the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 06 Feb 2026
Coherent Modes of Northern Hemisphere Wind Extremes and Their Links to Global Large-Scale Drivers
Abstract. Analogously to the well-known seasonal atmospheric circulation patterns that capture coherent large-scale regions of synchronous variability (e.g. North Atlantic Oscillation) we aim in this study to identify spatially coherent modes of stormy seasons over the Northern Hemisphere land regions. Locally, stormy seasons are defined here as seasons with a high number of days having daily wind speed higher than the local climatological 95th-percentile derived from ERA5 reanalysis data. Applying a principal component analysis (PCA) to seasonal October-through-March (ONDJFM) local storm indices reveals a leading mode of hemispheric variability characterised by a meridional dipole structure.
Regions north of 50° N (Europe–Asia) fluctuate coherently, in opposite phase to those farther south. Correlation analyses between the principal component time series and global spatial fields of sea surface temperature (SST), mean sea level pressure (MSLP), and skin temperature (i.e. surface temperature at radiative equilibrium; SKT) identify teleconnections to the North Atlantic Oscillation (NAO) and Pacific SST anomalies, indicating that known climate modes modulate the large-scale spatial coherence of seasonal extreme wind frequency. These teleconnections to large-scale modes arise in the months preceding the target season ONDJFM, suggesting potential predictability at seasonal timescales.
To explore physical causality between SKT and storminess modes related to the atmospheric response to SKT-anomalies, we use the atmospheric emulator ACE2 driven at the surface by the relevant patterns of SKT identified in the SKT–storm correlation analysis. The ACE2 emulator is a recently released artificial-intelligence emulator trained with ERA5 reanalysis. The emulator experiments reproduce the observed storm variability pattern and yield a split jet-stream response with both poleward and equatorward branches. ACE2 simulations driven by the relevant patterns of SKT do tend to produce more stormy seasons in the regions identified by the mentioned PCA analysis. These results support a causal link between coherent large-scale patterns of seasonal storminess and large-scale surface temperature gradients.
Our findings bridge statistical climate variability with physical processes, offering a framework for understanding how continental storm risks respond to changes in global surface temperature. We note that this framework can also be applied to other extreme events.
- Preprint
(9216 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-523', Anonymous Referee #1, 12 Mar 2026
-
AC1: 'Interim Reply on RC1', Kai Bellinghausen, 20 Mar 2026
Brief Reply to Comments of Reviewer 1
First of all we would like to thank the reviewer for the thorough feedback and appreciate the time invested for the fruitful and critical comments. We will address their evaluation point-by-point later, but in this comment, we would like to clarify a few important points that we think have been not fairly criticised .
In the following, the comments will be in normal text formatting, our answers in bold format.
Comment: The reviewer poses the question of whether monthly percentiles are useful for the definition of wind extremes and emphasizes the critique that „wind damage thresholds for buildings will depend on wind force, but be largely independent of the month, so that storm relevance will be overestimated for the autumn and spring events.“
Answer: The choice to define extreme winds as those surpassing an absolute global threshold or as those surpassing a local percentile of the local wind distribution is dictated by the research question, and we think that this choice cannot be absolutely labeled as correct or incorrect.A research question focused on damages caused by wind extremes would require the first definition (absolute threshold), whereas a research question, as ours, focused on the identification of coherent patterns of wind extremes would require the second definition (local threshold).
If we had chosen to set an absolute threshold, the spatial patterns would be dominated by regions where climatologically the mean wind is stronger. They would not give information about coherent patterns of wind extremes, as the absolute threshold would certainly neglect all regions with weaker climatological winds.
We agree that the choice of a local threshold with monthly percentiles leads to an overestimation of possible impact / damage related wind extremes in autumn and spring while pointing out that this study did not focus on extreme related damages or impacts. Rather in this study, we investigate if there are coherent modes in the NH regarding local wind-extremes as defined by local percentiles. We showed that the resulting spatial pattern of the PCA remains consistent when analyzing each month in isolation. This indicates that the spatial pattern of the PCA is robust. Additionally, we do not investigate absolute thresholds since we wanted to focus on the overall climate signal. A PCA applied to a storm index based on absolute thresholds would lose this signal and instead emphasize regions where strong winds are predominant.
