Evaluation of convective cloud microphysics in numerical weather prediction model with dual-wavelength polarimetric radar observations: methods and examples

: Summary This study examines a series of regional storm-resolving weather forecasting with different microphysics schemes against long-term observations of dual-wavelength polarimetric radars over Munchen area. Although there are many papers inter-compare microphysics schemes available in WRF, the noble aspect of this study is to utilize polarimetric radar simulator and cell-tracking algorithm for more consistent sampling of polarimetric radar observables and dual-wavelength ratio. However, there are some major questions/suggestions related to 1) missing citations, 2) separation of convective cells, and 3) actual rain drop-size distributions, and 4) uncertainties of the forward model. By improving these issues, this manuscript could be quite powerful. Thus, my suggestion is “major revisions” in order to publish in ACP. We like to thank the reviewer for the valuable comments which helped to improve the manuscript quality substantially. Please find below our point-by-point reply highlighted in blue. A marked-up manuscript version showing the changes made is provided along with the revised manuscript. for microphysics finger print from polarimetric radar signals. This describes various microphysics finger prints to microphysics processes.

Thank you for your suggestions. We included the proposed literature regarding polarimetric radar forward operators to evaluate cloud model microphysics in our introduction: There are some studies that directly use polarimetric radar forward operators to evaluate the performance of cloud microphysics schemes. For instance, Jung et al. (2010) and Snyder et al. (2017) each simulate idealized super cell events to test if the cloud microphysics schemes together with a polarimetric radar forward operator are able to reproduce known super cell radar signatures. Ryzhkov et al. (2011) and Putnam et al. (2017) compare simulated polarimetric radar signals with radar observations to evaluate microphysics schemes, but focus on one or two convective cases.
Furthermore, we have related the findings of Ryzhkov et al. (2011) and Putnam et al. (2017) to ours in the discussion. Jung et al. (2010) and Snyder et al. (2017) simulate idealized super cell events, which is a different approach than ours. That's why we think it is enough to relate our approach to theirs in the introduction, but we don't relate our findings to theirs in the discussion. Putnam et al. (2017), section 3.2: This is in agreement with Putnam et al. (2017) who compare radar signals simulated by 5 different microphysic schemes for two case studies and find that especially the Morrison scheme but to a lesser extent also the Thompson scheme produces too high Z. They attribute this to stratiform rain PSDs that contain too many large drops, to an overforecast of the precipitation coverage overall and in case of Morrison, to a high bias of wet graupel in convective regions. Given that the forward simulator applied in this study does not consider wet particles, we find the high bias in Z exists even without considering wet graupel and comes mostly from rain, suggesting PSDs that contain too many large rain drops compared to the observations.  Ryzhkov et al. (2011) for example evaluate radar signals simulated from a spectral bin scheme against a hailstorm case and find that their spectral bin scheme produces PSDs for rain that deviate from the gamma distribution. Bulk schemes would not be able to reproduce these PSDs and since radar signals strongly depend on the PSD, Ryzhkov et al. (2011) argue that spectral bin schemes are better suited to simulate polarimetric radar signals.
Finally, we referenced the Ph.D. Dissertation about fingerprints in polarimetric radar signals, also in our introduction. This Ph.D. Dissertation helps a lot in understanding the impact of rain processes on polarimetric radar signals: Kumjian (2012) demonstrate the impact of precipitation processes on polarimetric radar signals, though he focuses mainly on rain processes, such as raindrop evaporation or size sorting.
2) separation of convective cells One of the advantages of this paper is the large sampling volume, but it simultaneously induces ambiguity for analysis. During a long-term period, there must be various sizes of convections from shallow, congestus, and deep convective/stratiform cells. When you bundle all into a single CFAD, it tends to smear out important aspects of microphysics. Please check the following papers on how it separates cloud type and better evaluates different aspects of microphysics from long-term simulations/observations. You can use echo-top height from each cell-tracked target for separation. But if this type of separation is too difficult to implement (or too much effort), please just discuss and try it in the future.
We have added a part discussing this topic in the conclusions. This is a valuable comment, we plan to include a separation of convection type for the next steps of analysis of the data to be published soon after this manuscript. Apart from a possible separation using the echo-top height for different convection sizes, we think a separation into weak forcing / strong forcing situations could be interesting. We think that this is too much effort for the present work, that's why we are just discussing it in this manuscript and will add it for future publications.  Matsui et al., 2009). Furthermore, classifications into weak/strong forcing situations could be of interest, to analyze the effect of, e.g., frontal systems on the distribution of radar signals. This will be addressed in a future application of this framework.

