Bridging CORDEX and CMIP6: Machine Learning Downscaling for Wind and Solar Energy Droughts in Central Europe
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Bridging CORDEX and CMIP6".
Jane: Here's the gist of the paper. Reliable regional climate information is essential for assessing the impacts of climate change and for planning in sectors such as renewable energy; yet, producing high-resolution projections through coordinated initiatives like CORDEX that run multiple physical regional climate models is both computationally demanding and difficult to organize Machine learning emulators that learn the mapping between global and regional climate fields offer a promising way to address these limitations Here we introduce the application of such an emulator: trained on CMIP5 and CORDEX simulations, it reproduces regional climate model data with sufficient accuracy When applied to CMIP6 simulations not seen during training, it also produces realistic results, indicating stable performance Using CORDEX data, CMIP5 and CMIP6 simulations, as well as regional data generated by two machine learning models, we analyze the co-occurrence of low wind speed and low solar radiation and find indications that the number of such energy drought days is likely to decrease in the future Our results highlight that downscaling with machine learning emulators provides an efficient complement to efforts such as CORDEX, supplying the higher-resolution information required for impact assessments.
Tom: Pruning redundant reasoning tokens The study introduces a machine learning emulator trained on coupled CMIP5-CORDEX simulations to downscale coarse global climate model outputs to regional climate models The emulator is then applied to CMIP6 data, focusing on the occurrence of days with simultaneously low wind and solar radiation, defined as energy droughts.
Jane: This approach bridges the gap between methodological development and applied climate impact research by providing higher-resolution information required for impact assessments.
Tom: How the training signal is built The training data consists of daily averages of surface wind speed (sfcWind) and solar radiation (rsds) from CMIP5 and CORDEX simulations The ML model employed is CorrDiff, which was previously shown to successfully perform multivariate downscaling.
Title and authors: Tom: So, this paper is about using machine learning emulators to bridge the gap between global climate models and regional ones for wind and solar energy droughts in Central Europe <ref:2512.07429#pg1>.
Jane: It sounds like they're tackling that big problem of needing high-resolution climate information for planning renewable energy, which is hard because running those multiple physical regional models is just too computationally expensive and messy <ref:2512.07429#pg1>.
Lu: The core idea here is training an emulator on CMIP5 and CORDEX simulations to learn the mapping between those coarse global climate fields and the high-resolution regional ones, which essentially lets you reproduce regional model data with enough accuracy <ref:2512.07429#pg2>.
Meng: But how do you actually make this emulator useful at scale? We need to know if it can handle the real complexity when we apply it to newer models like CMIP6 <ref:2512.07429#pg1>.
Tom: The paper says that when they applied this trained emulator to CMIP6 simulations that they hadn't seen before, it still produced realistic results, which suggests the performance is stable <ref:2512.07429#pg1>.
Jane: That’s interesting because training on one set of models and testing on another set shows a kind of generalizability in the relationship they learned <ref:2512.07429#pg1>.
Lu: They built the training data using daily averages for surface wind speed, which they call sfcWind, and surface downwelling shortwave radiation, called rsds, from both CMIP5 and CORDEX simulations <ref:2512.07429#pg3>.
Meng: And they used a model called CorrDiff for the actual downscaling process, which they noted has successfully done multivariate downscaling before <ref:2512.07429#pg3>.
Tom: They trained this model over a long time span, from one thousand nine hundred seventy-one to two thousand ninety-nine using specific periods of one thousand nine hundred seventy-one to two thousand twenty-nine and two thousand forty to two thousand eighty-nine for the training <ref:2512.07429#pg3>.
Jane: During that training, they also fed in other daily climate variables like temperature, precipitation, surface wind speed, and solar radiation from both the global models and their corresponding regional models <ref:2512.07429#pg3>.
Lu: Now for the actual analysis part of the study on Central Europe, they focused on a region defined by longitudes from minus four point nine zero to nineteen point one eight and latitudes from forty point four four to fifty-five point nine seven <ref:2512.07429#pg3>.
Meng: They define an energy drought event as a day when both the daily spatial averages of wind speed, sfcWind, and solar radiation, rsds, fall below a threshold set at the twenty percentiles of their respective distributions <ref:2512.07429#pg3>.
