Regional sensitivity patterns of Arctic Ocean acidification revealed with machine learning

The study addresses how Arctic Ocean acidification varies regionally and what physical drivers control these differences. Ocean acidification (OA) is a clear, already-emerged anthropogenic signal in most of the global ocean and is projected to intensify. The Arctic is especially vulnerable due to low buffering capacity, strong stratification, sea-ice melt freshwater inputs, and changes in circulation and nutrient supply. Prior regionalizations often rely on subjective variable choices and geographic boundaries. The authors propose an objective, water-mass-based, unsupervised machine learning approach to identify robust sub-regions of Arctic OA across models and emissions scenarios, evaluate their consistency, and diagnose the dominant processes (temperature, salinity, alkalinity) and the role of sea ice in shaping OA trajectories and timing toward corrosive conditions.

The paper synthesizes prior findings that OA is widespread and rapidly emerging, with multiple stressors (warming, deoxygenation, nutrient changes) projected by CMIP-class models. Studies have highlighted the Arctic’s strong vulnerability to acidification driven by sea-ice melt dilution, increased stratification, altered mixed-layer depth, and freshwater inputs from Pacific inflow and rivers, which reduce total alkalinity and carbonate saturation states. Observations show undersaturation events and trends in Beaufort Gyre surface waters. Model biases in Arctic stratification, mixing, and Atlantic water penetration complicate regional projections; sea-ice simulation skill varies across CMIP6. Emergent constraints link surface density to anthropogenic carbon uptake among CMIP5/6 models, offering ways to reduce projection uncertainty. The biological significance of carbonate saturation (ΩAr, ΩCa) thresholds is species-dependent, with risk increasing as Ω falls below 1. Machine learning has been successfully applied to ocean biogeochemical datasets (e.g., pCO2 reconstructions and global eco-provinces), motivating an objective clustering framework for Arctic OA.

Models and experiments: Two NOAA-GFDL coupled models were used: GFDL-ESM4.1 (0.5° ocean, COBALTv2 biogeochemistry) and GFDL-CM4 (0.25° ocean, BLINGv2). Atmosphere uses AM4 variants; both include SIS2 sea ice. Scenarios include historical, SSP2-4.5 (ssp245), SSP5-8.5 (ssp585); ESM4 also includes an emission-driven esm-historical and esm-ssp585 configuration with interactive CO2. Model outputs were remapped from native tripolar grids to 1°×1°. Predictors and preprocessing: Predictors were differences in sea surface temperature (SST), sea surface salinity (SSS), and surface pH averaged over 2041–2060 relative to 1850–1949 for grid points poleward of 45°N. Linear trends from 500-year preindustrial control runs were used to detrend historical and future simulations. A World Ocean Atlas land-sea mask was applied. An Isolation Forest was used to remove outliers and inliers to suppress noisy extremes and emphasize broad regional signals. Clustering workflow (adapted SAGE): The three predictors were reduced to two dimensions via t-SNE, with perplexity 500, applied successively four times to enhance separation while preserving large-scale structure. Density-based clustering (DBSCAN) was then applied (minPts=500; epsilon tuned to yield a manageable cluster count). Due to t-SNE stochasticity, the workflow was repeated 40 times per model; the modal cluster assignment was retained. Only clusters with >50% areal overlap between CM4 and ESM4 were deemed robust; grid points with <80% ensemble agreement were hatched as uncertain. Clustering was performed independently for each model and scenario. Carbon metrics and validation: Anthropogenic carbon (Cant) inventories (1850–2011) were computed as historical minus preindustrial control, with a 16.6 GtC adjustment for 1791–1850 emissions. Patterns were compared to the 1°×1° GLODAPv2 2002 climatology via pattern correlations poleward of 65°N. An acidification potential metric was defined as Δ[H+] (2081–2100 minus 1850–1949) normalized by cumulative air–sea CO2 uptake (1850–2100), reported only where both Δ[H+] and cumulative uptake are positive. For CM4, ΩAr was computed offline with Mocsy using monthly surface fields. Time series and trends: Cluster-mean time series were analyzed for surface pH, total alkalinity (TA), SST, SSS, air–sea CO2 flux, and ΩAr. Uncertainties used a Monte Carlo approach with 100 synthetic series derived from 20-year-window-filtered, detrended residuals. Monthly trends (2015–2050) for ΩAr, pH, SST, SSS were calculated to diagnose seasonal sensitivities. Sea-ice extent (September, 15% threshold) was used to relate cluster boundaries to present-day (1979–2014) and mid-century (2041–2060) conditions. Data sources: CMIP6/ESGF archives for model outputs; GLODAPv2 mapped product for Cant; model source code public; analysis code available upon request.

