Introduction

Recognized by the IPCC Sixth Assessment Report1 as one of the most sensitive indicators of climate change, Arctic sea ice is a defining component of the high-latitude climate system, regulating exchanges of heat, moisture, and momentum between the ocean and atmosphere, and affecting Earth’s energy balance. Through its multiple feedbacks, it governs the magnitude of polar amplification2,3, second in importance only to temperature feedbacks3,4. Beyond being a player in climate change at high latitudes, Arctic sea ice decline has been linked to a weakening of the Atlantic Meridional Overturning Circulation5,6, while its associated shifts in atmospheric circulation can affect the frequency and persistence of extreme events7,8.

Arctic sea ice decline has been robustly linked to anthropogenic warming, with numerous attribution studies demonstrating a dominant contribution from rising greenhouse gas concentrations9 to the observed long-term loss of sea ice10,11,12,13. Superimposed on this forced response is a substantial contribution from internal variability14,15,16,17 via various mechanisms. The positive phase of the Atlantic Multidecadal Oscillation18 (AMO), associated with anomalously warm North Atlantic SSTs and enhanced poleward heat transport, has been linked to sea ice decline in regions such as the Barents–Kara Seas and Baffin Bay9,15. On shorter time scales, atmospheric variability in the North Atlantic modulates sea ice variability by altering storm tracks, surface winds, and the advection of cold and warm air masses, generating pronounced dipolar sea ice anomalies15,19 through dynamic and advective processes20,21,22. Atmospheric teleconnections from the El Niño–Southern Oscillation (ENSO) also influence Arctic conditions via planetary-wave propagation and modulation of the Aleutian Low23,24,25.

Although Arctic sea ice has declined in all seasons, the winter season (here defined as October–March) is of particular physical importance because Arctic amplification is strongest during the cold half of the year26,27,28,29. In winter, reduced sea ice weakens the insulating effect of the ice and enhances upward ocean-to-atmosphere heat fluxes, promoting strong lower tropospheric warming2,29. Sea ice changes also explain a substantial part of the historical winter Arctic surface warming trend, underscoring that winter sea ice is an active component of Arctic amplification rather than merely a passive indicator of climate change30. The climatic relevance of winter sea ice also extends beyond local thermodynamic feedbacks, because the atmospheric temperature and energy-budget response to sea ice loss is strongest in winter, even though the largest sea ice losses occur in summer and autumn31. At the same time, winter sea ice reflects the combined influence of several interacting factors, including atmospheric circulation, poleward heat and moisture transport, and ocean heat transport into the Arctic32,33, making it difficult to disentangle the distinct drivers. Furthermore, most previous observational studies of Arctic sea ice variability rely on statistical methods9,13,14,15,16,34 that do not necessarily imply causation, especially in a highly dynamic system such as the Arctic that is shaped by feedbacks, teleconnections, and co-varying influences. To address this hurdle, this study proposes a novel framework for separating the impact of anthropogenic influence from multidecadal and interannual internal variability on winter Arctic sea ice and associated 2-m air temperature (T2M) and sea level pressure (SLP) patterns by combining Canonical Correlation Analysis35 (CCA) with two complementary causality techniques, Convergent Cross Mapping36 (CCM) and Information Flow37 (IF).

Results

Anthropogenic footprints on Arctic coupled SIC-T2M-SLP fields

To link global climate change to changes in the Arctic, we investigate the co-variability between global SSTs and key Arctic climate fields using CCA, which identifies pairs of spatial patterns ordered based on the correlation between their associated time series (see Methods). The SST pattern of CCA1 (Fig. 1a) explains 40.4% of the variance in this field and displays an almost globally uniform warming structure consistent with long-term anthropogenic radiative forcing. Superimposed on this global signal are the well-known cold anomaly south of Greenland previously linked to anthropogenic influence38,39, as well as muted warming around Antarctica and localized cooling in the central North Pacific. The corresponding Arctic SIC pattern (35% of variance explained; Fig. 1c) shows a basin-wide reduction in winter SIC, with the strongest negative anomalies extending from the Greenland–Iceland–Norwegian Seas through the Barents–Kara, Laptev, Chukchi, and Beaufort sectors. This spatial structure is consistent with previous observational9,13 and modeling studies10,40 regarding the anthropogenic influence on Arctic SIC. The CCA1 T2M pattern (43%; Fig. 1d) is characterized by amplified warming over the exact same regions, while the associated SLP pattern (Fig. 1e) features a weak low-pressure anomaly over the central Arctic and weak positive anomalies over mid-latitudes, a configuration previously linked to anthropogenic warming41. The SST-SIC/T2M/SLP time series of CCA1 are significantly correlated (r = 0.97; Fig. 1b) and closely track the Global Warming Index (r = 0.94, 95% confidence level). Taken together, the quasi-uniform global SST warming structure, the basin-wide Arctic SIC decline accompanied by amplified T2M warming, and the close correspondence of the canonical time series with the Global Warming Index support the interpretation of CCA1 as the anthropogenically forced mode of Arctic winter variability.

