Rise and fall of sea ice production in the Arctic Ocean's ice factories

Arctic sea ice has experienced sustained declines in extent, thickness, and age over the satellite era, closely tied to anthropogenic CO2 emissions and Arctic amplification. Despite these losses, winter sea ice production has increased in recent decades, posing an apparent paradox given stronger wintertime warming. Multiple physical feedbacks can reconcile this: thinner and more mobile first-year ice promotes higher growth rates and divergence-driven open water formation; later freeze-up reduces insulating snow, enhancing conductive heat loss; and summer heat stored in the upper ocean is vented to the atmosphere during autumn and early winter. However, extreme warm winters have shown growth suppression or even winter melt, raising the question of when warming overwhelms these negative feedbacks. The study aims to provide a unified, process-based explanation for the observed rise and projected fall of winter sea ice production in the Kara and Laptev “ice factory” regions by developing and applying a simple linear model grounded in sea-ice thermodynamics and trained on internal variability from a large climate model ensemble.

Prior work identifies negative feedbacks that stabilize Arctic sea ice: inverse dependence of thermodynamic growth on thickness, enhanced growth following late-summer snow melt, and wind-driven divergence creating open water for rapid freezing. Observations and models have documented strengthened winter heat fluxes from ocean to atmosphere associated with thinner, seasonal ice and noted the limited compensation by increased outgoing longwave due to stable stratification (lapse-rate feedback). Event-based analyses (e.g., 2015/2016, 2016/2017 winters) show suppressed growth during extreme warming episodes. Modeling studies suggested a positive trend in winter growth ending in the early 2010s (CICE) and a mid-century transition from positive to negative temperature-growth correlations in CESM-LE, indicating warming eventually overwhelms negative feedbacks. The Kara and Laptev shelves are recognized for high ice production and significant roles in freshwater redistribution, halocline maintenance, and Arctic biogeochemistry, motivating a regional focus on these ice factories.

The study analyzes the Kara–Laptev shelf seas using 40 members of the CESM1.1 Large Ensemble (CESM-LE): a historical (1920–2005) 20th Century run and an RCP8.5 run (2006–2080). Winter is defined as October–April. Total winter thermodynamic ice production is calculated from ocean-to-ice freshwater fluxes (excluding melt). To interpret drivers, total production is decomposed into: (1) freezing area days (spatio-temporal extent of freezing; "usage") and (2) mean growth rate over freezing area ("efficiency"). Usage is primarily set by time to cool the upper ocean to freezing; efficiency follows the sea-ice growth balance between conductive flux through snow/ice and ocean-to-ice heat flux. A simple regression for freezing area days uses September sea surface temperature (SST, top 10 m) and Oct–Dec surface air temperature (SAT), achieving R² ≈ 0.74 (Sep SST only) and up to 0.78 (with SAT) in 20C, and 0.46→0.61 in RCP8.5. For total winter ice production, a multiple linear model is constructed based on internal variability (ensemble-member deviations from ensemble mean) with regressors reflecting thermodynamic controls: (1) ΔT/hs (inverse snow depth times temperature contrast), (2) ΔT × Asep (September open-water area), (3) ΔT × Anet (net winter divergence area), (4) ΔT × Acomp (compensated divergence area), and (5) September SST. Here ΔT = (freezing temperature of seawater − SAT), hs is mean winter snow depth over ice, Asep is September open-water area within the region, Anet is area-integrated net divergence, and Acomp is positive (divergent-only) area minus net divergence to represent divergence balanced by convergence. Ocean-to-ice heat flux is tested but excluded due to statistical insignificance in this region. Multiple linear regression is performed for 20C, RCP8.5, and the full 1920–2080 period, reporting standardized coefficients (using 20C standard deviations) and skill on internal variability: R² = 0.81 (20C), 0.76 (RCP8.5), 0.78 (full). Uncertainty in coefficients is assessed by regressions using each single ensemble member’s deviations (40 estimates per run), summarizing means and one-standard-deviation ranges. Forced changes are reconstructed by applying the internally derived coefficients to ensemble-mean regressor time series, with a constant offset to match mean production. Observation-based reconstructions apply the same linear model to datasets: ERA5 2 m SAT, NSIDC Polar Pathfinder ice motion (divergence), NASA GSFC passive-microwave sea ice concentration (for September ice area), SnowModel-LG snow depth, and SST from HadISST and NOAA OISSTv2. Variables are computed consistently with the model setup and area-weighted over the defined Kara–Laptev region.

