Improved simulation of Antarctic sea ice due to the radiative effects of falling snow

As historical simulations end in 2005 and satellite sea-ice series begin in the late 1970s, we select output from 1980–2005 and for each grid-cell take the mean of the CMIP5 simulations for a given month and given property. There is little difference in the average properties between the CMIP3 and CMIP5 ensembles so we only report CMIP5 results here (supplementary figure 1). All CMIP5 models together are reported the Multi-Model Mean (CMIP5 M MM, Taylor et al 2012), and we separately present the mean for those models both with and without falling-snow radiative effects (CMIP5-S and CMIP5 NoS respectively, see supplementary table 1).

For our intervention 'snow radiative effects on' (S) and 'snow radiative effects off' (NoS) simulations we use the Community Earth System Model version 1 (CESM1) managed by National Center for Atmospheric Research (NCAR) and Department of Energy (DOE) which is composed of four separate models that simultaneously simulate Earth's atmosphere, ocean, land surface, and sea-ice. Model code and documentation are available from www.cesm.ucar.edu/models/cesm1.0/. The atmospheric model is the Community Atmosphere Model version 5 (CAM5), and snow in this model represents large ice crystals diagnosed from the falling ice mass flux at each model level and time step [Morrison and Gettelman 2008, Neale et al 2012]. The model uses two-moment, stratiform cloud microphysics scheme described by Morrison and Gettelman (2008) which accounts for diagnosed snow mass. Snow is included in the radiation code [Gettelman et al 2010], using the diagnosed mass and effective radius of falling snow crystals [Morrison and Gettelman 2008]. The simulated prognostic ice and diagnostic snow are comparable against CloudSat-CALIPSO retrieved products [Gettelman et al 2010]. More detailed descriptions and related references are in supplementary information section 2.

We conduct sensitivity experiments using a fully coupled setup with two simulations over 1850–2005: one excludes the precipitating ice radiative effect (NoS); and the other includes the effect (S). Each simulation otherwise follows the CMIP5 historical protocol, including initialization from the same CESM1 C AM5 CMIP5 300 year preindustrial control (piControl) run.

The global area-weighted mean of the net radiative flux at top of atmosphere (TOA) over the full simulation is −0.17 W m⁻² and 0.12 W m⁻² for the NoS and S cases, respectively. The reflected SW TOA flux is −1.6 W m⁻² for NoS and 0.6 W m⁻² for S and the outgoing LW at TOA is 0.6 W m⁻² for NoS and −1.9 W m⁻² for S.

Changes in model-observation discrepancy between the S and NoS cases are determined from:

$$\begin{equation} \Delta x = \left| {x_{NoS} - x_{obs} } \right| - \left| {x_S - x_{obs} } \right| \end{equation} \tag{ 1 }$$

where x is some property, the subscript NoS represents the value from the NoS simulation, S is from the snow simulation and obs denotes the observation-based value. We use this metric in figures 5(c) and (f), and it is defined such that a positive value represents a decrease in the magnitude of the model-observation discrepancy when snow-radiative effects are included.

We begin by highlighting the importance of including falling snow to cloud properties by considering biases over Antarctica and the Southern Oceans in the CMIP5 NoS ensemble average. We first compare modelled cloud properties with those determined over 2007–2010 from CloudSat-CALIPSO measurements. CloudSat and CALIPSO are satellites in the Afternoon-Train (A-train) constellation that fly along the same reference ground track on a Sun-synchronous orbit with an ascending (northward) local equatorial crossing time near 13:36. CALIPSO and CloudSat are separated by less than a minute, allowing reliable collocation between the CALIPSO-mounted CALIOP LIDAR and the CloudSat-mounted Cloud-Profiling Radar. We use the CALIPSO-CloudSat 2 C-ICE product [Deng et al 2010, 2013] over the region 40–82 °S where the poleward edge is limited by the satellites' orbit. Cloud ice water path (CIWP) is identified using the FLAG method [Li et al 2012] which separates cloud-only (floating or suspended cloud ice) and precipitating + convective clouds, and we use both in our comparison. Because of negligible sampling biases of cloud water, cloud formation etc [Li et al 2012, Guan et al 2013], the uncertainties related to the diurnal cycle and sampling issues are not specifically considered in the model-observation comparisons in this study.

