Seasonal transition dates can reveal biases in Arctic sea ice simulations

Arctic sea ice experiences a dramatic annual cycle, and seasonal ice loss and growth can be characterized by various metrics: melt onset, break-up, opening, freeze onset, freeze-up and closing. By evaluating a range of seasonal sea ice metrics, CMIP6 sea ice simulations can be evaluated in more detail than by using traditional metrics alone, such as sea ice area. We show that models capture the observed asymmetry in seasonal sea ice transitions, with spring ice loss taking about 1.5–2 months longer than fall ice growth. The largest impacts of internal variability are seen in the inflow regions of melt and freeze 5 onset dates, but all metrics show pan-Arctic model spreads exceeding the internal variability. Through climate model evaluation in the context of both observations and internal variability, we show that biases in seasonal transition dates can compensate for other unrealistic aspects of simulated sea ice. In some models, this leads to September sea ice areas in agreement with observations for the wrong reasons.


Introduction
Metrics of seasonality have been under-utilized in evaluating sea ice in climate models, due to a lack of long-term observational products, the required daily model output and the complexities in defining seasonal Arctic sea ice transitions. However, new process-based metrics for model evaluation are much needed-the spread between climate model projections of sea ice has been on the order of millions of square kilometers in Coupled Model Intercomparison Project (CMIP) Phases 3, 5, and 6 SIMIP, submitted 2019), while the causes of the model spread remain largely unknown. Furthermore, the 15 sources of model biases can be obscured by models that show realistic sea ice areas for the wrong reasons. Seasonal sea ice transitions can provide additional process-based metrics to assess climate models. Newly available observational data  and model output from CMIP6 models (Notz et al., 2016) allow such model assessment for the first time. In this study, we assess how different metrics of seasonal sea ice transitions are represented in models and observations, and evaluate how these metrics can inform our understanding of simulated Arctic sea ice throughout the year. To do this, we utilize 20 observations and sixteen global climate models, including three sets of ensembles with at least 30 members. Using this rich data set, we evaluate model biases in the context of both the observed sea ice state and multiple simulated representations of internal variability.
In this study we use satellite data to evaluate the performance of fifteen CMIP6 models and the Community Earth System Model Large Ensemble (CESM LE) in terms of their seasonal sea ice transitions in the Arctic from 1979-2014. By utilizing model ensembles, we are able to account for the role of internal variability in modeling the seasonality of Arctic sea ice. As there is no single metric that fully describes seasonal sea ice changes, we utilize a variety of metrics that have been developed for both models and observations. All spatial medians and means of sea ice variables are calculated between 66-84.5°N in both 70 models and satellite data in order to exclude the largest polar hole in satellite data.

Global coupled climate models
CMIP establishes a set of common experiments for global climate model simulations to quantify how the Earth system responds to forcing, as well identify the sources and consequences of model biases . This study uses models from the most recent phase, CMIP6, in order to evaluate the current state of sea ice simulation. Models are selected for analysis 75 based on the availability of two daily sea ice variables: sea ice concentration (CMIP6 variable name: siconc) and the surface temperature of the sea ice or snow on sea ice (CMIP6 variable name: sitemptop). Our study utilizes all CMIP6 models that met this criteria by March 4th, 2020, which includes fifteen models from nine different institutions (ACCESS-CM2, BCC-CSM2-MR, BCC-ESM1, CanESM5, CESM2, CESM2-FV2, CESM2-WACCM, CESM2-WACCM-FV2, CNRM-ESM2-1, CNRM-CM6-1, EC-Earth3, IPSL-CM6A-LR, MRI-ESM2-0, NorESM2-LM, NorESM2-MM) ( Table S2). As the scope of the study is 80 limited to the satellite era, we use the historical forcing experiment from each model for the period that overlaps with satellite data . All models are kept on their native grids to minimize errors related to interpolation and regridding. Ocean and ice model component details for all models are provided in Supplementary Table S2.
The number of available ensemble members varies by model, with some models providing as few as three members and others as many as thirty-five with the required daily variables. Here we use the first ensemble member (r1i1p1f1 or r1i1p1f2) 85 from each model for inter-model comparisons and evaluation against satellite data. To assess the internal variability of the seasonal sea ice metrics, the two CMIP6 Models with at least 30 members (CanESM5 and IPSL-CM6A-LR, hereafter referred to as IPSL) are utilized, in addition to the CESM LE. All of the coupled global models have a nominal ocean resolution of 1°.
Relevant variables are available at a daily temporal resolution for 40 members in the CESM LE, 35 members in the CanESM5 CanESM5 for comparison to each other and IPSL in order to standardize the sample size. The results are insensitive to the subsetting of ensemble members (select figures using all available members are provided in the Supplement).
In previous work, the CESM LE was employed to compare multiple model definitions of melt and freeze onset (Smith and Jahn, 2019

