this broad temporal consistency, we find that EnScale can reproduce the intensity of 5-day RCM heatwaves. Selecting the hottest 2030–2039 heatwaves in Paris per physical RCM (8 in total), ML-generated analogs of EnScale closely mirror individual RCM events at both local and spatial scales (Figure 4). To formally test this alignment, we perform a permutation test measuring the 5-day temperature trajectory RMSE relative to the ensemble reference. Sampling random groups of 8 events from the EnScale distribution (N=10,000N = 10{,}000) reveals that the 8 original RCM heatwaves are statistically indistinguishable from random EnScale samples (for detailed results, see Figure A.3). We then find the closest matches for each of the 8 hottest RCM heatwaves (Figure 4). Selecting analogs based on minimizing the mean squared error for temperature across the hottest 5 days, we find that the corresponding solar radiation is high (middle row), and precipitation is zero or close to zero before and at the beginning of the heatwave (last row), suggesting similar physical properties of the heatwaves.

Figure 4. Temporal and spatial alignment of analog heatwaves in RCM and EnScale. Comparison of the 8 hottest 5-day heatwaves in Paris (2030–2039) across physical models (RCM-GCM combinations, black) and their closest EnScale analogs (blue), selected by minimizing the mean squared error (MSE) of Paris temperature over the 5-day heatwave period (red shading). Left panels: Averaged 15-day temporal trajectories centered around the heatwave for temperature (top), solar radiation (middle), and precipitation (bottom). Right panels: Corresponding composite spatial anomaly fields (deviations from climatology; Equation (4)) averaged across all 8 heatwave events during the 5-day peak period for the original RCM (left sub-column) and EnScale (right sub-column). The black ’x’ denotes the location of Paris.

Comparing the composite spatial fields averaged across all 8 heatwaves yields a high spatial correlation (0.988) and SSIM (0.980), confirming that EnScale accurately reconstructs the two-dimensional spatial structure and correctly centers the localized temperature maximum over Paris. Similarly, solar radiation is very high in Paris and the surrounding areas; precipitation is very low in Paris. However, surrounding areas show more rain in EnScale compared to the original RCM. Crucially, analogs were identified based on the temporal alignment of temperatures at the Paris grid point only, without spatial field constraints during selection. Therefore, the spatial fields and other variables are not expected to be identical.

To further evaluate intervariable consistency, we analyze whether EnScale variables behave as physically expected across the range of temperatures. Figure 5 reveals distinct patterns of climate parameter sensitivities across GCMs, RCMs, and EnScale. For the amount and probability of precipitation >0.1> 0.1mm/day, EnScale exhibits intermediate temperature sensitivities that fall between the physical RCMs and GCMs (for a figure separated by RCM, we refer to Figure A.4). During the hottest days, mean precipitation sinks to almost zero in both RCM and EnScale data. Specifically, for temperatures above 29°C, precipitation is usually less than 1 mm/day. However, overall precipitation is slightly overestimated in the ML samples for hotter temperatures. However, because our selection relies strictly on temperature thresholds, some identified heatwaves may include intense late-day convective storms [18], introducing rain on peak temperature days and making precipitation a noisy evaluation metric. For mean solar radiation, a WLS regression reveals that all three model classes demonstrate statistically indistinguishable sensitivities (Table A.2), increasing at approximately 6.59 to 7.26 W/m² per °C. Although solar radiation amounts vary between RCM, GCM and EnScale below 15°C, they increase at temperatures above 15°C across all three model classes, exhibiting similar slopes and absolute values.

Figure 5. Comparison of temperature sensitivities at the Paris grid point across the whole temperature ranges, binned in 1°C increments, for precipitation amount (left), precipitation probability (center), and mean solar radiation (right) across EnScale, RCM, and GCM. Shaded regions denote ±1 standard deviation of daily values within each bin.

Out-of-distribution heatwaves

Having established that EnScale reproduces in-distribution heatwaves, including realistic multivariate aspects and the temporal dynamics of the GCM, we now evaluate EnScale’s capacity to emulate out-of-distribution extreme events. Using anomaly datasets spanning the same 80 years of RCM simulations and 8,000 years of EnScale emulations, we evaluate whether EnScale can generate unprecedented extremes while maintaining physical realism. To better compare heatwaves of different RCM-GCM combinations, we compute standardized anomalies as explained in § 2.5.

