The Arctic has warmed nearly four times faster than the globe since 1979

Arctic amplification (AA) refers to the observation that the Arctic warms faster than the global average, a robust feature seen in observations, models, and paleoclimate proxies. Multiple mechanisms have been proposed, including sea-ice loss and associated ice–albedo feedbacks, cloud and temperature feedbacks, ocean heat transport, and atmospheric moisture transport, with possible contributions from aerosol changes. Despite agreement on the existence of AA, recent literature quotes a wide range for its magnitude (about twice, more than twice, or three times global warming), often based on older estimates. Given sea-ice loss as a key driver and known discrepancies between observed and modeled sea-ice sensitivity to global warming, the study asks: How large is AA over the satellite era using up-to-date datasets, and do current climate models reproduce the observed magnitude? The purpose is to quantify recent AA robustly and compare with CMIP5/CMIP6 and large-ensemble simulations to assess whether models capture the observed amplification and its variability.

The paper summarizes proposed drivers of AA: enhanced oceanic heating and ice–albedo feedback from diminishing sea ice; cloud feedbacks; near-surface inversion effects; ocean heat transport; and meridional moisture transport. Reductions in European aerosols may have contributed to recent Arctic warming, and future aerosol changes could affect AA. Previous work shows AA emerges rapidly in response to forcing via atmospheric feedbacks, with sea-ice feedbacks growing later. However, there is little consensus on the exact magnitude of recent AA; AMAP reported about threefold Arctic warming for 1971–2010. Studies have highlighted uncertainties in defining the Arctic domain (e.g., 60–75°N) and period length, and models (notably CMIP generations) often underestimate sensitivity of Arctic sea-ice loss to global warming or CO2 emissions. Internal variability has been argued to account for a substantial fraction (up to ~50%) of recent multi-decadal Arctic sea-ice decline, complicating attribution and model–observation comparisons. Comprehensive, updated evaluations of the AA ratio in observations versus modern model ensembles had been lacking.

Observations: Four datasets were used for near-surface temperature: GISTEMP v6, Berkeley Earth (BEST), HadCRUT5, and ERA5 reanalysis. Monthly mean 2 m temperatures were used; anomalies were computed relative to 1981–2010. ERA5 data extend back to 1950 using the preliminary extension, but pre-1979 uncertainties are noted. Observational validation against bias-adjusted GHCN-M stations north of 67.5°N with at least 39 years of data (1979–2021) indicated that the average of the four gridded datasets captures near-surface Arctic temperature trends reasonably well (median station–grid difference about −0.19 °C per decade). Arctic region: Primary Arctic domain is north of 66.5°N (the Arctic Circle). Sensitivity tests used alternative southern boundaries from 60°N to 75°N. Global means cover the full globe. Trends: Linear least-squares trends were computed for annual and monthly means over various windows, with a primary focus on the 43-year satellite era, 1979–2021. Arctic amplification (AA) was defined as the ratio of the Arctic trend to the global trend (AA = dT_A/dT_G). AA ratios where the global trend was not statistically significant (Mann–Kendall test) were rejected for modeled series. Seasonality: Monthly AA ratios were computed to assess seasonal dependence. Models: Four ensembles were analyzed: CMIP5, CMIP6 (one realization per model in each multi-model ensemble), and two single-model initial-condition large ensembles (SMILEs): MPI Grand Ensemble (MPI-GE) and CanESM5. Historical plus scenario simulations spanning roughly 1950–2040 were used (details in Supplementary). Comparison protocol: For each ensemble, all overlapping 43-year windows with start years ≥1970 and end years ≤2040 were extracted (e.g., 1970–2012, …, 1998–2040). For each window and realization, AA was computed. The likelihood of observing AA ≥ observed value (3.8 for 1979–2021) was estimated as the fraction of simulated 43-year AAs meeting or exceeding this value, weighting models equally in CMIP ensembles. An additional robust statistical test (details in Supplementary Methods) compared observed and simulated AA accounting for internal variability and model uncertainty. Spatial analysis: Local amplification ratios and spatial patterns of warming were mapped to identify regions of extreme AA (e.g., Barents–Kara–Novaya Zemlya sector). Sensitivity analyses examined dependence on Arctic boundary choice and trend length.

