Observation-based selection of climate models projects Arctic ice-free summers around 2035

The retreat of Arctic sea ice is a prominent consequence of global warming with significant climatic, ecological, and societal implications. Observations show that Arctic sea-ice area has declined by about 2 million km² (yearly average) over the last 40 years, with intensified summer losses, and central Arctic ice has thinned by 1.5–2 m since 1980, leading to an Arctic sea-ice volume decline of ~3000 km³ per decade since 1979. These losses are strongly tied to rising global temperatures and cumulative greenhouse gas emissions. While linear relationships between emissions and sea-ice changes have been used to estimate future sea-ice area, such extrapolations neglect key non-linearities and ocean–ice–atmosphere feedbacks that can cause substantial deviations from trends. Coupled global climate models (GCMs) provide a framework to include these interactions and project Arctic sea ice under different forcing scenarios, as organized within CMIP phases. CMIP6 projections under SSP scenarios inform AR6. This study introduces, to the authors’ knowledge, the first CMIP6 model selection based on a broad set of criteria targeting present-day Arctic sea-ice state and northward ocean heat transport (OHT), a major driver of recent sea-ice loss. By selecting models that best match observations, projected sea-ice loss is larger than the unweighted multi-model mean, advancing the timing of a nearly ice-free summer Arctic to as early as around 2035 (vs 2061 without selection).

Model simulations: The study analyzes CMIP6 models (historical 1850–2014 and future 2015–2100) under SSP1-2.6 (low emissions) and SSP5-8.5 (high emissions). Monthly sea-ice concentration and sea-ice volume per area (or thickness if volume per area unavailable) were used. Arctic sea-ice area is computed as sea-ice concentration times grid-cell area summed north of 40°N; volume is sea-ice volume per area (or thickness×concentration) times grid-cell area summed north of 40°N. Sea-ice area: 32 models for SSP1-2.6 and 33 for SSP5-8.5; total 166 ensemble members for each scenario. Sea-ice volume: 28 models for both SSPs; 155 (SSP1-2.6) and 154 (SSP5-8.5) members. Historical northward ocean heat transport (OHT) was available for 16 models. For each model, ensemble means over members were used for sea-ice area, volume, and OHT to represent the forced response. Observational/reanalysis references: Sea-ice area from OSI SAF (since 1979); sea-ice volume from PIOMAS (since 1979); basin-scale OHT estimates (Atlantic and Pacific) from Trenberth et al. (2000–2016) derived from TOA radiation (CERES), ERA-Interim atmospheric divergence, and ocean heat content (ORAS5); Atlantic OHT observations from RAPID-MOCHA at 26.5°N (2004–2018) and OSNAP around 57°N (2014, 2016). Selection criteria: Models closest to observations/reanalyses over 1979–2014 (sea ice) and 2000–2014 (OHT) were retained using the following diagnostics, combining March and September for sea-ice metrics and paired latitudes/basins for OHT. (1) Mean sea-ice area: 15 models closest to observed means (1979–2014). (2) Mean sea-ice volume: 15 closest. (3) Sea-ice area variability: 15 closest in detrended standard deviation (1979–2014). (4) Sea-ice volume variability: 15 closest. (5) Trend in sea-ice area: 15 closest to observed trends (1979–2014). (6) Trend in sea-ice volume: 15 closest. (7) Atlantic OHT: 8 models closest to observed mean OHT at 26°N and 57°N (2000–2014; OSNAP mean over 2014 and 2016). (8) Atlantic and Pacific OHT: 8 models closest to observed mean OHT at 70°N (Atlantic) and 60°N (Pacific) combined (2000–2014). Combined criteria intersect sea-ice mean state with OHT-based selections: (9) Atlantic OHT + mean sea-ice area: 6 models; (10) Atl/Pac OHT + mean sea-ice area: 4 models; (11) Atlantic OHT + mean sea-ice volume: 3 models; (12) Atl/Pac OHT + mean sea-ice volume: 5 models. (13) Minimum ensemble size: models with ≥5 members (10 models). Selection procedure: For criteria 1–8, rankings for the paired diagnostics (e.g., March and September means) were combined and the best-placed models were selected up to the target count. For each criterion, multi-model means (across selected models) of sea-ice area and volume were computed for both SSPs. Robustness: A bootstrap test randomly selected 10 models (without applying performance criteria), repeating 1000 times; averages and standard deviations from these realizations demonstrated that results from the targeted selections are not due to sampling alone. The analysis focused on the ensemble means per model to reduce internal variability influences and used the longest available comparison periods.