We first evaluate the distribution of 5-day standardized temperature anomalies (in σ\sigma). As shown in Figure 6 (left), EnScale covers the full range of anomaly distributions; however, its expanded 8,000-year sample size generates extreme intensities beyond RCM bounds and into the range of samples from ensemble boosting (Section 2.3). While the maximum RCM anomaly is 3.56σ\sigma, EnScale produces events of up to 6.08σ\sigma, compared to 4.64σ\sigma in ensemble boosting. The frequency of these very extreme tail events drops with increasing magnitude. Crucially, these high-magnitude anomalies are strongly RCM-dependent rather than GCM-driven: over 80% originate from RegCM4-6 and REMO2015. Conversely, CCLM4-8-17 produces almost no events in these upper bins, while ALADIN63 displays intermediate counts, mainly in the [4.6, 5.2σ\sigma) range (Table A.3). This RCM dominance over driving GCMs aligns with established literature: while winter climate is dictated by large-scale synoptic advection, summer heat extremes are governed by local land-surface feedbacks and convective parameterizations that differ across RCMs [9].

Histograms of maximum standardized 5-day anomalies and frequency of extreme events as a function of event duration

Figure 6. a) Histograms of maximum standardized 5-day anomalies in Paris of each summer in boosted data, GCMs, original RCMs and EnScale ensemble members. b) Frequency of 1σ1\sigma extreme events as a function of event duration. Vertical dashed lines mark the duration of maximum frequency for each model class.

To test whether EnScale alters heatwave duration characteristics relative to the physics-based baselines, we evaluate whether the model classes share the same daily probability decay slope (β\beta) using WLS regression. The distribution of heatwave lengths is visualized in Figure 6 (right) and shows a decay in the number of events with increasing heatwave length. The longest GCM, RCM, and EnScale 1σ1\sigma heatwaves are 18, 25, and 63, respectively. EnScale accurately captures the exponential temporal decay of multi-day 1σ1\sigma heatwave events, remaining consistent with both the raw RCM reference and the driving GCMs. The log-linear probability decay yields a daily slope of β=−0.2773\beta=-0.2773 for EnScale, corresponding to a 75.8% daily continuation probability (eβe^\beta). This translates to an atmospheric memory scale (−1β-\frac{1}{\beta}) of τ=3.61\tau=3.61 days. While EnScale exhibits slightly higher temporal persistence than the RCM (τ=2.98\tau=2.98 days) and GCM (τ=3.29\tau=3.29 days), the regression interaction terms confirm that these decay rate differences are statistically insignificant (Table A.4), suggesting EnScale retains physically consistent heatwaves.

Looking at the most extreme heatwaves per RCM-GCM combination reveals spatial similarities across all models (Figure A.5). The maximum values again depend on the RCMs, with ALADIN, RegCM, and REMO generating much higher temperatures than CCLM (5-day mean daily temperature in Paris >37∘>37^\circC vs. <35∘<35^\circC.) Most very strong heatwaves in EnScale (>4.6σ>4.6\sigma) occur in June; see Figure A.6.

Applications

Analog-based downscaling of ensemble boosting events

Analog-based downscaling of extremes relies on the assumption that the target search space contains high-resolution realizations that are physically plausible and representative of rare tail events. The previous section establishes this foundation by demonstrating that EnScale preserves spatial and temporal structures, multivariate physical relationships, and extreme-value decay rates under baseline conditions. We therefore conclude that EnScale’s sample space (N=8,000N=8,000 summers) functions as a physically credible library of high-resolution climate states. As it also spans the range of extremes observed in ensemble boosting (Figure 6 (left)), we use it to downscale ensemble-boosted GCM extremes via analog matching, as described in § 2.4. Specifically, we first select days when the temperatures in Paris are similar to those found in boosting (35∘35^\circC ±0.1∘\pm0.1^\circC) and then search for high-resolution spatial analogs corresponding to the peak heatwave periods in the boosted GCM time series.