Defining seasonal sea ice transitions
Establishing a set of metrics for studying the seasonality of Arctic sea ice is important for comparing models and observations, 110 as well as interpreting the relationships between transition times and other sea ice characteristics. Here we utilize a range of seasonal sea ice transition dates and periods to study multiple thermodynamic phases of the ice that may be relevant to our physical understanding of the sea ice. These metrics are summarized in Tables 1 and 2.

Melt onset, freeze onset and the melt season
The melt season length is defined as the number of days between melt onset and freeze onset. The melt season length has been 115 utilized as a parameter to investigate energy absorption of the Arctic surface ocean and relationships have been found between the melt season length and sea ice extent (Stroeve et al., 2014). The metrics of melt onset and freeze onset are used to describe the first date of continuous sea ice melt and freeze at each grid cell for each year. Melt and freeze onset are meant to capture a change of phase between water and ice. For melt onset, this means water on the surface of the ice or snowpack. For freeze onset, the change of phase refers to either congelation or frazil ice growth. 120 4 https://doi.org/10.5194/tc-2020-81 Preprint. In satellite retrievals, continuous melt and freeze onset are defined using the brightness temperature of the surface because brightness temperature is sensitive to the phase of water (Markus et al., 2009;Steele et al., 2019). Brightness temperatures are collected at the 19V and 37V polarizations from SMMR,SSM/I and SSMIS sensors. Melt and freeze onset dates are derived from weighted brightness temperature parameters to determine early melt and freeze onset and continuous melt and freeze onset, and the retrieval algorithm (known as PMW) is described fully in Markus et al. (2009). In this study we use continuous 125 melt and freeze onset because these dates are more representative of a seasonal transition in the sea ice compared to early melt onset. The AHRA dataset provides an alternative set of melt onset dates that are derived from passive microwave brightness temperatures using the AHRA retrieval algorithm instead of the PMW retrieval algorithm. However, the AHRA melt onset dates are more representative of early melt (Drobot and Anderson, 2001), so they are not utilized in this study. by the brightness temperature-based satellite data, recent work demonstrates that melt and freeze onset dates derived from surface temperature are comparable, particularly when considering the range of internal variability (Smith and Jahn, 2019).
We therefore utilize model definitions of melt and freeze onset developed in Smith and Jahn (2019), based on the surface temperature passing below/above a given threshold. For melt onset, a threshold of -1°C is used to minimize the impacts of 135 daily variability and maintain comparability with previous studies (Jahn et al., 2012;Mortin and Graversen, 2014). For freeze onset, a threshold equal to the freezing point of ocean water (-1.8°C) is used.

