Mitigation needed to avoid unprecedented multi-decadal North Atlantic Oscillation magnitude

nature climate change Article https://doi.org/10.1038/s41558-025-02277-2 Mitigation needed to avoid unprecedented multi-decadal North Atlantic Oscillation magnitude Received: 6 April 2024 Accepted: 5 February 2025 Published online: 12 March 2025 Check for updates D. M. Smith 1 L. Hermanson , N. J. Dunstone 1, R. Eade 1, S. C. Hardiman , A. A. Scaife 1,2 & M. Seabrook 1 , 1 1 The North Atlantic Oscillation (NAO) dominates winters in Western Europe and eastern North America. Future climate model projections of the NAO are highly uncertain due to both modelled irreducible internal variability and different model responses. Here we show that some of the model spread in multi-decadal NAO simulations is caused by climatological water vapour errors, and develop an emergent constraint that reveals a substantial response of the NAO to volcanic eruptions and greenhouse gases (GHGs). Taking account of the signal-to-noise paradox apparent in these simulations suggests that under the high-emissions scenario the multi-decadal NAO will increase to unprecedented levels that will likely cause severe impacts, including increased flooding and storm damage. This can be avoided through mitigation to reduce GHG emissions. Our results suggest that taking model projections at face value and seeking consensus could leave society unprepared for impending extremes. The North Atlantic Oscillation (NAO)1 is the leading mode of atmospheric circulation variability in the North Atlantic, reflecting changes in the pressure gradient between the Azores High and the Iceland Low. In its positive phase, an increased pressure gradient drives stronger mid-latitude westerly winds, increased storminess, a poleward shift of the North Atlantic jet stream and storm track, warm and wet conditions in Northern Europe and southeastern North America, and cold and dry conditions in northeastern North America and Southern Europe2,3. Conditions are opposite during the negative NAO phase. Hence, the NAO severely impacts society, including through water security4, flooding5, mortality due to cold weather6, transport7, energy demand8 and supply9, structural damage from storms10 and economic losses11. Understanding how the NAO will change in future decades is key for developing effective adaptation measures. However, there are three major challenges to overcome. First, model simulations of the NAO are highly chaotic, such that tiny changes that would be impossible to measure can lead to opposite trends12. If this is true for the real world, the future NAO will be highly uncertain due to irreducible internal variability13. However, there is mounting evidence that the real-world NAO is much more predictable than models suggest14–21. This model error has been called the ‘signal-to-noise paradox’ (SNP) because a climate model can predict the real world better than one of its own ensemble members despite perfectly representing itself22. This arises when the ratio of the predictable signal to unpredictable noise is too small in models. Consequently, the magnitude of the modelled ensemble mean is too small and must be inflated to obtain realistic and reliable predictions15,18,22. Although the causes of the SNP are currently unknown, there is evidence that weak atmospheric eddy feedback23,24 and/or errors in ocean–atmosphere interactions25,26 play a role. Both of these would be expected to