Figure 7 illustrates the heatwaves and the corresponding downscaling results for the four hottest events (in terms of σ\sigma) identified in the ensemble boosting simulations. For each coarse GCM boosted event, the EnScale search space yields matching high-resolution analogs (SSIM ≈0.88\approx0.88) that reconstruct fine-scale regional features, such as localized heat anomalies around Paris and topographic attenuation over the Alps, while adhering to the synoptic-scale circulation patterns of the original GCM data. The (average-pooled) RCM samples exhibit a cold bias over the Alps compared to the GCM data, consistent with e.g. Rudisill et al. [ [17]. Across all four cases, the high-resolution analogs yield higher temperatures than their low-resolution counterparts, ranging from 36.1 ∘C36.1\,^\circ\mathrm{C} to 37.6 ∘C37.6\,^\circ\mathrm{C}, underscoring the need for spatial downscaling to capture true localized severity.

Four rows of four temperature maps showing ensemble boosting, averaged ML data, and original ML data, with temperatures and similarity indices annotated

Figure 7. The four hottest events from ensemble boosting (top row), alongside their corresponding EnScale analogs at coarse average-pooled resolution (middle row) and native EnScale resolution (bottom row). Values above each panel show the temperature at the grid point closest to Paris. Annotations above the middle row report the Structural Similarity Index between the original boosted events and the coarsened EnScale fields. The black ’o’ denotes the location of Paris.

To evaluate whether the spatial error introduced by analogue matching is acceptable, we compare the downscaling mismatch against RCM multi-model variability. Spatially, the multi-model variability across the most extreme heatwaves per RCM-GCM combination (those from Figure 4) averages 1.80 ∘C1.80\,^\circ\mathrm{C}. While our two-step approach constrains the temperature error in Paris to <0.1 ∘C< 0.1\,^\circ\mathrm{C} before matching spatial patterns, the average root-mean-square error (RMSE) of the eight most extreme events averages 2.21 ∘C2.21\,^\circ\mathrm{C} (the difference between row 1 and 2 in Figure 7). As shown in Figure A.7, downscaling errors exceed RCM variability primarily over sea areas and mountainous regions such as the Alps. This mismatch is not surprising since coarse climate models smooth out mountain ranges [Gutowski Jr et al., 2020] and coastlines [6], whereas detailed regional models capture these local differences better. Over land, particularly across northern France and the broader Paris region, the mismatch remains lower than or comparable to the RCM spread. Within ±3\pm3 GCM grid cells around Paris (≈600×730\approx600 \times730 km), EnScale achieves an RMSE of 1.43 ∘C1.43\,^\circ\mathrm{C} against an RCM spread of 1.54 ∘C1.54\,^\circ\mathrm{C}. Because the spatial downscaling error stays within the range of RCM ensemble uncertainty over the larger Paris region, the analog-matching approach proves fit for purpose for localized climate impact modeling.

Sampling model variability

Beyond downscaling boosted extremes, EnScale can also sample alternative heatwave storylines. By selecting events in the emulator data that are closest to the most extreme boosted event with a peak temperature of 35∘C35^\circ\mathrm{C} over Paris, without constraining the remaining spatial field, we sample across regional model realizations to explore model uncertainty. The results are visualized in Figure 8 and highlight spatial differences: while some samples yield broad continental heatwaves (such as model 2), others feature more spatially constrained patterns (such as model 1). Notably, all global climate models downscaled with CCLM4-8-17 (models 3, 4 and 5) yield consistently lower temperatures for Paris that fail to reach 35∘C35^\circ\mathrm{C}.

Spatial fields of extreme events across eight EnScale climate models

Figure 8. Spatial fields of 35∘C35^\circ\mathrm{C} extreme events across EnScale climate models. For each RCM-GCM combination, the panel shows the event where the grid point nearest Paris (indicated by a black ’o’) most closely matches 35∘C35^\circ\mathrm{C} with temperature at this grid point indicated above each subplot.