For an observation-based estimate of surface radiation we use the satellite-based Clouds and the Earth's Radiant Energy System-Energy Balanced and Filled (CERES-EBAF) Surface products over 2000–2010, which are widely used and have been validated with surface measurements, with a monthly-grid-mean uncertainty of ± 10 W m⁻² for shortwave and ±14 W m⁻² for longwave [Kato et al 2011, Kato et al 2012, 2013]. We select the same 40–82 °S region and use each of the radiative energy-budget components: longwave up and down plus shortwave up and down.

We take gridded monthly-averaged sea-ice concentration data from the National Snow and Ice Data Center (NSIDC) over 1980–2005 derived from the NASA Team Algorithm [Cavalieri et al 2012, 1999]. These are based on measurements from passive microwave sensors beginning with the Scanning Multichannel Microwave Radiometer (SMMR) aboard the Nimbus-7 satellite and proceeding with the Special Sensor Microwave/Imager (SSMI) instruments aboard the Defense Meteorological Satellite Program (DMSP) F8, F11, F13, and F17 platforms. The data's native resolution of 25 km is remapped onto a 1°×1° latitude-longitude grid and we use data from 50–70°S, excluding areas closer to Antarctica to avoid discrepancies in mapping of the Antarctic coast and its ice shelves. 82% of the magnitude of the annual cycle of sea-ice area is captured within this area, so our results apply to the majority of the sea-ice area.

Figure 1 shows that the CMIP5 NoS models have little bias in a like-with-like comparison with CloudSat-CALIPSO Cloud Ice-Water Path (CIWP; cloud ice only) in both Austral winter (figure 1(a): June-July-August, JJA) and Austral summer (figure 1(e): December-January-February, DJF). However, the CMIP5 NoS substantially underestimates total IWP (TIWP) over and near Antarctica once precipitating clouds are included (figure 1(b): JJA; figure 1(f): DJF). This contributes to a multi-model mean (MMM) bias in total ice water path of −90.7 g m⁻² with relative bias of −72% averaged over the region 40–82°S, relative to the CloudSat-CALIPSO 2 C-ICE product [Deng et al 2010, 2013]. The spatial patterns and magnitudes of TIWP biases between CMIP3 and CMIP5 are similar with an absolute bias larger than 110 g m⁻² over the Southern Ocean, implying a zeroth-order deficiency embedded in both NoS CMIP3 and NoS CMIP5 GCMs (see supplementary figure 1). Figure 1 also shows cloud and net surface radiation biases between CMIP5 NoS and CERES-EBAF (figure 1(c): JJA; figure 1(g): DJF) and sea-ice concentration biases (figure 1(d): JJA; figure 1(h): DJF) between CMIP5 NoS and NSIDC.

Zoom In Zoom Out Reset image size

Given the correlation between patterns of bias in TIWP, surface fluxes and sea-ice concentration, we propose a causal link as follows: models without falling-snow radiative effects underestimate TIWP, reducing surface downward longwave and increasing surface downward shortwave fluxes, which drive changes in sea-ice concentration (see supplementary figures 2 and 3 for the radiation budget split into its components). We support our hypothesis using specialized simulations to isolate and quantify the contribution of falling-snow radiative effects. These specialized simulations are necessary as the comparison in figure 1 includes all potential biases in simulated physical processes between CMIP5 NoS and the real-world state.

For our quantification of the precipitating-ice radiative effects, we present results from our CESM1 simulations both with (S) and without (NoS) the radiative effects of falling snow. Figure 2 shows the model minus observed monthly values of surface radiation budget components and sea-ice concentration, presented for two different latitude bands. This illustrates that the S simulation is consistently closer than the NoS simulation to the observed energy-budget components (figures 2(a), (b), (d) and (e) and sea-ice concentration (figures 2(c) and (f), CMIP5 M MM also shown). The largest remaining bias in the S simulations is in shortwave fluxes during austral summer, which may be related to other cloud-simulation biases such as underestimated supercooled liquid fraction in mixed-phase clouds [Tan et al 2016, Kay et al 2016] and large-scale atmospheric conditions [Simmonds 2015, Holland et al 2017]. Snow radiative effects reduce the size of the JJA longwave bias from 58–70°S relative to CERES-EBAF, from −30 W m⁻² down to −10 W m⁻² (figure 2(d), and the DJF shortwave bias over the same region from approximately +55 W m⁻² to +35 W m⁻². We do not present a formal uncertainty estimate in these fluxes as there is a lack of CERES validation over this region. However, the reported monthly zonal mean uncertainty over oceans for the CERES data are ±10 W m⁻² in downward LW and SW [Kato et al 2012]. Our averaging over 10 year s would favor a smaller error, but the large uncertainties associated with surface albedo would favor a larger value. Taking the reported ±10 W m⁻² as an approximation of the error, the radiative effects of falling snow are similar in size to the observational error, and the NoS simulations are indeed in disagreement with the CERES-EBAF data over much of the year. Of the changes we see, improved model representation of downwelling longwave radiation restricts winter sea-ice growth leading to a decrease of 2.11 × 10⁶ km² in the September peak sea-ice area, whose lower level is maintained by increased shortwave absorption during the summer. This substantially reduces the bias relative to observations.