Fig. 1: First coupled pair of global SST (ERSSTv5) and Arctic SIC, T2M, and SLP (ERA5) October–March anomalies from 1950 to 2024.
Fig. 1: First coupled pair of global SST (ERSSTv5) and Arctic SIC, T2M, and SLP (ERA5) October–March anomalies from 1950 to 2024.The alternative text for this image may have been generated using AI.
Full size image

The SST spatial structure of the first CCA pair (a), explaining 40.4% of variance, together with coupled Arctic SIC (c; 35% of variance), T2M (d; 43% of variance), and SLP (e; 2.7% of variance) fields north of 50° N for SIC and T2M, and north of 40° N for SLP. The correlation between their corresponding time series (b) is 0.97. Their correlation with Global Warming Index (red line) is 0.94 (95% confidence level).

Multidecadal footprints on Arctic coupled SIC-T2M-SLP fields

The SST structure of CCA2 (16.1%; Fig. 2a) is characterized by pronounced positive anomalies in the North Atlantic and negative anomalies over most of the south Atlantic and tropical Pacific. This global-scale pattern has been widely associated with the positive phase of the AMO in both observations42,43,44 and climate model simulations45. The North Pacific SST pattern, which resembles the negative phase of the Pacific Decadal Oscillation (PDO) and a La Niña-like cooling pattern in the tropical Pacific, is also consistent with previous studies showing that the positive phase of the AMO can induce coherent Pacific SST responses through large-scale atmospheric teleconnections, including modulation of the Aleutian Low and tropical atmospheric bridge mechanisms42,45,46. However, the physical origin of the AMO signal itself remains debated, with proposed mechanisms involving ocean circulation47,48,49 and external forcing such as aerosols50, solar variability51, or volcanic activity52. Here, we interpret CCA2 as capturing multidecadal internal variability regardless of its ultimate mechanistic origin. The Arctic SIC pattern of CCA2 (12.4%; Fig. 2c) shows reduced SIC across most of the Arctic, more pronounced over the Barents–Kara Seas and Baffin Bay. This structure is in agreement with previous studies regarding the influence of multidecadal Atlantic variability on Arctic sea ice, in which increased oceanic heat transport into the Nordic Seas, Barents–Kara Seas and Baffin Bay lead to diminished winter sea ice growth9,32,53,54,55. This pattern closely resembles the basin-wide SIC decline identified in the CO₂-dominated CCA1 mode, but with a weaker amplitude. A notable exception is the Fram Strait, where CCA2 displays a slight increase in SIC, indicative of sea ice export from the central Arctic that is associated with AMO56,57. The T2M pattern of CCA2 (8.4%; Fig. 2d) is for most regions co-located with changes in sea ice and shows widespread winter warming with maxima over the Barents–Kara Seas and Baffin Bay, regions with the most intense negative SIC anomalies. The associated SLP structure (3.9%; Fig. 2e) features positive pressure anomalies over the subpolar North Atlantic and negative-pressure anomalies over the Nordic Seas, favoring sea ice export through Fram Strait56,57. The time series for CCA2 (Fig. 2b) exhibits variability on multidecadal timescales that aligns with AMO Index (r = 0.45, 95% confidence level). The SST structure, the Arctic SIC decline centered over the Barents–Kara Seas and Baffin Bay, the corresponding T2M warming pattern, and the multidecadal evolution of the CCA2 time series support the interpretation of this pair as an AMO-related mode of Arctic winter variability.

Fig. 2: Second coupled pair of global SST (ERSSTv5) and Arctic SIC, T2M, and SLP (ERA5) October–March anomalies from 1950 to 2024.
Fig. 2: Second coupled pair of global SST (ERSSTv5) and Arctic SIC, T2M, and SLP (ERA5) October–March anomalies from 1950 to 2024.The alternative text for this image may have been generated using AI.
Full size image

The SST spatial structure of the second CCA pair (a), explaining 16.1% of variance, together with coupled Arctic SIC (c; 12.4% of variance), T2M (d; 8.4% of variance), and SLP (e; 3.9% of variance) fields north of 50° N for SIC and T2M, and north of 40° N for SLP. The correlation between their corresponding time series (b) is 0.93. Their correlation with AMO Index (red line) is 0.45 (95% confidence level).

Interannual footprints on Arctic coupled SIC-T2M-SLP fields

The SST pattern of CCA3 (8.8%; Fig. 3a) shows negative SST anomalies over the subpolar gyre (Fig. 3a), while positive anomalies are observed in the western subtropical North Atlantic and near the eastern US coastline. This tripolar pattern has been previously linked to interannual atmospheric variability43,58,59,60. The Arctic SIC pattern of CCA3 (Fig. 3c) explains 7.7% of SIC variance and exhibits a dipole-like structure: reduced SIC across the Barents–Kara Seas, the Greenland and Norwegian Seas, accompanied by increased SIC over Baffin Bay and the Labrador Sea. Reduced SIC in the Barents–Kara region and along the Eurasian Arctic coast has been previously documented in connection with interannual variability60,61,62, while positive SIC anomalies in Baffin Bay and the Labrador Sea are consistent with earlier results linking Atlantic atmospheric variability to enhanced cold air advection and regional winter ice accumulation63,64.

Fig. 3: Third coupled pair of global SST (ERSSTv5) and Arctic SIC, T2M, and SLP (ERA5) October–March anomalies from 1950 to 2024.
Fig. 3: Third coupled pair of global SST (ERSSTv5) and Arctic SIC, T2M, and SLP (ERA5) October–March anomalies from 1950 to 2024.The alternative text for this image may have been generated using AI.
Full size image

The SST spatial structure of the third CCA pair (a), explaining 8.8% of variance, together with coupled Arctic SIC (c; 7.7% of variance), T2M (d; 5.0% of variance), and SLP (e; 3.0% of variance) fields north of 50° N for SIC and T2M, and north of 40° N for SLP. The correlation between their corresponding time series (b) is 0.91.

