CO₂ flux was calculated from the 10 Hz data as FCO2=ρa‾w′c′‾, where ρa‾ (mol m⁻³) is the mean dry air density, w (m s⁻¹) is the vertical wind speed, c is the CO₂ mixing ratio (µmolmol-1), primes indicate fluctuations about the mean, and the overbar corresponds to the time average (20 min for this study). As the product of measurements from different instruments, the accuracy of the FCO2 measurement is challenging to quantify without an independent validation, which was not performed. The LI-7200 has a measurement accuracy of ±1 % with an RMS noise of 0.11 ppm at 10 Hz, while the vertical wind speed of the CSAT3 is accurate within ±0.04 m s⁻¹ with an RMS noise of 0.0005 m s⁻¹. While the noise can occasionally be larger than the true environmental fluctuations, it has been found to minimally influence the calculated FCO2 because the noise from the separate instruments is uncorrelated and therefore filtered out by the flux calculation (Miller et al., 2010).

An investigation of FCO2 measurement uncertainties from ships indicated a detection limit for a dried, closed-path eddy covariance system of roughly |ΔpCO2|>35 µatm for the mean wind speed observed in this study (Blomquist et al., 2014). The ΔpCO₂ in the region often exceeds this value (Duke et al., 2021; Sims et al., 2023). Additionally, we expect some reduction in the detection limit (i.e., increased sensitivity) for this study compared to ship-based studies, because the measurements were from a stationary tower. Therefore, the observations avoid some common sources of uncertainty experienced from moving platforms, such as the needed for a complex wind vector motion correction and tilt effects that degrade the performance of the LI-7200 (Miller et al., 2010; Vandemark et al., 2023).

While we cannot perform a direct assessment of FCO2 uncertainty, we can estimate the order of magnitude of the uncertainty by assessing the variation in FCO2 measurements during periods expected to have stable fluxes. Here we do that by calculating the standard deviation for 6 h intervals during periods of full ice cover, when diurnal variations in FCO2 were expected to be minimal. The standard deviation across these winter periods had a mean of ±1.02 mmol m⁻² d⁻¹ and a median of ±0.75 mmol m⁻² d⁻¹. Spring and summer seasons were excluded from the estimate because standard deviation measured during those periods was expected to be a combination of measurement uncertainty and actual diurnal FCO2 trends.

The difference in partial pressure of CO₂ (pCO₂) between the air and the water dictates the direction of the flux (up or down), while the gas transfer velocity (k) describes the efficiency of transport. The latter incorporates all of the physical processes at the air-sea interface that affect gas exchange. Using existing parameterizations for k we can use the FCO2 measurements to estimate the partial pressure of CO₂ in water (pCO_2w). Viewing the flux data as pCO₂ provides context for the flux by showing the seasonal pattern of waterside carbon inventories, without the short-term variability caused by the impact of wind speed on k and therefore FCO2 magnitude. To estimate pCO_2w we set our measured FCO2 equal to the open water bulk formula for CO₂ flux: 1FCO2=ks[pCO2w-pCO2air], where k is estimated from wind speed using Wanninkhof (2014), s is the solubility of CO₂ in seawater (calculated using satellite-derived sea surface temperature [SST] and salinity [SSW] data from the mooring), and pCO_2air was partial pressure of CO₂ in air. Because pCO_2w was the single unknown in the equation we were able to solve for it.

