Lute et al. present simulations of snow over a large domain for reasonably long periods and at high spatial resolution. To achieve this, they have made some quite severe approximations to limit computational expense. Neglecting forest effects seems like a limitation for water resource applications in the western US. Simple canopy models, and snow models with a few layers to allow a more physical representation, need not be very expensive. Has code optimization to offset the cost of improved model physics been considered? MATLAB is convenient for parallel computation, but otherwise might not be the best language for this. Having said that, the authors have adopted a particular development strategy and reporting the results is worthwhile.

I was able to access the code and run the MATLAB Online example easily (a quickstart guide might have saved me a couple of minutes).

I suggest some corrections and clarifications, identified by line number:

24

There are too many to list, but the abstract should give some idea of what kind of “snow metrics” are offered.

52

Should note that this is earlier but slower snowmelt

106

Equation (1) is not the surface energy balance because G is not a surface flux. I would describe it as the energy balance of the snowpack.

127

Note that separate visible and near-infrared radiation fluxes are not supplied as inputs; the albedo is simply averaged between these bands. The illumination angle dependence should only be applied to the direct-beam component of incoming radiation.

133

Equation (3) is incorrect. If the emissivity is not 1, the upward longwave radiation includes a reflected fraction of the downward radiation (Kirchoff’s law).

179

Constant G is not a very common (or realistic) feature of modern models. Etchevers et al. (2002) included an experiment with fixed G as a sensitivity study, but did not report the results.

189

Iterative solution of the surface energy balance to find snow surface temperature is not the only possibility. Best et al. (2011, https://gmd.copernicus.org/articles/4/677/2011/), for example, linearizes the surface energy balance equation to find a surface temperature solution without the expense of iteration. In either an iterative or linearised solution, the aim is to find a surface temperature that is consistent with the surface energy balance. I don’t see how this can be achieved with the pragmatic but ad hoc approach in SnowClim.

281

How were the 210 m and 1050 m resolutions selected? How many points are there in this domain?

330

SNOTEL sites may not be representative of larger areas. Could this calibration skew the model performance? shttps://agupubs.onlinelibrary.wiley.com/doi/10.1029/2005WR004229

362

Any ideas why model performance would improve from 12 to 24 hour timesteps?

369

There is no evaluation of the large-scale snow simulations for the western US. MODIS snow cover extent products would have a convenient resolution for this.

Figure 2

There are a lot of sites to balance in the calibration, but errors up to +/- 50 days in duration and +/- 50% in maximum SWE even after calibration seem large for practical applications. Other datasets used to demonstrate poorer agreement in the discussion were not calibrated to SNOTEL sites.

Table 1

Given temperature, pressure and one of dewpoint temperature, relative humidity and specific humidity, the other two can be calculated – they are not all required forcing data.

Table 2

As it is abbreviated as Q, I guess that the WRF mixing ratio output is taken as the SnowClim specific humidity input (they are not the same thing, but will have nearly identical values).

Table 5

The monthly SWE and snow depth are minimum, mean and maximum.

This looks like it might be a useful addition to the suite of datasets of snow simulations available in the western United States. The combination of a reasonable physical basis, validation against observations in diverse climates, and an underlying high resolution provides, at least conceptually, a reasonably satisfying sense that the estimates provided near the end in Figures 5 and 6 are sound.  It seems appropriate to publish this presentation of the details of the methods used in producing these data sets. At the same time, there are improvements that could be made in the writing to provide a more constructive contribution.

The paper gives a distinct sense of critique of the existing set of snow models available. Admittedly some are related to questions about run times rather than performance, but sometimes the criticisms come across more like aspersions than measured or tested issues.  In this context, I look at all the plots from SNOWMIP2 (Rutter et al. 2009), and I’m left wondering where this model would fit into the cloud of lines. The bottom line is that this is another snow model on top of an already long list, with several new approximations, assertions, and assumptions, and we are left wondering what exactly we are learning from this particular modeling exercise.  The focus was on improving model speed so that larger areas could be evaluated at finer resolutions in a reasonable time, but the net gain in understanding at these larger scales is not really highlighted so much as alluded to.  This is a problem because there are several products out there and in use for larger scales, and while this model conceptually critiques these products, some uncited, it only has a somewhat philosophical basis for that critique.  At a basic level, the claim for superiority of this modeling approach relates to a reductionist philosophy that if we can just do our simulations at finer and finer scales, we can resolve everything.  The paper ultimately bumps into the fact that as we get into finer scale simulations, we find processes that depend on adjacent areas, and the upshot of this hazard appears in line 416 – near the end of the paper – where they note that the model is “most realistic in minimally vegetated areas with relatively little snow redistribution”, a rare piece of ground in the western US. 

So, the question remains about what we learn from this modeling exercise. The central claim seems to be that when we have more resolution, we can resolve effects of things like aspect and solar radiation, so we have a better representation than we have had in the various existing undisclosed products.  This may be true, but at the same time the paper’s central argument is fundamentally recommending further reduction in element scale with addition of vegetation and redistribution effects.  Is there simply an argument for just doing the grind of getting complete lidar coverage over the western US, including data that can tell us about accumulation enhancement/loss patterns, and running a hyper-resolution snow model everywhere? Is this the only way that we can ultimately get a satisfying snow data set?  At what scale does a reductionist argument end?  Is there any utility in considering subgrid parameterizations of the effects of solar radiation, redistribution, and forests to a grid scale that is useful … or is that just a side-track that gets swallowed up by research in computing, storage, and remote sensing? My point is that this paper – though currently unstated – takes a stand with respect to these philosophical discussions in the literature. Being unstated, it is a strongly one sided stand being taken, and it’s possible the authors actually have a more nuanced perspective and this paper is simply an exploration in one dimension.  In that case they might want to say something.