Tom: They then compared the drought days across two time periods: two thousand thirty to two thousand thirty-nine and two thousand ninety to two thousand ninety-nine and found that in both cases, drought days were absent in spring and summer across all projections <ref:2512.07429#pg3>.
Jane: That’s a key finding for us because it suggests that the climate patterns they are looking at are pretty consistent across the future scenarios they tested <ref:2512.07429#pg3>.
Lu: The emulator's performance was validated against CORDEX data, and it agreed quite closely with both the original regional model data and their emulated results regarding key seasonal patterns and the magnitude and direction of projected changes in drought days <ref:2512.07429#pg4>.
Meng: What I see here is that even though they used an emulator instead of running a full RCM, the emulated regional model data agreed well with the original regional model data on those patterns <ref:2512.07429#pg4>.
Tom: And they also mentioned that the spread of their eighty emulated regional model samples was generally larger than that of the eight original RCM samples <ref:2512.07429#pg4>.
Jane: That’s a little caution, maybe indicating that the emulator introduces some variability in its predictions compared to running just one specific RCM <ref:2512.07429#pg4>.
The paper's summary: Tom: So, we've been talking about how hard it is to get high-resolution climate data for energy planning, and now we're looking at this paper on machine learning emulators <ref:2512.07429#pg6>.
Jane: Right. The main takeaway from the summary is that these emulators are trained on older models like CMIP5 and CORDEX to learn how to turn those big global climate model outputs into the more detailed regional ones, which is what we need for planning <ref:2512.07429#pg6>.
Lu: Specifically, they train a machine learning model called CorrDiff on daily wind speed and solar radiation data from CMIP5 and CORDEX to reproduce regional climate model data with enough accuracy <ref:2512.07429#pg3>.
Meng: And the paper shows that when you test this trained tool on completely new CMIP6 simulations, it still gives realistic results, which is a big deal because those are the models we're actually using for future projections <ref:2512.07429#pg1>.
Tom: That stability across different model generations is what I like; it suggests the mapping they learned isn't just lucky for one set of simulations <ref:2512.07429#pg6>.
Jane: They then use this emulator to look at energy droughts, which they define as days when both the wind speed and the solar radiation are simultaneously below a certain threshold <ref:2512.07429#pg3>.
Lu: The analysis focuses on Central Europe, using a specific geographic area bounded by those longitudes and latitudes <ref:2512.07429#pg3>.
Meng: They looked at the comparison between drought days in the two thousand thirty to two thousand thirty-nine period versus the two thousand ninety to two thousand ninety-nine period, and found that for spring and summer, those drought days just disappeared in all projections <ref:2512.07429#pg3>.
Tom: That's a pretty strong pattern they found; it means these specific conditions are unlikely to happen in the future summers for that region <ref:2512.07429#pg3>.
Jane: And they validated this against CORDEX data, and it showed that the emulated regional data agreed closely with the original regional model data on those key seasonal patterns <ref:2512.07429#pg4>.
Lu: But they also noted that the spread of their eighty emulated regional model samples was actually bigger than the eight original RCM samples they used for comparison <ref:2512.07429#pg4>.
Meng: That difference in spread means there's more variability in the emulator's predictions than we might expect from just running one specific model <ref:2512.07429#pg4>.
Tom: So, when they applied this to CMIP6 data, they found that downscaling to the regional resolution actually increased the number of identified drought days compared to just looking at the underlying global models <ref:2512.07429#pg5>.
Jane: That's an important nuance; it means we might be missing more of these events if we only look at the coarser global data <ref:2512.07429#pg5>.
Lu: What’s interesting is that even though there were more drought days in the high-resolution data, the overall trend—that decrease in drought days toward the end of the century—was consistent across all model types <ref:2512.07429#pg5>.
Meng: So, regardless of whether you start with CMIP5 or CMIP6, the long-term outlook for these energy droughts seems pretty consistent <ref:2512.07429#pg5>.
Tom: And they also checked how sensitive their results were to the number of samples they used; comparing ten samples to fifty showed that the mean number of drought days per season only shifted a little when using fifty <ref:2512.07429#pg6>.
Jane: But drawing more samples did bump up the extremes, showing that the maximum number of drought days went from twenty with ten samples to twenty-three with fifty in autumn and from sixty-six to seventy-two in winter <ref:2512.07429#pg6>.
