the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 04 May 2026
Spin-up in humidity and temperature and its consequences for convective diagnostics: a Model Uncertainty Model Intercomparison Project experiment
Abstract. We analyse the evolution of convective diagnostics such as mixed-layer convective available potential energy (CAPE), level of neutral buoyancy and precipitation rate as a function of lead time in the model uncertainty model-intercomparison project. Four model physics packages are exposed to common dynamics to form a large single-column model dataset. We analyse tendencies in an equatorial band over the Indian Ocean out to 6 hr lead time over one month. We prescribe dynamics and initial conditions from an ICON-DYAMOND simulation after coarse-graining to 0.2 degrees. The physics suites represent state-of-the-art global numerical weather and climate prediction models.
Correlation analysis shows that the spatial mean change of CAPE is not associated with precipitation rate, but it correlates very well with mean mixed-layer drying across our suites. This systematic drying occurs below 700 hPa in some suites, especially in the first hour. The sub-grid physics adjusts the initialised ICON state towards the native climate of each physics suite, in particular at low levels.
We apply a column-by-column empirical orthogonal function (EOF) analysis to a two-layer representation of physics and dynamics tendencies, CAPE tendency and precipitation rate. The first EOF is associated with free-tropospheric tendencies and nearly all precipitation variability, with neat compensation between physics and dynamics tendencies. The second and third EOFs of each suite indicate that a imbalance between these terms in the mixed-layer correlates with the CAPE change at least one of them, which are explained by temperature and humidity adjustments, but with little imprint on precipitation.
- Preprint
(5850 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-1445', Anonymous Referee #1, 23 May 2026
- RC2: 'Comment on egusphere-2026-1445', Anonymous Referee #2, 03 Jun 2026
-
RC3: 'Comment on egusphere-2026-1445', Anonymous Referee #3, 29 Jun 2026
The manuscript describes a series of experiments where various physics schemes were coupled to a single dynamical simulation in an offline setting. While this is an interesting experiment, I had severe difficulties following the logic of the manuscript. In addition to lacking clarity, the author's interpretation of their EOF analysis is questionable and needs to be revised.
### overall comments
- It was not very clear to me why exactly such an experiment (coupling physics to a common, non-native model) is relevant in the first place. Perhaps you can give some more motivation, what exactly do we hope to learn about either these physics suites or about tropical convection in the real world?
- There is a lot of text describing relatively basic diagrams like pdfs or profiles in verbose detail. Consider shortening section 3 wherever possible.
- The discussion and interpretation of the EOF analysis is confused and probably misleading in some ways (see comments below). In particular 1) the entries in the PCs are neither correlations nor do they represent the physical magnitude of some tendency and 2) physical interpretation of the higher EOF modes is usually impossible because of the orthogonality constraints.
- I find it hard to draw any general conclusions from this paper. What have we learned beyond "different physics suites behave differently"? Perhaps a section 5 "Conclusions" could be added?
### specific comments
l.13f unclear grammar ("at least one of them"? who or what "are explained"?) please rephrase
l.17 "have" or "serve"?
l.25 "persist in datasets with increasing sample sizes" unclear what you mean, which sample size, the length of the simulation?
l.29 unclear grammar, I feel like the verb is missing?
l.50f unclear what the brackets ("observations, validation of MUMIP") refer to
l.60 unclear what you mean by "more accurate". Clearly a free-running model has no correlation with the actual weather observed at any time (no skill), but that's probably not what you mean?
l.63 "supposes more rigorous uncertainty sampling" what is that supposed to mean?
l.70 unclear what "them" refers to, the subject of the previous sentence is singular.
l.110 unclear in what way a subset is extracted and what it means to "mimic ICON dynamically"
l. 127 which is it, 30 or 60? or were both used?
l.136f what exactly does it mean that you "mixed a layer of 500m depth"? does that mean the parcel to be lifted has the average conditions of that layer? from where is the ascent initiated?
Figure 1: lines are hard to distinguish, consider (a) trying a log x-axis, (b) showing the frequency bias with respect to ICON (ratio between the four physics suites and ICON)
Figure 1: unclear what you mean by "ICON & initial conditions", are those two separate things?
Figure 1 (last comment): why was the y axis cut off at 10? for some applications (especially weather forecasting), the tail of the distribution might be the most interesting part.
l.155 and elsewhere: figure 1 (also 2) does not show pdfs (densities) but raw counts. That won't make a difference to the figure unless the sample sizes are different for some reason.
l.178 what is "clearly stable", is that defined thing or just a turn of phrase?
l.183 these two peaks (at least the first one) are not so easy to see in the figure.
Figure 2: naively, I would expect that LNB height and CAPE are connected and at least looking at the tails of the distribution ist seems to me that the two panels actually show similar things with a non-linearly compressed x-axis. Did you look at the joint distribution of CAPE and LNB? Are they closely linked?
l.233 where have we seen anything about the tendencies of cape? figure 2 only showed the pdf of cape itself, not of its tendencies?
Figure 3 and 4 why not include the ICON reference here?
Figure 4 why is cape not shown in terms of w_max here? wouldn't that make it easier to interpret and compare with the pdfs for before?
l.258 all the curves look pretty much linear near the end of the 6h, where to you see a reduction in rain rate?
Figure 5 left: the areas are very hard to distinguish, for most lines we cannot see whether the mean is within the quartile-range (see below). Is this the distribution over the whole domain or just the equatorial band? If the former, wouldn't it make more sense to show the same region as in the right panels, so we could compare them?