The processes affecting FCO2 from sea ice are different from open water, but a similar bulk flux formula can be applied. During periods of full sea ice cover, we can use this formula to estimate the partial pressure of CO₂ in ice (pCO_2ice). This value describes the concentration of CO₂ in the ice, which in this context could represent any ice surface interacting with the atmosphere including snow crystals, the sea ice surface, or the sea ice volume (including brine). Like pCO_2w in open water, pCO_2ice in ice can vary over time and its difference from pCO_2air is still expected to dictate the direction of flux. In the laboratory study of Kotovitch et al. (2016), FCO2 was measured in a tank over periods of forming, thickening, and melting sea ice. Supporting measurements of pCO_2air and pCO_2ice enabled the derivation of a gas transfer coefficient (Kice) using the following bulk formula: 2FCO2=Kice[pCO2ice-pCO2air]. The Kice parameter encapsulated both the gas transfer velocity and solubility of CO₂ in ice. This was done to avoid estimating solubility using seawater-based functions of temperature and salinity outside the range for values for which they were designed. Kice during periods of ice growth was 2.5 mol m⁻² d⁻¹ atm⁻¹, while for periods of ice decay it was 0.4 mol m⁻² d⁻¹ atm⁻¹ (Kotovitch et al., 2016).

Because we did not collect in situ pCO_2ice measurements we could not use Eq. (2) to calculate Kice for independent verification. Instead, we estimated pCO_2ice during periods of full ice cover using Eq. (2) with measured FCO2 and pCO_2air and the Kice values for ice growth and decay found by Kotovitch et al. (2016). Comparisons of estimated pCO_2ice to previous in situ measurements were used to determine if the laboratory-derived Kice values were applicable in field conditions.

For periods where the surface was a mix of open water and sea ice we estimated pCO_2w by scaling FCO2 linearly to the fraction of open water (f=1-SIC) in front of the tower (Butterworth and Miller, 2016a). In these cases, we omitted the influence of air-ice gas exchange in the calculation of pCO_2w due to the fact that Kice is much lower than its equivalent (ks) for the air-water interface (Wanninkhof, 2014; Kotovich et al., 2016). Under the environmental conditions (e.g., temperatures, salinity, etc.) in this study we estimated that FCO2 from open water is roughly 20 times more efficient. Therefore, the omission of air-ice gas exchange is expected to have a minimal influence on the pCO_2w calculation. For this work, seasons were defined by in situ observations (i.e., by visits to the station, and from camera images) rather than standard astronomical definitions, with spring being broadly represented by ice melt, summer by open water, and fall by ice formation.

For this study, we mark the beginning of the spring season as the moment when mean daytime temperature rises above 0 °C (Fig. 3c, d). In the two years presented, this spring start date shifted by about 3 weeks. This difference appeared to play a role in the differences in CO₂ flux direction and magnitude throughout the remainder of each season, which will be discussed in Sect. 3.4.1.

The spring season is marked by distinct periods (Table 1). In early spring, the surface is characterized by freeze–thaw cycles (e.g. Hanesiak et al., 1999). While there may be leads during this period, the ice is landfast, with typically 100 % coverage. During mid spring, standing water melt ponds form on the ice surface, still with 100 % ice coverage. The late spring season is marked by a break-up of the sea ice, where the ice concentration decreases from 100 % to 0 % coverage.

During early spring the behavior of snow melt/refreezing appears to be the key factor affecting FCO2. On early spring days in which air temperature oscillated around the melting point, FCO2 oscillated with a mean range of 1 mmol m⁻² d⁻¹ on a diurnal cycle negatively correlated with air temperature (Fig. 5b). During the day, positive temperatures caused melt, resulting in a negative FCO2 (uptake). At night, when negative temperatures caused water to refreeze and expel CO₂ gas, FCO2 was positive (outgassing). Incoming solar radiation did not have an immediate impact on FCO2 (Fig. 5a, b), though was correlated once lagged to temperature. During this period the sea surface was characterized by an average ice coverage of 99 %. In photographs from the 19 d included in the freeze–thaw analysis, the surface showed a slight darkening during the daytime, consistent with Hanesiak et al. (1999) who observed diurnal albedo patterns caused by increased water content during the day, and freezing overnight. At this time, no discernable standing water melt ponds had formed.

