|Review of Philip et al., Sensitivity of Chemistry-Transport Model simulations to the duration of chemical and transport operators: A case study with GEOS-Chem v10-01|
The manuscript provides an analysis of the growth of errors from using longer model timesteps to achieve faster model execution with the GEOS-Chem chemical transport model. In particular, the paper focuses on the growth of error as the timestep used for the time split solution of individual processes in the model is increased. The paper presents an analysis of a fundamental choice that chemical transport modellers must make, that has not received much analysis to the best of my knowledge. The paper is well organized and the use of the ‘CPU-time adjusted Composite Normalized Error’ provides a nice summary of the trade-offs of runtime versus accuracy, the paper has a few omissions that hinder a full understanding of the results. I would find the paper acceptable for publication after these concerns are addressed.
I should mention at the outset that this is my first review of the manuscript – I was not one of the two original reviewers – and that I have read and taken into consideration the comments and replies made during the first round of reviews.
My most significant concern of the manuscript in its present form is the interpretation of the differences between the C10T05 and C10T10 simulations that are shown in Figure 2 and discussed on lines 215-227. In particular, in interpreting the differences between the C10T05 and C10T10 simulations the authors state ‘Increasing the transport operator duration tends to increase concentrations of emitted species like CO and NOx over source regions since species are more uniformly mixed by long operator durations before loss processes such as deposition and chemistry occur.’ (lines 218-221). The differences between C10T05 and C10T10 are due to a change in going from an operator order of C-T-T-C-T-T... to C-T-C-T-C-T..., if I have correctly understood how the model timestepping works. In both cases, the transport operator must still calculate a full 10 minutes of model time to keep pace with the chemistry operator for each model timestep. The differences you have shown in the middle column of Figure 2, I would argue, are more due to the different time truncation errors associated with each transport operator calculation than a change in the errors due to time splitting. For a single timestep of 10 minutes, advection will produce different solutions depending on the timestep used (2x5 minutes or 10 minutes). Even a simple calculation such as dry deposition can produce different results whether called twice or once to cover 10 minutes, particularly depending on the details of the numerical treatment (implicit, explicit, analytic). The discussion seems to suggest the differences are due to time splitting errors (‘long operator durations’) but each individual operator in both the C10T05 and C10T10 case has the same operator duration. The discussion presented in this section would be much more applicable to the differences between C20T05 and C10T05, as the operator duration has actually changed in this case. If you keep in mind the idea that there are two sources of error (time truncation and operator splitting) it may help in the interpretation of the differences between the C=T and C=2xT simulations, in particular the saw-tooth pattern of error growth in O3 and CO shown in Figure 4.
A second concern is the presentation of the operator structure of GEOS-Chem, found around lines 113-124. In equation (2), there is a photolysis operator that is separate from the chemistry operator. Is the photolysis operator really a separate operation that modifies the model solution (the chemical fields) and is calculated separately from the chemistry? I am guessing this is the calculation of photolysis rates, but if so this calculation would not strictly be counted as a separate model operator as the photolysis rates only affect the model solution through the calculation of chemistry. In that case, photolysis should be removed from equation (2), perhaps with a note that photolysis rates are calculated as part of the chemistry operator.
On lines 129-130 there is a short mention that ‘GEOS-Chem uses an internal integration timestep of 5 min for convective mixing.’ I assume that no matter what the chemical and transport operator durations are set to, the convection operator shown in equation (2) is calculated with an internal 5 minute timestep. This would seem to be an important caveat that should be included in the preceding paragraph discussing the time splitting.
On lines 127-129, there is the statement that ‘Transport by convection is coupled with gas-aerosol wet deposition.’ yet equation (2) shows separate operators for cloud convection (Z) and wet deposition (W). I would assume the wet deposition operator is only from large-scale (stratiform) precipitation. If so, it should be made more transparent to the reader that wet deposition from convective precipitation is calculated as part of the convective transport operator and stratiform wet deposition is in the separate W operator.
Lastly, regarding the form of the operators in GEOS-Chem, there is no mention of how aerosol (and cloud?) chemistry and gas-phase chemistry interact. Are they two separate operators? And if so they should probably be presented as separate operators in equation (2). This could be an issue as mention is made on lines 234-236 of differences in in-cloud sulphate oxidation due to H2O2 differences with longer chemical operator durations.
Other minor comments are given below.
Lines 48-50: very generally the form of CTMs is introduced in the first paragraph and discussion of operator duration is broached, but there is no mention of why operator splitting in CTMs is required.
Lines 64-65: a minor quibble but here it is stated that ‘Fine vertical resolution can better represent convection’. To be more exact, this statement should probably say ‘can better represent the effects of convection...’.
Lines 67-79: here the continuity equation is discussed and advection by resolved winds along with chemical production and destruction are explicitly discussed but there is no mention of sub-grid scale transport – turbulent mixing, convection – or other terms such as emissions, dry deposition.
Line 119-120: The sentence ‘The traditional chemical operator durations have varied from either 60 min or twice the transport operator duration...’ is overly complex. On re-reading the reference to 60 minutes is twice the maximum of 30 minutes given for the transport operator, but the construction is a bit convoluted and it took a bit of rereading to figure out what was being expressed. How about ‘The traditional chemical operator duration is set to twice the transport operator duration...’
Lines 159-160: Would you have the range of execution times for the five repetitions of each simulation handy? A passing mention of the range would help the reader understand the robustness of the timing.
Lines 290-291: In comparing the errors and execution time of the C30T15 and C20T10 simulations you say that ‘the CPU time increases by 20% by the decrease in operator duration.’ If it was completely linear, going from C30T15 to C20T10 should increase the model execution time by 50% because you now need 50% more chemical and transport operations to complete a given amount of model time. Do you have an idea of why this particular case does not seem to scale linearly?