Figure 5 left and l.264f: since precipitation has a skewed distribution, the mean can lie outside the interquartile range (see last timestep for GFS). If you want to use that range to gauge the uncertainty, maybe it makes more sense to show the median instead of the mean? The text discusses this behavior but doesn't mention the simple explanation that the distributions are non-gaussian.
l.272f and elsewhere, you discuss what the ICON benchmark "would" look like. Why not just include it in the figures?
l.307 I was very confused for a minute because DCAPE usually refers to the energy available for convective downdrafts (https://glossary.ametsoc.org/wiki/dcape/), but from the units in figure 6 I'm guessing you mean the time derivative of regular CAPE? Please chose a different abbreviation.
figure 6 isn't the most striking observation that ARPEGE has only positive tendencies and all other models have only negative tendencies? I don't understand that at all, how can CAPE tendencies have unique sign? Is that because the models always adjust the initial profile to their own climate? Also the correlations with precipitation have opposite signs, how can we interpret that?
l.312 f: do I understand correctly that you correlate spatial mean dCAPE/dt with spatial mean precipitation? If so, why? Wouldn't it make a lot more sense to correlate them for each gridpoint individually? As it is, you might have precipitation cells feeding on CAPE in some regions while CAPE is still being generated elsewhere and the two effects average out?
figure 7 if you want to define what you mean by delta, do it somewhere in the text or just refer to delta T here since there is only one variable with a delta in this figure.
l. 337f just because area mean tendency does not coincide with area mean CAPE change, we cannot conclude that other processes must be at work: the tendencies can be non-zero at every gridpoint, but CAPE tendencies can remain zero in many locations, depending on the overall temperature and moisture profile (for example if there is no LFC).
Figure 8: unclear what you mean by "covers 6 out of 8 initialisations".
Figure 8: Solid and dashed lines are hard to distinguish, and why is there no uncertainty bound for hour 1?
Section 3.5 please clarify whether these EOFs were estimated from area means of (as I assume) from individual gridbox values. If the latter, especially don't understand why area means were used in figures 6 and 7 because now we can't relate those results to this EOF analysis.
l.410f the formulation "we assume the following" makes it sound like these are assumptions that somehow go into the EOF analysis (are require for it to work or something). Is that the case or are these more like your "expectations"? Either way I have no idea how assumptions 1 and 2 could be related to an EOF analysis since they refer to the mean values whereas EOFs depend only on the covariance matrix. In particular l.417 makes no sense to me, the EOFs are literally invariant under changes of the mean.
l.438: I know what you mean but the values of the principal components are not correlations. If you want to know whether these things are correlated you have to go back to the matrix.
l.444f: similarly, the entries of the principal components cannot be directly interpreted as the physical magnitude of those terms. In particular, you standardised all variables, which means that the physics tendencies could theoretically be an order of magnitude larger than the dynamics and still have the same amplitude values here. Even if you had not standardised, the EOFs still say nothing about the overall mean tendencies: if one of those tendencies has a near-constant positive value it will contribute next to nothing to the EOFs but might be a big part of the overall balance.
l.451 this is one of the main things I would actually conclude from this EOF analysis, that ML and FT are relatively independent wrt these tendencies for most of your models. The other thing would be that precipitation shows up only in EOF1 and thus has little to do with the ML variables.
sec 3.5.2 I did not read this in detail because the discussion of higher order EOFs, apart from the most general characteristics, is often futile: the shape of EOF N is dictated by the requirement of orthogonality to all N-1 higher EOFs, interpreting them in terms of physical consequences is therefore dangerous at best.
l.507 unclear what you mean by "increases", increases compared to what?
l. 541f if the second point is virtual temperature, it is not "separated" from humidity.
l.554f how do you know that Tv adjustment "will be dominated by T adjustment", do you have a reference for this claim?
l.572 I have no idea what you mean here or whether this was "argued" in section 3.4. What does "scales well with most samples" mean?
594 again I'm afraid you cant interpret the fact that EOF1 is most similar among models in such a way: small, random discrepancies in one EOF will propagate into bigger differences at higher order because each additional EOF is orthogonal to all others.
l. 605f again, I don't know how you think that the EOFs make any statements about any kind of mean when they are derived exclusively from the covariances.
l. 608 "orthogonality between CAPE and precip" please clarify which part of your paper you refer to here. The EOF analysis included only CAPE tendency, not CAPE itself. Did Buschow (2024) discuss CAPE or CAPE tendency?Citation: https://doi.org/10.5194/egusphere-2026-1445-RC3
Data sets
Spin-up in the Model Uncertainty Model Intercomparison Project: humidity and temperature adjustment and its consequences for convective diagnostics Edward Groot, Hannah Christensen, Xia Sun, Kathryn Newman, Wahiba Lfarh, and Romain Roehrig https://doi.org/10.5281/zenodo.18174141
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 269 | 41 | 20 | 330 | 28 | 25 |
- HTML: 269
- PDF: 41
- XML: 20
- Total: 330
- BibTeX: 28
- EndNote: 25
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Convective physics are a vital part of climate models but lead to lots of uncertainty and biases. As convective processes tend to occur on spatial scales smaller than can be resolved by most models, they must be parameterized. These parameterizations are vital to understand as they control lots of model results, but are enormously complex and have effects that are hard to isolate.
In this manuscript, the authors place data from a full GCM into single column models to see how the physics model interacts with the column when it is isolated from the dynamics. They find that all the physics models tend to decrease CAPE through boundary layer drying, and that there are free troposphere changes associated with precipitation.
The manuscript is interesting and the framework for analyzing model biases is valid. However, the analysis could be much more thorough and the results are not very general. Additionally, the manuscript has serious grammatical errors and inappropriately cites “In preparation” references.
Specific Comments