the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Submitted as: development and technical paper 16 Feb 2021
Submitted as: development and technical paper | 16 Feb 2021
A machine learning-guided adaptive algorithm to reduce the computational cost of atmospheric chemistry in Earth System models: application to GEOS-Chem versions 12.0.0 and 12.9.1
- 1John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA
- 2Computational Health Informatics Program, Boston Children’s Hospital, Boston, MA, USA
- 3Department of Pediatrics, Harvard Medical School, Boston, MA, USA
- 4Center for Functional Nanomaterials, Brookhaven National Laboratory, Upton, NY 11973, USA
- 1John A. Paulson School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA
- 2Computational Health Informatics Program, Boston Children’s Hospital, Boston, MA, USA
- 3Department of Pediatrics, Harvard Medical School, Boston, MA, USA
- 4Center for Functional Nanomaterials, Brookhaven National Laboratory, Upton, NY 11973, USA
Abstract. Atmospheric composition plays a crucial role in determining the evolution of the atmosphere, but the high computational cost has been the major barrier to include atmospheric chemistry into Earth system models. Here we present an adaptive and efficient algorithm that can remove this barrier. Our approach is inspired by unsupervised machine learning clustering techniques and traditional asymptotic analysis ideas. We first partition species into 13 blocks, using a novel machine learning approach that analyzes the species network structures and their production and loss rates. Building on these blocks, we pre-select 20 submechanisms, as defined by unique assemblages of the species blocks, and then pick locally on the fly which submechanism to use based on local chemical conditions. In each submechanism, we isolate slow species and unimportant reactions from the coupled system. Application to a global 3-D model shows that we can cut the computational costs of the chemical integration by 50 % with accuracy losses smaller than 1 % that do not propagate in time. Tests show that this algorithm is highly chemically coherent making it easily portable to new models without compromising its performance. Our algorithm will significantly ease the computational bottleneck and will facilitate the development of next generation of earth system models.
- Preprint
(421 KB) -
Supplement
(2475 KB) - BibTeX
- EndNote
Lu Shen et al.
Status: final response (author comments only)
-
RC1: 'Comment on gmd-2020-425', Mathew Evans, 11 May 2021
The computational cost of atmospheric chemistry with atmospheric chemistry transport or an Earth System Models is large and methods to reduce these costs are so useful. The authors have over the years presented a series of papers (Santillana et al., 2010; Santillana et al., 2016; Shen et al., 2020) that attempt to reduce this computational burden by means of (simplifying here) separating the chemistry into fast species for which the differential equations need to be explicitly solved and slow species which can be solved analytically. The complexity of the approach has increased over the years and this paper represents the current incarnation of the methodology.
These technical advances in the numerical methods for atmospheric chemistry transport model (and more generally for geophysical models) are not seen as being sexy science, however, they are essential if these models are to be useful for the wider community. I am thus supportive of this paper and would suggest publication after some clarifications and corrections suggested below.
Major comments.
The figures from the supplementary material should be included in the main text of the paper. For me, it is hard to understand some of the more complex mathematical aspects of the paper without a diagram to support it (e.g. S1 etc). I don’t really see why all of the figures from the supplementary material can’t be included in the main text. They are not particularly repetitive and I think it would be useful to the reader to see them all.
It would be useful at the end of the introduction to provide a context for what is coming up in the rest of the paper. The lines around 185 provide this commentary about what has been done in the past, the problems associated with those and the method of addressing those which is discussed in the rest of the paper. This would help to contextualize Section 2 which seems like a rather remote set of definitions at the moment. In a few places, the paper feels like it is rather disjointed with one section not necessarily rolling into the next with much cohesion. Perhaps a re-read and a re-think of some of the structure would be beneficial for much of the paper.
The available code is included in a RAR format. This doesn’t seem to decompress on my Mac. Can we have the data in a standard zip format?
Minor Comments.
Line 60. This sounds like the end of an abstract rather than an introduction. At this point, the methodology hasn’t been explained or tested so how can it be called chemically coherent, accurate or
Line 87. Can the value of 10 molecules cm-3 s-1 be put into some context? Why was this chosen? Which reactions fall into this category?
Line 99. Giving the number of species (228?) in the reaction mechanism would be useful for contextualizing the 3400 other numbers?