We estimated pCO_2ice using Eq. (2) and found a diurnal range of 600 µatm during this period, corresponding in sign to the direction of the flux (Fig. 5b, c). Mean diurnal minimum pCO_2ice was roughly 0 µatm and occurred in the afternoon, coinciding with the warmest air temperatures and greatest active melting. The mean diurnal maximum was 600 µatm and occurred shortly after sunrise, when mean air temperature was at a minimum at -2 °C.

As temperatures increased during mid spring (Table 1), standing water melt ponds began to form on the landfast ice. During this period the magnitude of FCO2 increased, showing more strongly negative fluxes (i.e., uptake), with occasional outgassing events (Fig. 4a, b). The positive to negative FCO2 diurnal oscillations seen in early spring (Fig. 5b) were no longer evident (Fig. 6).

During late spring (Table 1) the landfast ice begins to break up and the ocean surface in front of the tower is characterized by varying concentrations of sea ice in the form of ice floes. During this period (lasting several weeks) FCO2 becomes even more strongly negative. This increased CO₂ uptake was likely due to the exposure of seawater that had low pCO_2w relative to atmospheric pCO_2air. The pCO_2w (calculated using Eq. 1) decreased from a mean of 394 µatm (ΔpCO₂ of -10 µatm) during mid spring to 373 µatm (ΔpCO₂ of -29 µatm) during late spring (Fig. 7). This decrease in pCO_2w acts in opposition to the water temperature effect on pCO_2w during this period. In both years, water temperature (both satellite SST and 13 m mooring) increased by roughly 1 °C over the period, which independently should cause a roughly 20 µatm increase in pCO_2w, based on the direct positive relationship between water temperature and pCO_2w (Takahashi et al., 1993).

Springtime FCO2 is characterized by the distinct physical processes related to freeze–thaw, melt ponds, and ice break-up. These processes likely all occur to some degree throughout the spring period, but they generally progress sequentially along with the advance of warming over the spring.

The observation of diurnal cycles in early spring influenced by active melting and freezing has implications for sampling design for instruments not intended for continuous deployment (e.g., chambers) – namely that measurement biases could arise based on collection time (e.g., cold morning measurements would predict a CO₂ source and warm afternoon measurements would predict a sink). Despite the diurnal variability, the ice acts as a weak sink during this period with mean flux of -0.35 mmol m⁻² d⁻¹.

While the negative mean FCO2 suggested mean pCO_2ice was below mean pCO_2air, the diurnal oscillations in FCO2 indicated diurnal changes in the magnitude of pCO_2ice. The diurnal range of pCO_2ice was quite large (0–600 µatm; Fig. 5c) and was likely due to physical processes associated with the phase change of water. That is, the expulsion of CO₂ gas as water freezes and then the subsequent melting of low pCO_2ice (Nomura et al., 2006; Rysgaard et al., 2011; Kotovitch et al., 2016). During active melting we found a diurnal minimum in our flux-estimated pCO_2ice of 0 µatm, which corresponds in magnitude to previous directly-measured, in situ melt pond pCO₂ of 36 µatm (Geilfus et al., 2015). The low pCO₂ of melt ponds are expected to immediately begin to equilibrate toward atmospheric values (Geilfus et al., 2015). However, the diurnal change in flux direction from uptake to outgassing indicates that pCO_2ice rose above atmospheric values. This suggests that the CO₂ gas expelled during freezing accumulated in a thin, supersaturated layer near the surface. This is in line with the laboratory experiment of Kotovich et al. (2016), who also observed outgassing during freezing due to supersaturation in the top 5 cm of ice, while the underlying water remained undersaturated with respect to the atmosphere.