affect all timescales, including climate projections for which there is already some evidence20,26,27. Hence, in this study, we tested for the SNP and made appropriate adjustments. Second, simulations of atmospheric circulation depend on the model used, leading to large uncertainties28–30. However, model differences can potentially be exploited to estimate the real world using an emergent constraint31 if a robust physical relationship exists between model differences in projected changes and model differences in something that can be observed. The resulting regression enables uncertainties in future projections to be narrowed through a weighted model average, where the weights depend on how well each model Met Office Hadley Centre, Exeter, UK. 2Department of Mathematics and Statistics, Exeter University, Exeter, UK. 1 Nature Climate Change | Volume 15 | April 2025 | 403–410 e-mail: 403 Article https://doi.org/10.1038/s41558-025-02277-2 Table 1 | Model simulations and ensemble sizes Model Hist-nat Historical Historical + SSP2-4.5 Historical + SSP1-2.6 Historical + SSP5-8.5 ACCESS-CM2 3 (r[1-3]1ip1f1) 10 (r[1-10]i1p1f1) 10 (r[1-10]i1p1f1) 10 (r[1-10]i1p1f1) 10 (r[1-10]i1p1f1) ACCESS-ESM1-5 3 (r[1-3]1ip1f1) 40 (r[1-40]i1p1f1) 40 (r[1-40]i1p1f1) 40 (r[1-40]i1p1f1) 40 (r[1-40]i1p1f1) BCC-CSM2-MR 3 (r[1-3]1ip1f1) CanESM5 50 (r[1-25]1ip[1-2]f1) 65 (r[1-25]i1p1r[1-40]i1p2]f1) 50 (r[1-25]i1p[1-2]f1) 50 (r[1-25]i1p[1-2]f1) 50 (r[1-25]i1p[1-2]f1) 50 (r[1-25]i1p[1-2]f1) 20 (r[1-10]i1p[1-2]f1) 2 (r1i1p[1-2]f1) 20 (r[1-10]i1p[1-2]f1) 3 (r[1-3]i1p2f1) 3 (r[1-3]i1p2f1) 3 (r[1-3]i1p2f1) 3 (r[1-3]i1p2f1) 11 (r[1-11]1ip1f1) 3 (r[4,10,11]i1p1f1) 3 (r[4,10,11]i1p1f1) 4 (r[1,4,10,11]i1p1f1) 3 (r[1-3]i1p1f1) 3 (r[1-3]i1p1f1) 1 (r1i1p1f1) 3 (r[1-3]i1p1f1) 30 (r[1-30]i1p1f2) 6 (r[1-6]i1p1f2) 6 (r[1-6]i1p1f2) 6 (r[1-6]i1p1f2) 10 (r[1-10]i1p1f2) 10 (r[1-10]i1p1f2) 5 (r[1-5]i1p1f2) 5 (r[1-5]i1p1f2) EC-Earth3 21 (r[1,2,4-7,9,11-19,21-25]i1p1f1) 21 (r[1,2,4-7,9,11-19,21-25] i1p1f1) 6 (r[4,6,9,11,13,15]i1p1f1) 7 (r[1,4,6,9,11,13,15] i1p1f1) EC-Earth3-Veg 10 (r[1-6,11-14]i1p1f1) 7 (r[1-4,6,12,14]i1p1f1) 6 (r[2-4,6,12,14]i1p1f1) 7 (r[1-4,6,12,14]i1p1f1) 3 (r[1-3]1ip1f1) 3 (r[1-3]1ip1f1) 3 (r[1-3]1ip1f1) 3 (r[1-3]1ip1f1) 6 (r[1-6]1ip1f1) 4 (r[1-4]1ip1f1) 4 (r[1-4]1ip1f1) 4 (r[1-4]1ip1f1) CanESM5-1 CanESM5-CanOE CESM2 3 (r[1-3]1ip1f1) CESM2-WACCM CNRM-CM6-1 10 (r[1-10]i1p1f2) CNRM-ESM2-1 E3SM-2-0 5 (r[1-5]1ip1f1) EC-Earth3-Veg-LR FGOALS-g3 3 (r[1-3]1ip1f1) GFDL-CM4 3 (r[1-3]1ip1f1) GFDL-ESM4 3 (r[1-3]1ip1f1) 3 (r[1-3]1ip1f1) 3 (r[1-3]1ip1f1) 1 (r1i1p1f1) 1 (r1i1p1f1) GISS-E2-1-G 10 (r[1-10]i1p1f3) (only to 2014) 38 (r[1-10]i1p1f1,r[1-10]i1p1f2, r[1-6, 8-10]i1p3f1, r[1-4,6-10]i1p5f1) 24 (r[1-10]i1p1f2, r[1-5] i1p3f1, r[1-4,6-10]i1p5f1) 14 (r[1-5]i1p1f2, r[1-5]i1p3f1, r[1-4]i1p5f1) 14 (r[1-5]i1p1f2, r[1-5] i1p3f1, r[1-4]i1p5f1) GISS-E2-2-G 11 (r[1-6]i1p1r[1-5]i1p3f1) 5 (r[1-5]i1p3f1) 5 (r[1-5]i1p3f1) 5 (r[1-5]i1p3f1) HadGEM3-GC31-LL 60 (r[1-60]i1p1f3) 55 (r[1-5,11-60]i1p1f3) 55 (r[1-5,11-60]i1p1f3) (r11-60 only to 2040) 1 (r1i1p1f3) 4 (r[1-4]1ip1f3) IPSL-CM6A-LR 10 (r[1-10]i1p1f1) 32 (r[1-32]i1p1f1) 11 (r[1-6,10,11,14,22,25]1ip1f1 6 (r[1-4,6,14]1ip1f1 6 (r[1-4,6,14]1ip1f1 3 (r[1-3]1ip1f1) 3 (r[1-3]1ip1f1) 3 (r[1-3]1ip1f1) 3 (r[1-3]1ip1f1) KACE-1-0-G MIROC6 50 (r[1-50]1ip1f1) 50 (r[1-50]1ip1f1) 50 (r[1-50]1ip1f1) 50 (r[1-50]1ip1f1) MIROC-ES2L 50 (r[1-50]1ip1f1) 30 (r[1-30]1ip1f2) 30 (r[1-30]1ip1f2) 10 (r[1-10]1ip1f2) 10 (r[1-10]1ip1f2) MPI-ESM1 (...truncated)