Lu: The spatial patterns and the trends they found matched what you got using CorrDiff, which means their main conclusions aren't really dependent on which specific machine learning architecture they chose <ref:2512.07429#pg6>.
Meng: That’s good; it means we can trust this downscaling approach more because it doesn't seem tied to one specific ML setup <ref:2512.07429#pg6>.
Tom: So, to wrap up the summary, they proved that these machine learning emulators can effectively downscale those coarse global climate model projections into regional climate models <ref:2512.07429#pg6>.
Jane: And because they reproduce the regional signals and variability so well, it really highlights how much more detailed information we get for energy impact studies <ref:2512.07429#pg6>.
Lu: This tool is a computationally efficient way to bridge those global and regional climate projections, providing more detailed insights for impact studies <ref:2512.07429#pg6>.
Meng: If we can get this high-resolution info reliably, it supports more detailed insights for impact studies that used to be too hard to get because of the computational cost <ref:2512.07429#pg3>.
Tom: It’s a practical tool that helps us plan better using climate data that is actually useful for our energy needs <ref:2512.07429#pg3>.
Jane: We've seen how it works and what it does for regional drought events, but the authors also pointed out that they need to incorporate more elaborate definitions of compound energy droughts like capacity factors <ref:2512.07429#pg7>.
Lu: That means the next step is refining those definitions so the downscaling is truly ready for direct energy system applications <ref:2512.07429#pg7>.
The paper's improvements: Tom: So, we've seen how the paper uses an emulator trained on older models to reproduce data from newer CMIP6 simulations <ref:2512.07429#pg1>.
Jane: Right, and what they suggest next is that this whole emulator thing can be used to efficiently downscale those coarse global models to regional climate model resolution at a fraction of the computational cost <ref:2512.07429#pg2>.
Lu: They're saying this saves you from having to run a full regional climate model every single time, which is a huge win for speed and organization <ref:2512.07429#pg3>.
Meng: If we can skip the massive computational cost of running an RCM, that's what matters for production; it makes high-resolution info accessible <ref:2512.07429#pg6>.
Tom: And they're also pointing out that this approach allows you to analyze the co-occurrence of low wind speed and low solar radiation across multiple model ensembles, which helps with uncertainty <ref:2512.07429#pg3>.
Jane: They're basically saying this method is a way to bridge that gap between the big global pictures and the specific local conditions we need for impact assessments <ref:2512.07429#pg6>.
Lu: The implication is that you get higher-resolution information required for impact studies without needing to coordinate multiple physical regional climate models <ref:2512.07429#pg3>.
Meng: I like that because coordination is a nightmare, and if this AI tool handles the mapping, it might actually make these assessments feasible for us at scale <ref:2512.07429#pg6>.
Tom: They also mentioned that they should incorporate more elaborate definitions of compound energy droughts, including capacity factors and where renewable power installations are located <ref:2512.07429#pg7>.
Jane: That means the next step for this research is to make the drought definition more specific to actual energy systems, not just general weather patterns <ref:2512.07429#pg7>.
Lu: They even pointed out that approaches like EnScale-t show promise in handling those kinds of detailed calculations and could be used later <ref:2512.07429#pg7>.
Meng: And if they re-train a machine learning model with additional regional model global model pairs and capacity factors, that would really align the downscaling approach with actual energy system applications <ref:2512.07429#pg7>.
Tom: So, it's moving from just climate science to being directly useful for energy infrastructure planning <ref:2512.07429#pg6>.
Jane: That’s the big picture; using this tool means we get a much more realistic look at potential risks for renewable energy systems based on these high-resolution details <ref:2512.07429#pg6>.
Conclusion: Tom: So we've covered how this paper uses an emulator trained on older models to reproduce data from newer CMIP6 simulations <ref:2512.07429#pg1>.
Jane: Exactly, and the core finding is that these machine learning emulators can effectively downscale those coarse global climate model projections into regional climate model resolution <ref:2512.07429#pg6>.
Lu: It’s a way to capture key characteristics of drought events relevant for energy systems by reproducing the signals and variability <ref:2512.07429#pg6>.
Meng: It’s efficient, which is good because running a full regional climate model every time is just not feasible at scale <ref:2512.07429#pg3>.