None of this detracts from the likelihood that this model product probably provides greater insights for some purposes about changes that are occurring than some existing products, but it is important to examine and state in the introduction what the central argument is and its philosophical history.  At the very least, this makes it easier for others to follow up on the conceptual advance of the paper.  In this case, the argument seems to be that some aspects of model physics can be sacrificed in order to better incorporate the effects of spatial heterogeneity in solar radiation, elevation, and temperature on snowpack over the same area, and potentially be able to display some of the heterogeneity in snowpack at finer scales for some purposes.  This is a good question, and not entirely certain in its outcome, particularly when other subgrid processes are set aside.  I don’t know that this tradeoff has been assessed in this paper so much as asserted, and it leaves on the table the question of whether a subgrid paramaterization coupled with earlier products could potentially generate greater benefits in time and accuracy … at least for purposes not needing to directly display fine-scale heterogeneity.  It also leaves open questions about which particular process for heterogeneity are most critical to incorporate if following a reductionist approach.

It is worth noting here that we have seen images very similar to Figures 5 and 6 (which display fairly coarse features) in multiple recent publications that are not necessarily mentioned or cited (including by the first author).  At some level, though, these earlier coarser resolution papers must be doing something somewhat correctly to get essentially the same images.  The abstract and introduction begin by pointing out that there are no existing data sets that are “based on physical principles, simulated at high spatial resolution, and cover large geographic domains”, but they do not reflect on applications of these existing somewhat similar data sets, and clearly define “physical principles” or “high resolution” and explain for what purposes they are valuable.  These definitions probably relate to the purposes of the simulation and the scale/resolution of the display.  It is very likely that some of these existing products are adequate for many tasks, and rather than suggesting a position that this new data set makes all preceding simulations obsolete, perhaps discussing new applications the higher resolution data allow or note specific improvements they can demonstrate.  This is partially addressed in Figure 7 and references, but only for one specific application where details of heterogeneity are important, although arguably could be focused on very small areas where redistribution may be important.  On the net, if this data set is intended as replacement for the other data sets, it would be more polite to acknowledge them and their utility in the conversation so far rather than to ignore them.  If in no other way, they could be acknowledged for their corroboration of images like Figures 5 and 6.

There is a similar set of discussions about philosophical subjects with respect to the framing set up in lines 55-65.  There is a significant discussion in the hydrology literature at the moment, and spilling into geomorphology and snow as well, about the utility of machine learning approaches.  Some argue that these are not “physically based”, but others note that the calibration needs of some models renders them effectively as non-physically based as some of the machine learning approaches.  Quite a bit of work has been done in snow with ML and it is currently an area of active research.  I’d recommend the authors take a look at a brief paper by Fleming et al 2021 to see how aspects of that framing might be unhelpful to furthering the goals of their work.  I’m certainly in agreement that we all need to understand 1) the physics of the real world and 2) how our models represent that physics, but there are sound arguments that a priori assertions about the details of the physics in model formulation that could detract from our learning about the physics of the real world.  There are others who deliver that argument better than I, and hopefully after looking at some of those you might find a better way to describe the meaning of “physics-based” in the paper. 

The above points are for general framing and are important to address so that the contributions of this specific modeling exercise can be better placed among much that has occurred before and much going on concurrently.

A more specific point begins at line 180 in the characterization of “single” layer snowpack models and shortcomings, particular with respect to the UEB model.  The UEB model deals with the issue in two ways. During non-melt periods, it simply uses a conduction estimate based on a high frequency and a low frequency component – essentially using a Fourier decomposition approach to calculating the conduction rather than a finite difference approach – which would use “layers”.  UEB also makes further computations for heat transfer when the surface is melting.  Insofar as you would like to emphasize the physical basis of snowclim, you may want to reconcile the “tax” approach and framing with the physical basis of conduction in the frequency domain.  As it stands, when the surface is cooling, it looks like only a portion (as low as 10%) of the net flux at the surface is being removed from the pack – the remainder isn’t accounted for at all.  It is not clear what errors this accounts for, and you may want to add some content (text and appendix figures?) to the paper to explain how equations 16 work.  There are at least 4 parameters involved, all of which are calibration parameters.  This seems potentially important, but without some validation against temperature data, its difficult to tell.

Note that the snow surface temperature calculation in UEB takes into account this conduction heat flux from the surface along with the several other fluxes at the surface (e.g. solar, longwave, latent, sensible) to find the T at which all fluxes balance. This was not a particularly computationally burdensome iteration and the uncertainty cited by Raleigh (2013) was in parameter estimates for conduction, which Snoclim just does not consider.  In the broad scheme of uncertainties for energy balance, I don’t know that these are big numbers, but the criticisms in line 189-190, which are drawn mostly from Raleigh (2013) are not really warranted and don’t bear repeating in the context of this paper.  There remains some chance that the UEB calculations might be nearly as parsimonious and generate less uncertainty.  No assessment has been done in this paper to evaluate these issues.

Reference (other than already cited in the paper):

Fleming, S. W., Watson, J. R., Ellenson, A., Cannon, A. J., & Vesselinov, V. C. (2021). Machine learning in Earth and environmental science requires education and research policy reforms. Nature Geoscience, 14(12), 878-880.