Zoom In Zoom Out Reset image size

Figure 3 shows a spatial anti-correlation in the NoS minus S (NoS-S) change in JJA. When snow radiative effects are excluded, the underestimated IWP results in increased net surface shortwave flux (RSDS-RSUS: figures 3(a), (b)) and decreased surface downward longwave flux (RLDS: figure 3(c)). Longwave dominates in austral winter, meaning smaller net radiation (figure 3(e)), cooler surface temperature (see supplementary figures 4 and 5) and increased sea-ice concentration (figure 3(f)).

Zoom In Zoom Out Reset image size

Figure 4 shows that during austral summer the situation is more complex. In some regions, such as off the coast of Dronning-Maud land, there is a net decrease in downward radiation (figure 4(e)) along with a net decrease in sea-ice concentration (figure 4(f)). This is counterintuitive, but can be understood from the seasonal cycles in figure 2. In these regions, the increased winter longwave heating restricts sea-ice growth and leads to a reduced surface albedo. Although less sunlight reaches the surface due to atmospheric reflection by falling snow, the surface is darker due to reduced winter sea-ice growth caused by increased wintertime longwave heating. This increases absorption by enough to offset the reduction in downward shortwave radiation at the surface, helping to prevent summertime sea-ice growth in these regions.

Zoom In Zoom Out Reset image size

We consider these effects in more detail and determine their robustness in figure 5, which maps changes in the bias in sea-ice concentration during JJA and DJF for the S and NoS simulations and the relative difference in bias from equation (1). In this case, positive means a reduction in bias magnitude. The inclusion of snow radiative effects leads to a widespread decrease in sea-ice concentration that brings the simulated sea-ice case closer to observations in most regions, with the exception of a band centered near 45°E in JJA.

Zoom In Zoom Out Reset image size

From 50–70 °S the NoS simulations of sea-ice area have a relative bias of 3.01 × 10⁶ km² in DJF and 3.88 × 10⁶ km² in JJA, which is reduced in the S simulations to 1.84 × 10⁶ km² (by 39%) in DJF and 1.75 × 10⁶ km² (by 55%) in JJA. The greatest absolute change is in March-April-May (MAM) where the discrepancy is reduced by 65%. September-October-November (SON) shows a reduction of 54%. Using t-tests applied to the monthly simulated time series, the differences are statistically robust with, for example, sea-ice concentration changes significant at p < 0.01 over much of the Southern Ocean (supplementary figures 6 and 7). The surface radiation budget and surface or near-surface temperature biases are also reduced in most regions (supplementary figures 8 and 9).

Figure 6 shows mean SIC seasonal cycles from a range of sources. Figures 6 (a), (d) and (g) show the 40 m ember full CMIP5 ensemble along with the multi-model mean (MMM) and standard deviation. Figures 6(b), (e) and (h) show where the observations and CESM-1 S and NoS simulations fall within the CMIP5 distribution. Figures 6(c), (f) and (i) show the CMIP5 MMM versus observations along with the mean for CMIP5 NoS models (N = 14) and CMIP5-S models (N = 26). The CMIP5 mean annual cycle is similar to observations, with slight underestimates in maximum extent, as reported elsewhere [Zunz et al 2013, Lenaerts et al 2016, Boucher et al 2013, IPCC 2013]. Consistent with our CESM1 results, the CMIP5-S models show less sea-ice than the CMIP5 NoS models. The CMIP5-S underestimate of sea-ice extent indicates that other aspects of the model design likely overcompensate for the inclusion of snow radiative effects.

Zoom In Zoom Out Reset image size