Tom: And the paper shows that this method reveals more frequent droughts in high-resolution data than what the driving global models indicated <ref:2512.07429#pg6>.
Jane: That suggests we might be underestimating the risk based on coarser simulations, which is important for planning <ref:2512.07429#pg6>.
Lu: The overall potential here is using machine learning emulators as a computationally efficient tool to bridge global and regional climate projections <ref:2512.07429#pg6>.
Meng: If they can re-train a machine learning model with additional regional model global model pairs and capacity factors, that's how we get it ready for production <ref:2512.07429#pg7>.
Tom: So the final word on "Bridging CORDEX and CMIP6: Machine Learning Downscaling for Wind and Solar Energy Droughts in Central Europe," is that machine learning emulators offer a computationally efficient way to get high-resolution climate projections <ref:2512.07429#pg6>.
Jane: It's a solid piece of work that shows the added value of using these emulators for impact studies <ref:2512.07429#pg6>.
Lu: This sets a clear direction for how we can use AI to handle the massive computational demands of climate modeling <ref:2512.07429#pg3>.
Meng: We'll keep an eye on these next steps, especially how they integrate those real-world energy factors <ref:2512.07429#pg7>.
Institute for Atmospheric and Climate Science ETH Zurich · Swiss Data Science Center EPFL and ETH Zurich
stat.AP, physics.ao-ph
Submitted: 2025-12-08
Updated: 2025-12-08
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 85/100
The gist: Reliable regional climate information is essential for assessing the impacts of climate change and for planning in sectors such as renewable energy; yet…
Key concepts
- Machine Learning Emulator
- This is a trained artificial intelligence model designed to learn the relationship between large-scale global climate fields and smaller, regional climate models. It acts as a fast tool to create high-resolution regional data without needing to run computationally expensive physical simulations for every scenario.
- Energy Drought Days
- These are specific weather events defined as days where both surface wind speed and solar radiation fall simultaneously below a certain low threshold. This combination indicates periods of energy scarcity relevant to renewable energy planning in regions like Central Europe.
- Downscaling
- This process involves taking large-scale climate model outputs (like global models) and refining them to provide more detailed, localized information relevant to specific areas, such as a particular region. The emulator achieves this by mapping the coarse global data onto the finer resolution of regional models.
Terminology
Summary
Reliable regional climate information is essential for assessing the impacts of climate change and for planning in sectors such as renewable energy; yet, producing high-resolution projections through coordinated initiatives like CORDEX that run multiple physical regional climate models is both computationally demanding and difficult to organize Machine learning emulators that learn the mapping between global and regional climate fields offer a promising way to address these limitations Here we introduce the application of such an emulator: trained on CMIP5 and CORDEX simulations, it reproduces regional climate model data with sufficient accuracy When applied to CMIP6 simulations not seen during training, it also produces realistic results, indicating stable performance Using CORDEX data, CMIP5 and CMIP6 simulations, as well as regional data generated by two machine learning models, we analyze the co-occurrence of low wind speed and low solar radiation and find indications that the number of such energy drought days is likely to decrease in the future Our results highlight that downscaling with machine learning emulators provides an efficient complement to efforts such as CORDEX, supplying the higher-resolution information required for impact assessments. p. 0 ch
Maxim Samarin
Swiss Data Science Center
EPFL and ETH Zurich
Maybritt Schillinger
Seminar for Statistics
ETH Zurich
Reto Knutti
Institute for Atmospheric and Climate Science
ETH Zurich
Abstract
Reliable regional climate information is essential for assessing the impacts of climate change and for
planning in sectors such as renewable energy; yet, producing high-resolution projections through coordinated initiatives like CORDEX that run multiple physical regional climate models is both computationally demanding and difficult to organize. Machine learning emulators that learn the mapping
between global and regional climate fields offer a promising way to address these limitations.Where the paper says it
Pruning redundant reasoning tokens
The study introduces a machine learning emulator trained on coupled CMIP5-CORDEX simulations to downscale coarse global climate model outputs to regional climate models The emulator is then applied to CMIP6 data, focusing on the occurrence of days with simultaneously low wind and solar radiation, defined as energy droughts. This approach bridges the gap between methodological development and applied climate impact research by providing higher-resolution information required for impact assessments. p. 1 Importantly, we present both validation and application, providing new