The T2M pattern of CCA3 (5.0% of variance explained; Fig. 3d) also exhibits a pronounced dipole across the Arctic domain. Strong positive anomalies extend over northern Eurasia and the Barents–Kara sector, while negative temperature anomalies dominate the western Arctic, including Greenland, the Canadian Arctic Archipelago, and the Beaufort–Chukchi region. Unlike the thermodynamic responses seen in CCA1 and CCA2, the strongest T2M loadings in CCA3 are not co-located with the regions of strongest SIC anomalies. Instead, this pattern reflects large-scale continental advection, with enhanced westerly flow transporting warmer air into Eurasia and cooler Arctic-origin air toward Baffin Bay and the Canadian Arctic Archipelago. The SLP field associated with CCA3 (Fig. 3e) displays strong positive anomalies extending over the mid-latitude North Atlantic and western Europe, while negative anomalies dominate the Arctic and subpolar regions. The North Atlantic SST structure, the SIC dipole with intense loadings over the Atlantic Arctic sector and Baffin Bay, and the corresponding T2M dipole suggest that this pair is related to Atlantic atmospheric variability. The time series of this pair (Fig. 3b) shows mainly interannual variability, but also a weak trend during some periods. To assess the temporal relationship between CCA3 and Atlantic atmospheric variability, we compared the CCA3 time series with the North Atlantic Oscillation59 (NAO), Arctic Oscillation65 (AO) and East Atlantic Pattern66 (EA) indices (Fig. S1). Over the full 1950–2024 period, the PC3 detrended correlations are strongest for NAO (r = 0.40), followed by AO (r = 0.32), while EA is weaker (r = 0.18). These relationships change across subperiods: AO has the strongest correlation with CCA3 during 1950–1975, EA becomes relatively more prominent during 1975–2000, and NAO becomes the strongest correlate during 2000–2024 (Fig. S1b). Therefore, CCA3 is not interpreted as a single-driver mode, but rather as a broader Atlantic atmospheric circulation mode.

With CCA4, we identify another coupled mode dominated by interannual variability. The SST pattern of CCA4 (9.7% of variance explained; Fig. 4a) is characterized by positive anomalies across the central and eastern equatorial Pacific together with negative anomalies in the western tropical Pacific and North Pacific subtropics, closely matching the canonical El Niño-Southern Oscillation67,68 (ENSO) pattern. The Arctic SIC pattern of CCA4 (4.0% of variance explained; Fig. 4c) shows strong negative anomalies around Ellesmere Island, the Canadian Arctic Archipelago, and Baffin Bay, extending into the Greenland Sea. This is consistent with findings that show El Niño conditions are linked to decreased SIC and increased T2M across the Canadian–Greenland sector24. In contrast, positive SIC anomalies appear over the North Pacific Arctic sector, matching the tendency for increased SIC in the Pacific sector due to changes in the Aleutian Low69,70. The SIC pattern of this pair also exhibits a dipole-like structure similar to that identified in the CCA3 SIC, with opposite loadings over Baffin Bay and the Barents–Kara Seas, which is consistent with previous work showing that ENSO can influence Barents-Kara Seas winter sea ice through an Atlantic pathway involving atmospheric circulation anomalies71. The T2M pattern of CCA4 (5.0%; Fig. 4d) exhibits a pronounced Pacific–North American–Arctic temperature dipole25, with strong positive anomalies over northern Canada, Ellesmere Island, and Hudson Bay, co-located with regions of reduced SIC. The SLP field of CCA4 (10.9%; Fig. 4e) displays a marked Arctic dipole structure, with a deepened Aleutian Low over the North Pacific and positive SLP anomalies centered over the Canadian Arctic and Greenland. This pattern is a well-established extratropical response to ENSO24,69,70. The time series of this pair (Fig. 4b) are significantly correlated with Niño3 Index (r = 0.54, 95% confidence level). The tropical Pacific SST structure of this pair, the Arctic SIC–T2M-SLP footprint consistent with previous studies, and the significant correspondence with the Niño3 Index support the interpretation of CCA4 as an ENSO-related mode of Arctic winter interannual variability.

Fig. 4: Fourth coupled pair of global SST (ERSSTv5) and Arctic SIC, T2M, and SLP (ERA5) October–March anomalies from 1950 to 2024.
Fig. 4: Fourth coupled pair of global SST (ERSSTv5) and Arctic SIC, T2M, and SLP (ERA5) October–March anomalies from 1950 to 2024.The alternative text for this image may have been generated using AI.
Full size image

The SST spatial structure of the fourth CCA pair (a), explaining 9.7% of variance, together with coupled Arctic SIC (c; 4.0% of variance), T2M (d; 5.0% of variance), and SLP (e; 10.9% of variance) fields north of 50° N for SIC and T2M, and north of 40° N for SLP. The correlation between their corresponding time series (b) is 0.89. Their correlation with the Niño3 Index (red line) is 0.54 (95% confidence level).

