Clouds Increasingly Influence Arctic Sea Surface Temperatures as CO2 Rises

As Arctic sea ice retreats during the melt season, the upper ocean warms in response to atmospheric heat fluxes. Overall, clouds reduce these fluxes in summer, but how the radiative impacts of clouds on ocean warming could change as sea ice declines has not been documented. In global climate model simulations with variable CO2, the timing of sea ice retreat strongly influences the amplitude of cloud‐induced summer cooling at the ocean surface. Under pre‐industrial CO2 concentrations, summer clouds have little direct effect on maximum annual sea surface temperatures (SST). When CO2 concentrations increase, sea ice retreats earlier, allowing more solar radiation to warm the ocean. Clouds can counteract this summer warming by reflecting solar radiation back to space. Consequently, clouds explain up to 13% more variability in maximum annual SST under modern‐day CO2 concentrations. Maximum annual SST are three times more sensitive to summer clouds when CO2 concentrations are four times pre‐industrial levels.

relationship is particularly important for interior Arctic seas where vertical air-sea fluxes have a larger influence on the surface energy budget than lateral advection (Carmack & Wassmann, 2006). Steele and Dickinson (2016) showed that the timing of when sea ice completely melts strongly influences annual SST maxima in the Pacific Basin of the Arctic. In observations, the dates of when sea ice begins and then completely melts are occurring earlier, on the order of 5-10 days earlier per decade (Bliss et al., 2019;Peng et al., 2018), and Arctic oceans are already absorbing more solar radiation now than in previous decades (Sledd & L'Ecuyer, 2021). These trends are predicted to continue in coming years (Crawford et al., 2021;Lebrun et al., 2019), which will further alter the surface energy balance as the ocean continues warming (Carton et al., 2015).
Despite the known importance of clouds for the Arctic energy budget, for example, Intrieri et al. (2002), Sedlar et al. (2011), andKay andL'Ecuyer (2013), the influence of clouds on upper ocean warming has yet to be fully explored. Clouds warm the surface by trapping and re-emitting longwave (LW) radiation and cool the surface by reflecting shortwave (SW) radiation that would otherwise be absorbed (Shupe & Intrieri, 2004). Since clouds influence the surface energy balance, and the surface energy balance impacts Arctic SST, it seems reasonable to conclude that clouds could impact SST. Previous works have focused on the influence of clouds on sea ice, for example, spring melt onset (Mortin et al., 2016) and the 2007 record minima (Kay et al., 2008), but the timing of when radiative anomalies have the largest impact on SST maxima likely differs from that for sea ice due the insulating effects of sea ice on the ocean (Perovich et al., 2007). In the few studies that have considered the influences of clouds on Arctic SST, clouds were not actually the focus, so the results were either speculative based on limited data (Minnett, 1999) or only briefly mentioned using reanalysis data (Carvalho & Wang, 2020), which have known biases for Arctic clouds, for example, Lindsay et al. (2014).
In this work we investigate the extent to which clouds influence Arctic SST annual maxima using new experiments in a state of the art coupled climate model. We first ask if clouds can affect ocean surface warming in the Arctic, and, if so, is it through warming or cooling effects? Second we ask, do the impacts of clouds on SST change with a warming climate? We use model experiments with variable CO 2 concentrations to answer these questions and quantify how the sensitivity of annual SST maxima to clouds could change with rising CO 2 .

Data
We use the Community Earth System Model version 2.1.3 (CESM2; Danabasoglu et al., 2020) for this work with the following components: Community Atmosphere Model version 6 (CAM6), Parallel Ocean Program (POP) version 2, Community Land Model (CLM) version 5.0, Los Alamos Sea Ice Model (CICE) version 5, OAA WaveWatch-III ocean surface wave prediction model (WW3), Community Ice Sheet Model (CISM) Version 2.1, and Model for Scale Adaptive River Transport (MOSART). All model simulations are fully coupled between components. The atmosphere grid has nominal resolution of 0.9 × 1.25°, and the ocean and ice grids have 1° resolution. Additional output was saved in each simulation from the Cloud Feedback Model Intercomparison Project Observational Simulator Package version 2.0 (COSP2; Swales et al., 2018) lidar simulator (Chepfer et al., 2008) with 250 sub-columns. This output allows scale-aware and definition-aware comparisons to monthly cloud fraction observations from CALIPSO GOCCP for June 2006 through March 2018, available up to 82°N (Chepfer et al., 2010).