Line 104. “Ti,j is the is the number of reactions that include both species i and j.” Is that t as a reactant, products or either?
Figure S1 should appear here to help explain the definition of Di,j. It would be useful to give an explanation for the numbers which are obtained given the chemical mechanism linking Toluene, xylene and Glyxoyl.
The methods used to calculate “distance” appears to refer to the mechanism without any chemistry occurring. In the case of A+B-->C and A+D-->C, the “distance” between A and C doesn’t care about the 2 rate constants or the concentration of the reactants. It would appear that chemically minor routes or channels are given the same weight as the dominant routes or channels. Presumably, the ideal way of working out the distance between species would be to explore the mechanism with some chemistry occurring and the weigh the distances between vertices by the flux between species or something like that? This obviously has a significant downside of being much more complex to implements and contextual (the distances would change with the chemical environment). It might be useful to describe this as being an optimal approach, but the approach taken is a simplification of this. Otherwise, it is rather hard to understand why the fluxes have not been used to represent the distance between species?
Line 116. It might be worth including the definition of “long-lived” and fast here.
Line 130. It would be worth pulling in the description of f from the supplementary information.
Line 134. Bring in the figure S2 into the main body.
Line 144. By only considering values greater than 1e6 cm-3 from your statistical analysis you are not excluding many values for typically high concentration species (O3 etc) but you will be excluding a significant fraction of the values for OH shown in Figure S6. It is not appropriate to do this for OH. A significant fraction the grid boxes will have OH concentrations less than 1e6 (the canonical global mean value). I appreciate that there might be increased errors at lower concentrations but if a large number of the global grid boxes have lower concentrations than this, this metrics is being rather selective in the grid boxes that it is using the analyse the model. From Figure S8 it would appear that this excludes evaluation of the model error in large chunks of the surface of the globe for OH?
It’s also not clear why the value is set at 1e7 for NO2 in Figure S6? Is this alternative cut off used in Figure S8? Presumably, this is the reason why the polar regions are excluded from Figure S6 for NO2?
Line 158. Can Figure 1 include information about the definition of slow and long-lived for reader clarity otherwise they are having to flick back to find the definitions used?
Line 182. “different blocks based on similarity of chemical behaviours using a machine learning clustering method.” I think this is described in the supplementary material. This should be brought into the main text of the paper or a reference to where this is described included in the text.
Line 185. It would be very useful to have had the contextual information given here much earlier in the paper so the reader can understand what didn’t work in the past, what the proposed solution is and how this will be implemented.
Line 209. I’m not sure I understand the comment that iodine reservoirs are inert? Many of them are highly photolabile. This doesn’t seem to make sense?
Line 243. When talking about the error this is the RRME? This should probably be clarified. It would also be useful to discuss the implications of the >1e6 cm-3 limit here on the statistics. Is the NO2 value with the 1e7 cm-3 limit as indicated in the supplementary material?
Line 245. Being able to update the chemical mechanism is an important aspect of maintaining the viability of the model in the long term. If there was a significant change to the mechanism the whole process of running the model without the chemical mechanism splitting would need to be done again? The method outlined here is for “on the fly” updates and relies upon the changes being small and the chemical intuition of the person doing the update. It would be useful to explain that the approach described here is for “patching” the mechanism etc. rather than updating the whole mechanism.
It's not clear where new “biogenic VOC” degradation products would go (blocks 7,8 or 9 etc). If an exact mechanism for working out the placement hasn’t been found and new species were randomly allocated to 7,8 or 9 (or another mechanism) it would be useful to have that documented.
Line 272. I don’t think that it has been demonstrated that it can be ported to different atmospheric models “easily”. Relatively minor changes to the chemical mechanism within one model were shown to be able to be incorporated without re-running the whole tuning procedure but I don’t think it can be demonstrated that it is easy to move into a different model (CESM etc).
Figure 3. Could be greyed out area showing the number of grid boxes in that category be split into some subcategories: marine boundary layer, continental boundary layer, free troposphere, stratosphere etc to provide some additional information?
Figure 4. This is the performance over what timescale? All of the 1-year timesteps for all grid boxes? It's not clear that the bars indicate the simulation speed up and the symbols represent the accuracy.