insights into future risks for renewable energy systems. Further, we demonstrate that the gap between methodological
development and applied climate impact research can be bridged.Where the paper says it
How the training signal is built
The training data consists of daily averages of surface wind speed (sfcWind) and solar radiation (rsds) from CMIP5 and CORDEX simulations The ML model employed is CorrDiff, which was previously shown to successfully perform multivariate downscaling. The training period for the models spans 1971-2099, with 1971-2029 and 2040-2089 serving as the training periods. p. 2 1971-2029 and 2040-2089 serve as the
training periods. 2030-2039 and 2090-2099 are used for testing purposes in Schillinger et al. [2025] and are analyzed
here. During training, we consider daily 2 m temperature (tas), total precipitation (pr), surface wind (sfcWind), and
surface downwelling shortwave (solar) radiation (rsds) from both, the driving GCMs and the corresponding downscaled RCMs.Where the paper says it
Where the paper says it
Where the gains come from
The analysis focuses on Central Europe, selecting the region bounded by longitudes [−4.90, 19.18] and latitudes [40. p. 1 We focus
on Central Europe, selecting the region bounded by longitudes [−4. 18] and latitudes [40. We focus
on Central Europe, selecting the region bounded by longitudes [−4. 18] and latitudes [40. As in their study, we adopt a seasonal approach,
dividing the year into four seasons: December, January, and February (winter in our study region); March, April, and
May (spring); June, July, and August (summer); and September, October, and November (autumn). To identify critical
low-renewable days, we first compute the daily spatial averages of wind speed sfcWind and solar radiation rsds across
the region and then flag days as drought days when both averages simultaneously fall below a pre-defined threshold
p. Emulated and original RCM data agree closely on both the key seasonal patterns and the magnitude and direction
of projected changes in drought days, as shown in Figure 1 for daily spatial averages. Our seasonal analysis reveals
that drought days are absent in spring and summer across all projections.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Emulator validation and analysis of CORDEX data
The emulator validates its performance against CORDEX data, showing that emulated and original RCM data agree closely on both key seasonal patterns and the magnitude and direction of projected changes in drought days. The analysis reveals that drought days are absent in spring and summer across all projections Overall, the spread of the 80 emulated RCM samples is generally larger than that of the 8 original RCM samples. p. 3 Our seasonal analysis reveals
that drought days are absent in spring and summer across all projections. Overall, the spread of the 80 emulated
RCM samples is generally larger than that of the 8 original RCM samples; we examine the spread in more detail in
Section 3.Where the paper says it
Application to CMIP6
When applied to CMIP6 data, the analysis highlights two key findings: first, downscaling to regional resolution increases the number of identified drought days compared to underlying GCMs Second, the general trend—a decrease in drought days toward the end of the century—is consistent across original RCMs, emulated RCMs, and GCMs in both CMIP5 and CMIP6.
Sample size sensitivity and robustness across ML models
Comparing results based on 10 samples with those from 50 samples shows that the mean number of drought days per season differs only slightly when using 50 instead of 10 samples. Drawing more samples, however, yields higher extremes, as the maximum number of drought days rises from 20 (10 samples) to 23 (50 samples) in autumn and from 66 to 72 in winter. p. 6 more drought days: the maximum number of drought days rises from 20 (10 samples) to
23 (50 samples) in autumn and from 66 to 72 in winter. In both 10 and 50 samples, more and higher extremes appear
than in the original RCM data, which essentially represent a single realization (one sample). , 2025], on the
same CMIP5–CORDEX pairs. The trends in CMIP5 and CMIP6, as well as the spatial patterns, closely match those
obtained with CorrDiff (see Figure B.Where the paper says it
Where the paper says it
Conclusion
In this study, we demonstrate that machine learning emulators can effectively downscale coarse global climate model projections to the resolution of regional climate models, capturing key characteristics of drought events relevant to energy systems The emulator reproduces RCM signals and variability, and our analysis highlights the added value of high-resolution modeling by revealing more frequent droughts in high-resolution data than indicated by the driving GCMs. Overall, our study highlights the potential of machine learning emulators as a computationally efficient tool to bridge global and regional climate projections, supporting more detailed insights for impact studies. p. 8 Our results emphasize the importance