- Without model selection (SSP5-8.5; multi-model mean over 33 CMIP6 models, 166 members): March sea-ice area and volume decline by 45% and 78%, respectively, by 2096–2100 relative to 2015–2019; September area and volume decline by 90% and 98%, respectively. The first almost ice-free September (<1 million km²) occurs in 2061 for the multi-model mean, with large inter-model spread over the 21st century. - SSP1-2.6 (no selection): March area and volume decline by 8% and 28%, respectively, by century’s end; September area declines by 49% (no multi-model mean ice-free event in this century), and September volume declines by 69%. - With model selection, end-of-century sea-ice area and volume are generally lower than without selection for both SSPs. Strongest additional losses arise when selecting models best reproducing historical ocean heat transport (OHT), alone or combined with mean sea-ice metrics. Under SSP5-8.5 and OHT-based selections, March area falls below ~7 million km² and March volume below ~5000 km³ by 2096–2100; September sea ice disappears by century’s end. - First ice-free Arctic in September under SSP5-8.5: For five of six OHT-including criteria, the multi-model mean reaches ice-free conditions before 2040 (range 2032–2039), up to 29 years earlier than the no-selection mean (2061), implying ice-free conditions could occur around 2035. The ≥5 members selection yields 2043. Selections based solely on sea-ice metrics give 2047–2052, except sea-ice area variability (2066). - Model spread and outliers: Four models do not reach ice-free September conditions by 2100 in SSP5-8.5, delaying the no-selection multi-model mean ice-free date to 2061, although 64% of models cross the threshold before 2050. Removing the four non-ice-free models advances the multi-model mean ice-free date to 2048. Model selection excludes such outliers, narrowing projection spread. - Under SSP1-2.6: OHT-including selections and the ≥5 members criterion yield first ice-free September conditions sometime before 2100, but sea-ice area remains near the 1 million km² threshold thereafter. Relative to SSP5-8.5, first ice-free dates are delayed by 4–16 years for five of six OHT criteria, by 33 years for the ≥5 members criterion, and by 36 years for the remaining OHT criterion. Selections based only on sea-ice area/volume do not reach ice-free conditions by 2100. Despite 45% of models becoming ice-free before 2050, the no-selection multi-model mean does not become ice-free this century (12 models never reach ice-free). - Additional quantitative result: Selecting models with ≥5 members leads to stronger losses under SSP5-8.5, with March sea-ice area and volume reduced by about 60% and 87%, respectively, and no September sea ice by 2100. - Mechanistic implication: Selected models often have higher sensitivity to anthropogenic warming and better representation of OHT, a key driver of long-term Arctic sea-ice decline.

By filtering CMIP6 models using observation-based criteria that target the present-day Arctic sea-ice state and northward ocean heat transport, the study reduces projection uncertainty and removes outliers that bias the unweighted multi-model mean. The approach indicates that the unweighted ensemble likely underestimates future Arctic sea-ice loss. Ocean heat transport emerges as a pivotal constraint: selections based on Atlantic and combined Atlantic/Pacific OHT produce the earliest and most pronounced future sea-ice reductions, consistent with OHT’s role as a dominant long-term driver beyond decadal timescales. Consequently, the timing of an almost ice-free Arctic in summer advances substantially—often by more than two decades—relative to the no-selection mean, with plausible occurrence around 2035 under high emissions. The narrowing of spread, particularly by excluding models that retain substantial summer ice through 2100, provides a more internally consistent projection set aligned with observed mean state, variability, and trends. The findings underscore that ensemble-mean aggregation without performance screening can mask earlier emergence of critical thresholds (like ice-free summers), which has implications for climate risk assessments, adaptation planning, and attribution of Arctic amplification feedbacks. Incorporating performance-based selection complements scenario-based uncertainty analyses by highlighting the sensitivity of projections to model fidelity in key Arctic processes.

The study demonstrates that observation-based selection of CMIP6 climate models—using metrics of Arctic sea-ice mean state, variability, trends, and especially northward ocean heat transport—yields projections with stronger Arctic sea-ice losses than the unweighted multi-model mean. Under SSP5-8.5, this advances the projected timing of a nearly ice-free September Arctic by up to 29 years, potentially as early as around 2035, and indicates lower end-of-century sea-ice area and volume in both winter and summer. Results suggest that including all CMIP6 models likely underestimates future Arctic sea-ice loss. Future work should expand to process-based selection criteria that directly test whether models reproduce mechanisms linking OHT and other drivers to sea-ice changes, enabling further reduction of uncertainties and improved physical credibility of projections.

- Scenario, model, and internal variability uncertainties remain; while ensemble means per model reduce internal variability, they do not eliminate it. - The selection does not prove model correctness; good historical agreement may result from compensating errors, and selected models may have higher climate sensitivity, potentially biasing projections toward greater loss. - Observational constraints have limitations: PIOMAS is a reanalysis product; OHT observations are temporally limited (e.g., OSNAP years 2014 and 2016) and geographically sparse; basin-scale OHT estimates are derived indirectly and span 2000–2016. - Some diagnostics (e.g., variability) are influenced by atmospheric variability that can be realistic even in models with biased mean ice states, complicating selection. - The selection reduces the number of models, which may limit representation of model structural diversity; bootstrap tests mitigate but cannot fully resolve this. - Sea-ice metrics are aggregated north of 40°N and may obscure regional process differences relevant to future change.