Comment: The reviewer has concerns with the application of ACE2 understanding it as a „re-shuffling of SST and weather situations in ERA5, combining such situations which fit to each other“ and poses concerns on the circularity of the results in the ACE2 section, arguing that ACE2 has learned the spatial modes of wind extremes from the ERA5 data.
Answer:The ACE2 model is not merely reshuffling, i.e. searching its memory for similar SKT patterns that it learned based on ERA5 and using the associated wind field pattern as a simulation output. ACE2 is an auto-regressive machine-learning model, trained on ERA5 as the ground-truth and seems to have learned the atmospherically non-linear dynamical relationships within the ERA5 climate and produces its output in an autoregressive manner. As explained in the manuscript, multiple ACE2 runs with slightly different initial conditions diverge in their simulated trajectory, similar to full-fledged climate models, and thus the ACE2 output is not merely a ‘repetition or reshuffling’ of ERA5. We think that the criticism of the reviewer is unfair because, if it were true, it would apply to all Machine-Learning weather predictions models that have been recently presented in the literature. Even the renowned ECMWF has built its own machine-learning emulator. The title of the main publication on the ACE2 model is “ACE2: accurately learning subseasonal to decadal atmospheric variability and forced responses” , published in one of the Nature Publishing Group journals. Thus, their authors do believe that the ACE2 model can be used to investigate responses to forcing.
ACE2 will see forcing input data that it might have never seen during its training period (especially during our sensitivity experiments, where we introduce artificial forcings M). In this study, we use ACE2 for sensitivity analysis to filter out the effect of possible drivers. By perturbing the SKT field with a specific pattern, we test whether the model produces a consistent atmospheric response. This is a standard methodological approach used to isolate the influence of specific drivers. The experiment showed that the artificial forcing (which – again - is non-existent in the ERA5 training set and hence unknown to ACE2) leads not only to a recovery of the ERA5 wind extreme patterns, but it increased the spatial modes discovered by the PCA.
Additionally, the submitted version of the manuscript includes a paragraph on the issue of possible circularity, explaining why we believe it does not affect our study (see lines 64ff).
Comment: The reviewer has concerns that AI was used to revise the literature and formulate text. We reply to some of the comments with our specific intent of the citation and clearly dismiss the reviewer’s suggestions of using AI text generation for this study.
1: „For example, Zheng et al discuss mechanisms for wind stilling with respect to energy production, but not hazardous storms.“
Answer: We agree that the study does not focus on extreme windspeeds and their connection to the climate modes but only analyses mean windspeeds. Nevertheless, their study shows that climate modes are connected to wind speeds on decadal timescales; and in lines 91-99 of our manuscript we only summarize the literature that used climate modes as precursors for wind characteristics. Not all studies mentioned in this paragraph are focused on wind extremes but rather on the connection between climate modes and wind characteristics.
2: „Delworth is referenced in terms of a link between the NAO and westerly flow as if it was a statistical coincidence, while geostrophic wind or thermal wind are certainly useful approaches in this respect.“
Answer: We agree that the word "show" might be misinterpreted and will replace it with "show an association" similar as they state in their paper (see quote below). Nevertheless they show an association between the NAO and the westerly wind flow, which could be a possible explanation for the coherent modes captured in our study given the correlation between the NAO-Index and the PCA-storm index as well as the resulting MSLP correlation pattern with the PCA storm index that seems to resemble the NAO structure.
- "There has been pronounced multidecadal variability of the North Atlantic Oscillation (NAO) over the past century, with a negative phase of the NAO from the 1960s into the 1970s (Fig. 1a), associated with weakened westerly winds and cold ocean surface temperatures in the subpolar North Atlantic. This was followed by a rapid switch to a positive NAO phase from the 1980s into the early 2000s, along with strengthened westerly winds and rapid warming in the subpolar North Atlantic." p.509 Delworth et al 2016Citation: https://doi.org/10.5194/egusphere-2026-523-AC1
-
AC1: 'Interim Reply on RC1', Kai Bellinghausen, 20 Mar 2026
-
RC2: 'Comment on egusphere-2026-523', Anonymous Referee #2, 16 Apr 2026
Review of ‘Coherent modes of Northern Hemisphere wind extremes and their links to global large-scale drivers’ by Bellinghausen et al.