The large range of pCO_2ice in this study has some analogies in the literature. This includes the range measured by Delille et al. (2014) in Antarctic pack ice (roughly 50–900 µatm) and the range observed by Geilfus et al. (2015) in Arctic springtime ice (36–380 µatm). While these studies represent daytime-only pCO_2ice measurements over longer time frames (seasonal and sub-week, respectively), they show that pCO_2ice of these magnitudes (0–600 µatm) are plausible. The agreement between our estimated pCO_2ice and previous direct in situ measurements of pCO_2ice suggests that the gas transfer coefficient for melting ice measured by the laboratory experiment of Kotovitch et al. (2016; which we used to estimate pCO_2ice from our flux measurements) may be reasonably applicable to the real-world environment. However, it is worth noting that pCO_2ice (Fig. 5c) occasionally dropped below zero, which is a physically impossible value. Such instances may indicate that the Kice value used to calculate pCO_2ice was too small. Because Kice combines both gas transfer velocity and solubility, inaccuracies in either term could be responsible. However, it is also possible that the negative values of pCO_2ice are simply due to the random error inherent in eddy covariance systems. Because random error can cause both positive and negative deviations in measured flux, these data points were retained to avoid biasing the average.

To further constrain the gas transfer coefficient over melting sea ice (Kmelt) we ran an additional test using our flux measurements. We assumed that pCO_2ice was zero during periods of time in early spring when temperatures were positive. This represents the lowest possible pCO_2ice and therefore the most negative ΔpCO₂ that was physically possible at the site. Using Eq. (2) with this prescribed pCO_2ice and mean FCO2 we calculated a Kmelt value of 0.36 mol m⁻² d⁻¹ atm⁻¹. This is nearly identical to the Kmelt of 0.4 mol m⁻² d⁻¹ atm⁻¹ found by Kotovitch et al. (2016). This is a rough estimate for several reasons. First, the pCO_2ice is not expected to be zero for this entire period. Past studies (e.g., Geilfus et al., 2015) have shown that pCO₂ of freshly melted ice approaches zero, but that value is expected to rise quickly as the water equilibrates with the atmosphere. A pCO_2ice value of zero is therefore theoretically possible for a rapidly melting surface, but it would be a transient state. An average pCO_2ice higher than zero would result in a higher Kmelt. Secondly, this calculation assumes that 100 % of the surface is decaying ice – which may not be true. With a lower fraction of the surface actively decaying we expect the estimated Kmelt to increase. Overall, however, it provides a constraint on the lower limit of Kmelt and suggests that the laboratory value proposed by Kotovitch et al. (2016) is the correct order of magnitude in the natural environment. That the laboratory value aligns with the lower limit measured in the field makes sense, given that some of the natural factors that are known to increase fluxes (e.g., wind) are absent in laboratory settings.

Mid spring (Table 1) was characterized by the formation of large standing melt ponds on the landfast ice (Fig. 2b). During this period, we observed a discontinuation of the diurnal cycles observed during early spring (i.e., negative correlation between FCO2 and temperature). This was likely due to consistently positive air temperatures eliminating the potential for refreezing, ending freeze–thaw related forcings on the flux. Mid spring also had more strongly negative FCO2 than early spring. This suggests that melt ponds were acting as a sink for CO₂. This is in line with previous studies which have found that low pCO_2w concentrations in melt water cause melt ponds to be a net sink of CO₂ (Semiletov et al., 2004, Geilfus et al., 2015).

A quantitative analysis of pCO₂ during the melt pond period was not attempted due to uncertainties in gas transfer coefficients. The laboratory-derived Kice values for ice growth and decay that were applied during the freeze–thaw period were not expected to be applicable over flux footprints that contained both ice and standing water melt ponds. And while we assume that melt ponds are exchanging gas with the atmosphere with physics more closely aligned to air-water gas exchange than air-ice gas exchange, there are reasons to believe that open ocean parameterizations of gas transfer velocity are not entirely suitable to melt ponds, due to the expected differences in wind-wave fields and waterside turbulence between the two environments.