of validating emulators on impact-relevant metrics, and the probabilistic nature of the emulator enables the generation
of large ensembles, facilitating robust statistical analyses. Overall, our study highlights the potential of machine
learning emulators as a computationally efficient tool to bridge global and regional climate projections, supporting
more detailed insights for impact studies.Where the paper says it
Future work
Future work should incorporate more elaborate definitions of compound energy droughts, including capacity factors and locations of renewable power installations Alternative approaches, such as EnScale-t, show promise in this regard and could be exploited in future work. Re-training a machine learning model with additional RCM-GCM pairs and capacity factors directly could further align the downscaling approach with energy system applications. p. 8 , 2025], show promise in this regard and could be exploited in future work. Re-training a machine
learning model with additional RCM-GCM pairs and capacity factors directly could further align the downscaling
approach with energy system applications.Where the paper says it
Acknowledgements
Nina Effenberger, Maybritt Schillinger, Maxim Samarin, and Reto Knutti are part of SPEED2ZERO, a Joint Initiative co-financed by the ETH Board. The paper is available as Preprint. p. 0 , 2021] and solar radiation fields [Ha et al. However, because
high-resolution regional CMIP6-based data are not yet available, impact studies often face a trade-off between using
the most up-to-date coarse-resolution CMIP6 projections and higher-resolution regional CORDEX data [Giorgi and
Gutowski Jr, 2015] from CMIP6’s predecessor, CMIP5 [Rummukainen, 2016, Gibba et al.Where the paper says it
References
Luna Bloin-Wibe, Robin Noyelle, Vincent Humphrey, Urs Beyerle, Reto Knutti, and Erich Fischer. Estimating return periods for extreme events in climate models through ensemble boosting Cameron Bracken, Nathalie Voisin, Casey D Burleyson, Allison M Campbell, Z Jason Hou, and Daniel Broman. p. 9 EGUsphere, 2025:1–40, 2025. Cameron Bracken, Nathalie Voisin, Casey D Burleyson, Allison M Campbell, Z Jason Hou, and Daniel Broman. Renewable Energy, 220:119550, 2024. Ui-Yong Byun, Eun-Chul Chang, Joowan Kim, Joong-Bae Ahn, Dong-Hyun Cha, Seung-Ki Min, and Young-Hwa
Byun. David Carvalho, Alfredo Rocha, Xurxo Costoya, Maite DeCastro, and Moncho Gómez-Gesteira. Wind energy resource over Europe under CMIP6 future climate projections: What changes from CMIP5 to CMIP6. Francesco De Marco, Jacob Mannhardt, Alfredo Oneto, and Giovanni Sansavini. Climate-resilient energy systems
planning via system-informed identification of stressful events. Francesco De Marco, Jacob Mannhardt, Alfredo Oneto, and Giovanni Sansavini. Climate-resilient energy systems
planning via system-informed identification of stressful events. Nina Effenberger and Nicole Ludwig. Turbine location-aware multi-decadal wind power predictions for germany
using cmip6. Nina Effenberger, Nicole Ludwig, and Rachel H White. Mind the (spectral) gap: how the temporal resolution of wind
data affects multi-decadal wind power forecasts. Wiley Interdisciplinary Reviews: Climate Change, 6(3):301–319, 2015. Veronika Eyring, Sandrine Bony, Gerald A Meehl, Catherine A Senior, Bjorn Stevens, Ronald J Stouffer, and Karl E
Taylor. 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. Annual review
of environment and resources, 40(1):467–490, 2015. Aleksander Grochowicz, Koen Van Greevenbroek, and Hannah C Bloomfield. Aleksander Grochowicz, Koen Van Greevenbroek, and Hannah C Bloomfield. Using power system modelling outputs
to identify weather-induced extreme events in highly renewable systems. Our
results highlight that downscaling with machine learning emulators provides an efficient complement to efforts such as CORDEX, supplying the higher-resolution information required for impact
assessments. 1 Introduction
Global or regional climate models provide the foundation for assessing future changes in weather and climate. David Carvalho, Alfredo Rocha, Xurxo Costoya, Maite DeCastro, and Moncho Gómez-Gesteira. Wind energy resource over Europe under CMIP6 future climate projections: What changes from CMIP5 to CMIP6. effenberger@env. ch
Maxim Samarin
Swiss Data Science Center
EPFL and ETH Zurich
Maybritt Schillinger
Seminar for Statistics
ETH Zurich
Reto Knutti
Institute for Atmospheric and Climate Science
ETH Zurich
Abstract
Reliable regional climate information is essential for assessing the impacts of climate change and for
planning in sectors such as renewable energy; yet, producing high-resolution projections through coordinated initiatives like CORDEX that run multiple physical regional climate models is both computationally demanding and difficult to organize. However, there is work bridging this gap: Strnad et al. [2025] recently applied a generative deep
learning framework to downscale CMIP6 climate simulations, assessing future Dunkelflaute events in Germany under
various emission scenarios, and Lewis et al. Traditional downscaling approaches include