Causal links from anthropogenic and internal drivers to Arctic SIC

CCA does not provide any information about causality between different patterns and their attributed drivers. To move beyond correlation-based inferences and assess whether the large-scale drivers inferred from CCA exert a causal influence on Arctic SIC, we therefore analyze CCM together with IF, two methods that explicitly quantify causality between different pattern changes. To ensure consistency with the CCA framework, and increase the signal to noise ratio, all causal maps are computed using the Arctic winter SIC field reconstructed from the leading 8 EOFs. In order to support the attribution of the CCA pairs we first use CCM to test for a causal relationship from GW Index to CCA1 (Fig. 5a), AMO Index to CCA2 (Fig. 5b), and Niño3 Index to CCA4 (Fig. 5c). For all three indices, we find a significant causal signal, where cross-map skill increases monotonically with library size and stabilizes well above the 95% confidence level envelopes of both Ebisuzaki72 and bootstrap surrogates (see Methods for details). We find the strongest convergence for the GW-related CCA pair and for ENSO, followed by AMO. Because CCA3 is interpreted as a broader Atlantic atmospheric circulation mode, we compared the CCA3 time series with several interannual atmospheric indices, including NAO, AO, EA, and Niño3 (Fig. S2). These tests show detectable but weaker convergence for NAO (Fig. S2a) and AO (Fig. S2b), while EA (Fig. S2c) and Niño3 (Fig. S2d) remain within the surrogate envelopes and do not show robust convergence. Similar CCM analyses reveal significant causal links from the same indices to the CCA SST time series (Fig. S3). Time-delay CCM analysis supports the inferred direction of causality, with peak cross-map skill occurring at negative or near-zero lags (Fig. S4), consistent with a driver to response relationship rather than a reversed association.

Fig. 5: CCM evidence for causal links between major climate drivers and coupled SIC-T2M-SLP pairs derived from CCA.
Fig. 5: CCM evidence for causal links between major climate drivers and coupled SIC-T2M-SLP pairs derived from CCA.The alternative text for this image may have been generated using AI.
Full size image

Cross-map skill as a function of library size for Global Warming (GW) Index → CCA1 (a), AMO Index → CCA2 (b), and Niño3 Index → CCA4 (c). Red curves denote CCM cross-map skill, while cyan and blue shaded envelopes indicate the 95% confidence ranges estimated from Ebisuzaki phase-randomised surrogates and bootstrap resampling, respectively.

In a second, CCA-independent step, we apply CCM directly between each of the four indices and the Arctic winter SIC anomalies, thereby mapping regions where each driver exerts a robust causal influence on SIC variability. The GW Index shows widespread and spatially coherent cross-map skill along both the Atlantic and Pacific marginal ice zones (Fig. 6a), particularly over the Greenland Sea, Barents–Kara Seas, Chukchi Sea, and Beaufort Sea. The AMO Index shows a similar but more regional footprint (Fig. 6b), with the strongest and most consistent cross-map skill over the Greenland Sea and Baffin Bay. The difference between Fig. 6a and Fig. 6b, in line with CCA1 (Fig. 2a) and CCA2 (Fig. 2c) SIC patterns, supports the attribution and separation of anthropogenic warming (CCA1) from multidecadal Atlantic variability (CCA2). The IF map from the Global Warming Index to winter Arctic SIC field (Fig. 6c) also reveals strong and statistically significant directed information transfer across much of the Arctic marginal ice zone, confirming the spatially coherent and basin-scale imprint of the anthropogenic mode identified in CCA1. In contrast, the IF map for AMO (Fig. 6d) shows significant but spatially confined influence, primarily over the Greenland Sea and Baffin Bay. For Niño3 Index, the CCM map (Fig. S5b) shows significant influence over the North Pacific sector, the Canadian–Greenland sector, and Hudson Bay, a pattern consistent with the main SIC loadings of CCA4 but only partly reproduced by the IF map (Fig. S5d). Thus, while the causal maps are less spatially robust than those for GW and AMO, they remain physically consistent with the CCA4 SIC structure (Fig. 4c) and with the Niño3 → CCA4 CCM result (Fig. 5c). For transparency, we also retain the corresponding CCM and IF maps obtained from the SIC field without the 8-EOF reconstruction in Fig. S6.

Fig. 6: Spatial causal maps of major climate drivers on Arctic winter SIC.
Fig. 6: Spatial causal maps of major climate drivers on Arctic winter SIC.The alternative text for this image may have been generated using AI.
Full size image

a, b CCM cross-map skill from the Global Warming (GW) Index and AMO Index to Arctic winter sea-ice concentration (SIC), respectively. Colors denote CCM cross-map skill, and stippling marks grid points exceeding the 95% confidence level based on surrogate testing. c, d Corresponding IF maps showing directed information transfer from the GW Index and AMO Index to SIC. Colors indicate the magnitude of information flow (non-negative), and stippling denotes grid points significant at the 95% confidence level. SIC field was first reconstructed from the leading eight EOFs prior to the causal analysis.