During the transition to ice breakup in late spring we measured consistently negative FCO2, which indicated pCO_2w values during this period were below atmospheric values. Because increasing water temperatures during this period should have led to increased pCO_2w, the observed decrease in pCO_2w suggests that other processes were driving the low pCO_2w values observed during this period. For example, hyperspectral transmitted irradiance measurements made in spring 2017 on the nearby mooring revealed ice algal and under-ice phytoplankton blooms occurring from 5 March to 21 May and 1 to 10 June, respectively (Yendamuri et al., 2024), that could have drawn down pCO_2w. However, primary production in the area is relatively low compared to other Arctic regions due to nitrogen limitation (Kim et al., 2020; Back et al., 2021) and thus, may not have significantly contributed to the low pCO_2w observed. An alternative process is simply ice melt, which has been shown to lower pCO₂ both through simple mixing of low-pCO₂ melt water, and due to non-linearities in carbonate system chemistry (Yoshimura et al., 2025). The salinity mooring data was inspected to determine whether melt water dilution was observed. At 13 m depth there was a small decrease in SSW (-0.2) over the late spring period. This would correspond to a small decrease in pCO_2w (-3 µatm), a relatively small FCO2 forcing. However, because the water was stratified at this period (i.e., SST >T13m), it is possible (and likely) that the change in SSW at the surface was greater, resulting in a larger FCO2 forcing.

Here we define summer as the open water period, which spans from mid-July to mid-October (Table 1). In 2017, the difference between SST (derived from satellite) and 13 m water temperature (T13m, from the mooring) showed that the sea was stratified from May to August (Fig. 8). On 13 August the SST peaked for the season at 9.8 °C. For the remainder of August, SST decreased while T13m increased, indicating a growing mixed layer, which reached 13 m depth on 2 September when equivalence between SST and T13m was reached. The two temperatures tracked together until early October when sea ice began to form. Temperature measurements obtained at 22 and 39 m depths showed that by 1 October the water in the region became mixed from at least the surface to a 39 m depth, which is close to the charted bottom depths for most of the area within the flux footprint. A similar story unfolded in 2018, with SST also reaching its peak on 13 August. However, compared to 2017 its maximum temperature was much lower at 4.4 °C, presumably due to the delayed onset of melt, providing a shorter window for the absorption of incoming solar radiation by the sea surface. Because the mooring data stopped on 14 August mixed layer depths during the second half of summer 2018 were not available.

Figure 8 shows the FCO2 dependence on water temperature as it varies across seasons. During the open water period in 2017 the mean FCO2 tracks with SST, rising together in July and August, peaking in mid-August, and decreasing through late August to October. We separated the summer season into three subseasons (early, mid, and late) corresponding to changes in environmental conditions. The early season (13 July 2017–13 August 2017) was from the beginning of open water until peak SST and showed increasing FCO2 (Fig. 8). The mid season (13 August 2017–1 October 2017) was from peak SST until the water profile became unstable and showed decreasing FCO2 (Fig. 8). The late season (1 October 2017–14 October 2017) was the period immediately preceding the onset of freezing in which the mixed layer deepened. We then investigated the role of thermodynamic processes on the observed seasonal FCO2 changes. Figure 9 shows a pCO_2w estimate derived from FCO2 using Eq. (1) and a pCO_2w projection calculated using established temperature and salinity relationships (1pCO2w∂pCO2w∂SST≈0.0423 °C⁻¹ from Takahashi et al., 1993; SSWpCO2w∂pCO2w∂SSW≈1 from Sarmiento and Gruber, 2006). For the thermodynamic projection, the FCO2-derived pCO_2w estimate for the first day of early summer was used as a starting pCO_2w, then projected forward for each flux interval through the end of summer using only the above SST and SSW relationships. In both early and mid summer, the two pCO_2w estimates track well, indicating that changes in SST and SSW are important drivers of FCO2 changes during these seasons. In late summer, the curves show greater divergence with the FCO2-derived pCO_2w estimate showing larger values (over 25 µatm greater) than the thermodynamic projection. While the FCO2 increased from its seasonal low during this late summer period (Figs. 4 and 8; due to the reduced wind speed (Fig. 3e)), the FCO2-derived pCO_2w estimate continued to drop in magnitude (Figs. 7 and 9). This was in opposition to the SST forcing, but coincided with deepening of the mixed layer and increased SSW values.