dynamical downscaling, which uses high-resolution RCMs nested within GCMs, and statistical downscaling, which
links large-scale model outputs to local-scale variables [Ekström et al. ML emulators can be trained directly on
RCM outputs, thereby linking physical modeling and statistical approaches [Van Der Meer et al. Using
CORDEX data, CMIP5 and CMIP6 simulations, as well as regional data generated by two machine
learning models, we analyze the co-occurrence of low wind speed and low solar radiation and find
indications that the number of such energy drought days is likely to decrease in the future. Our
results highlight that downscaling with machine learning emulators provides an efficient complement to efforts such as CORDEX, supplying the higher-resolution information required for impact
assessments. Uneven sample sizes among GCMs, RCMs, and emulated data also complicate the comparison of mean changes
reported in Table 2. Following the recommendation of Von Storch and Zwiers [2013], we refrain from applying formal statistical tests.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
Bridging CORDEX and CMIP6: Machine Learning Downscaling for Wind and Solar Energy Droughts in Central Europe Nina Effenberger Maxim Samarin Maybritt Schillinger Reto Knutti. The paper is 2512. p. 0 Under review. arXiv:2512.Where the paper says it
The single most informative sentence about the paper's subject and finding is: Reliable regional climate information is essential for assessing the impacts of climate change and for planning in sectors such as renewable energy; yet, producing high-resolution projections through coordinated initiatives like CORDEX that run multiple physical regional climate models is both computationally demanding and difficult to organize
Reliable regional climate information is essential for assessing the impacts of climate change and for planning in sectors such as renewable energy; yet, producing high-resolution projections through coordinated initiatives like CORDEX that run multiple physical regional climate models is both computationally demanding and difficult to organize Machine learning emulators that learn the mapping between global and regional climate fields offer a promising way to address these limitations Here we introduce the application of such an emulator: trained on CMIP5 and CORDEX simulations, it reproduces regional climate model data with sufficient accuracy When applied to CMIP6 simulations not seen during training, it also produces realistic results, indicating stable performance
The training data consists of daily averages of surface wind speed (sfcWind) and solar radiation (rsds) from CMIP5 and CORDEX simulations The ML model employed is CorrDiff, which was previously shown to successfully perform multivariate downscaling The training period for the models spans 1971-2099, with 1971-2029 and 2040-2089 serving as the training periods During training, daily 2 m temperature (tas), total precipitation (pr), surface wind (sfcWind), and surface downwelling shortwave (solar) radiation (rsds) from both driving GCMs and corresponding downscaled RCMs are considered.
The analysis focuses on Central Europe, selecting the region bounded by longitudes [−4.90, 19.18] and latitudes [40.44, 55.97] The definition of energy drought events involves flagging days when both daily spatial averages of wind speed sfcWind and solar radiation rsds simultaneously fall below a pre-defined threshold p, set at the 20th percentiles of their respective distributions. p. 2 As in their study, we adopt a seasonal approach,
dividing the year into four seasons: December, January, and February (winter in our study region); March, April, and
May (spring); June, July, and August (summer); and September, October, and November (autumn). To identify critical
low-renewable days, we first compute the daily spatial averages of wind speed sfcWind and solar radiation rsds across
the region and then flag days as drought days when both averages simultaneously fall below a pre-defined threshold
p.Where the paper says it
The emulator validates its performance against CORDEX data, showing that emulated and original RCM data agree closely on both key seasonal patterns and the magnitude and direction of projected changes in drought days
When applied to CMIP6 simulations, the analysis highlights two key findings: first, downscaling to regional resolution increases the number of identified drought days compared to underlying GCMs Second, the general trend—a decrease in drought days toward the end of the century—is consistent across original RCMs, emulated RCMs, and GCMs in both CMIP5 and CMIP6
Comparing results based on 10 samples with those from 50 samples shows that the mean number of drought days per season differs only slightly when using 50 instead of 10 samples Drawing more samples, however, yields higher extremes, as the maximum number of drought days rises from 20 (10 samples) to 23 (50 samples) in autumn and from 66 to 72 in winter The trends in CMIP5 and CMIP6, as well as the spatial patterns, closely match those obtained with CorrDiff, demonstrating that the main conclusions are not dependent on the specific ML architecture used.