We ran three simulations in CESM2 with variable CO 2 concentrations. The control run (hereafter called PI-control) uses pre-industrial forcing, nominally year 1850 with 284.7 ppm CO 2 concentration. Additional runs were branched from a separate pre-industrial run with 1% annually increasing CO 2 . Branches were started after years 40 and 140, corresponding to CO 2 levels of 424.0 and 1193.3 ppm, respectively, with CO 2 levels kept constant after branching. The former run represents approximately modern-day conditions (global average CO 2 concentrations were 417.2 ppm in 2022 (Friedlingstein et al., 2022)), called CO 2 -modern, and the latter is roughly four times the pre-industrial CO 2 concentration, called CO 2 -high. Each simulation was run for 100 years. The annual mean temperatures over the Arctic (>60°N) are 264, 265, and 275 K for the simulations, from lowest to highest CO 2 concentrations respectively. In PI-control and CO 2 -modern simulations there are small drifts in temperature (Figure S1a in Supporting Information S1). CO 2 -high does not reach equilibrium in 100 years, but for the purposes of this study the CO 2 -high Arctic is in a different state with respect to sea ice, clouds, and SST that still provides contrast to the lower CO 2 simulations. In particular, the Arctic is seasonally ice-free in CO 2 -high ( Figure S1b in Supporting Information S1).
While model biases exist, CESM2 is an effective tool for investigating the changing cloud influence on Arctic SST over a large range of Arctic climates. Overall CESM2 reasonably simulates the coupled Arctic Ocean system, for example, DeRepentigny et al. (2020), DuVivier et al. (2020), andMcIlhattan et al. (2020). Summer sea ice retreat is excessive when compared to modern satellite observations, but this increased summer retreat does not change the fundamental relationships identified in this work. Annual cycles of total cloud fraction (CF tot ) from COSP2 in PI-control and CO 2 -modern simulations broadly agree with the seasonal cycle from CALIPSO GOCCP observations (Figure 1b).

Season Definitions
To study the impacts of clouds on annual maximum SST, we use several descriptive dates and time periods as defined in Steele and Dickinson (2016). The key dates for sea ice melting and upper ocean warming are determined using daily time series of sea ice concentration (SIC) and SST for individual grid boxes with a 15-day running mean boxcar smoother applied. In a given year, the day of melt onset for a grid box is defined as the last day SIC is greater than or equal to 0.95. This date is the start of the melt season. The melt season ends on the last day SIC is at or above 0.15, termed the day of heat onset. The heat season lasts from the day of heat onset until the day of maximum SST (SST max ). Defining these seasons allows for more physical comparisons than monthly lag-correlations that are often used. Additional variables, for example, total cloud fraction (CF tot ) and cloud liquid water path (LWP), are averaged over each season to correlate with SST max . Time series are linearly de-trended prior to statistics being calculated. In addition to melt and heat seasons, we also reference meteorological summer, defined as June, July, and August.

Results
In PI-control, Arctic sea ice persists through most of summer ( Figure 2). The mean date of melt onset occurs in April at latitudes below 70°N (exterior Arctic) and in June poleward of 70°N (interior Arctic) (Figure 2a). In the interior Arctic, heat onset typically occurs in August or September if at all (Figure 2d). While the heat season can last 3-4 months in the exterior Arctic, it lasts only ∼2 weeks on average in the interior Arctic (Figure 2g). Considering all grid cells that experience heat onset leads to an Arctic-wide average of 43 days for the heat season length.
With higher CO 2 levels, sea ice melts earlier and the ocean warms for longer. In CO 2 -modern, heat onset occurs at the beginning of August in the interior Arctic, on average ( Figure 2e). In this simulation, over 50% more grid cells have sea ice that completely melts, typically further north and later in the year, than in PI-control. Because more grid cells become ice free later in the year further north (Figure 2e vs. Figure 2d, Figure 1f), the date of 10.1029/2023GL102850 5 of 10 heat onset averaged over the entire Arctic is actually later in the year compared to PI-control (means given in Figures 2d and 2e). However, for grid cells that experience heat onset in both simulations, melt and heat onsets typically occur earlier in CO 2 -modern than in PI-control (Δs in Figures 2b and 2e). The length of the heat season is also longer in CO 2 -modern than in PI-control when comparing only grid cells that experience heat seasons in both simulations. The changes in heat season length are greatest around the interior Arctic coast where locally differences are on the order of a few weeks (Figure 2h vs. Figure 2g). The largest changes for these dates occur when CO 2 is quadrupled. In CO 2 -high, all sea ice completely melts in summer, on average by June 4 (Figure 2f). Because much of the ocean is ice-free earlier in the year, the heat season can last up to several months even toward the North Pole in CO 2 -high (Figure 2i), as compared to just a few weeks in PI-control (Figure 2g). The interannual variability of SST max also increases by approximately a factor of two from PI-control to CO 2 -high (not shown).