- CESM-LE ensemble mean shows a rise-then-fall in total winter ice production in the Kara–Laptev seas: gentle increase from ~1970 to ~2010, peak around ~2020, followed by decline through 2080 under RCP8.5. - Decomposition: freezing area days ("usage") decline from the 1990s, accelerating after ~2050; mean growth rate ("efficiency") increases from ~1970 to ~2030, then declines from ~2040 onward. - Seasonal cycle shifts: October production collapses towards zero by the 2030s; peak production shifts later (to November), intensifies until the 2040s, then weakens to mid-20th-century levels by 2080. - Linear model skill on internal variability: R² = 0.81 (20C), 0.76 (RCP8.5), 0.78 (full). Coefficients are robust in sign and similar magnitude across periods; dominant standardized contributors to internal variability are ΔT×Anet, ΔT×Asep, and ΔT/hs. All term p-values < 1e-30; model p-values < 1e-300. - Forced trends decomposition: Up to ~2020, the ΔT×Asep term (larger September open water) is the largest positive contributor, despite warming SAT reducing ΔT. Once September sea ice area approaches zero in the region (~2020), Asep saturates and continued warming drives this contribution negative. Snow thinning (ΔT/hs) contributes positively until ~2030, after which warming dominates, reducing its contribution despite continued snow thinning. - The ΔT×Anet and ΔT×Acomp terms contribute modestly to forced trends due to offsetting increases in divergence by warming SAT until mid-century; later, warming dominates, turning their contributions negative. - The September SST term is the dominant driver of the forced decline: strong late-summer upper-ocean warming delays freeze-up, reducing freezing area days and total production. Ocean-to-ice heat flux remains small (ensemble mean ~0.5→~2 W m⁻² by 2080) and is statistically insignificant as a regressor in this region. - Observation-based reconstruction (1980s–2010s) indicates an increase in winter ice production by roughly 150–200 km³ from 1983 to 2000 (10-yr mean), then a slight decline from mid-2000s onward. Divergence-related terms show stronger increases than in CESM-LE, reflecting steeper observed trends in net and compensated divergence; September SST exerts a negative influence from ~2010; ΔT×Asep increase is partially offset by concurrent SAT warming (strong anticorrelation between Sep ice area and SAT). Overall, observations suggest the region is currently passing peak production.

The study resolves the apparent paradox of increasing winter sea ice production amid Arctic amplification by showing that multiple negative feedbacks—greater September open water, enhanced divergence, and reduced snow cover—initially intensify mid-winter thermodynamic growth even as the freeze season shortens. However, with continued greenhouse forcing, these feedbacks are overwhelmed by warming of both the atmosphere (reducing ΔT) and, critically, the upper ocean at the end of summer (raising September SST), which delays freeze onset and reduces freezing area days. The timing of peak production is primarily governed by when September sea ice area in the Kara–Laptev region reaches near-zero: once Asep saturates, its capacity to contribute additional production vanishes and warming ensures decline. Application to observations indicates a similar rise and recent peak, supporting the model-based interpretation. These findings are significant for Arctic climate dynamics and biogeochemistry: reduced Kara–Laptev production will alter the redistribution of sediments, pollutants, and nutrients by sea ice and may impact the stability of the Arctic halocline by modifying shelf brine rejection and the supply of cold, fresh surface waters to the Eurasian Basin. Regionally, as the central Arctic transitions to seasonal ice, interior basins may still experience increasing winter production for a time as their September open water and divergence increase, even while the Kara–Laptev share of Arctic production decreases.

A simple, physically informed linear model trained on internal variability in CESM-LE explains most of the variance in winter sea ice production and successfully reconstructs the forced rise-then-fall in the Kara–Laptev seas. The analysis attributes the historical rise (~1970–2010/2020) to negative feedbacks (increasing September open water, enhanced divergence, decreasing snow depth) and the subsequent decline to continued atmospheric warming and, especially, late-summer upper-ocean warming that shortens the freeze season. Applying the model to observation-based datasets suggests the region is presently passing peak production, implying a coming decline under continued warming. The approach is simple, extensible to other regions and models, and provides clear physical diagnostics of drivers. Future work should extend the framework spatially across the Arctic and Antarctic, incorporate additional variables where relevant (e.g., initial ice thickness, ocean-to-ice heat flux in regions with higher Fw), explore the timing and spatial patterns of divergence, and assess implications for halocline stability, ocean-ice-atmosphere coupling, and biogeochemical transports.

- Reliance on a single climate model framework (CESM1.1/CESM-LE) with known Arctic biases (e.g., Atlantic Water temperature, halocline salinity) may affect quantitative results and regional sensitivities. - Linear model simplicity neglects timing and spatial structure of divergence events, potential nonlinearities, and feedbacks beyond the chosen regressors; ocean-to-ice heat flux is insignificant in this region but may matter elsewhere. - Observation-based reconstructions inherit uncertainties from input products: ERA5 winter SAT warm biases over ice, passive microwave SIC retrieval errors, divergence estimates from ice motion vectors, and modeled snow depth (SnowModel-LG) rather than direct observations. - Sea-ice thickness observations are too short/uncertain to include as a robust regressor; direct satellite-based production estimates were not performed regionally due to thickness data limitations. - The forced future projection uses RCP8.5, assessed as low-likelihood in IPCC AR6, though useful for signal-to-noise; timing and magnitude of peak/decline may differ under other scenarios. - Regional focus (Kara–Laptev) limits generalizability; other regions may have different balances of atmospheric/oceanic heat fluxes and dynamics.