The overall pattern of FCO2 in 2018 was similar to 2017, with downward fluxes predominating during spring melt and break-up, then increasingly upward fluxes as SST increased during early summer (26 July 2018–13 August 2018). Like 2017, FCO2 began to decrease as soon as the maximum SST was reached at the start of mid summer (13 August 2018–N/A (not available)). However, the first two weeks of September showed a turn towards increasingly positive fluxes (Figs. 4b, 7b). In contrast, during this same period in 2017 the fluxes were becoming increasingly negative.

In summer, thermodynamic drivers appear to be the most important contributors to the direction and magnitude of FCO2. For most of the summer, the trend in FCO2 corresponds to the trend in SST. Both increase in early summer, both decrease in mid summer (Figs. 4 and 8). The mechanism causing this pattern is the direct positive relationship between SST and pCO_2w (Takahashi et al., 1993). As SST increases, it causes pCO_2w to increase, which results in increased outgassing of CO₂ to the atmosphere. SSW also has a direct positive relationship with pCO_2w (Sarmiento and Gruber, 2006). In this instance, steady reductions in SSW over the course of the early and mid summer periods (28 down to 25) partially offsets the projected peak magnitude of pCO_2w by the SST effect alone. The projection of pCO_2w using both SST and SSW effects tracks well with the FCO2-derived pCO_2w estimate (Fig. 9). This suggests that SST and SSW are the main drivers of changes to FCO2 in the early and mid summer periods. The one period of the summer in which the thermodynamic pCO_2w projection most noticeably diverges from the FCO2-derived pCO_2w estimate is late summer (1 October 2017–14 October 2017). During this period the SST continues to drop, but SSW begins to increase (25 up to 27). This corresponds to a reduced (but still negative) slope to both the FCO2-derived pCO_2w estimate and the thermodynamic projection. The cause of the increased SSW was the reversal of the temperature profile from stable to unstable (i.e., SST <T13m<T22m<T39m) resulting in greater upward mixing of higher salinity water from depth. While the similar trends in both the pCO_2w estimate and the thermodynamic projection suggest that SST and SSW are still important drivers of FCO2 during late summer, the higher magnitudes of the pCO_2w estimate compared to the thermodynamic projection suggest an additional source of increased pCO_2w. One possibility is that the increased mixing of water from depth during this late summer period may have, in addition to increasing SSW, brought CO₂-rich waters to the surface, thus slightly offsetting some of the pCO_2w reductions expected by the thermodynamic processes alone.

As stated above, the pattern of FCO2 in 2018 was similar to 2017, with the exception of 2018 showing increasing positive fluxes and increasing pCO_2w in the first two weeks of September, running in opposition to the SST forcing. One explanation is that the lower SST during 2018 enabled mixed layer deepening a month earlier than the previous year, causing mixing to increase pCO_2w (e.g., due to SSW and CO₂ concentration effects) earlier in the season. Unfortunately, the mooring temperature and salinity data were not available during this period to confirm. However, an inspection of the flux cospectra during this period showed no reason to discount this upward trend on the grounds of flux measurement error.

The fall season was defined by the occurrence of sea ice formation. Photographs from the camera at the top of the tower confirmed that freezing began on 14 October 2017 and continued to increase until consistent, full ice cover was reached on 14 November 2017. This initial freeze-up period was defined as early fall, followed by a late fall period of thickening landfast ice, during which the site continued to measure active FCO2. At the outset of freezing, the fluxes were downward (uptake), but transitioned upward (outgassing) shortly after ice formation (28 October 2017). They remained upward until they reached zero at the end of December 2017.