General Comments
This paper deals with the important problem of linking societally impactful wind extremes and stormy seasons over land to large-scale atmospheric drivers. The topic of this paper is definitely within the scope of the WCD. Unfortunately I cannot support the publication of the paper in its current form. I believe there are several interesting points within the paper, but these are obscured by the approach and methodology. Attempting an POD/EOF decomposition of the storm index is an interesting approach. However, given that the first EOF explains only 8% highlight that this alone might not be the best methodology to follow. Perhaps keeping 52 modes to retain 90% and combining this dimensionality reduction methodology with clustering analysis might lead to results that are easier to explain. The use of the ACE2 emulator needs further support. There are several details that are left incomplete. For example, there are several mentions of the data that comes with the emulator as it is downloaded, but there are no details as to what kind of climate those data represent. How the emulator works should also be explained in detail. Otherwise the last part of the analysis that consists on creating a factual and a counterfactual climate is just not clear enough. For example, there is restriction on the dates to be used as initial conditions so one single date is used for several years. However, no explanation is given as why it is not possible to use e.g. ERA5 to provide initial conditions for the emulator. I also found poor connection between the different sections in the paper. For example, the correlation analysis in Section 5.1 focuses on the correlation between stormy October-March seasons and SSTs the leading months (January-September). This analysis led, according to the authors, to the identification of signals in SSTs and MSLP related to stormy seasons in the months ahead. However, in the emulator-based analysis this is ignored and instead the focus is on the October-March period only. I provide a more comprehensive list of comments below.
Specific Comments
L84-85: While I agree that extratropical cyclones are primarily driven by baroclinic instability the same is not true for tropical cyclones, which develop over tropical oceans with much weaker horizontal temperature gradients and therefore much weaker baroclinicity.
L86-87: The predominance of extratropical cyclones in boreal winter rather than summer is not due to a strong land-sea temperature gradient, but an enhanced temperature gradient between the subtropical and polar regions.
L186: Comment on the validity of extrapolating the trend found on one period to a different period. How strong is the assumption that the trend will remain constant?
L189: ‘Due to increased computing speed…’ Do you actually mean ‘To increase computing speed…’?
L250: Is one mode sufficient to represent variability given the small values of the variance explained by each mode (only 24% with 5 modes)?
L282-283: You talk about robust or not robust signals but it is not clear what you mean by that. Do you mean that they are persistent in time? All the correlations are below 0.4, which can be considered low.
L297: I don’t see a signal over Europe and Asia, but over Africa and Asia in January.
L298: There is a north-south dipole but I’m not sure how much this should be identified with the NAO. In January the dipole is not at all over the Atlantic, and in summer the NAO itself is less useful.
L326-327: ACE2 might be a dynamical system sensitive to initial conditions, but how much is that sensitivity similar to that of the atmosphere or that of ERA5?
L341: What are the standard forcings obtained through downloading ACE2? Do they represent some climatic conditions in particular? Are the SSTs based on observations? If so, on what observations?
L367: What CO2 concentrations are used in ACE2?
Figure 6: Why does the time series in the left panel both end in a minimum completely discordant to the rest of the time series or their tendencies? And what is the red dashed line? In general, add labels, such as (a), (b), etc., to panels in figures to distinguish between them.
L387: Why was it necessary to compute the NAO index using boxes when it was already computed using an EOF-based approach?
Figure 7: It is not clear what is the difference between the second and third column? What are the ERA5 statistics?
Figure 8: Why is the ERA5 correlation never negative? This leads to a big difference between ERA5 and ACE2 around 1980 involving a difference in sign. And what does the green curve represent and how is it different to the orange curve?
L420-421: From the description of ACE2 it sounds as it just requires SSTs (or surface temperature) as boundary conditions. Why can these not be obtained e.g. from ERA5 for the months previous to the extended winter season?
L428: Where are the initial conditions taken from?
L432: What is the source of the ACE2 emulator’s boundary and initial conditions?
L441: I don’t understand the restriction leading to use the same 1 October 1940 for all years. Why is it not possible to use ERA5 as before? What are the implications for the emulator’s climatology? How would it differ from e.g. the more realistic ERA5 climatology?
L446: How is SKT related to MSLP?
L453: Why is M added to F and not to climatological conditions?