The timing of when a given area of ocean begins warming is important because the radiative effects of clouds vary throughout the year: clouds cool the surface in summer but warm the surface the rest of the year. The cloud radiative effect (CRE) is defined as the difference between all-sky and clear-sky fluxes: where fluxes are positive down. In PI-control and CO 2 -modern clouds cool the surface (net CRE is negative) for fewer than 3 months during summer (Figure 1d). In CO 2 -modern SW CRE is stronger by almost 20 Wm −2 in July compared to PI-control because there is less sea ice in CO 2 -modern; clouds are more reflective compared to open ocean than compared to sea ice, for example, Alkama et al. (2020). During summer the warming effect of clouds is slightly reduced in CO 2 -modern compared to PI-control ( Figure 1c) despite similar cloud cover in these two runs (Figures 1a and 1b). Because the atmosphere is warmer with higher CO 2 , clear-sky LW increases and the relative warming impact of the clouds decreases, for example, Sedlar and Devasthale (2012). Over September through December LW CRE is a few Wm −2 greater in CO 2 -modern than in PI-control. Larger LW CRE is consistent with greater CF tot and LWP compared to PI-control over the same months (Figures 1a and 1b). This increased cloud cover is also in agreement with observed fall cloud-sea ice feedbacks (Morrison et al., 2019).
These changes in CRE are amplified in CO 2 -high. During June through September LW CRE is about 12 Wm −2 less positive than PI-control (Figure 1c), consistent with a much warmer atmosphere and greater clear-sky downwelling LW compared to lower CO 2 experiments. SW CRE is up to 60 Wm −2 more negative during the same months as sea ice declines (Figure 1e). SW CRE also becomes stronger before summer in CO 2 -high because LWP and CF tot increase in March through June compared to PI-control (Figures 1a and 1b). Greater cloud cover allows more SW to be reflected by clouds and prevented from being absorbed at the surface. Additionally, increased LWP and CF tot in CO 2 -high increase LW CRE from October to June by upwards of 20 Wm −2 . The combined impact is that net CRE is negative from May to September, meaning clouds have a larger cooling effect for longer at the surface in CO 2 -high than in the lower CO 2 experiments (Figure 1d).
The earlier an ocean region begins warming, the greater the cooling effect of clouds during the heat season. As a result, in PI-control and CO 2 -modern clouds cool the ocean when it is free of sea ice along the interior Arctic coasts and further south (Figure 3). In contrast, poleward from the coasts average net CRE during the heat season is positive and clouds warm the ocean. In PI-control (CO 2 -modern), regions where heat onset occurs before July (Figures 2d and 2e) have more negative average net CRE, −67 (−73) Wm −2 , compared to grid cells that melt during or after July, −13 (−4) Wm −2 (Figures 3a and 3b). In CO 2 -modern, 24% more grid cells experience negative net CRE during the heat season, on average, than in PI-control. This increase suggests that clouds may already be cooling the surface more in our current climate than in the pre-industrial era. Meanwhile in CO 2 -high, the average heat onset occurs before July in all grid cells (Figure 2f), and the average heat season net CRE is strongly negative (Figure 3c). Not only do clouds have a net cooling effect over entire Arctic Ocean during the heat season, but the strength of the cooling effect is also greater. In CO 2 -high the average net CRE is −91 Wm −2 during the heat season, a 34% increase compared to the lower CO 2 runs.
Despite clouds influencing surface radiation under all CO 2 concentrations, they have limited influence on SST max during the heat season in PI-control and CO 2 -modern. In these simulations, SST max is instead strongly controlled by the date of heat onset (Figures 4a and 4b). Negative correlations mean earlier heat onset leads to warmer SST max . This relationship is consistent with the findings of Steele and Dickinson (2016) that earlier heat onset allows more energy to be absorbed by the ocean. Averaged over the full Arctic, heat onset explains 60% (69%) of SST max variability in PI-control (CO 2 -modern). Around the interior Arctic coast, average cloud LWP and CF tot during the heat season (LWP heat , CF tot, heat ) and SST max have weak negative correlations in PI-control (Figures 4j and 4m), suggesting clouds have a net cooling effect on annual maximum SST, consistent with Figure 3a. Farther poleward, clouds have fewer statistically significant correlations with SST max (stippling based on t-test with 95% confidence). The extensive sea ice in these simulations shields the ocean from the atmosphere and clouds since most incoming energy goes toward melting the sea ice.
In PI-control and CO 2 -modern simulations, clouds primarily impact SST max during the melt season. The melt season occurs in summer when CRE is most strongly negative. Average CF tot and cloud LWP over the melt season (CF tot,melt , LWP melt ) negatively correlate with SST max (∼−0.5; Figures 4d, 4e, 4g, and 4h). Overall, clouds during the melt season (typically June through August) cool the surface so that it takes longer for sea ice to retreat. In turn, the longer sea ice persists in a given region, the later heat onset occurs, if at all. Because the date of heat onset is so strongly correlated to SST max its delay leads to less warming.