Ultimately, storyline sampling with multi-model ML emulators provides a practical way to explore climate risks beyond a single model simulation. By combining boosted GCM events with high-resolution RCM analogs, the framework shows that extreme heatwaves can unfold in different ways, from broad continental heatwaves to localized hot spots. Capturing these different spatial patterns and model-specific differences is essential for realistic uncertainty and risk assessment [23]. Generative machine learning models trained on a multi-model ensemble provide a computationally efficient way to sample model uncertainty and physically realistic extremes beyond what is possible with standard climate model runs.

Discussion and Conclusion

While physical ensemble boosting enables targeted sampling of extreme climate events in coarse GCMs, high-resolution dynamical downscaling remains computationally unfeasible across large ensemble sets. Conversely, traditional statistical downscaling often fails due to a lack of historical analogs for unprecedented extremes. Our hybrid methodology overcomes these limitations by combining physical ensemble boosting with ML-driven analog downscaling across several emulated RCMs.

By applying the temporally consistent ML emulator EnScale, pre-trained on physics-based GCM–RCM simulations, we generate a high-resolution dataset encompassing 8,000 simulated summers across 8 RCMs. We show that the ML samples span the range of extremes from ensemble boosting and are physically consistent. Our approach is applicable to a variety of extremes. For instance, it could easily be expanded to cold spells, and by working with two variables, renewable energy droughts could be analyzed [3]. Application to new cases requires validating the physicality of the ML model for the specific task. Analog matching can further align with our two core objectives: downscaling to introduce spatial detail and identifying alternative storylines by selecting samples based on local events without spatial field constraints. Furthermore, analog matching is flexible with respect to the similarity metric used and can, for instance, also be performed on standardized anomalies (ZZ-scores) rather than absolute temperature fields, controlling for model biases or matching extremes across differing background climate states. Our hybrid setup does not only increase spatial resolution to more impact-relevant scales. It also offers major efficiency gains by eliminating dynamical RCM runs, effectively capturing internal variability and, by emulating several regional models simultaneously, model uncertainty. Because the emulator is trained on GCM-RCM pairs, the samples also include GCM model uncertainty.

Despite our promising results, several limitations offer directions for future work. A primary constraint is that the emulator datasets are not aligned with the ensemble-boosted data; future efforts could address this by leveraging regional CESM2 data for machine learning training. Additionally, the limited sample size of the original RCM data makes robust significance testing and cross-method comparisons difficult, as long return periods and very extreme events fall outside the range of what is captured by the original RCM data. Finally, it remains unclear whether the very extreme ML events that go beyond events observed in ensemble boosting (>4.6σ> 4.6\sigma) are statistical artifacts or plausible scenarios. These extreme values may stem from our normalization, or because we can sample 10 full summers rather than just a few individual events, we may be capturing extreme scenarios with different spatial fields than those found in boosting.

Acknowledgments

Nina Effenberger, Maybritt Schillinger, Maxim Samarin, and Reto Knutti are part of SPEED2ZERO, a Joint Initiative co-financed by the ETH Board.

Appendices

Supplementary material

Average daily temperature time series in the MPI model, corresponding physical RCMs, and EnScale for summer 2033

Figure A.1. Average daily temperature time series in the MPI model, corresponding physical RCMs, and EnScale for summer 2033. Thick lines show mean values for each model class, while thin blue lines display individual EnScale realizations. Subpanels are separated by their driving GCM (from left to right: CNRM-CM5, MPI-ESM1-2-LR, and MIROC5).

CNRM-CM5MPI-ESM1-2-LRMIROC5
RCM vs GCM0.84970.76040.8119
Enscale Members vs GCM (Avg)0.73700.60960.7075
Enscale Ensemble Mean vs GCM0.93390.81870.8902

Table A.1. Temporal Pearson correlation coefficients of daily mean temperature at the Paris grid point between downscaled models and their parent GCM in summer 2033.

ModelPrecipitation Amount: Slope2Precipitation Amount: pp-value3Probability of Precip.: Slope4Probability of Precip.: pp-value3Solar Radiation: Slope5Solar Radiation: pp-value3
ML Emulator−0.1094-0.1094-−1.9567-1.9567-+6.5894+6.5894-
RCM−0.1646-0.16460.00169−3.1648-3.16480.01798+7.2609+7.26090.61598
GCM−0.0144-0.01440.00167−0.8457-0.84570.19794+6.8547+6.85470.90798

Table A.2. Temperature sensitivity of hydrological parameters and surface solar radiation across model classes at the grid point closest to Paris.