Break-up, freeze-up and the open water period
The open water period, also known as the inner ice-free period , is defined as the number of days between ice break-up and freeze-up (also commonly referred to as ice retreat and advance). The open water period has been utilized as 140 a metric to study variability and trends in the sea ice (Serreze et al., 2016;Barnhart et al., 2016) and seasonal predictability of the ice .
Of the seasonal sea ice transition dates investigated here, the definitions of break-up, freeze-up and the open water period vary the most across the literature (Table S1). In the models, we use the definitions for break-up, freeze-up and the open water period used in Steele et al. (2019) to allow for comparison with observations. Of the definitions identified and described in Break-up is defined as the last day that sea ice concentration passes below the threshold of 15% between March 1 and the annual sea ice concentration minimum date . Freeze-up is defined as the first day that sea ice concentration passes above the 15% threshold between the sea ice concentration minimum date and February 28 of the following year. The open water period is defined as the number of days between break-up and freeze-up. The outer ice-free period has been used the least frequently as a metric of Arctic sea ice seasonality, and it is based on the dates of opening and closing defined by Steele et al. (2015). The Steele et al. (2015) definitions are applied in the Steele et al.
(2019) dataset . The date of opening is defined as the first day that sea ice concentration passes below the threshold of 80% between March 1 and the annual sea ice concentration minimum date. Likewise, the date of closing is defined 155 as the first day that sea ice concentration passes above the 80% threshold between the sea ice concentration minimum date and February 28 of the following year. By definition, opening must occur before break-up and freeze-up must occur before closing. However, dates of opening and closing are not limited solely by the existence of break-up and freeze-up dates: if ice concentration falls below/above 80% but not below/above 15%, there will still be an opening/closing date. This means that the areal coverage of opening/closing dates is generally larger than those of break-up/freeze-up dates. period is defined as the number of days between freeze onset and freeze-up, and is designed to describe the rate of transition between initial ice growth water and when an area stops being "ice-free".

Seasonal loss-of-ice period and seasonal gain-of-ice period
The The seasonal loss-of-ice period is defined as the number of days between sea ice opening and break-up, and the seasonal gain-170 of-ice period is defined as the number of days between freeze-up and closing . The seasonal loss-of-ice period and the seasonal gain-of-ice period can only be calculated at grid cells where both of their respective dates exist for that year (e.g. both a date of opening and break-up are needed for a valid seasonal loss-of-ice period). The seasonal loss-of-ice period describes how quickly the ice concentration transitions from 80% to 15%, while the seasonal gain-of-ice period describes the rate of transition between 15% and 80% ice concentration.

Spring transitions 195
We find that the transition from sea ice melt onset to break-up takes two to three months in both satellite data and models.
Satellite data show that melt onset generally occurs between April and June over most regions of the Arctic (Fig. 3a), with the 9 https://doi.org/10.5194/tc-2020-81 Preprint. Discussion started: 3 April 2020 c Author(s) 2020. CC BY 4.0 License. median date of melt onset occurring on May 30. The median date of opening (July 8) occurs about 40 days after melt onset and the median break-up date (July 28) occurs 20 days after opening (Table 3). This indicates that the most time-consuming aspect of the observed spring ice loss is the transition between water formation on the ice or snow surface and a decline in ice 200 area (the melt period). Once open water is present in the grid cell, the transition between 80% ice concentration and 15% ice concentration is faster, due to energy absorption from the change in the surface albedo (Perovich et al., 2002).  3) shifts earlier toward the satellite data. Melt onset dates in the inflow regions that fall between January-March demonstrate that melt onset at the surface of the snow pack can occur while the ice area is still expanding in those regions. This highlights that surface temperature based definitions, such as melt onset, capture different physical processes than sea ice concentrationbased definitions, as was previously shown (Smith and Jahn, 2019). In agreement with observations, all models project the median length of the melt period to be longer than the seasonal loss-of-ice period. The median time between melt onset and 210 opening (the melt period) is 28-53 days in models and 31 days in the satellite data, while the median time between opening and break-up (the seasonal loss-of-ice period) is 13-33 days in models and 20 days in observations.  We find that for all spring transition metrics, the model spread exceeds estimations of internal variability, which show a maximum of 6 days between ensemble members (Table 3). Of the spring sea ice transition dates (melt onset, opening and break-up), the sea ice melt onset dates show the largest spread in satellite-era medians between the models (33 days) (Table 3).

215
This range is skewed late by the CNRM-ESM2-1 and CNRM-CM6-1. If the two CNRM models are excluded, the spread in median melt onset dates is 19 days instead of 33 days, still larger than the other two spring metrics (17 days for opening and 13 days for break-up). As the CNRM melt onset dates are more than a week later than the other models, this also means that differences between the CNRM models and the other models are unlikely explained by internal variability alone. The CNRM models are further discussed in Sect. 4.5. Of the spring transition dates, the internal variability is highest for the melt onset 220 12 https://doi.org/10.5194/tc-2020-81 Preprint. Discussion started: 3 April 2020 c Author(s) 2020. CC BY 4.0 License.
dates, particularly in the marginal ice zones (Fig. 9). High variability between ensemble members in the marginal ice zones is likely related to the interannual variations in the position of the ice edge. Additionally, modeled melt onset is defined using daily surface temperature, which exhibits greater variability than daily ice concentration (Smith and Jahn, 2019).  sixteen models fall within the maximum range of internal variability (10 days) of the satellite data (Table 3). Freeze-up is also generally delayed in models compared to satellite data, with seven of the sixteen models falling within the maximum range of internal variability (9 days) of the observations. The observed time between median freeze onset and freeze-up (the freeze 230 period) is ten days, and this is twenty-nine days shorter than the time between melt onset and opening (the melt period) ( Table   3).