3) rain drop-size distributions
Probably the most robust finding in this study is the variability of rain-DSD related radar signals (ZDR and DWR) among different microphysics schemes as seen in Figures 5, 6, and 7. Abovemelting-zone evaluations tend to have more uncertainties in the forward model (described in next).
To augment your finding in radar signals and discussion, it's much better to directly examine simulated rain drop size distribution profiles (like CFAD format) from different microphysics schemes. This should not be a difficult task. ( it's much better if you have disdrometer observations!) We provided a CFAD of rain drop size distributions for convective cells (Appendix B): The following passage was added to section 3.3: In order to separate the analysis into reasons due to differences in the underlying modeled microphysics and due to different processing in the forward simulator, we examined rain particle size distributions directly produced by the NWP model (

4) uncertainties of forward model
Details and uncertainties in assumptions of the forward model (CR-SIM) are not discussed. In order to represent simulated microphysics in polarimetric observables, one must assume particle shape and orientation simultaneously in the forward model, because these are not "explicitly" simulated in most of the microphysics schemes in WRF. Following paper discusses and tests different We have not been clear on the assumptions of the radar forward model (CR-SIM). The section of CR-SIM (2.4) has been appended with the assumptions that CR-SIM is making regarding particle shapes, particle orientation and dielectric constants for each microphysics schemes: The dielectric constant of water is 0.92. Solid phase hydrometeors are assumed to be dielectric dry oblate spheriods and are represented as a mixture of air and solid ice. The refractive index hence depends on the hydrometeor density and is computed using the Maxwell-Garnet (1904) mixing formula. There are no mixed phased particles simulated. This means mixed phase radar signatures (for example the "bright band", Austin and Bemis, 1950) will not be reproduced by the simulation. In order to simulate polarimetric radar observables, a radar forward simulator must assume particle shapes and particle orientation. The particle orientation assumptions are the same for all schemes. It is assumed that the particle orientations are 2D Gaussian distributed with zero mean canting angle as in Ryzhkov et al. (2011). The width of the angle distributions is specified for each hydrometeor class: 10° for cloud, rain, and ice and 40° for snow, unrimed ice, partially rimed ice, and graupel. Regarding the shape assumptions, cloud droplets are simulated as spherical (aspect ratio of 1) and raindrops are simulated as oblate spheriods with a changing axis ratio dependent on the drop size according to Brandes et al. (2002) in all schemes. For ice hydrometeor classes, the same aspect ratio assumptions are applied for all schemes except the P3 scheme: cloud ice is assumed as oblate with a fixed aspect ratio of 0.2. Snow is assumed as oblate with a fixed aspect ratio of 0.6. Graupel is assumed to be oblate with an aspect ratio that is changing from 0.8 to 1, dependent on the diameter and according to Ryzhkov et al. (2011) there are no large cloud ice particles observed, 2) that the signal is dominated by other more spherical particles in the observations, or 3) that the assumed aspect ratio of 0.2 by the radar forward operator is unrealistic and the observed particles are more spherical in nature.

Section 4:
This could either be a result of simulated cloud ice particles being too large or too many, but this could also be a result of the assumed flat cloud ice shape with an aspect ratio of 0.2.

