We rely on blowing snow mass flux and friction velocity measurements obtained by from 6 September 2012 to 7 October 2012 to evaluate the blowing snow routine at S10. A snow particle counter SPC; was kept at approximately 1.00±0.15 m above the surface using a vertically adjustable frame connected to an SR50 snow height sensor Table ;. To save energy, the SPC was turned off at wind speeds below 5 ms-1, when blowing snow is not expected to occur . The set-up also included an 8 m high profile tower equipped with a CSAT3 sonic anemometer (Table ) at approximately 5 m height, which was used to derive the observed friction velocity.

The SPC is self-steering using a wind vane and measures horizontal particle number fluxes for 64 radius classes with a super-luminescent diode sensor, ranging from 18 to 245 µm . As blowing snow particles approach perfect ice spheres more closely than other forms of snow , we assume that the blowing snow particles are perfectly rounded. Therefore, the integrated horizontal mass flux q [kgm-2s-1] is given by: 1q=∑cr=164qr=∑cr=164nr43πr3ρice, where qr is the mass flux per radius class [kgm-2s-1], cr is the index of the 64 particle radius classes with median radius r [m], nr is the measured number flux [m-2s-1], and ρice=917 kgm-3 is the density of ice.

The mass flux from the SPC includes the effect of snowfall as the instrument counts particles from both blowing snow transport and snowfall . Because RACMO2.4p1 modelled concurrent snowfall during all major blowing snow events within the measurement period, the method of is used to identify and exclude the events in which the measured mass fluxes were severely impacted by snowfall. For blowing snow events without concurrent snowfall, it is known from theory that: (i) blowing snow transport fluxes exhibit a power law with windspeed , (ii) observed particle sizes increase with windspeed , and (iii) snow particle size follows a gamma distribution .

The SPC detected 15 blowing snow events in various weather conditions during its operational period , of which we label the largest five events from A–E during our analysis. Events C and D are excluded from analysis (15 % of datapoints) based on visual inspection of the snowfall detection method of . The measured snow particle size distributions during the five largest blowing snow events are approximately gamma distributed, as expected from theory , except for an anomalous peak in the largest particle radius class (245 µm, not shown). This particle radius class therefore makes a major contribution to the total blowing snow mass transport flux (up to 76 %), whereas contributions from other large particle radius classes were generally below 1 % during blowing snow events without concurrent snowfall. As we could not identify an alternative physical explanation for this observed peak, we surmise that these large particles are precipitating snow and exclude this largest particle radius class from the analysis, following .

We use measurements of the four broadband radiation components from S10 and KAN_U to evaluate modelled radiative fluxes between 18 August 2010 and 22 January 2016, the period of available SRB data from the AWS operated by the IMAU. The downwelling shortwave radiation (SW_↓) measurements from S10 and KAN_U have been corrected for tilt , and the longwave radiation measurements of S10 have been corrected for the window heating effect due to the absorption of shortwave radiation. The temperature and relative humidity observations of S10, which have been corrected for radiation heating effects , are used to evaluate RACMO2.4p1. Both datasets are averaged to an hourly resolution and linearly interpolated to the model output timestamps.

We use stand-alone versions of the blowing snow and radiation subroutines from the Regional Atmospheric Climate Model RACMO2.4p1 (hereafter R24). The polar version (“p”) of this hydrostatic model is developed and maintained at the IMAU and combines the atmospheric dynamics package of the High Resolution Limited Area Model (HIRLAM) version 5.0.3 and the physical processes representation of the Integrated Forecast System (IFS) cycle 47r1 of the European Centre for Medium-Range Weather Forecasts . ERA5 data are used every three hours to force the model at the lateral boundaries and to nudge the model at the upper boundary . R24 was run on the Greenland domain on a 5.5×5.5 km2 horizontal resolution with 40 atmospheric levels and a variable model time step of one to five minutes depending on a numerical stability criterion .