Even increasing CO 2 from pre-industrial to present day levels leads to clouds having greater impacts on SST max . In PI-control, SST max is relatively insensitive to LWP and CF tot during the heat season (Figures 3d and 3g). In CO 2 -modern, SST max is sensitive to clouds during the heat season outside of the interior Arctic basin on the order of −15°C/gm −2 or −0.1°C/%, notably north of the Atlantic Ocean, the Barents and Bering Seas, as well as the Canadian coast (Figures 3e and 3h). Clouds in these regions also explain more SST max variability during the heat  season in CO 2 -modern than in PI-control. Correlations between LWP heat and SST max grow stronger, that is, more negative, most often in areas that consistently become ice free by the end of summer compared to PI-control, such as the Barents and Bering Seas (Figure 4k). Averaged over grid cells with mean heat onset dates in June or earlier (pink and purple areas in Figure 2e), LWP heat explains 13% more variability in SST max than in PI-control ( Figure  S2 in Supporting Information S1).
Notably, as CO 2 increases to levels well beyond present-day values, fundamental changes occur in the relationships between clouds, heat onset, and SST max . In CO 2 -high, clouds are more important to SST max than the timing of heat onset (Figures 4l and 4o vs. Figure 4c), a significant shift from PI-control and CO 2 -modern simulations. LWP heat explains 35% more variability in SST max while the date of heat onset explains 36% less variability than in PI-control ( Figure S2 in Supporting Information S1). Nearly all ocean grid cells in CO 2 -high completely melt before the end of June, near the summer solstice, which reduces the previously large variability of how much downwelling SW could be absorbed by the ocean based on the date of heat onset. On average, most grid cells begin warming at least a month earlier than in PI-control, closer to the time when net surface flux peaks and SW CRE is at its most negative (Figure 1e). As a result, the ocean is warming when clouds are most strongly cooling the surface.
When CO 2 is quadrupled from PI-control, the sensitivity of SST max during melt and heat seasons changes in both magnitude and spatial pattern. SST max in CO 2 -high becomes three times more sensitive to clouds during the heat season averaged over the Arctic than in PI-control (−0.06°C/% compared to −0.02°C/%). This increased sensitivity of SST max to clouds during the heat season corresponds with decreased sensitivity to clouds during the melt season ( Figure S3 in Supporting Information S1). With lower CO 2 concentrations, the sensitivity of SST max to clouds is largely dictated by latitude and the timing of sea ice melt; areas that melt earlier in the summer are more sensitive to clouds. In CO 2 -high, SST max is most sensitive to clouds north of Siberia and the Canadian Arctic Archipelago, upwards of −50°C/gm −2 or −0.2°C/% locally (Figures 3f and 3i). This pattern is likely because ocean surface warming on these "interior shelves" (Carmack & Wassmann, 2006) is mostly influenced by atmospheric surface heat fluxes, as opposed to northward-flowing warm currents from the North Atlantic and Pacific Oceans (e.g., Polyakov et al., 2020;Steele et al., 2010).

Discussion and Summary
A novel contribution of this work is quantifying the seasonal impacts of clouds on Arctic SST under increased CO 2 concentrations. We find that under PI-forcing, clouds have little impact on SST max , regardless of season, because sea ice persists through summer and the underlying ocean is largely shielded from cloud radiative effects. Instead of clouds, the timing of when the ocean begins warming largely controls SST max , explaining 60% of its variability. This result is consistent with observations from Steele and Dickinson (2016) that the date of sea ice retreat explains the majority of SST max variability.
The most important finding of this study is that as sea ice melts earlier with higher CO 2 levels, clouds increasingly influence SST max . Even with a relatively moderate increase in CO 2 to present day levels, clouds are more important to SST max than during the pre-industrial climate. Clouds influence SST max during the melt season by cooling the surface and delaying the start of upper ocean warming. Where sea ice retreats earlier, clouds also explain more SST max variability during the heat season because the ocean warms when cloud radiative effects are most negative. In lower latitude regions that begin warming before July, cloud LWP explains 13% more SST max variability. When CO 2 is quadrupled, the sensitivity of SST max to summer clouds increases by a factor of three, averaged over the whole Arctic. Overall, clouds explain 35% more SST max variability during the heat season than in the pre-industrial climate over the full Arctic.
In future work, observations should be used to test if the relationships between clouds and SST are currently changing in the Arctic. While reliable satellite observations of LWP are not currently available across the Arctic, cloud fraction exhibits many of the same relationships to radiative fluxes and SST max as LWP, and it is available. While we might expect clouds to have similar relationships to maximum annual SST as in the simulation with modern-day CO 2 concentrations, future work could assess to what extent this relationship exists outside of a climate model. In particular, the Arctic may be experiencing the transition of clouds influencing SST max , as in the CO 2 -modern run, in some Arctic basins. Based on this, work we expect clouds to grow in their importance to SST as Arctic sea ice continues to decline.