We compare the absolute temperature distributions of the peak 5-day heatwaves occurring in each 10-year period (note that these events are not temporally aligned across the different datasets). Although the RCM ensemble mean exhibits a slight negative temperature bias during peak heatwave days relative to the emulator (see Figure A.2a), the histogram shows that all RCM heatwaves are fully spanned by the EnScale ensemble spread (Figure A.2b). Further, a two-sided permutation test (N=104N = 10^{4}) suggests that this discrepancy is not statistically significant (p=0.94p = 0.94; Figure A.3).

model 1model 2model 4model 6model 7model 8
Threshold Bin (σ\sigma)
[4.6, 4.8)33316810
[4.8, 5.0)010766
[5.0, 5.2)220142
[5.2, 5.4)000541
[5.4, 5.6)100211
[5.6, 5.8)010110
[5.8, 6.0)000100
[6.0, 6.2]000100
Relative Share (%)6.32%7.37%3.16%35.79%26.32%21.05%

Table A.3. Distribution of extreme daily temperature anomaly events at the Paris grid point during summer 2033 across 0.2σ\sigma threshold bins. Standardized anomalies (σ\sigma) are calculated for daily mean temperature relative to the summer reference climatology. Columns denote individual downscaling models (model 1–model 8); models 3 and 5 contained zero events exceeding 4.6σ\sigma and are omitted. Percentages in the final row represent the relative share of total extreme heat events contributed by each run.

Figure A.2. Temporal evolution of most extreme heatwaves at the Paris grid cell during 2030–2039. The red shaded background highlights the hottest 5-day average temperature window, along with the 30 preceding and 5 subsequent days. Thick lines display the ensemble mean daily temperature for each model class (GCM, RCM, EnScale and EnScale-t), while thin lines depict individual ensemble realizations. Overall, the mean heatwave profile generated by EnScale-t closely aligns with that of the physical RCMs.

Figure A.3. Histogram of permutation test assessing the similarity of heatwave profiles between physical RCMs and the ML emulator at the Paris grid point. The histogram shows the distribution of average RMSE (°C) relative to the mean ML emulator heatwave, computed from 10,000 random samples of 8 ML realizations. The vertical black line indicates the observed mean RMSE across the 8 physical RCMs, demonstrating that the RCM heatwave deviation is statistically indistinguishable from random emulator sampling.

Figure A.4. Temperature dependence of daily precipitation characteristics at the Paris grid cell during summer across model classes, subset by RCM configuration. Top subpanels depict mean daily precipitation amount (mm/day) binned in 1.0 °C temperature intervals; bottom subpanels show precipitation probability (percentage of days with >0.1> 0.1 mm/day).

Decay Slope (β\beta)pp-value
EnScale−0.2773
RCM−0.33560.601
GCM−0.30400.898

Table A.4. Weighted log-linear probability decay rates (β\beta) of 1σ1\sigma heatwave event durations at the Paris grid cell across model classes. Heatwave duration is defined as the number of consecutive days where daily mean temperature exceeds 1σ1\sigma above the reference summer climatology. Slopes (β\beta, in day−1^{-1}) are estimated via weighted least-squares (WLS) regression, using raw event counts as weights. Two-sided pp-values assess slope difference interaction terms relative to the reference EnScale model.

Spatial patterns of peak 5-day rolling mean temperature across eight model panels

Figure A.5. Spatial patterns of the peak 5-day rolling mean temperature (°C) during the absolute hottest summer event in Paris (black ’x’) per EnScale model. Subpanel titles indicate the peak 5-day average temperature in Paris. The general patterns look similar with the heatwave covering a larger area.

Figure A.6. Temporal distribution of extreme 5-day heatwave events exceeding a 4.6σ4.6\sigma standardized anomaly threshold at the Paris grid cell during summer, indexed by day of summer. Standardized anomalies (σ\sigma) are calculated using 5-day centered rolling mean temperatures relative to the reference pooled summer climatology mean and variance.