The blowing snow scheme of R24 is based on the double-moment bulk version of the PIEKTUK-model hereafter PIEKTUK-D;. The model calculates horizontal blowing snow transport fluxes and sublimation from horizontal wind speed, temperature, and humidity profiles . Blowing snow is simulated when the friction velocity u* [ms-1] exceeds the threshold friction velocity u*t [ms-1], which depends on characteristics of the upper snow layer such as density, grain size, dendricity, and sphericity . In R24, dendricity and sphericity are not explicitly modelled and are instead prescribed as constant values of 0.5 . The snow mixing ratio in the saltation layer qsalt [kgkg-1] is calculated using a parametrisation by : 2qsalt=esaltghsalt(u*2-u*t2), in which esalt=1/(3.25u*) is the dimensionless saltation efficiency which is derived emperically , g=9.81 ms-2 is the gravitational acceleration, and hsalt=0.08436u*1.27 is the height of the saltation layer [m].

The distribution of blowing snow particles in the saltation layer is described by a two-parameter gamma function : 3F(r)=Nr(α-1)e(-r/β)βαΓ(α), in which N is the number concentration of snow particles [m-3], α=4 and β=100/α [µm] are the shape and scale parameters of the gamma distribution, and Γ(α)=(α-1)! is the gamma function. A relation between the bulk (i.e. integrated over the particle size distribution) blowing snow mixing ratio qb [kgkg-1], snow number concentration N, and mean particle radius rm [m] is found by assuming that blowing snow particles are perfectly spherical : 4rm=α[3ρqbΓ(α)4πρiceΓ(α+3)N]1/3, where ρ is the density of air [kgm-3]. The distribution of snow particles can be calculated from the snow mixing ratio qb and number concentration N due to the assumption that α=4 is constant in PIEKTUK-D.

The evolution of qb and N is governed by eddy diffusivity, settling velocity, and sublimation : 5∂qb∂t=∂∂z(Kb∂qb∂z+vbqb)+Sb6∂N∂t=∂∂z(KN∂N∂z+vNqN)+SN, in which z is height above the surface [m] and Kb,KN=Km=u*l are the eddy diffusivities for qb and N respectively [m2s-1]. Mixing length l=(1/(κ⋅(z+z0m))+1/lmax)-1 [m] is given by κ(z+z0m) close to the surface and converges to a maximum of lmax=20 m at increasing heights. vb and vN are the bulk settling velocities weighted by the third and first moment respectively [ms-1], and Sb and SN are the rate of change in the snow mixing ratio [kgkg-1s-1] and number concentration [m-3s-1] due to sublimation.

The sublimation mass change is given by the formulation of : 7dmdt=(2πrσ-QrKNNuTa[LsRvTa-1])/(LsKNNuTa[LsRvTa-1]+RvTaNShDei), in which σ is the water vapour deficit over ice, Ta is the air temperature [K], K is the thermal conductivity of air [Wm-1K-1], Ls is the latent heat of sublimation [Jkg-1K-1], Rv is the gas constant for water vapour [Jkg-1K-1], D is the molecular diffusivity of water vapour through air [m2s-1], Qr is the net radiation transferred to the ice particles [W], ei is the vapour saturation pressure over ice [Pa], and NNu and NSh are the Nusselt and Sherwood numbers.

Steady-state profiles of blowing snow transport and sublimation are computed by convergence of Eqs. (), (), and () with the mixing ratio and particle concentration in the saltation layer as boundary conditions. Finally, the sublimation rate (-Sb) and latent heat from blowing snow sublimation (LsSb/cp) are added in the prognostic equations of R24 of atmospheric water vapour and temperature.

EcRad-1.4.1 is the stand-alone version of the IFS radiation scheme , which is incorporated in R24 to calculate radiation fluxes every hour . The scheme uses the Rapid Radiative Transfer Model for Global Climate Models RRTM-G; to estimate gas optical properties. The IFS cycle 47r1 includes updated greenhouse gas and aerosol concentrations that vary each month but not between years. The 3-D aerosol data are derived from the Copernicus Atmospheric Monitoring Service (CAMS) reanalysis and trace gases from the Global Environmental Monitoring System (GEMS) and the Monitoring Atmospheric Composition and Climate (MACC) reanalysis . The optical properties of liquid clouds are calculated using the Suite Of Community RAdiative Transfer codes based on Edwards and Slingo SOCRATES;, and the optical properties of ice clouds by . The shortwave and longwave radiation fluxes are computed with the RRTM scheme and are embedded within the Monte Carlo Independent Column Approximation (McICA) framework using an exponential-exponential cloud overlap scheme to calculate the cloud-radiation interactions.