• Objective clustering consistently identified four Arctic surface water-mass OA regions across models and scenarios: Central Arctic Surface Waters (CASW), Arctic Surface Waters (ASW), North Atlantic Surface Waters (NASW), and Sub-Arctic Surface Waters (SASW).
• Cluster–sea-ice linkage: ASW boundaries roughly coincide with present-day September sea-ice extent (1979–2014). CASW boundaries align with projected mid-century September sea-ice extent (2041–2060), indicating regions retaining multi-year ice the longest are most shielded until mid-century but then experience rapid OA.
• Magnitude of pH decline by 2100: CASW and ASW exhibit larger declines (~0.6–0.8) than the pan-Arctic average; NASW and SASW decline more modestly (~0.5–0.6).
• Drivers: • CASW: Strong late-summer surface freshening (−0.3 to −0.4 psu decade−1; SST trends <0.2 °C decade−1), rapid reductions in SSS (−2 to −2.5 psu by mid-century) and TA (−0.1 to −0.2 mol m−3), diminishing buffer capacity; catches up to ASW in pH decline by late century. • ASW: Earlier pH declines in early 21st century, then slower; stronger SST warming (0.3–0.6 °C decade−1) with minimal SSS change (<0.02 psu decade−1) across the annual cycle. • NASW: Strong SST warming (4–6 °C under ssp585), near-constant SSS and slight TA increases (+0.05 mol m−2), enhancing buffering and delaying corrosive conditions. Largest mid-century air–sea CO2 flux increase (peaks ~1–2 kg m−2 yr−1) before declining as waters warm.
• Corrosive conditions (ΩAr < 1): Under ssp585/esm-ssp585, most clusters cross the corrosive threshold around mid-century; NASW lags by ~2–3 decades relative to pan-Arctic average. Transition from ~20% to ~100% cluster area being corrosive occurs abruptly over 1–2 decades in CASW and ASW.
• Seasonal sensitivity (2015–2050 monthly trends): CASW shows the strongest declines in ΩAr (−0.2 to −0.25 decade−1 in CM4 ssp585; −0.1 to −0.2 in other runs) and pH (−0.04 to −0.07 decade−1), peaking in late summer/early autumn. ASW and pan-Arctic trends are also strong, especially in early summer; NASW/SASW trends are weaker than pan-Arctic.
• Acidification potential metric: Very high efficiencies (>100 mol [H+] per µmol CO2) occur at Arctic margins, especially along the North American coastline (Beaufort Sea, Queen Elizabeth Islands), and in other global transition regions (tropics’ edges, ACC), underscoring the Arctic’s heightened vulnerability.
• Anthropogenic carbon inventories and validation: CM4 Cant 1850–2011 = 153.1 ± 1.6 GtC; ESM4 historical = 143.8 ± 0.2 GtC; ESM4 esm-historical = 178.8 ± 2.8 GtC. Compared to observed 134.9 ± 24.0 GtC and CMIP5 mean 136.6 ± 14.0 GtC (with 1791–1850 adjustment). Pattern correlations with GLODAPv2 (≥65°N): ESM4 historical r=0.78; ESM4 esm-hist r=0.79; CM4 r=0.87, reflecting better Atlantic water penetration and carbon transport in higher-resolution CM4.

The unsupervised, water-mass-based clustering robustly delineates Arctic OA regimes that are consistent across two models with different resolutions and biogeochemistry and across low/high emissions scenarios. The analysis shows that regional variability in OA is governed primarily by physical changes—freshwater balance (salinity) and total alkalinity—rather than by temperature alone. Sea ice emerges as a key differentiator: present-day sea-ice extent demarcates regions of rapid near-term OA (ASW), while areas projected to retain multi-year ice longer (CASW) undergo the most intense late-summer freshening and fastest progression to corrosive conditions mid-century once exposed. The high acidification potential and abrupt shift to ΩAr < 1 over just 1–2 decades pose significant risks to Arctic biota ill-suited to adapt on such timescales. The findings suggest that constraining sea-ice projections and reducing model biases in Arctic mixing and stratification can directly improve OA projections. The clustering framework, using only SST, SSS, and pH, captures key sea-ice-related information content and offers a scalable, objective approach to cross-model OA evaluation, with potential for emergent constraints linking sea ice and OA.

This study introduces an objective, unsupervised machine learning framework to identify four robust Arctic OA sub-regions (CASW, ASW, NASW, SASW) from coupled climate model projections using only SST, SSS, and pH. The cluster boundaries align with present and projected sea-ice extents, revealing distinct drivers and timelines of acidification. Central Arctic regions experience the largest and most abrupt transitions to corrosive conditions, governed by freshening and alkalinity declines, while sub-Arctic inflow regions warm strongly but retain greater buffering and delay corrosivity. The approach highlights sea ice as a powerful observable to constrain OA projections and underscores the need to reduce biases in Arctic mixing/stratification. Future work should extend the clustering across broader CMIP ensembles and scenarios, develop sea-ice-based emergent constraints on OA, connect surface regimes to deep-ocean acidification pathways, and compare with long-term observations and large ensembles to refine timing and uncertainty of ecological impacts.

• Projections are based on a single model family (NOAA-GFDL CM4/ESM4) and limited scenarios, which may narrow the range of structural uncertainties; inclusion of more CMIP-class models and ensembles would broaden uncertainty bounds.
• Internal climate variability significantly affects the timing of the abrupt transition to corrosive conditions, making decadal timing uncertain and model-dependent.
• Uncertainties in future Arctic sea-ice projections propagate directly into OA regionalization and timing.
• The clustering relies on three surface predictors (SST, SSS, pH); other variables (circulation, nutrients, biology) may add nuance but were omitted to maintain generality and data availability.
• Method performance can be sensitive to model resolution, data density structures, treatment of outliers, and mean-state biases (e.g., sea ice), potentially affecting portability across models without methodological tuning.
• ΩAr timing estimates use annual means; seasonal extremes may occur earlier or vary regionally.