Spatial comparison of RCM ensemble variability against EnScale analogue-matching error across Europe

Figure A.7. Spatial comparison of RCM ensemble variability against EnScale analogue-matching error across Europe. Left: RCM model variability, computed as the pixel-wise standard deviation across the regional climate model ensemble for the most extreme heatwave event per model. Center: EnScale spatial downscaling mismatch, calculated as the pixel-wise root-mean-square error between the coarse target fields and the analogs selected via a two-step matching procedure. Right: Spatial difference field (RMSE−RCM Variability). Blue regions highlight areas where the downscaling error is lower than the multi-model RCM spread. The black marker denotes the location of Paris, and the black rectangle outlines the regional 7×77 \times7 grid-cell evaluation window (±3\pm3 GCM cells around Paris).

References

  1. [1]Venkatramani Balaji, Eric Maisonnave, Niki Zadeh, Bryan N Lawrence, Joachim Biercamp, Uwe Fladrich, Giovanni Aloisio, Rusty Benson, Arnaud Caubel, Jeffrey Durachta, et al. CMIP: measurements of real computational performance of Earth system models in CMIP6. Geoscientific Model Development, 10(1):19–34, 2017. doi: https://doi.org/10.5194/gmd-10-19-2017.
  2. [2]Timothy J Barrett. Computations using analysis of covariance. Wiley Interdisciplinary Reviews: Computational Statistics, 3(3):260–268, 2011. doi: https://doi.org/10.1002/wics.165.
  3. [3]Luna Bloin-Wibe, Erich Fischer, Leonard Göke, Reto Knutti, Francesco De Marco, and Jan Wohland. Climate change impacts on net load under technological uncertainty in European power systems. Environmental Research Letters, 21(8):084004, 2026. doi: https://doi.org/10.1088/1748-9326/ae5727.
  4. [4]Ben Clarke, Friederike Otto, Rupert Stuart-Smith, and Luke Harrington. Extreme weather impacts of climate change: an attribution perspective. Environmental Research: Climate, 1(1):012001, 2022. doi: https://doi.org/10.1088/2752-5295/ac6e7d.
  5. [5]Gokhan Danabasoglu, Clara Deser, Keith Rodgers, and Axel Timmermann. CESM2 Large Ensemble, September 2020. URL https://gdex.ucar.edu.
  6. [6]Alejandro Di Luca, Ramón De Elía, and René Laprise. Potential for small scale added value of RCM’s downscaled climate change signal. Climate Dynamics, 40(3):601–618, 2013. doi: https://doi.org/10.1007/s00382-012-1415-z.
  7. [7]Veronika Eyring, Sandrine Bony, Gerald A Meehl, Catherine A Senior, Bjorn Stevens, Ronald J Stouffer, and Karl E Taylor. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geoscientific Model Development, 9(5):1937–1958, 2016. doi: https://doi.org/10.5194/gmd-9-1937-2016.
  8. [8]E. M. Fischer, U. Beyerle, L. Bloin-Wibe, C. Gessner, V. Humphrey, F. Lehner, A. G. Pendergrass, S. Sippel, J. Zeder, and R. Knutti. Storylines for unprecedented heatwaves based on ensemble boosting. Nature Communications, 14(1):4643, 2023. ISSN 2041-1723. doi: 10.1038/s41467-023-40112-4.
  9. [9]Erich M Fischer, Sonia I Seneviratne, Daniel Lüthi, and Christoph Schär. Contribution of land-atmosphere coupling to recent european summer heat waves. Geophysical Research Letters, 34(6), 2007.DOI
  10. [10]Claudia Gessner, Erich M Fischer, Urs Beyerle, and Reto Knutti. Very rare heat extremes: quantifying and understanding using ensemble reinitialization. Journal of Climate, 34(16):6619–6634, 2021. doi: https://doi.org/10.1175/jcli-d-20-0916.1.