We first assess the ability of R24 to simulate the relevant input variables of the blowing snow routine at S10. Figure  shows the observed and simulated hourly average 2 m temperature, 2 m relative humidity, 10 m horizontal wind speed, and friction velocity during the operational period of the SPC from 6 September 2012 to 7 October 2012. R24 simulates 2 m temperature and 10 m horizontal wind speed with high R2 scores of 0.78 and 0.73, and low biases of -1.38 °C and -0.68 ms-1, respectively (Fig. a, c). report lower biases of -0.63 °C for 2 m temperature and -0.39 ms-1 for 10 m wind speeds when evaluating R24 against multiple IMAU and PROMICE AWSs for a longer observational period. The observed variability of 2 m relative humidity and friction velocity is captured less well by R24 at S10, with lower R2 scores of 0.42 and 0.55, respectively (Fig. b, d).

The stand-alone version of PIEKTUK-D accurately replicates the integrated horizontal blowing snow transport fluxes and sublimation rates of R24, with high R2 scores of 0.97 and 0.99, and low biases of 0.35 gm-2s-1 and -0.02 mmd-1, respectively (Fig. a, b). Figure a shows the measured and simulated horizontal blowing snow mass transport fluxes at S10. Most blowing snow events are accompanied by concurrent snowfall rates above 10-4 kgm-2h-1 in R24. During events C and D, observations do neither exhibit a power law between horizontal blowing snow transport and wind speed (Fig. b), nor a linear relation between mean particle radius and wind speed (Fig. c). The snowfall detection method of therefore indicates that the SPC measurements of events C and D are severely impacted by snowfall, whereas the influence of snowfall during events A, B, and E is modest. During these blowing snow events (A, B, and E), the peak simulated horizontal mass transport fluxes are underestimated by 69 % to 79 % compared to observed transport fluxes, consistent with the findings of during austral summer using the same model over Antarctica. Moreover, the blowing snow scheme overestimates horizontal transport fluxes during the onset of events B, D and E, resulting in an R2 score of 0.50 and a bias of -0.53 gm-2s-1 (Fig. c) even though the model captures the timing of the blowing snow events well.

The baseline experiment (BL) of the offline ecRad-1.4.1 version, which does not include the radiative effects of blowing snow, correctly replicates the surface radiation fluxes of R24, with high R2 scores of 1.00 for shortwave fluxes and 0.99 for longwave fluxes, and biases below 1.5 Wm-2 (Fig. ). Figure  shows the performance of the BL experiment compared to observations at S10 when blowing snow is and is not simulated by PIEKTUK-D, using a threshold of QT=10-3 kgm-1s-1. The BL experiment simulates SW_↓ well, with high R2=0.97 in both cases and low biases of -0.4 Wm-2 without and -3.2 Wm-2 with blowing snow (Fig. a, b). On the other hand, the BL experiment underestimates LW_↓ with lower R2 scores of 0.45 and 0.38 and larger model biases of -22.9 and -22.6 Wm-2 without and with blowing snow, respectively (Fig. c, d), which is a well-known deficiency of R24 .

Next, we show the impact of forcing ecRad-1.4.1 with PIEKTUK-D output using a case study from 12 October 2012 to 17 October 2012. This event was selected due to the large vertically integrated horizontal blowing snow transport fluxes, reaching up to 0.4 kgm-1s-1, and small background liquid and ice water paths (LWP, IWP) from simulated clouds in R24. The impact of blowing snow on the SRB is more pronounced during clear-sky conditions when SW↓ is large and atmospheric emissivity is near its lower limit, allowing blowing snow to both substantially lower SW↓ and increase LW↓. Figure  shows hourly input (panels a–c) and output (panels d–g) of ecRad-1.4.1 during this blowing snow event. The largest background LWP and IWP are observed during short episodes of snowfall in the afternoon of 14 October 2012 and the morning of 15 October 2012 (Fig. b). The BL experiment simulates SW_↓ fluxes accurately during the overcast and snowfall conditions, but underestimates SW_↓ during clear-sky conditions on 12 October 2012, 13 October 2012, and 16 October 2012 (Fig. d). Moreover, the BL experiment consistently underestimates LW_↓ (Fig. f), which is in line with Fig. c and d, and findings of .

