Ocean acidification (OA), a consequence of the absorption of anthropogenic carbon emissions by the ocean, is a critical concern highlighted by the IPCC. The Arctic is particularly susceptible due to its low buffering capacity and high stratification. While global OA patterns are well-documented, the Arctic's complex interplay of processes—sea ice melt, freshwater input, changes in stratification, and varying carbon dioxide (CO2) uptake—makes it challenging to model accurately. Model biases in upper ocean dynamics and the resolution-dependent representation of water masses can lead to inconsistencies in regional OA projections. This research uses a machine-learning approach to objectively identify and analyze sub-regions within the Arctic experiencing different OA sensitivities, aiming to transcend model biases and provide a robust framework for evaluating future OA projections across different models and emission scenarios. The high signal-to-noise ratio of OA makes it an ideal indicator of anthropogenic effects, which have already surpassed natural variability in most of the world's oceans. The potential negative impacts on marine ecosystems, particularly organisms relying on calcium carbonate for exoskeletons, emphasize the need for a comprehensive understanding of regional OA variations in the Arctic.

Existing research demonstrates the vulnerability of the Arctic to OA. Studies have highlighted the role of melting sea ice in increasing the ocean's surface area exposed to the atmosphere and in reducing sea surface salinity, thus accelerating pH decline. Freshwater from ice melt and rivers contributes to reductions in total alkalinity, further reducing the ocean's buffering capacity. However, inconsistencies exist in model projections due to variations in how models represent processes such as ocean mixing, stratification, and the transport of carbon-rich Atlantic waters. The impact of model resolution on the simulation of these processes and, consequently, on the accuracy of OA projections, is significant. Previous studies have attempted to divide the Arctic into sub-regions based on surface properties, but these approaches often rely on subjective selection of variables. The authors posit that machine learning offers a more objective approach to identifying distinct regions based on their unique OA responses.

This study employs an adapted version of the Systematic AGgregated Ecoprovince (SAGE) methodology, an unsupervised machine learning technique, to objectively classify Arctic regions based on their OA responses. Data from two state-of-the-art coupled climate models from NOAA's Geophysical Fluid Dynamics Laboratory (GFDL-ESM4.1 and GFDL-CM4) were used. These models differ in horizontal resolution and biogeochemical model complexity, offering a valuable comparison. Simulations included historical runs, a lower emission scenario (SSP2-4.5), and a higher emission scenario (SSP5-8.5). The analysis focused on 20-year averages (2041-2060) of sea surface temperature (SST), sea surface salinity (SSS), and pH anomalies relative to a centennial average (1850-1949). Model data were remapped to a 1° x 1° grid. The SAGE method clusters data points in a multi-dimensional space based on their similarity, thereby identifying distinct regions based on shared patterns in SST, SSS, and pH. Anthropogenic carbon (Cant) uptake was calculated and compared to observational data (GLODAPv2). An 'acidification potential' metric was defined to quantify the efficiency of CO2 uptake in driving OA at each location. The clustering workflow was performed independently for each model and scenario, focusing on clusters with >50% areal agreement between CM4 and ESM4 to ensure robustness.

The study identified four distinct Arctic sub-regions based on their responses to future climate forcing: Central Arctic Surface Water (CASW), Arctic Surface Water (ASW), Sub-Arctic Surface Water (SASW), and North Atlantic Surface Water (NASW). These regions were consistent across both models and emission scenarios. The CASW and ASW clusters exhibited larger pH declines than the pan-Arctic average, driven by reductions in sea surface salinity and total alkalinity. The boundary of the ASW cluster roughly corresponds to the present-day September sea ice extent, while the CASW boundary aligns with the projected mid-century September sea ice extent. This indicates that the presence of sea ice plays a crucial role in determining the regional sensitivity to OA. The CASW shows the most rapid decline in salinity and alkalinity, while the ASW shows initially faster pH decline, indicating distinct trajectories despite similar final pH changes. The NASW cluster showed more modest pH declines due to reduced CO2 solubility in warmer waters and increased total alkalinity. Analysis of the acidification potential metric revealed that the Arctic, particularly the coastal regions of North America (Beaufort Sea and Queen Elizabeth Islands), displays significantly higher acidification efficiency than other global regions. Projections of aragonite saturation state (Ωₐ) showed a rapid transition to corrosive conditions (Ωₐ < 1) in the CASW and ASW clusters by mid-century under higher emission scenarios, while the NASW cluster remained above the threshold for longer. The strongest trends in Ωₐ were observed in the CASW cluster during late summer and early fall, linked to surface freshening driven by sea ice melt. The ASW and pan-Arctic average show stronger early-summer declines in Ωₐ, correlating with the seasonal loss of sea ice.

The consistent identification of distinct Arctic OA regions across models and scenarios underscores the robustness of the methodology and the importance of considering regional variations. The study highlights the critical role of sea ice in shaping OA patterns in the central Arctic. The observed correspondence between regional boundaries and sea ice extent provides a valuable constraint for model evaluation and projection refinement. The rapid transition to corrosive conditions in the central Arctic poses a significant threat to marine ecosystems, given that the timescale of the change may be shorter than the adaptation timescales of many organisms. The study also underscores the need for improvements in model representations of Arctic mixing, stratification, and sea ice dynamics to reduce uncertainty in future OA projections. Furthermore, the differing responses of the identified clusters highlight the limitations of using pan-Arctic averages to assess OA impacts, emphasizing the importance of regional-scale analysis.

This research provides a novel and objective approach to identifying and characterizing regional patterns of Arctic Ocean acidification using machine learning techniques. The study reveals distinct sub-regions with varying sensitivities to OA, influenced by sea ice dynamics, freshwater input, and alkalinity changes. The rapid transition to corrosive conditions in some areas underscores the urgent need for continued efforts to refine climate models and better understand the complex processes influencing Arctic OA. Future research should focus on further narrowing uncertainty in OA projections by incorporating more models, ensemble members, and additional climate change scenarios. Exploring the connection between surface OA changes and deep-water formation regions, as well as extending this methodology to other ocean regions, will further enhance our understanding of global OA impacts.

The study’s findings are based on two GFDL climate models. While this offers a comparison between models with differing resolutions and complexities, the results may not be fully representative of the entire CMIP6 ensemble. Further, the analysis focused on a specific set of variables (SST, SSS, and pH). Including additional variables might reveal additional insights and potentially refine the identified regions. Finally, the use of a single realization from each model limits the ability to fully quantify the impact of internal climate variability on the timing of transitions to corrosive conditions. Future work should expand this analysis to the wider CMIP6 ensemble to assess the robustness and generalizability of the findings.