Fall transitions
The sea ice closes when it passes the 80% ice concentration threshold, and the median closing date occurs on October 5 in satellite data and between October 4-November 5 in the models (Table 3). Ice freeze-up occurs before the date of closing by definition, as both are defined using ice concentration, but areas closer to the Central Arctic (that fall below 80% but not 15%)  Like the spring transition metrics, we find that Pan-Arctic model differences in fall transition metrics are unlikely due to internal variability alone. Of the fall sea ice transition dates (freeze onset, freeze-up and closing), freeze onset shows the largest 240 spread in satellite-era median between models (35 days) (Table 3) is highest for marginal ice zone freeze onset dates (Fig. 9). As described for melt onset, this large internal variability is due to 245 the changing interannual position of the ice edge and the variability of surface temperature. The fall transition dates do not always occur in the expected order at each grid cell, with freeze up occurring at the same time or earlier than freeze onset. In satellite data, simultaneous freeze onset and freeze-up dates may in part be explained by the satellite retrieval algorithms: the PMW retrieval algorithm for freeze onset uses an ice concentration metric to derive freeze onset at locations where the date can not be reliably derived using the weighted brightness temperature scheme (Markus 250 et al., 2009), but the data does not contain information on how often this back-up method is employed. Hence, the use of ice concentration by both the freeze onset and freeze-up retrieval algorithms may contribute to cases where the dates are equal.
In models, the definitions of seasonal sea ice metrics aim to capture thermodynamic changes in the sea ice, but the similar dates for freeze onset and freeze-up highlight that dynamic sea ice changes influence the ice concentration-based seasonal sea ice transition metrics as well. While a particular grid cell may not register a persistent change in surface temperature below 255 the threshold for freeze onset (-1.8°C), it is possible that the ice concentration of the grid cell surpasses 15% due to dynamic transport into the grid cell, triggering the detection of freeze-up. In multiple models, freeze onset occurs later than freeze-up in   some parts of the Central Arctic. All models show freeze onset later than freeze-up occurring in the marginal ice zones, which are particularly susceptible to dynamic ice changes due to generally lower ice concentrations.

Inter-seasonal transition periods 260
Out of the three inter-seasonal periods of transition (the melt season, the open water period and the outer ice-free period), the outer ice-free period is the only one that is consistent with satellite data. The outer ice-free period (80% ice concentration thresholds) has an observed median length of 81 days and model medians falling between 72-95 days (Table 4).
In contrast, the melt season length and the open water period are too long in models compared to observations. Generally, the greatest contribution to the differences between the observed and modeled open water period is from later than observed 265 freeze-up dates. The open water period has a median of 64 days in the satellite data and medians ranging between 73-113 days in the models (a spread of 40 days) (Table 4). Likewise, modeled melt seasons that are too long compared to observations are largely driven by later freeze onset dates. The observed median melt season length is 110 days and the model medians range between 115-165 days (Table 4). Therefore the melt season length exhibits the largest model spread of all the inter-seasonal periods (50 days). This is due to larger model ranges in both melt onset and freeze onset than the other transition dates, and the 270 contribution of each date to the model spread in melt season length is approximately equal ( Table 4). The melt season length model range is also skewed high by the IPSL model, which has a median melt season length of 165 days in its first ensemble member. This is 11 days longer than the next longest model median, and the choice of ensemble member likely plays a rolethe IPSL model has a particularly large range of internal variability in the median melt season length (17 days compared to 9 days in the other two model sets) (Table 4).