L495: I don’t see the point in comparing only the magnitudes of the zonal wind disregarding the sign. The wind could be moving in opposite directions but the effect will be obscured by the authors’ approach.
L501: The authors talk about a dipole but I see a tripole.
Section 8.4.5: The visual analysis of the sort of ‘higher signal in red regions of the loading pattern and lower signal in the blue regions’ is very subjective. I don’t really see the signals described by the authors. Perhaps a more in depth correlation analysis would be useful here.
Technical Comments
L143: ERA5 does not only hold daily estimates of atmospheric variables. In fact it has an hourly temporal resolution.
L306: Why not mention the unmentioned variables?
L340: Check whether the journal has a specific format for dates.
L390: It should read ‘ratio’ rather than ‘ration’.
L466-469: Why are those very specific pressure values used and not simply 600 hPa (or 500 hPa) or 250 hPa?
L475: Add that the metric shows the mean differences between factual and counterfactual.
L480: Nomenclature: Why are the variables named wind_speed_5 and wind_speed_3? Where do 3 and 5 come from?
Figure 1: Figures in general require work. Increase the size of the map relative to the colour bar.
Figure 9 (and others): It is not necessary to repeat the colour bar.
Figure 15: Adjust the colour scale to the values of the field. Otherwise the colour bar might get saturated as it is happening in the right column.
Citation: https://doi.org/10.5194/egusphere-2026-523-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 224 | 91 | 23 | 338 | 45 | 41 |
- HTML: 224
- PDF: 91
- XML: 23
- Total: 338
- BibTeX: 45
- EndNote: 41
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
A review of egusphere-2026-523
“Coherent Modes of Northern Hemisphere Wind Extremes and Their Links to Global Large-Scale Drivers”
by Kai Bellinghausen, Birgit Hünicke, and Eduardo Zorita
General comments
The manuscript reports on research efforts regarding the connection of a hemispheric variability pattern of winter storm occurrence with known large scale atmospheric variability modes like NAO, PNA and ENSO. The idea is that the particular areas affected by more than the average number of damaging storms in a specific winter could be associated with other areas less affected. The phase of such a storm occurrence variability pattern may be related to large scale atmospheric modes and other drivers that could be predictable. Using daily mean 10 m wind speeds from ERA5 reanalysis as a proxy for storm damage, the number of days with stronger wind than the local 95th percentile for each month of the year is counted per winter. The resulting data for the Northern Hemisphere are put into a PCA. The resulting index is then related to indices of the known patterns and correlated to the spatial distribution of parameters like the global SST, also considering that there may be indications for a predictability of the storm occurrence pattern from the anomalies. The authors also make use of a recently released statistical climate emulator called ACE2 which had been trained with ERA5 reanalysis. ACE2 simulation is driven by patterns of SST (or of skin temperature?), comprising both land and ocean areas. The paper claims to have identified connections between the storminess index and SST-patterns, suggesting a relation to ENSO and the AMOC without an explicit testing. It tests the predictive potential of the global anomaly patterns found, and finds no significant skill (chapter 6). It explores a surrogate climate simulations with the statistical climate emulator ACE2, losing the focus on the specific approaches of the previous sections (e.g, looking now into other seasons than winter, different aggregations, spatial extents of analyzed regions). In this section statistical relationships with variability patterns like PNA and NAO which was missing in the earlier part. Here the paper turns into an evaluation effort for ACE2, with multiple aspects mentioned.
While the overall initial idea of the paper is interesting, I see a couple of aspects of the approach and its realization that must be re-thought by the authors. I wonder, for example, why the authors regard it a good idea to leave the role of extratropical cyclones out in their study, which are the local and regional link between windstorms and large-scale pressure pattern variability. Another question is the definition of the threshold wind speed for months of the year. Wind damage thresholds for buildings will depend on wind force, but be largely independent of the month, so that storm relevance will be overestimated for the autumn and spring events. While the authors claim that the resulting patterns look similar (line 259), I still see no good reason for the chosen monthly separation.