Physical links from anthropogenic and internal drivers to Arctic SIC

To explicitly diagnose how Arctic SIC, T2M, SST, and atmospheric circulation interact within each CCA mode, we analyze composite differences (high minus low PC years; Oct–Mar, 1950–2024) of upward turbulent heat flux (sensible + latent), downward longwave radiation, and 850-hPa geopotential height (Z850) with winds (Fig. 7). For CCA1, the regions of strongest winter SIC decline, especially Baffin Bay, Greenland Sea, Barents–Kara Seas, and Chukchi Sea, are co-located with enhanced upward turbulent heat flux and increased downward longwave radiation (Fig. 7a, e), indicating strong local thermodynamic coupling. This coupling is consistent with previous work showing that winter Arctic amplification is tightly linked to SIC loss and associated ocean-to-atmosphere heat release, with surface fluxes and longwave feedbacks maximizing in the cold season rather than at the time of maximum summer SIC retreat29,31. Over Baffin Bay, the Z850 and wind anomalies (Fig. 7i) indicate cold air advection from the Arctic, which in isolation would favor increased SIC. However, SIC still declines there, consistent with the strong local surface-energy anomalies seen in the flux and radiation diagnostics. Thus, in CCA1, thermodynamic forcing appears to dominate the SIC response even where the circulation anomaly would otherwise favor increased sea ice.

Fig. 7: Thermodynamic and dynamical coupling diagnostics for CCA 1–4 (Oct–Mar, 1950–2024).
Fig. 7: Thermodynamic and dynamical coupling diagnostics for CCA 1–4 (Oct–Mar, 1950–2024).The alternative text for this image may have been generated using AI.
Full size image

a–d Upward turbulent heat flux composite difference (high minus low PC years; sign convention: ocean→ atmosphere positive). Stippling denotes grid points significant at the 95% confidence level. e–h Downward longwave radiation (STRD) composite difference (atmosphere→surface positive). Stippling denotes grid points significant at the 95% confidence level. i–l Z850 composite difference with 850-hPa wind anomalies. Stippling denotes grid points significant at the 95% confidence level; vectors are plotted south of 84°N for clarity.

For CCA2, increased upward turbulent heat flux and downward longwave radiation are confined mainly to the Barents–Kara Seas and Baffin Bay (Fig. 7b, f), consistent with the more regional and lower-amplitude SIC and T2M anomalies relative to CCA1. This suggests that thermodynamic coupling remains important in CCA2 but is spatially restricted. The composite difference map of upper-ocean heat content for CCA2 (Fig. S7) shows positive anomalies across much of the North Atlantic–Arctic SIC sector, including the Barents–Kara Seas and Baffin Bay, consistent with multidecadal oceanic influence contributing to the regional SIC decline in this pair53,55. By contrast, SIC anomalies in the Beaufort Sea and Canadian Arctic Archipelago sectors are not accompanied by comparably strong local surface-flux amplification. Instead, a broad positive Z850 center over the North Pacific (Fig. 7j) favors warm air advection into these regions, supporting regional SIC decline73. For CCA3, turbulent heat-flux anomalies are weak and spatially limited (Fig. 7c), while downward longwave radiation exhibits a dipole structure that mirrors the SIC dipole (Fig. 7g). The Z850 and wind anomalies (Fig. 7k) instead indicate strong horizontal temperature advection, with warm air transported into the Greenland and Barents–Kara Seas and cold Arctic air advected into Baffin Bay and the Labrador sector60,61,62. This circulation-driven forcing explains the opposing SIC and T2M anomalies without the strong local thermodynamic amplification characteristic of CCA1. In this sense, the CCA1–CCA3 contrast illustrates that similar circulation patterns can produce fundamentally different Arctic SIC responses depending on whether local surface-energy feedbacks or dynamical advection dominate the mode. For CCA4, turbulent heat-flux anomalies (Fig. 7d) show positive values over the North Pacific, contrasting with the positive SIC anomalies (Fig. 4c), while downward longwave-radiation anomalies are weak across most of the Arctic, including regions where pronounced SIC anomalies occur (Fig. 7h). However, the Z850 and 850-hPa wind anomalies (Fig. 7l) show significant Z850 anomalies over the North Pacific/Aleutian Low region and the opposite Arctic–North American sector. This circulation pattern generates warm air advection into the Canadian Arctic Archipelago and Hudson Bay, where SIC is reduced (Fig. 4c), and cold air advection toward the Chukchi/North Pacific sector, where SIC in CCA4 slightly increases. SIC anomalies in this pair are therefore interpreted primarily as a circulation-driven ENSO response24,69,70 rather than as a response to strong local surface-energy feedbacks.

Discussion

Using a coupled SST–Arctic SIC–T2M–SLP framework, we separate observed winter Arctic variability into a dominant global warming mode, a multidecadal Atlantic mode, and two modes linked primarily to interannual variability from Atlantic and Pacific realms. The attribution of each CCA mode is not based on a single diagnostic, but on convergent evidence from the global SST structure, the temporal behavior of the CCA time series in relation to the associated index and the consistency of these patterns with known physical mechanisms. CCA1 combines a quasi-uniform global SST warming pattern, basin-wide Arctic SIC decline with amplified T2M warming and a time series that closely follows the Global Warming Index. CCA2 is distinguished from CCA1 by its Atlantic-centered multidecadal SST structure and its regional Arctic footprint, with strongest SIC and T2M anomalies over the Barents–Kara Seas and Baffin Bay. In contrast, the third and fourth CCA pairs, which are governed primarily by large-scale circulation anomalies, generate opposing SIC and T2M responses across the Arctic. CCA-pair attribution is independently supported by the causal diagnostics. CCM analysis between each driver index and the corresponding CCA time series indicates robust convergence for all four modes, confirming that the temporal evolution of the CCA patterns is consistent with directional forcing from the corresponding drivers. The clearest causal footprint on winter Arctic SIC is found for the Global Warming Index, which shows widespread and spatially coherent influence along the Arctic marginal sea ice zone whereas the AMO-linked pair shows a more regional Atlantic-sector footprint, particularly over the Barents–Kara Seas and Baffin Bay. For ENSO, the causal signals are strong at the index to CCA time series level, but weaker and less spatially extensive. However, they remain physically consistent with the SIC structures of CCA4. The weaker and less spatially extensive ENSO signal in the SIC causal maps likely reflects the intermittent and circulation mediated nature of ENSO teleconnections into the Arctic.