Seasonal transitions affect sea ice area and thickness year-round
Model representations of seasonal sea ice transitions are expected to impact sea ice area and thickness because seasonal transitions are strongly linked to the ice-albedo feedback (Perovich et al., 2008;Timmermans, 2015;Kashiwase et al., 2017;Perovich, 2018;Lebrun et al., 2019). Ice loss earlier in the spring has been related to later ice gain in the fall (Stroeve et al., 2014Lebrun et al., 2019), and a weaker relationship has been described between later ice gain and earlier spring loss Here we demonstrate year-round relationships using seasonal transition dates, March mean ice thickness and summer (June-September) mean ice area. We show that pan-Arctic relationships between seasonal transitions and other ice characteristics are most discernible using seasonal transition metrics with extensive spatial coverage (Fig. 10). Summer mean ice area is evaluated 285 instead of the ice area of a single month in order to better represent the integrated surface energy absorption as ice area declines.
Ice area and seasonal ice transition dates are practical for assessing sea ice in a pan-Arctic sense, as they are reliably available for both models and observations. Discussion of the sea ice thickness here is limited to model projections, since observations of Arctic sea ice thickness are temporally limited and contain large uncertainties (Bunzel et al., 2018).
We find that in satellite data, mean summer ice area (June-September) and the median timing of freeze onset are strongly 290 anti-correlated (R=-0.93) ( Table 5 and Fig. 10). Lower summer ice area corresponds to a lower surface albedo, allowing for greater shortwave absorption by the surface ocean and increasing ocean heat content (Timmermans, 2015), delaying the freeze onset (Stroeve et al., 2014). Slightly weaker relationships exist between mean observed summer sea ice area and freeze-up (R=-0.64) and closing (R=-0.81). In models, the greatest agreement on the correlation between mean summer ice area and fall transition metrics is seen using the freeze onset dates, where all of the correlation coefficients that are statistically significant 295 at the 95% level (13 out of 16 models) are more negative than -0.73 (Table 5). Models tend to show later freeze onset than observed, as discussed in Sect. 4.2, and despite this offset the observed relationship between summer ice area and freeze onset  Figure 10. Scatter plots of median seasonal sea ice transition metrics versus other ice characteristics for CMIP6 models (colors), CESM LE (gray) and satellite data (black). Each scatter point represents one year in one ensemble member from 1979-2014. Panels (a-d) show relationships with median melt and freeze onset dates, panels (e-h) show relationships with median opening and closing dates and panels (i-l) show relationships with median break-up and freeze-up dates. Metrics are scattered against mean summer (June-September) ice area in the first and fourth columns and March mean ice thickness in the second and third columns. All metrics are scattered against ice characteristics from the same year, except those in the second column, in which fall transition metrics are scattered against the next year's mean March ice thickness. All available ensemble members are used for CESM LE, CanESM5 and IPSL. All metrics are calculated between 66-84.5°N. All models are represented in each panel (a)-(l), but the labels are distributed across panels (h), (k) and (l).
19 https://doi.org/10.5194/tc-2020-81 Preprint. Discussion started: 3 April 2020 c Author(s) 2020. CC BY 4.0 License. is captured well by the models (Fig. 10). Summer ice area is generally larger in satellite data than in the models, but falls within the model spread. Relationships between fall transition dates and mean summer ice thickness are similar but slightly weaker than found with ice area (Supplementary Table S3).