In this study, we demonstrate that machine learning emulators can effectively downscale coarse global climate model projections to the resolution of regional climate models, capturing key characteristics of drought events relevant to energy systems The emulator reproduces RCM signals and variability, and our analysis highlights the added value of high-resolution modeling by revealing more frequent droughts in high-resolution data than indicated by the driving GCMs. Overall, our study highlights the potential of machine learning emulators as a computationally efficient tool to bridge global and regional climate projections, supporting more detailed insights for impact studies
Future work should incorporate more elaborate definitions of compound energy droughts, including capacity factors and locations of renewable power installations Alternative approaches, such as EnScale-t, show promise in this regard and could be exploited in future work. Re-training a machine learning model with additional RCM-GCM pairs and capacity factors directly could further align the downscaling approach with energy system applications
Nina Effenberger, Maybritt Schillinger, Maxim Samarin, and Reto Knutti are part of SPEED2ZERO, a Joint Initiative co-financed by the ETH Board The paper is available as Preprint.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Machine learning emulators can be used to reproduce regional climate model data with sufficient accuracy by learning mapping between global and regional climate fields, which addresses the computational demands of high-resolution projections like CORDEX.
-
The improved system can efficiently downscale coarse global models to RCM resolution at a fraction of the computational cost of running an RCM, providing higher-resolution information required for impact assessments and avoiding coordination demands.
-
This system can analyze co-occurrence of low wind speed and low solar radiation events across multiple model ensembles, enabling the assessment of uncertainty and extremes across various climate realizations.
-
The emulator can provide insights into future risks for renewable energy systems by revealing more frequent droughts in high-resolution data than indicated by the driving GCMs, supporting the conclusion that
higher resolution can improve the representation of extremes.
-
The system demonstrates robustness across different machine learning architectures, as
the main conclusions regarding the frequency and spatial distribution of co-occurring wind and solar droughts are not dependent on our specific choice of ML architecture.
Abstract
Reliable regional climate information is essential for assessing the impacts of climate change and for planning in sectors such as renewable energy; yet, producing high-resolution projections through coordinated initiatives like CORDEX that run multiple physical regional climate models is both computationally demanding and difficult to organize. Machine learning emulators that learn the mapping between global and regional climate fields offer a promising way to address these limitations. Here we introduce the application of such an emulator: trained on CMIP5 and CORDEX simulations, it reproduces regional climate model data with sufficient accuracy. When applied to CMIP6 simulations not seen during training, it also produces realistic results, indicating stable performance. Using CORDEX data, CMIP5 and CMIP6 simulations, as well as regional data generated by two machine learning models, we analyze the co-occurrence of low wind speed and low solar radiation and find indications that the number of such energy drought days is likely to decrease in the future. Our results highlight that downscaling with machine learning emulators provides an efficient complement to efforts such as CORDEX, supplying the higher-resolution information required for impact assessments.
Sources
- Long-duration electricity storage needs for coping with Dunkelflaute events in Europe
- Generative AI-Downscaling of Large Ensembles Project Unprecedented Future Droughts
- Meteorological conditions during Dunkelflauten in Germany: Characteristics, the role of weather regimes and impacts on demand
- Heavy-Tailed Diffusion Models
- EnScale: Temporally-consistent multivariate generative downscaling via proper scoring rules
Related papers
- Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization
- Strategic Causal Policy Learning with Manipulable Eligibility
- BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments
- Automating structural reliability analysis with a multi-agent large language model framework
- Improving Minority Population Sampling with BISG Probabilities: Evidence from a Survey of Jewish Americans
- When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data