Minor Comments/Suggestions
Line 25: "the huge number" -> "a large number" Changed as suggested.
Line 26: "on scales of Om to mm and" -> "on scales of mm or smaller" for consistency. In fact, microphysics processes occur less than the scale of micron, such as ice crystallization processes.
Changed as suggested.
By sound statistical basis we mean a large sample size. We changed the phrase for clarification: 2. Evaluate multiple state-of-the-art cloud microphysics schemes for current generation numerical weather prediction models in a common model framework against observations with a large sample size.
Line 89: "separate the microphysical impacts from possible feedbacks." I agree. But more bottom line, I would argue whether your set of numerical weather model resolved dynamics or not with 2km horizontal grid spacinig.
There is a misunderstanding here. The middle domain of our model setup has a 2 km horizontal grid spacing. The inner domain that we used for the analysis has a grid spacing of 400 m. We assume you refer to the cloud dynamics which we believe are resolved at a grid spacing of 400 m. That's why we left this sentence as it is. Perhaps the comparison to current operational weather models (line 150) led to the confusion of our horizontal grid spacing. We added a sentence clarifying that our grid spacing is better than current operational numerical weather prediction models and is rather representing the future generation of NWP models in section 2.2: Currently, operational limited area weather models operate at 2 km grid spacing (e.g., 2.8 km in COSMO-DE of the German Weather Service; Baldauf et al., 2011) which means our inner domain has a resolution that is effectively about 5 times higher and should be representing the future generation of operational limited area weather models.
Line 149 and repeat the same: "a horizontal resolution of 10 km," must be replaced by "horizontal grid spacing of 10km". Numerical atmospheric model does not have 10km resolution with 10km of horizontal grid spacing. Effective dynamic resolutions are x5 ~ x10 of horizontal grid spacing in numerical dynamic core. Apply this correction elsewhere in the manuscript. Thank you for this input. Even though we think that 'resolution' is a widely accepted term instead of grid spacing, we understand the logic behind your comment and adapted the definition of the reference throughout the manuscript concerning the model grid descriptions.
Line 159: Please briefly describe other physics options, such as land surface, PBL, and radiation schemes.
The namelist of WRF is added as a supplement where all options used can be seen. The requested information about land surface, PBL and radiation scheme has been added to section 2.2: Other physics options include the Noah Land Surface model (Ek et al., 2003;Chen and Dudhia, 2001), the MYNN2 planetary boundary layer scheme (Mellor-Yamada scheme by Nakanishi and Niino;Nakanishi and Niino, 2006) and the RRTMG radiation scheme (rapid radiative transfer model for general circulation models; Iacono et al., 2008). For any other options, please refer to the WRF namelist that is provided as a supplement to this manuscript.
Line 203: Did you store and use all 33bin of hydrometeor classes to calculate radar observables in CR-SIM?
Yes, we stored and used all bins of all hydrometeor classes to calculate the radar observables for the fast spectral bin simulations. We also stored the aerosol bins (43), but these are not used by CR-SIM. However, the spectral bin scheme uses shared bins for rain / cloud droplets (First 17 bins for cloud droplets, second 16 bins for rain) and cloud ice / snow (First 17 bins for cloud ice, second 16 bins for snow). The output for graupel consists of the full 33 bins. The data is saved at our institute and available on request. We don't think this information is relevant for the reader which is why we did not change the phrasing at this point.
Line 286: "but none of them as pronounced as in the observations." Well, this is typical situations that relatively coarse-resolution model won't be able to resolve tiny cells. So you are running with 2km horizontal grid, meaning that you can resolve convective features in 10km or 20km well, but never be able to resolve 2km-size of convection, which tend to have shallower echo-top heights. So, don't blame to microphysics, but model dynamic core and grid spacing you chose. This is again connected to a misunderstanding. We use a 400 m horizontal grid spacing. That means we are able to resolve convective cells at 2 km or 4 km in size. However, the point still stands: it is likely that we miss the very small convective cells anyways which correlate to the lower echo-top heights. It was not our intention to blame the cloud microphysics for this, as this is a feature in all simulations independent of the cloud microphysics. We slightly rephrased the sentence to clarify the meaning: All NWP simulations independent of the microphysics scheme are able to reproduce a peak at a similar altitude but none of them as pronounced as in the observations.
Furthermore, we added another sentence in the following abstract to emphasize that we don't blame the microphysics for this effect: This is independent of the chosen cloud microphysics scheme and mainly a result of the missing small-scale cells in the simulations which is indicative of a resolution effect: the very small cell heights correspond to tiny cells that we might not be able to resolve even with our 400 m grid spacing.
Line 313: "Contoured frequency by altitude distributions" -> "Contoured frequency of altitude diagram" Changed as suggested, also applied elsewhere in the manuscript.
Changed as suggested.