The downward FCO2 at the beginning of the fall occurred when sea ice concentrations were lowest. This likely represents dominant flux between the atmosphere and open water areas, since ΔpCO₂ between the water and air was negative at the onset of freezing. The upward fluxes that follow this period (November–December) coincide with sea ice concentration nearing 100 %. This suggests that these fluxes were dominated by the freezing process, whereby CO₂ gas is expelled into the brine channels, where it can then exchange with both the water below and the air above (Nomura et al., 2006). The mean flux during the initial freeze-up of early fall was 0.1±3.8 mmol m⁻² d⁻¹. During late fall the mean was 1.1±1.5 mmol m⁻² d⁻¹. However, because the quality-controlled data are not a perfectly continuous record (due to data gaps), in order to gain a measure of the seasonal flux we integrated the area under the local regression line (Fig. 4) and divided by time. For the entire fall period this gave a flux of 0.38 g C m⁻². When excluding the two weeks of downward flux in October this value rose to 0.73 g C m⁻².

While the freeze-up period appears to have distinct parts separating air-water exchange from freezing-related flux, it is expected that both processes are occurring throughout. If so, the fluxes are simultaneously acting in opposite directions, acting to reduce the magnitude of the total flux (Fig. 10).

To isolate the flux due to freezing in our dataset we estimated the flux through the open water areas using 3FCO2=fFBULK, where f is fraction of open water and FBULK is CO₂ flux calculated using the bulk formula (Eq. 1). This value was then subtracted from the measured fluxes to obtain a freeze-only flux estimate. Because we did not measure pCO_2w we used a constant value of -49 µatm in the calculation of the bulk flux, which was the pCO_2w estimate based on FCO2 for 14 October, under open water conditions just prior to freeze-up. The assumption that pCO_2w remains constant after the onset of freezing is based on there being minimal biological activity (Yendamuri et al., 2024), minimal water temperature changes to influence pCO_2w during this season, and mixed layer depths approaching the sea floor. This assumption is supported by a yearlong dataset of under-ice pCO_2w measured from an autonomous, underwater sensor platform in nearby Cambridge Bay, which showed only minor variations in pCO_2w after the onset of freezing (Duke et al., 2021).

The bulk flux estimate suggests that without the influence of freezing the measured flux would have been consistently downward or zero (Fig. 11). Subtracting the bulk flux from the measured flux we get an estimate of the freeze-only flux. Diurnal variations in both wind speed and freeze–thaw during this freeze-up period complicate the assessment. This can be seen by the large variance in FCO2 over short timescales in Fig. 11. However, by smoothing diurnal variations we observed that outgassing from freeze-only flux starts at the onset of freezing and continues through the fall season (Fig. 11 – green line). Integrating the area under this curve we estimate a freeze-only flux of 1.07 g C m⁻² for the season.

Previous measurements of freezing-related outgassing from the initial formation of sea ice have been limited to laboratory studies. Laboratory tank experiments have found FCO2 over forming sea ice ranging from 0 to 1.0 mmol m⁻² d⁻¹ (Nomura et al., 2006) and -0.4 to 0.75 mmol m⁻² d⁻¹ (Kotovitch et al., 2016). Previous field studies have measured FCO2 over young sea ice soon after it formed and have found slightly larger (though still small) upward fluxes. Nomura et al. (2018) measured FCO2 of 3.7±2.0 mmol m⁻² d⁻¹ for young ice and 0.7±0.7 mmol m⁻² d⁻¹ for older ice. These ?uxes are the same order of magnitude as other chamber-based measurements over land fast ice like Nomura et al. (2013) and Delille et al. (2014), whose measurements over Antarctic pack ice showed a temperature dependence (i.e., T<-8 °C → no flux; -8 °C <T<-6 °C → 1.9 mmol m⁻² d⁻¹).