300
In the models, the timing of fall transition dates are strongly correlated with the March mean ice thickness (Table 6 and  Table S3). The differences in correlation coefficients indicate that increased heat absorption and delayed freeze onset reduce the March thickness of the ice but have a much smaller impact on the ice area. This supports past work on the Canada Basin, showing that anomalous solar heat input (Perovich et al., 2008) reduced ice thickness over the winter of 2007-2008 by 25% (Timmermans, 2015). The strongest 305 correlations between March ice thickness and the previous year's fall transition metrics are found between freeze onset and March ice thickness, with statistically significant correlation coefficients ranging between -0.54 and -0.92 in the models ( Table   6). Because of sea ice thickness uncertainties discussed earlier (Bunzel et al., 2018), we are unable to confidently evaluate 20 https://doi.org/10.5194/tc-2020-81 Preprint. Discussion started: 3 April 2020 c Author(s) 2020. CC BY 4.0 License.
whether model biases in freeze onset impact the simulated relationship between freeze onset and March mean ice thickness compared to observations. With respect to the other fall transition metrics, we find that statistically significant correlations 310 between March ice thickness and freeze-up/closing (which are both based on ice concentration) are less consistent between models, and generally stronger for the closing dates rather than freeze-up dates (Table 6).  (Table 6). For March ice thickness and opening dates, statistically significant correlations range from 0.14-0.65 (Table 6). One might expect that thinner ice would correspond to earlier breakup dates, because thinner ice is easier to melt out or split apart. However, models do not agree on the sign or statistical significance of any relationship between break-up (which are defined using ice concentration, like opening dates) and March mean ice thickness. This lack of relationship is a strong indication that the spatial coverage of break-up dates is not sufficient for describing pan-Arctic sea ice feedbacks. Relationships between spring transition dates and March ice area are weaker than those between spring transition dates and March ice thickness (Supplementary Table S4).
Melt onset and opening are related to mean summer ice area in both satellite data and models (Table 5 and Fig. 10) (excluding the CNRM models, which are discussed in Sect. 4.5). Earlier melt is correlated with lower mean summer ice area with a correlation coefficient of 0.81 in satellite data and a range of 0.37-0.85 in statistically significant model correlations (Table 6).

325
Earlier opening is slightly less correlated with lower mean summer ice area with a correlation coefficient of 0.72 in satellite data, and the models range between 0.41-0.89 in statistically significant correlations (Table 6). Both earlier melt onset and opening dates decrease the surface albedo-the former though the formation of melt ponds and the latter through the presence of more open ocean. This once again facilitates greater surface absorption, which has been shown to increase the ocean heat content and decrease the summer sea ice cover (Stroeve et al., 2014). Relationships also exist between melt onset/opening and 330 summer sea ice thickness (Supplementary Table S3), but since the summer sea ice is already quite thin, greater ocean heat content is more likely to affect the ice area than it is in March, when ice is much thicker overall. Models do not agree on the sign or magnitude of the correlation between break-up and summer ice area, again indicating that the spatial coverage of break-up dates is insufficient for describing pan-Arctic sea ice processes. days later than the next latest model and considerably later (14-18 days) than those found in satellite data (Table 4). The largest differences in melt onset between the CNRM models and both satellite data and the other models are found in the Central Arctic ( Fig. 3k, l). While melt onset dates fall late in the CNRM models, their September ice areas are overall realistic (Voldoire et al.,340 2019) and fall within the spread of available models ( Supplementary Fig. S1). The CNRM models are the only two models (out of sixteen) that the lack statistically significant correlations between later melt onset and larger summer ice area seen in most models and observations (Sect. 4.4). Furthermore, mean ice thickness in the CNRM models from 1979-2014 is too low ( Supplementary Fig. S1). Thus, the models' ability to produce realistic September sea ice areas likely relies on the biased seasonal transition: late melt onset acts to retain thin ice that would otherwise be lost over the summer by shortening the length 345 of the melt season. Following melt onset, the CNRM models have median opening and break-up dates that fall fully within the model spread, indicating that the impact of seasonal transition biases can be be large, even if the biases exist only in one metric.
The cause of delayed melt onset in CNRM models is not currently clear. Melt onset is the only transition metric that captures changes at the surface of the snowpack rather than a change in ice concentration. Recent work suggests that the winter snow 350 on the sea ice is too thick in the CNRM models (Voldoire et al., 2019), over-insulating the sea ice and preventing it from reaching realistic ice thicknesses (Voldoire et al., 2019). We find that the over-insulation in CNRM models may be more related to September-November snow thickness, since the CNRM models show the largest area of 15-30 cm deep snow of all the models across this time frame (Supplementary Fig. S3), but show similar snow thicknesses compared to other models during December-February ( Supplementary Fig. S4). Delayed melt onset could also be related to the use of the GELATO 355 sea ice model, as the CNRM models are the only models used in this study that use the GELATO model (Supplementary   Table S2). Since GELATO has a single snow-on-sea-ice layer and fixed albedos for dry snow and melting snow (0.88 and 0.77 respectively) (Voldoire et al., 2019), simplified processes in GELATO may contribute to snow biases.