- Observed amplification (1979–2021): Arctic mean trend 0.73 °C per decade; global mean trend 0.19 °C per decade, yielding AA43 ≈ 3.8 (multi-dataset mean). Individual datasets give AA43 from about 3.7 (ERA5) to up to ~4.1 (BEST); tabled values list BEST 3.8, HadCRUT5 3.8, ERA5 3.7. - Spatial extremes: Large portions of the Arctic Ocean warmed at least 4× the global rate. The Eurasian Arctic, especially near Novaya Zemlya/Barents Sea, shows local ratios of ~6–7, linked to strong cold-season sea-ice loss and circulation changes. - Sensitivity to Arctic boundary and trend length: For southern boundaries between 60°N and 75°N and trend lengths ≥20 years, AA generally exceeds 3, increasing with higher latitude. - Seasonality: AA is strongest in late autumn (Oct–Dec; roughly fivefold) and weakest in summer (June–August; around twofold). April exhibits anomalously high observed AA. In every month, observed AA lies in the upper quartile of CMIP6 distributions and in the top 5% for several spring and summer months. - Model–observation comparison: Ensemble-mean AA is underestimated: CMIP5 ≈ 2.5 (−34% vs observed 3.8), CMIP6 ≈ 2.7 (−29%). The observed AA lies outside most of the CMIP5 spread and is rare in CMIP6. - Likelihoods from 43-year windows (1970–2040): Probability of AA ≥ 3.8 is p ≈ 0.006 in CMIP5 and p ≈ 0.028 in CMIP6. In SMILEs, MPI-GE shows p = 0.00 (no realization reaches observed AA), while CanESM5 shows p ≈ 0.054 (noting CanESM5’s high climate sensitivity and faster overall warming). - Additional statistical test: p-values indicate inconsistency between observed and simulated AA43 at the 5% level for CMIP5 (p = 0.00), CMIP6 (p = 0.027), and MPI-GE (p = 0.000); CanESM5 p = 0.091. - Dependence on period: When including 1950–2021, observations and models agree better (higher p-values), but pre-1979 observational uncertainties and non-linear temperature evolution reduce interpretability.

The study addresses whether modern models reproduce the magnitude of recent Arctic amplification. It finds that since 1979 the Arctic has warmed nearly four times faster than the globe, substantially exceeding the commonly cited “twice as fast” estimate. The observed AA is larger than ensemble-mean CMIP5/CMIP6 responses and is rare within their spreads, indicating either an extremely unlikely realization of internal variability or systematic model underestimation of AA. Seasonal diagnostics reinforce this underestimation, particularly during the melt season and in spring (e.g., April). Sensitivity analyses show AA remains >3 for reasonable Arctic boundaries and multi-decadal periods, though AA estimates depend on the chosen start year due to non-linear temperature evolution and the late-20th-century “hiatus” affecting global trends. Longer periods (e.g., 1950–2021) yield closer model–observation agreement, but such comparisons are hampered by pre-1979 observational uncertainties and changing forcing/variability regimes. Potential reasons for model underestimation include biases in sea-ice and snow-cover feedbacks, cloud processes, distribution of forced heating among atmosphere/cryosphere/ocean, and representation of internal variability (e.g., AMV). The finding that many models cannot simultaneously match observed Arctic and global warming rates suggests implications for projections of circulation, storm tracks, and regional climate linked to polar–midlatitude contrasts.

Using four modern observational datasets, the Arctic’s 1979–2021 warming rate is about 3.8 times the global mean, with regional maxima up to 6–7 in the Barents–Kara sector. This exceeds many prior literature claims and establishes a new benchmark for recent AA. State-of-the-art climate models (CMIP5/6) generally underestimate this amplification; the observed fourfold AA is extremely rare in simulations, even when internal variability is sampled across many realizations. While longer analysis periods reduce apparent discrepancies, they also increase observational uncertainty and can mask non-linear recent warming. The study highlights the need to improve model representations of key Arctic processes (sea-ice/snow feedbacks, cloud/radiation, ocean–ice–atmosphere coupling) and internal variability to better capture AA. Future research should refine observational constraints in the pre-satellite era, diagnose process-level model biases governing seasonal AA (especially spring melt), and evaluate whether improving polar processes yields models that reproduce both realistic Arctic and global warming concurrently.

- Observational uncertainty is larger before 1979; ERA5 shows a cold bias relative to other datasets in the pre-satellite period due to sparse assimilated observations. - Sparse in situ data over the ice-covered Arctic Ocean may introduce common biases across gridded datasets despite modern interpolation; regional inconsistencies (e.g., north of Greenland in early ERA5) exist. - AA estimates depend on the chosen start/end years; non-linear temperature evolution (e.g., late-1990s to mid-2010s global warming slowdown) affects the global trend denominator and thus AA. - Model comparisons use one realization per CMIP model in some analyses, potentially under-sampling extremes; however, large ensembles suggest internal variability alone cannot explain the discrepancy. - Scenario details and ensemble composition vary across models; while the focus period limits strong scenario divergence, residual differences may contribute small biases. - Local amplification and monthly diagnostics can be sensitive to sea-ice data/model biases, particularly in spring melt season.