Our measurements span the range of these different ice regimes, and importantly include the period of initial ice formation. During the week when SIC first reached 100 % occurred (1–8 November) the mean measured FCO2 was at a fall maximum at 2.6±3.6 mmol m⁻² d⁻¹. This outgassing agrees well with previous measurements over young ice, but is roughly a factor of 3 higher than FCO2 measured by previous tank experiments. This may indicate the effect that wind has on increasing FCO2, a process which is absent from tank experiments. Additionally, this higher magnitude flux is seen in the freeze-only estimate of FCO2, which peaked for a month (22 October–22 November) at 1.7±0.1 mmol m⁻² d⁻¹. This peak period includes earlier periods of ice formation (i.e., before ice concentration reached 100 %), meaning that the freeze-only portion of the flux was positive and of a similarly high magnitude, but was competing with downward air-sea FCO2.

Outgassing over the entire late fall period was lower, with a mean FCO2 of 1.1±1.5 mmol m⁻² d⁻¹. Additionally, we found a seasonal/temperature trend, with fluxes decreasing from their highest magnitudes (2.6±3.6 mmol m⁻² d⁻¹) during the first week of November towards their lowest flux magnitudes (0.5±1.5 mmol m⁻² d⁻¹) during the last two weeks of December, when temperatures were colder and the ice was thicker. This fits previous findings that gas migration is more effective in warmer sea ice compared with colder sea ice, where the formation of brine is significantly reduced (Gosink et al., 1976; Delille et al., 2014). In practical terms this means that full, solid, cold ice cover acts as a barrier to gas exchange.

To put the freezing-related fluxes from this study into context we estimated an Arctic-wide flux from freezing. The area of Arctic first-year sea ice was estimated to be 9.4 million km², calculated as the average annual range of sea ice area over a five-year period from 2014–2018, based on the NSIDC monthly sea ice area for the Northern Hemisphere (Fetterer et al., 2017). Using the cumulative flux for the entire fall season at our site to extrapolate, we estimate the total Arctic CO₂ outgassing from freezing for 2017 (14 October–31 December) was 6.8 Tg C. If we use our freeze-only estimate (which removes the influence of downward air-sea gas exchange) that increases to 9.9 Tg C. Bates and Mathis (2009) estimated an annual Arctic Ocean CO₂ exchange of -66 to -199 Tg C yr⁻¹, a net sink. Our estimate for outgassing from the freeze-up period represents a counterbalance equivalent to 3.5 % to 10 % of this total Arctic sink, or 5 % to 15 % if we use our estimate for the “freeze-only” component of the measured flux. While this is a rough estimate, it suggests that outgassing from freezing represents a small, but non-negligible portion of annual flux, which is not typically considered in Arctic CO₂ budgets.

Another aspect of freeze-up that we were able to address is the previous hypothesis that gas exchange is enhanced in the presence of forming sea ice (Anderson et al., 2004; Else et al., 2011). To do so, we assumed that ΔpCO₂ remained constant during the marginal ice conditions at the beginning of the freeze-up period (i.e., early fall). As described above, this assumption is rooted in evidence for minimal biological activity or temperature changes. We then set measured FCO2 equal to Eq. (1) to estimate gas transfer velocity normalized to a Schmidt number of 660 (k660) and weighted it to the fraction of open water.

The magnitudes of k660 during the freeze-up period with marginal ice conditions (14–28 October) stayed relatively close to open water relationships of k660 and 10 m wind speed (Fig. 12). The scatter in the figure was most likely due to the fact that ΔpCO₂ was held constant at -49 µatm, which was unlikely to have been rigidly the case through this period. Without ΔpCO₂ measurements we have no way to determine whether fluxes were enhanced in minor ways (e.g., say 20 %). But our data does contradict previous findings of large enhancements (e.g., orders of magnitude) to gas exchange in the vicinity of sea ice. Such a scenario would have been characterized by our k660 far surpassing open water parameterizations.