Conclusions
Seasonal sea ice transitions can be characterized by various metrics (melt onset, opening, break-up, freeze onset, freeze-up 360 and closing), and each metric represents a distinct stage of sea ice loss or gain. As such, seasonal transitions provide unique insights into Arctic sea ice processes, but they have so far been under-utilized in evaluating climate models due to a lack of long-term observational products and daily model output, as well as the complexities of defining seasonal transitions. Taking advantage of newly available daily model output (Notz et al., 2016) and observational data of seasonal transitions (Steele et al., 2019), we show that models capture the observed asymmetry in seasonal sea ice transitions, with spring ice loss taking about 365 1.5-2 months longer than fall ice growth (Figs. 3-8). Models also generally agree with satellite data on the timing of spring transitions, but eleven out of sixteen models show median freeze onset dates later than observed, such that the differences between each model's median freeze onset date and the observed date exceed the largest estimations of internal variability (Table 3). Likewise, in almost half of the models (seven out of sixteen), the difference between the median freeze-up date and the observed date exceed the largest estimations of internal variability. Delayed freeze onset and freeze-up extend simulated 370 melt seasons and open water periods respectively, making the outer ice-free period (the time between ice opening and closing) the only inter-seasonal period in which models consistently agree with satellite observations.
We find that differences in seasonal transitions between models are unlikely due to internal variability alone, and are hence likely a reflection of model differences. Sea ice metrics are each impacted differently by internal variability: models do not agree on a metric most affected, and no single model exhibits the greatest internal variability across all metrics. Despite 375 the uncertainty associated with internal variability, all metrics show pan-Arctic model spreads exceeding even the largest estimations of internal variability in seasonal sea ice transition metrics (Tables 3 and 4). The largest standard deviations between ensemble members are seen in the inflow regions of melt and freeze onset dates (Fig. 9), and this is due to the changing interannual position of the ice edge and the variability of surface temperature.
Because differences in seasonal sea ice transition metrics between models are unlikely due only to internal variability, these 380 metrics can be used for evaluating differences between models in terms of other sea ice characteristics. We show that pan-Arctic relationships between transition metrics and sea ice area and thickness depend on the spatial coverage of the metric (Fig. 10).
Out of the six transition dates, melt and freeze onset dates consistently cover the largest area of the Arctic, and they are most closely related to pan-Arctic ice area and mean thickness. Low mean summer ice area delays freeze onset (Table 5), which in turn leads to lower March ice thickness (Table 6). Thinner March ice leads to earlier melt onset and, again, low mean summer dates, but almost indistinguishable using break-up and freeze-up dates (Fig. 10). Since the differences in relationship strengths are seen across definitions that use both surface temperature and ice concentration-based definitions, these differences are more likely related to the spatial coverage of the seasonal sea ice transition dates rather than their defining variables (Tables 5 and 6, Fig. 10). While models tend to show later freeze onset than observed, this offset does not impact the ability of the models to 390 produce the observed relationship between lower summer ice area and later freeze onset.
Finally, we demonstrate how seasonal sea ice transition metrics can provide context to sea ice changes that otherwise lack quantified explanations. We find that CNRM-ESM2-1 and CNRM-CM6-1 exhibit biases in both melt onset (late) and ice thickness (thin), exemplifying how seasonal ice transitions can compensate for other unrealistic aspects of the sea ice. Late melt onset helps retain thin ice throughout the summer such that both CNRM models exhibit realistic September sea ice areas 395 for the wrong reasons. Seasonal sea ice transitions metrics therefore provide a process-based constraint on model simulations in addition to the commonly used September and March sea ice areas Rosenblum and Eisenman, 2017).
To conclude, routinely saved daily sea ice variable output (in particular sea ice concentration and surface temperature) will be critical for using seasonal transitions as a new metric to assess and quantify model uncertainties associated with Arctic sea ice simulations. Since a new observational data product for these seasonal sea ice transition now exists , 400 seasonal sea ice transition dates should be used routinely in the future to better identify model biases in sea ice evolution as well as the sources of these biases. Author contributions. AS and AJ conceived the study, and AS analyzed the data and prepared the manuscript, with guidance and edits from AJ and MW.