Physical diagnostics provide the final line of support for this interpretation. In CCA1 and, to a lesser extent, CCA2, regions of winter SIC decline are co-located with enhanced upward turbulent heat flux and increased downward longwave radiation, indicating strong thermodynamic coupling between SIC loss and lower-tropospheric warming. In contrast, CCA3 and CCA4 are dominated more clearly by lower tropospheric circulation and horizontal temperature advection, with comparatively weak local surface flux anomalies. This contrast explains why similar large-scale circulation patterns can produce different SIC and T2M responses across modes, and why weak SLP variance in the leading modes should not be interpreted as evidence of weak physical coupling. Rather, it reflects the fact that the anthropogenic and AMO-related influence are expressed primarily through local surface-energy feedbacks, whereas the interannual modes operate more strongly through dynamical atmospheric forcing. Our results are broadly consistent with other studies of Arctic SIC variability based on similar statistical methods9,15,16,19,74, including Low-Frequency Component Analysis33, that also identified a dominant forced component together with distinct regional internal modes.

Regional trend quantification suggests where the leading coupled modes account for the observed Arctic winter SIC decline (Fig. S8). Prior to about 1980, regional SIC anomalies fluctuate substantially but did not exhibit a persistent, pan-Arctic declining tendency (Fig. S8a). After 1980, CCA1 and CCA2 together explain a large fraction of the observed 1980–2024 winter SIC trend across more regions, including about 81% in the Beaufort Sea–Canadian Arctic Archipelago, 77% in the Greenland Sea, 71% in the East Siberia–Laptev Seas, 63% in the Chukchi Sea, and 58% in the Baffin Bay–Labrador Sea sector (Fig. S8b). The main exception is the Barents–Kara Seas, where the observed decline is very large but CCA1 and CCA2 together capture only about 31% of the trend, suggesting that additional regional feedbacks or forcings amplify the SIC trend there beyond what is represented by the two leading CCA pairs. Over the Greenland Sea, the trend spatial pattern (Fig S8b) suggests a superposition of strong negative trend signals with localized weak positive trend, likely due to sea ice advection through Fram Strait associated with AMO variability, which can partially counteract thermodynamically driven losses from CCA1. The relative impact of the two CCA pairs is also different in each SIC sector (Fig. S8c), with CCA1 exceeding CCA2 in all regions except Baffin Bay–Labrador Sea, where the AMO-related contribution is slightly larger, indicating an especially strong influence of multidecadal Atlantic variability there.

A growing body of work has linked Arctic sea ice decline to intensified mid-latitude extreme events7,75,76,77,78, showing that large-scale reductions in SIC weaken the meridional temperature gradient, alter the jet stream, and promote the development of persistent summer blocking over Europe78,79,80 that favor clear-sky, subsident, and radiatively driven summer heatwaves81. However, in most of these studies the decline in Arctic SIC is treated as a single, monolithic forcing, irrespective of its origin and the nature of the causal link to external or internal drivers. As a downstream implication, we examined whether the winter coupled Arctic modes project onto subsequent European summer daily maximum temperature (TX) variability (Fig. S9). The anthropogenically induced SIC decline (CCA1) is associated with a broad warm-season TX signal over much of Europe (Fig. S9a), whereas the AMO-related mode (CCA2) exhibits a much weaker and confined response over Eastern Europe (Fig. S9d). Consistent with this contrast, causality maps indicate a clearer downstream influence for CCA1 (Fig. S9b, c) than for CCA2 (Fig. S9e, f). Both CCA3 and CCA4 are associated with SIC patterns characterized by alternating positive and negative anomalies across the Arctic and therefore lack a coherent and spatially integrated signal and they do not generate a consistent downstream response in summer European TX (Fig. S10). We do not interpret TX loadings over Russia, Turkey, and Northern Africa due to known data limitations over these regions82. We also emphasize that these results should be interpreted as the projection of the winter coupled SST–SIC–SLP Arctic state onto later warm season variability, rather than as evidence that SIC decline alone drives European summer extremes. A full mechanistic attribution of this lagged connection would require dedicated analyses of the intervening circulation pathways, such as atmospheric blocking, jet variability, and stratosphere–troposphere coupling28,78,83.

The separation of anthropogenic influence from interannual and multidecadal internal variability is essential for interpreting recent trends, understanding regional patterns of winter Arctic amplification and downstream influence on European extreme events. These results highlight the importance of viewing Arctic winter sea ice not as a single, uniformly declining field, but as the superposition of multiple physical processes operating on different spatial and temporal scales. This framework can in principle also be applied to output from climate simulations, thereby exploiting the ability of climate models to extend analysis beyond the relatively short observational record, and to trace changes in sea ice via statistical and causal relationships in distant past or future climate states. Upcoming generations of both highly efficient and high-resolution climate models will provide valuable opportunities to further investigate the mechanisms identified here, particularly as very high-resolution simulations become available through CMIP7 and related model intercomparison efforts.