  11. [11]Cristian Giardina, Jorge Kurchan, Vivien Lecomte, and Julien Tailleur. Simulating rare events in dynamical processes. Journal of statistical physics, 145(4):787–811, 2011. doi: https://doi.org/10.1007/s10955-011-0350-4.
  12. [13]Daniela Jacob, Juliane Petersen, Bastian Eggert, Antoinette Alias, Ole Bøssing Christensen, Laurens M Bouwer, Alain Braun, Augustin Colette, Michel Déqué, Goran Georgievski, et al. EURO-CORDEX: new high-resolution climate change projections for European impact research. Regional environmental change, 14(2):563–578, 2014. doi: https://doi.org/10.1007/s10113-013-0499-2.
  13. [15]Neelesh Rampal, José González-Abad, Henry Addison, Jorge Baño-Medina, Maria Laura Bettolli, Valentina Blasone, Ben Booth, Erika Coppola, Serafina Di Gioia, Joshua Oldham-Dorrington, et al. CORDEX-ML-Bench: A Benchmark for Data-Driven Regional Climate Downscaling-Experiment Design and Overview. arXiv preprint arXiv:2606.29172, 2026.arxiv.org/abs/2606.29172
  14. [16]Keith B Rodgers, Sun-Seon Lee, Nan Rosenbloom, Axel Timmermann, Gokhan Danabasoglu, Clara Deser, Jim Edwards, Ji-Eun Kim, Isla Simpson, Karl Stein, et al. Ubiquity of human-induced changes in climate variability. Earth System Dynamics, 2021:1–22, 2021. doi: https://doi.org/10.5194/esd-12-1393-2021.
  15. [17]William Rudisill, Alan Rhoades, Zexuan Xu, and Daniel R Feldman. Are atmospheric models too cold in the mountains? The state of science and insights from the SAIL field campaign. Bulletin of the American Meteorological Society, 105(7):E1237–E1264, 2024. doi: https://doi.org/10.1175/bams-d-23-0082.1.
  16. [18]Christoph Sauter, Jennifer L Catto, Hayley J Fowler, Seth Westra, and Christopher J White. Compounding heatwave-extreme rainfall events driven by fronts, high moisture, and atmospheric instability. Journal of Geophysical Research: Atmospheres, 128(21):e2023JD038761, 2023. doi: https://doi.org/10.1029/2023jd038761.
  17. [19]Maybritt Schillinger, Maxim Samarin, Xinwei Shen, Reto Knutti, and Nicolai Meinshausen. EnScale: Temporally-consistent multivariate generative downscaling via proper scoring rules. arXiv preprint arXiv:2509.26258v2, 2025.arxiv.org/abs/2509.26258
  18. [20]Maybritt Schillinger, Maxim Samarin, Xinwei Shen, Reto Knutti, and Nicolai Meinshausen. EnScale: Temporally-consistent multivariate generative downscaling via proper scoring rules. Journal of Advances in Modeling Earth Systems, 18(8):e2025MS005593, 2026. doi: https://doi.org/10.1029/2025ms005593.
  19. [21]Uwe Schulzweida. CDO User Guide, October 2023. URL https://doi.org/10.5281/zenodo.10020800.
  20. [22]Karl E Taylor, Ronald J Stouffer, and Gerald A Meehl. An overview of CMIP5 and the experiment design. Bulletin of the American meteorological Society, 93(4):485–498, 2012. doi: https://doi.org/10.1175/bams-d-11-00094.1.
  21. [23]Claudia Tebaldi and Reto Knutti. The use of the multi-model ensemble in probabilistic climate projections. Philosophical transactions of the royal society A: mathematical, physical and engineering sciences, 365(1857):2053–2075, 2007. doi: https://doi.org/10.1098/rsta.2007.2076.
  22. [24]Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004. doi: https://doi.org/10.1109/tip.2003.819861.
  23. [25]Heini Wernli, Marcus Paulat, Martin Hagen, and Christoph Frei. SAL—A novel quality measure for the verification of quantitative precipitation forecasts. Monthly Weather Review, 136(11):4470–4487, 2008. doi: https://doi.org/10.1175/2008mwr2415.1.

Paper details

vlnc:W_b8abswfh
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 2

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents