Evaluation of coastal Antarctic precipitation in MAR3.9 regional and LMDz6 global atmospheric model with ground-based radar observations

In the current context of climate change in the poles, one of the objectives of the APRES3 (Antarctic Precipitation Remote Sensing from Surface and Space) project is to characterize the vertical structure of precipitation in order to better simulate it. Nowadays, the precipitation simulated by models in Antarctica is very widespread and overestimated the data. Sensitivity studies have been conducted using two models and compared to the observations obtained at the Dumont d’Urville coast station, 5 obtained by a Micro Rain Radar (MRR). The MAR meso-scale model specifically developed for the polar regions and the LMDz/IPSL general circulation model, with zoomed configuration over Dumont d’Urville, have been considered for this study. These models being different in resolution and physical configuration, performing an inter-comparison required numerical, dynamic and physical adjustments in LMDz. A sensitivity study was conducted on the physical and numerical parameters of the LMDz model and on the resolution of the MAR with the aim of estimating their contribution to the precipitation simulation. 10 Sensitivity tests with MAR revealed that this model is well adjusted for precipitation modeling in polar climates, this confirming that this model is a reference in polar climate modeling. Regarding LMDz, sensitivity experiments revealed that modifications in the sedimentation and sublimation parameters do not significantly impact precipitation rate. However, dissipation of the LMDz model, which is a numerical process that dissipates spatially excessive energy and keeps the model stable, impacts 1 https://doi.org/10.5194/tc-2020-167 Preprint. Discussion started: 26 August 2020 c © Author(s) 2020. CC BY 4.0 License.

by more than 100%. And even though the simulated surface precipitation is compared to an observation level at an altitude of 1200 meters above the local surface, the discrepancy between data and models is large, and questionable for the future prediction of precipitation. In addition, the agreement between data and models is even worse for the simulation of precipitation on the plateau than over the peripheral regions (Palerme et al., 2017;Roussel et al., 2019).
Since November 2015, during a field campaign at the French base in Dumont d'Urville, instruments have been installed, 5 including a Micro Rain Radar (MRR) observing clouds and precipitation particles from surface (Grazioli et al., 2017a). This instrument has provided a continuous vertical structure of precipitation and its climatology. Among other results, this has highlighted the sublimation of precipitation by katabatic winds, as well as providing information on the mean sedimentation rate of precipitation (Grazioli et al., 2017b;Durán-Alarcón et al., 2019). This vertical profile is also an excellent tool for evaluating the simulated vertical structure of precipitation.

10
In this study, we propose to evaluate the vertical structure of precipitation at Dumont d'Urville, simulated by two different models with the MRR dataset. The first model is the general circulation model LMDz, an atmospheric component of the coupled IPSL model. The second model is the mesoscale model Modèle Atmosphérique Régional (MAR). Each of these models having different degrees of complexity because of different uses, it is important to verify how precipitation is simulated by these two models, and especially to verify if the vertical profile of precipitation is in agreement with the observed profile. In section 15 2, each model configuration and the ground radar observations are presented to do this study. The sensitivity experiments performed on each model and their results are discussed in section 3. Then, an exploration of numerical dissipation in the LMDz model applied to temperature and its impact in simulated precipitation is discussed in section 4. Finally, we conclude this study in section 5.

The LMDz-IPSL climate model
The LMDz dynamical core is based on finite difference and finite volume discretization of the primitive equations of meteorology and transport equations, coupled to a set of physical parameterizations (Hourdin et al., 2013). The radiative transfer scheme is the Rapid Radiative Transfer Model (RRTM) from Mlawer et al. (1997), also used in MAR. The microphysical cloud scheme is from statistical type and includes large scale condensation. A fraction f iw of the condensed water q c is assumed to 25 be frozen, depending on the temperature between 273.15 K where f iw = 0 and 243.15 K where f iw = 1. Then a fraction of the condensed water is partially precipitated according to Zender and Kiehl (1997). The associated sink of cloud water is: where w iw = γ iw × w 0 , w 0 = 3.29(ρq iw ) 0.16 being the characteristic sedimentation rate of ice crystals given by Heymsfield and Donner (1990) depending on the solid cloud water and γ iw being a tunable parameter. Precipitation is then re-evaporated and included into the vapor water following : where P is the precipitation flux and β is a tunable parameter.
This model configuration only admits the atmospheric model, without taking into account vegetation or ocean circulation models. However, there is a surface scheme. It is composed of four categories: oceans, continental surfaces, sea-ice and glaciers.

5
The surface fluxes are calculated by taking into account the parameters of each type of surface. It is important to note that for desert surfaces such as ice caps, a skin effect model is used to describe surface flows. In order to have the better resolution possible above Dumont d'Urville with a GCM, the model is stretched longitudinally and latitudinally, reaching a horizontal resolution of ∼25 km. We nudged the LMDz model with wind, temperature and humidity ERA-Interim reanalysis outside the zoom. It is nudge-free inside the zoomed area (Coindreau et al., 2007). This makes it possible to represent the processes that 10 are part of the LMDz model in atmospheric situations that are close to the real condition of the atmosphere. It has 79 vertical levels in its current configuration, with refinement in the boundary layer and troposphere. The vertical precipitation profile studied at Dumont d'Urville in the LMDz model is selected over continental surface. A spin-up of 4 months is necessary to balance the model, then each simulation is conducted for one month corresponding to our dataset period.

15
MAR is a primitive equations hydrostatic model, developed for polar regions studies. Its dynamical core (Gallée and Schayes, 1994) and its turbulent scheme (Duynkerke, 1988) are designed to replicate classical linear mountain waves such as katabatic winds. The hydrological cycle includes a cloud microphysical model, with conservation equations for cloud droplet, rainfall, cloud ice crystal and snow flake distributions. Blowing snow is included in this scheme and these particles are considered as snow flakes. The sublimation of snow flakes is function of the ice relative humidity, according to (Lin et al., 1983). The MAR 20 model represents accurately the atmospheric boundary layer, blown snow processes and their interactions with katabatic winds (Agosta et al., 2019). However, we have not enabled this in our study to save computing time because the processes of interest occur at higher altitudes. The representation of the cloud microphysical processes is essentially based on the parameterizations of Kessler (1969). The atmospheric component of MAR is coupled to a snow pack model (Gallée and Duynkerke, 1997), enhanced by metamorphism laws of the CROCUS snow model (Brun et al., 1992).

25
MAR is forced in wind, temperature and humidity with ERA-Interim reanalysis outside the domain of study. The optimal configuration of MAR is a 5 km horizontal resolution with a considered domain of 1000x1000 km, as well as 40 vertical levels between the surface and the top of the troposphere (about 8000 m above sea level). MAR is accurate on the surface and in the boundary layer, the first level being 15 cm from the surface. The vertical precipitation profile studied at Dumont d'Urville in MAR is selected over continental surface. The MAR model requires a shorter spin-up than the LMDz, however to facilitate 30 comparison with the LMDz, the same duration is chosen.

Micro Rain Radar (MRR) observations
The MRR is a vertically profiling Doppler radar operating at a frequency of 24.3 GHz (K-band) and having a beamwidth of 2°(around 50 m in diameter at 3000 m). The vertical resolution is set to 100 m per bin ranging from 300 -first valid available measurements -to 3000 m (Grazioli et al., 2017a). The MRR's raw measurement -Doppler spectral densities -are available at 10s temporal resolution then minute averaged. The collected data are processed using the IMProTool developed by 5 Maahn and Kollias (2012). The radar reflectivity derived from MRR was calibrated by comparison with a colocated X-band polarimetric radar over the period from December 2015 to January 2016 (for more details, see Grazioli et al. (2017a)). Through this calibration with the second radar, the reflectivity (at X-band) is converted into snowfall rates using a radar reflectivity Ze / snowfall rate Sr relation (Grazioli et al., 2017a) : with Z e the radar reflectivity (in dBZ) and S r the snowfall rate (in mm/hr). Grazioli et al. (2017a), proposed a range of values of [69][70][71][72][73][74][75][76][77][78][79][80][81][82][83] for the prefactor and [0.78-1.09] for the exponent corresponding to a confidence interval of 95 %.
The period selected for this study is February 2017. During this period precipitation events are particularly frequent with different amplitudes and durations. Rather than studying a particular event, we focus on the monthly accumulation of precipitation at each vertical level of the MRR. The monthly accumulation of precipitation is presented in figure 1. The sublimated 15 part of the precipitation can be clearly observed below 1000 meters, due to katabatic winds (Grazioli et al., 2017b).

Description of sensitivity experiments
Mesoscale models are very sensitive to horizontal resolution, as the consideration of many parameterizations will strongly depend on it. We used different dimensions of the grids as well as the size of the domain under study. Two domains are presented: the first is 250 km x 250 km, the second is 1000 km x 1000 km, and they are called respectively SMALL and BIG.
These areas are presented in figure 2. Two horizontal resolutions are also used. The first one is 5 km and the second is a coarser 5 resolution of 25 km. Three simulations have been conducted. The first one, which corresponds to the standard simulation, has a BIG domain and a fine resolution of 5 km, the second has a BIG domain and a resolution of 25 km and the third has a SMALL domain and a resolution of 25 km. The second simulation aims at differentiating the impact of the horizontal resolution of the domain size. The purpose of the two last simulations is to get as close as possible to the LMDz horizontal configurations in order to compare both models. Although the last two of these configurations are difficult to achieve in previous versions of the 10 MAR model (Franco et al., 2012), the latest version of this model allows these simulations to be run as "critical" cases of MAR use.
As for MAR on figure 2, we evaluated the horizontal resolution of LMDz by performing simulations on two zoomed domains of different sizes. Indeed, when zooming with the LMDz model, the zoomed region can be widened, so we were able to reproduce the two domains evaluated for MAR. The size of the "SMALL" zoom domain in LMDz allows the model to adapt its 15 own physics inside the zoom in an environment where large-scale wind, temperature and humidity advections are controlled by ERA-Interim reanalyses. The second configuration with a BIG domain is larger, so the model can have its own mesocyclonic circulations within the zoom. The center of the zoom is in this case not very affected by the ERA-Interim reanalysis.
The first sensitivity experiment is evaluating the feedback of the LMDz model to the extent of the nudged-free zoomed domain. Indeed, in the case where the zoom area is restricted in size, the center of the zoom is very sensitive to forcing outside this area. This case is similar to a regional climate model. Inversely, when the zoom area is large, the center of the zoom area 5 is less affected by the forcings imposed on it from the outside and the model is more like a global climate model in a free configuration.
The second experiment studies the sensitivity of solid precipitation to sedimentation velocity rate. To do so, we have tested different values of the parameter w iw in the equation 1 through its parameter γ iw . The different imposed values are summarized in table 1. It is important to note the difference between experiment SedEx 02 whose sedimentation rate tends towards 1 m.s −1 10 and the experiment SedEx 03 whose sedimentation rate is equal to 1 m.s −1 (see equation 1). Indeed, the value of w 0 is varying with q iw and the air density as a function of pressure and temperature. In the SedEx 03 experiment, this variation is not taken into account. Green dashed line corresponds to best MAR configuration with a 5 km horizontal resolution and a BIG domain is in good agreement with MRR vertical observed profile. In this configuration, the model reproduces very well the sublimated part of the precipitation due to the katabatic winds, as already studied by Gallée and Pettré (1998). The inversion point is located at the same altitude as the observed profile. However, too much precipitation is simulated at high altitude, this being a characteristic bias of the models, highlighted by Grazioli et al. (2017b).

10
Green solid line corresponds to 5 km horizontal resolution with SMALL domain. Although the position of the precipitation inversion due to katabatic winds is in agreement with the observed precipitation profile, the amount of simulated precipitation is petite. This is expected since on this type of configuration, the MAR model is very sensitive to ERA-Interim fields outside the simulation domain. Surface precipitation rate is in good agreement with the best MAR configuration with a 5 km horizontal resolution and a BIG domain simulation. 15 The 25 km horizontal resolution with BIG domain represented by green dotted line shows a vertical evolution similar to the standard simulation at high altitude. From 8000 m above the ground level to 2000 m, precipitation accumulation is coherent with the green dashed line simulation but precipitation accumulation still increases when reaching the surface. There does not appear to be any sublimation by katabatic winds. This is also expected, since a grid so deteriorated in resolution does not allow to correctly simulate small scale processes in MAR. Despite the configuration of a small domain in the simulation "25km SMALL", there is an evaporation of precipitation, which is absent in the simulation "25km BIG". We therefore made a comparison between two transects for the particular precipitation  figure 4. In both cases, the katabatic winds are simulated (along the topographical slope). In the "5km BIG" precipitation flux decreases sharply near the surface at Dumont d'Urville. There is a dry air pool just above the ocean surface which seems to have a strong sublimating potential. In the "25km BIG" this cold air masses could be more 5 horizontally spread so it is not thick enough to sublimate precipitation. Moreover, the precipitation seems to be more diffuse.

Horizontal resolution in LMDz
We have evaluated two horizontal configurations of LMDz with different sizes of the zoomed domain. The SMALL configuration is a zoomed domain with a size of 250 x 250 km and the BIG configuration is a zoomed domain with a size of 1000 x 1000 km. It is important to note that there is the same horizontal resolution inside the zoom. Figure 5 shows the accumulation profiles 10 at Dumont d'Urville resulting from this experiment. The BIG simulation simulates a high precipitation accumulation on the surface with 130 mm compared to 55 mm for the SMALL simulation. The two simulated precipitation profiles overestimate the observed accumulation profile. The maximum before inversion of the BIG simulation is below 1000 m, which is in accordance with the observations. The maximum precipitation of the SMALL simulation is at a higher altitude, at 1200 m.  10 https://doi.org/10.5194/tc-2020-167 Preprint. Discussion started: 26 August 2020 c Author(s) 2020. CC BY 4.0 License.

LMDz microphysical parameterizations
Considering that the SMALL configuration of the LMDz model is in better agreement with observations than BIG configuration (see figure 5), and that the large-scale advected fields are well known thanks to ERA-Interim reanalysis, we performed this experiment in order to evaluate the physics of the model only. Figure 6 presents sensitivity experiments summarized in tables 1 and 2, in comparison with MRR vertical observed precipitation accumulation profile. The surface precipitation rate appears to   LMDz, like many GCM, contains a dissipation scheme to prevent the accumulation of energy at scales close to the grid resolution. These accumulations of energy appear when GCM is not resolving turbulent scales at the grid resolution (Jablonowski and Williamson, 2011;Spiga et al., 2018). In the LMDz model, it involves a spatial displacement of dynamic or thermal fields, which can induce, for example, local warming or a variation in dynamics created by purely numerical processes. Thus, a model 5 that is too dissipative may generate precipitation that has no physical relevance.
The dissipation is expressed in LMDz as an iterated Laplacian term on a given variable ψ as follows: where q d is the order of dissipation and τ ψ the damping timescale associated with the variable ψ at the smallest spatial scale l min , depending on the horizontal resolution of the model. q d is an iterative operator, it acts as a filter on the spatial resolution.

10
When q d = 1, the process is overly dissipative on circulations at large scales and at higher values, dissipation occurs more at the grid scale than at the large scale. Large values of τ ψ means weaker dissipation. Indeed, τ ψ represents the time to dissipate a perturbation on variable ψ developing at the spatial scale l min . The three variables designed by ψ are vorticity and divergence of winds, and potential temperature. They are chosen to set horizontal dissipation on the rotational component of the dynamic flows (q rot d and τ rot , i.e. Rossby waves), its divergent component (q div d and τ div , i.e. gravity waves) and the diabatic perturbations (q h d and τ h , i.e. latent heat of condensation, rain re-evaporation, snow sublimation, ...). In LMDz, and more generally in the GCMs methodology, q d and τ ψ are determined empirically. A trade-off between model stability, damping energy at the smallest scales and minimizing impact on the large-scale flows is sought. There are general rules for refining the dissipation parameters for LMDz, with q d ranging between 1 and 4, and τ ψ taking values ranging between one and two hours for a 0.5 • -1 • GCM simulation. The standard configuration of the LMDz model uses as dissipation values 20 q div d = 1, q rot d = 2, q h d = 2 as operators and τ div = 600 s, τ rot = 1200 s, τ h = 1200 s as timescales.

Sensitivity experiments results
In order to study and understand the impact of the different dissipation parameters on precipitation, we have performed sensitivity tests that are summarized in the table 3. For all sensitivity tests, the resulting simulations are less dissipative than the control simulation. The corresponding vertical precipitation accumulation profiles are shown in the figure 8. These experiments 25 were performed on the two configurations of the LMDz under consideration, the results and behaviours are similar but we will only present those performed on the SMALL configuration, which has a precipitation profile closer to the observed profile (see figure 5).
In a general way, sensitivity experiments on q div d and q rot d parameters have little impact on precipitation. The same applies to the τ div and τ rot parameters. However, the dissipation applied to the parameters q h d and τ h has a strong impact on the 30 dissipation profile, as observed on the simulations D03, D06 and D11. For the D07 simulation, where all q d parameters are modified, it can be deduced that the excellent agreement between the simulated and observed precipitation is due mainly to the modifications on diabatic perturbations. Finally, the D09 experiment best reproduces the MRR observations. Indeed, the simulated profile is very close to the observed profile and within the confidence range of the instrument.

Discussion on the dissipation adjustment
In order to study and understand how dissipation affects precipitation, we have investigated the time series of temperatures of the control simulation and the D09 simulation with the best results relative to the MRR observations. They are presented in the figure 9. The impact of the dissipation is mainly visible at low altitude, where the control model is about 3°C warmer than the D09 simulation. In addition, when a precipitation event occurs (e.g., February 1, 10, 14, and 21), the control simulation is 5 warmer than the D09 simulation, which can result in higher precipitation rates being triggered by higher temperature gradients and moister atmospheric masses. When time series are averaged and projected over a larger spatial scale, there is a geographic reorganization of temperature in the less dissipative simulation. In the D09 simulation, the area above Dumont d'Urville is on average colder than in the control simulation. This is due to warmer temperature fields over ocean regions that are less laterally diffused over Antarctic coastal regions. The colors range from blue to red. When the control configuration of the model overestimates temperature compared to the D09 simulation, the color used is red.
As shown in figure 11, as the atmosphere cools over the peripheral regions of Antarctica, air masses become less humid and 5 this has a strong impact on precipitation by concentrating it over ocean regions. Thus, the variation in precipitation observed in the figure 8 corresponds to a horizontal redistribution of precipitation in a less dissipative configuration of the LMDz model.
When comparing the MAR model in its optimal configuration with the D09 simulation of the LMDz model, as shown on figure 12, the average vertical evolution of precipitation is consistent between the two models. This result is interesting because it shows that a model whose microphysics is simplified to satisfy a global issue can correctly simulate solid precipitation in the 10 Antarctic region. Indeed, the LMDz model only contains a precipitation autoconversion equation and a snowfall resublimation equation, but this allows the climate in Dumont d'Urville to be accurately represented during the month of February 2017, and in particular for the katabatic inversion of precipitation. In the case of the LMDz model, which is too dissipative in its control version, the dissipation adjustment takes priority over the microphysics adjustment and this allows precipitation to be redistributed over oceanic rather than continental regions. Thus, there is no excess precipitation of purely numerical origin over 15 Dumont d'Urville and having no physical relevance.  profile, the MAR model seems to simulate too much precipitation. MAR with the same fine grid in a small domain (250 x 250 km) does not simulate enough precipitation because its meso-cyclonic activity is too dependent from the reanalysis fields to be fully effective. With a coarser grid (25 km of horizontal resolution) in the same domain size, MAR does not efficiently 10 generate the processes related to katabatic winds and dry cold shallow layers above the oceanic surface, which generates too much snow due to an absence of sublimation.
Variations in microphysical parameters related to LMDz precipitation have a small impact on the simulated precipitation profile. However, LMDz is very sensitive to the size of its zoomed region as well as to the advections of large-scale fields of winds, temperatures and humidity of ERA-Interim reanalysis. Indeed, in a large domain, analogous to MAR standard 15 configuration, where the model is able to generate its own mesoscale circulation, moisture is concentrated above Dumont d'Urville area and warm and moist bias is generated over the continent near the coasts (blue patterns on fig. 7.d). This is not an expected outcome. When a correct general circulation is forced by configuring a small zoomed region where the centre of the zoom remains influenced by the ERA-Interim reanalysis and by improving the GCM dissipation adjustment in a less dissipative way, the model generates a precipitation profile at Dumont d'Urville that is in excellent agreement with the observed profile. 20 Numerical parameters that guarantee the stability of a model, such as dissipation, often require empirical adjustments.
Dissipation being applied in cases of excess energy to be diffused at the mesh scale, the large-scale currents are not significantly affected by this numerical setting. Thus, the use of observations such as local precipitation rather than large-scale field can be an excellent tool for the fine-tuning of the dissipation of a model, as illustrated here with the LMDz model. This study showed that a better adjusted GCM model such as LMDz, or a mesoscale model such as MAR are correct for assessing the climate in 25 Antarctica and provide an additional element to the major problem of calculating the mass balance in Antarctica.