Methods

Data

We use monthly Extended Reconstructed Sea Surface Temperature v584 (ERSSTv5; NOAA), a global, homogenized SST product suited for long-term variability and low-frequency climate analysis (https://psl.noaa.gov/data/gridded/data.noaa.ersst.v5.html). ERSSTv5 is based on in situ measurements and statistical reconstruction techniques and provides a stable record from the mid-twentieth century onward. Sea ice concentration data is obtained from the fifth generation European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis (ERA5), at 0.25° × 0.25° spatial resolution(https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels-monthly-means?tab=form). The ERA5R is the state-of-the-art reanalysis produced by the ECMWF and provides a variety of atmospheric and climate variables85. From ERA5, we also used sea level pressure, two-meter temperature (T2M), geopotential height (Z850) together with zonal and meridional wind at 850 mb, upward turbulent heat flux (sensible + latent) and downward longwave radiation. Daily maximum temperature (TX) is taken from the E-OBS82 observational dataset (version 29.0e), covering the 1950–2023 period at 0.1° × 0.1° resolution(https://cds.climate.copernicus.eu/datasets/insitu-gridded-observations-europe?tab=overview). The dataset is produced by the Copernicus Climate Change Service and is based on station measurements that are quality-controlled and interpolated onto a regular grid. All fields are first converted to monthly anomalies relative to the 1980–2010 climatology. To construct a single winter value for each year, we use October–March means, where the October–December months belong to the preceding calendar year. For example, the winter of 1950 is defined from October–December 1949 and January–March 1950. For TX, we use May–September (MJJAS) averages to characterize summer conditions.

Climate indices

The Global Warming (GW) Index is constructed from ERA5 data by averaging the global T2M anomalies over all grid points. The Atlantic Multidecadal Oscillation (AMO) index is obtained from NOAA (https://psl.noaa.gov/data/timeseries/AMO/) and is defined as the detrended SST anomaly averaged over the North Atlantic basin (0°–70° N). The Arctic Oscillation, North Atlantic Oscillation and East Atlantic Index are taken from the NOAA Climate Prediction Center (https://www.ncei.noaa.gov/access/monitoring/). The Niño3 Index used here is retrieved from NOAA (https://psl.noaa.gov/data/timeseries/month/Nino3/) and represents the monthly SST anomaly averaged over 5°S–5°N and 150°W–90°W. For all indices, we compute October–March means to match the winter definition used for SST, SIC, T2M and SLP.

Canonical Correlation Analysis (CCA)

CCA35 is a multivariate statistical technique used to identify coupled modes between global SST and the Arctic coupled SIC, T2M and SLP fields. CCA provides pairs of spatial patterns whose associated time series (canonical variates) exhibit maximum correlation, while successive pairs remain mutually orthogonal. CCA estimates two sets of canonical vectors, solved through the generalized eigenvalue problem of the cross-covariance matrices. Before applying CCA, each field is reduced through EOF analysis to avoid degeneracy of the covariance matrices. We retain the leading eight EOF modes explaining ~70% of variance in each field86. Their PC time series form the two input matrices (X for SST and Y for SIC–T2M–SLP). CCA then solves the generalized eigenvalue problem for the cross-covariance matrices of X and Y, yielding two sets of canonical vectors (one for each field). The canonical time series are obtained by projecting the reconstructed fields onto their corresponding canonical vectors. The resulting pairs represent distinct, maximally coupled modes of co-variability. For each canonical mode k, spatial loading patterns are obtained by regressing the standardized canonical time series \({{\rm{a}}}_{{\rm{k}}}\left({\rm{t}}\right)\) onto the original anomaly field \({\rm{F}}({\bf{x}},{\rm{t}})\). At each grid point, the local fraction of variance explained by mode kis computed as the squared Pearson correlation coefficient:

$${{\rm{R}}}_{{\rm{k}}}^{2}({\bf{x}})={{\rm{corr}}}^{2}\left({\rm{F}}({\bf{x}},{\rm{t}}),{{\rm{a}}}_{{\rm{k}}}({\rm{t}})\right)$$
(1)

The resulting pairs represent distinct, maximally coupled modes of co-variability. This quantity represents the proportion of temporal variance at grid point xthat is linearly associated with the canonical mode. The regional percentage of variance explained reported in the manuscript is then obtained by spatially averaging \({{\rm{R}}}_{{\rm{k}}}^{2}\left({\bf{x}}\right)\) over the analyzed domain. This measure quantifies the mean local variance explained by a given coupled mode and is computed separately for each variable. The canonical modes are ordered according to decreasing canonical correlation between the paired time series, not according to the variance explained in any individual field. CCA has been previously successfully used in climate research to identify links between large scale patterns in key climate variables, like atmospheric circulation, temperature, sea ice or clouds14,41,81,86,87 and can provide similar results to Low-Frequency Component Analysis (LFCA) analysis33, a method that maximizes low-frequency variance within a combined state vector. However, our CCA framework maximizes coupling between global SST and the Arctic SIC–T2M–SLP state and therefore the two approaches are complementary and need not yield an identical mode correspondence.