With respect to the evaluation of ACE2, it appears that the authors conduct a couple of analysis approaches without a clearly founded and explained hypothesis behind it. One specific question is if a different version of the emulator ACE2 is used than what is cited. According to the publication of Watt-Meyer et al. (2025) and many locations in the text, ACE2 is using SSTs as input. However, the authors state that they use skin temperature because this parameter is needed by ACE2 (section 8.2). The authors could not convince me that the emulator can provide additional information on the searched links of storminess pattern and pressure variability patterns, and thus avoid the effort of running full numerical models. It is clearly not independent from the ERA5 reanalysis, and the fact that the emulator’s training with 6 hourly reanalysis data is no guarantee that it will produce independent data (avoiding circularity) for a quantification of relationships in the aggregated winter seasons. Isn’t the essential result or ACE2 merely a re-shuffling of SST and weather situations in ERA5, combining such situations which fit to each other? Given the steering of ACE2 by SSTs (or is it global skin temperatures?) and other patterns of variables (section 7.1.1) , I would not consider this a basis for independent free climate simulations.
Given my above comments, I think that the work required to come up with a manuscript which is acceptable for WCD clearly exceeds what can be done during a revision. I thus recommend rejection of the present submission, but would like to encourage the authors to follow the initial idea mentioned, which I regard interesting and relevant.
Specific Comments
1. Definition of the storm variability index
It is not clear if the different signs of variable storm activity in Fig. 2 represent a real dipole, or if it is really a monopole of variable activity which appears as a dipole because of the applied PCA. A correlation map (teleconnection analysis) showing the areas with strongest negative correlations may be elucidating in this respect.
2. Given the hemispheric scale of the PCA applied, different regional variability patterns (for example, related to the NAO or alternatively to the PNA) enter the PCA. The methodology can have the effect of putting different variability patterns of storm activity into one (see Ambaum, Hoskins and Stephenson, 2001, for a related discussion on NAO, PNA and AO). Thus, it is not clear if the hemispheric approach hides the existence of regional patterns which could have a much stronger relationship to the pressure patterns. Fig. 8 points at an intermittent link between the patterns, which has been documented in some studies some time ago, for example in conjunction with the so-called “storm track”.
3. Chapter 5 produces some handwaving reasons for variability without actually looking into the relevant data (e.g., AMOC, SSTs and baroclinic instability). These could be explicitly checked as the respective data are available.
4. I would have expected that physical concepts like geostrophic or thermal wind as well as baroclinicity are used in the analysis and the interpretation.
Other comments
A couple of the cited references are apparently used without having read them, and this makes me speculate if the authors could have simply used generative AI for producing parts of the text. For example, Zheng et al discuss mechanisms for wind stilling with respect to energy production, but not hazardous storms. Cited work on extratropical storms is erroneously also linked to tropical storms in the text. Delworth is referenced in terms of a link between the NAO and westerly flow as if it was a statistical coincidence, while geostrophic wind or thermal wind are certainly useful approaches in this respect. Pfleiderer’s and Klotzbach’s papers on hurricanes are cited in the context of some “predictive skill from favourable climatic conditions and established teleconnections” as a basis of statistical models, far from the mid-latitude issues which are mainly in the paper’s focus.
Line 82-86: There seems to be a surprising misconception of the authors with respect to the basic mechanisms of tropical and extratropical cyclones. I wonder how this can have made it into the submitted text.
Line 119: What is the basis of the speculation that “regions with similar extreme wind variability may experience comparable shifts due to climate change”?
Section 2.3: What is the pattern of the NAO-pattern associated with the downloaded index? For which months was it computed from the monthly indices at https://www.cpc.ncep.noaa.gov/products/precip/CWlink/pna/nao.shtml? What is the role of the seasonal changes for the analysis conducted?
Section 3: The section is very difficult to understand, and the sense of the procedure applied with linear trend subtraction and polynomial fit does not make apparent sense to me. How does the pattern of exceedances and the time series typically look like for subsections of the time series? It is necessary to have an idea of the data eventually going into the correlations. It is not clear in how far known storms are found in the time series.
Line 259: This is a manuscript related to meteorology. The word "trough" is linked to a specific atmospheric feature and should not be used here for describing a time series.
Section 4: It appears that the outcome of the procedure from section 3 and 4 results in rather useless patterns, with just 8% of the variance explained by the largest 5 (!) modes.
The citation of Bellomo et al on line 292 is obviously misleading, as these authors look into 4*CO2” experiments which do not represent small AMOC reductions.