Including blowing snow in the forcing of ecRad-1.4.1 decreases SW_↓ and increases LW_↓ at the surface (Fig. e, g), with exceptions of SW_↓ on 14 October 2012 at 16:00 UTC and LW_↓ on 13 October 2012 at 20:00 UTC. The increase in IWP due to blowing snow particles during this case study is proportional to the vertically integrated horizontal blowing snow transport QT for both the RG and IG experiment, and an order of magnitude smaller than the background LWP and IWP from clouds and snowfall (Fig. a, b, c). Therefore, the impact of blowing snow on surface radiation fluxes is best observed during the evening of 12 October 2012 or the morning of 14 October 2012, when background LWP and IWP are smallest.

The hourly input and output profiles of the IG experiment during the case study are shown in Fig. . The blowing snow layer (qblsn>10-6 kgkg-1) extends up to ∼ 175 m, and blowing snow mixing ratios and effective radii are largest near the surface, and decrease with height (Fig. a, b). The total ice mixing ratio and ice effective radii increase above the blowing snow layer during the three episodes of snowfall. Blowing snow increases the shortwave heating rates of the atmospheric layer near the surface from ∼ 1 Kd-1 in BL (not shown) up to 35 Kd-1 in IG due to the absorption of shortwave radiation, and decreases the longwave cooling rates from ∼ 1 Kd-1 in BL (not shown) up to 150 Kd-1 in IG due to increased atmospheric emissivity (Fig. c, d).

The RACMO grid (RG) and integrated grid (IG) experiments are compared to observations at S10 during blowing snow events (QT>10-3 kgm-1s-1) during the AWS operational period from 2012 and 2016 (Fig. ). Forcing ecRad-1.4.1 with blowing snow in the IG experiment results in an average decrease in SW_↓ of 1.2 Wm-2 and an average increase in LW↓ of 5.8 Wm-2 during blowing snow events compared to the BL experiment. Differences between the RG and BL experiments are smaller (ΔSW↓=-0.5 Wm-2, ΔLW↓=+2.4 Wm-2) because the blowing snow grid levels close to the surface, which contain the largest blowing snow mixing ratios, are neglected in this experiment (Fig. ). For SW_↓, the RG and IG experiments perform similarly to the BL experiment, with R2=0.97 and slightly higher biases of -3.7 and -4.6 Wm-2, respectively (Fig. a, b). The performance in simulating LW_↓ improves for both the RG and IG experiments with larger R2 scores, smaller root mean square error (RMSE), and smaller biases (Fig. c, d), where the IG performs best with an R2=0.5 and a LW_↓ bias of -16.7 Wm-2. Overall, the SRB improves most in the IG experiment, with a 3.4 Wm-2 reduction in RMSE and a 4.6 Wm-2 reduction in bias during blowing snow events compared to the BL experiment, as the improvement in LW_↓ exceeds the deterioration in SW_↓.

The simulated radiation fluxes of the BL, RG, and IG experiments are compared to observations at S10 in Fig.  during blowing snow events when R24 correctly simulates little to no clouds or fully overcast conditions. Datapoints are selected when both the simulated cloud fraction of R24 and the observed cloud fraction are below 0.2 for clear-sky conditions and above 0.8 for the fully overcast conditions. The cloud fraction is estimated from observations following the method of and . The method assumes that the clear-sky (cloud fraction = 0) and fully overcast (cloud fraction = 1) conditions are represented by a second-order polynomial fitted to the 5th and 95th percentile levels of LW_↓ plotted against 2 m temperature. Cloud fractions in between these lower and upper bounds are obtained by a linear interpolation between the fitted polynomials. Note that the resulting estimates are subject to uncertainty. Again, forcing ecRad-1.4.1 with PIEKTUK-D output decreases SW_↓ and increases LW_↓ during both experiments, but the impact of coupling is most clearly visible in LW_↓ during clear-sky conditions (Fig. c), as discussed above.