Convergent Cross Mapping (CCM)

CCM36 is a causality-detection method grounded in dynamical systems theory. CCM reconstructs the system’s state space using time-lagged embeddings of one variable and evaluates whether this reconstructed manifold can reliably estimate another variable. If X causes Y, then the reconstructed states of Y contain recoverable information about X. Cross-map skill is quantified as the correlation between predicted and observed values and must increase with library length (convergence) for the causal link to be distinguished from mere correlation. In this study, CCM is applied in two ways. First, we test causality between each climate index and the canonical time series associated with the CCA modes by cross-mapping the index from the corresponding PC series. A robust causal relationship is confirmed if the cross-map skill peaks at zero or negative lag, clearly identifying the direction of influence, the cross-map skill consistently increases and stabilizes as the library size increases (convergence criterion) and cross-map skill significantly exceeds the surrogate-based 95% confidence intervals. For this analysis, the optimal embedding dimension (E) and lag were determined using Time-Delay CCM88 (TDCCM). The selected parameters are E = 5 and lag = −1 for GW–PC1, E = 5 and lag = -4 for AMO–PC2, and E = 5 and lag = −2 for ENSO–PC4. All selected lags satisfy the requirement |lag | ≤ E − 1 and are used solely to maximize cross-map skill. Second, we perform CCM grid-point-wise on the winter SIC field: for each index, the index time series is cross-mapped from the SIC anomaly time series at every grid cell, yielding spatial maps of causal influence. The embedding lag is chosen to be 1 for all analyses. Statistical significance is assessed using surrogate time series built under two models: (1) Ebisuzaki72 phase shift, which randomizes the phases of the time series, but keeps the same frequency spectrum and (2) Bootstrap surrogates, which resample a contiguous randomly chosen 10-point block of the original effect time series.

Information flow

The Information Flow37 (IF) method estimates potential causal links between two time series by quantifying how much information is transferred from one to the other. The central quantity of interest is the information flow from X1 to X2, denoted \({T}_{1\to 2}\), which is interpreted as a measure of the influence of X1 on the evolution of X2. Conceptually, \({T}_{1\to 2}\), compares the total information in X2 with the portion attributable to X2’s own independent fluctuations; if \({T}_{1\to 2}=0\), then X1 provides no causal contribution to X2. When \({{T}}_{1\to 2}\ne 0\), and is statistically significant, its sign indicates the following: a positive\(\,{T}_{1\to 2}\), implies that X1 increases the uncertainty of X2 (i.e., makes X2 less stable), whereas a negative \({T}_{1\to 2}\), implies that X1 reduces the uncertainty of X2 (i.e., stabilizes X2). Under this framework, the observed variability in X2 can be viewed as arising from the contribution of information flow from X1, X2’s intrinsic dynamics, and stochastic noise. Dividing \({T}_{1\to 2}\), by the total variability of X2 yields the relative information flow \({{\rm{\tau }}}_{1\to 2}\) (in %). It has been shown to perform well in certain nonlinear settings and has been widely applied to infer causal structure from observational data across fields such as climate science and economics.

Here we are interested in the spatial pattern of causality from one variable, X1 to the field of a second variable X2. To this end, we calculate the \({T}_{1\to 2}\) (and \({\tau }_{1\to 2}\)) from the time series of X1 to the time series at each grid point in the field of X2 and thus obtain a causality map.

For both IF and CCM, the selected lags should be interpreted as statistical optima rather than physical propagation timescales. For CCM, optimal embedding dimension (E) is chosen as the Embedding which maximizes contemporaneous interactions and the lag is determined through time-delay CCM, the optimal lag is required to lie within the embedding window (lag ≤ E–1) and is chosen to maximize cross-map skill, but likewise it is not interpreted as a physical delay. Before applying CCM and IF, SIC field is reconstructed using the leading 8EOFs to increase the signal-to-noise ratio.

Regional trend decomposition and integrated pattern-impact metric. To quantify the contributions of individual CCA modes to regional winter (Oct–Mar) SIC change, we reconstruct the SIC anomaly field associated with each mode k as

$$\Delta {{\rm{SIC}}}_{k}(t,{\bf{x}})={{\rm{PC}}}_{k}(t){{\rm{regSIC}}}_{k}({\bf{x}}),$$
(2)

where \({{\rm{regSIC}}}_{k}\) is the SIC regression/canonical pattern (expressed in SIC units) and \({{\rm{PC}}}_{k}(t)\) is the corresponding standardized canonical time series. For each predefined region, we compute the area-weighted mean SIC time series from observations and from each reconstructed mode, and estimate linear trends over 1980–2024 (expressed as % per decade). Mode contributions to the regional trend are then given by the trends of the reconstructed regional series. The fraction of the observed trend captured by CCA1 + CCA2 is computed as \(| {T}_{1+2}| /| {T}_{{\rm{obs}}}|\) (reported as a percentage). In addition, we compute an integrated “pattern-impact” metric designed to summarize the regional leverage of each SIC canonical pattern on the 1980–2024 change. For each region and mode k, we calculate a regional structure strength

$${S}_{k}={\sum }_{{\bf{x}}\in r}w({\bf{x}})\left|{{\rm{regSIC}}}_{k}({\bf{x}})\right|,$$
(3)

using spherical/area weights \(w({\bf{x}})\). We then multiply this structure strength by the magnitude of the linear trend in \({{\rm{PC}}}_{k}(t)\) over 1980–2024 to obtain a trend-weighted impact \({I}_{k}={S}_{k}| {\rm{trend}}({{\rm{PC}}}_{k})|\). This metric retains the physical dimensions of a regional pattern leverage (up to a constant factor that is defined by PC normalization), and is used by us to compare the relative roles of CCA1 and CCA2 across regions.