Supplementary material. Much if not all of this should be in the main body of the paper.
Figure S4. Is this mislabelled as “anthropogenic blocks”, think it should be biogenic blocks? How is the decision made about fast/slow when there are multiple blocks? This could be explained.
Figure S5. I’m not sure that the figure actually shows the “mechanism complexity needed”? It shows the sub-mechanism at each location. I found the % fast labelling slightly confusing as it wasn’t that obvious whether the colour scale was describing the Regime or the % fast? Perhaps the % fast could just be removed for simplicity?
Figure S7. How do these curves look if the <1e6 cm-3 restriction in calculating the RRMS is removed?
Figure S8. Can the location where the RRMS has not been calculated due to the 1e6 cm-3 restriction be indicated (grey the area?).
Figure S9. What is the H / L notation indicating?
-
AC1: 'Reply on RC1', Lu Shen, 23 Aug 2021
The comment was uploaded in the form of a supplement: https://gmd.copernicus.org/preprints/gmd-2020-425/gmd-2020-425-AC1-supplement.pdf
-
AC1: 'Reply on RC1', Lu Shen, 23 Aug 2021
-
RC2: 'Comment on gmd-2020-425', Anonymous Referee #2, 21 May 2021
This paper present some clearly very well thought out and well executed methodologies that are shown to be effective at minimising computational costs for solving a complex chemical mechanism in the GEOS-Chem model. These tool are likely to have substantial benefits for the air quality and Earth System modelling communities. Unfortunately, the writing of the paper is confusing at times, understandable given the complexity of the subject matter but it does distracts from messages it is trying to convey. With a few small changes to the structure and presentation, this will be an excellent paper well suited for publication in GMD.Major CommentsThe biggest issue I find with the paper is the it goes into the nitty gritty of how it partitions different species/reactions into different categories (sections 2.2-2.5) before it explains the overall design philosophy and what these different categories are used for (in section 3.1). The result is that on first reading, I got very confused reading through section 2 and was only able to make sense of it on second reading. This could be improved substatially if either Section 2.1 were expanded upon to give an overview of the whole design philosophy and what each of the the different subcategories (fast, slow species, unimportant reactions etc.) are going to be used for before they are described in detail, or a new Section 2.2. is added giving an overview of all of the developments. Some of this description is in the first paragraph of Section 3.1. Also section 3.1 repeats a lot of the text currently in section 2.1 about the model description and is inapropriate here, therefore section 3.1 should also be editted to avoid repition.In short, Section 2 should contain all of the descriptions and definitions for the model and the adaptive algorithm for the chemical operator. Section 3 should focus on testing, evaluating and optimising the algorhithm in the 3D model. If these two sections were more clearly defined, the paper would be much easier to follow.There are also some inconsistancies in language and definitions - I found the use of "slow species" and "slow reactions" (each of which use different threshold definitions) particularly confusing. It would help to be consistent and use "unimportant reactions" for those <10 molecules cm-3 s-1. I have specific comments below to help with this.In terms of the statistical analysis (Section 2.6) it is important to have a measure of bias as well as error - I would generally be more concerned about a species that has an error of 1% and bias of ~1% (which would imply a consistant error in one direction), than one which has an error of 1% and a bias of ~0% (which would implpy more random error). Looking at Figure S8, the errors in surface Ozone seem to be biased in one direction. That is potentially concerning if the bias is enough to significantly affect tropospheric ozone burden - as ozone is a radiatively important species this could have consequences in an Earth System model. Please also include a normalised measure of bias in Section 2.6.Finally, there are a number of figures in the supplement which would have been useful in the main paper and I did not understand why they were not in the main paper. I would recomend moving at least figures S1, S2, S5, S6, S7 and S8 into the main paper.Specific commentsln 26. Ambigous "it" in: "because it exerts strong forcing and feedbacks...". change to "because chemical and aerosol species extert strong forcings and feedbacks..."ln 49. Change "guide us build" to "guide us to build".Ln 88. here and elsewhere, please be consistent and use the term "unimportant" instead of "slow" to describe those reactions that are removed with fluxes <10 molecules cm-3 s-1, otherwise it is easily confused with the "slow species".Section 2.3. If I am to understand this correctly, when species are defined as being "slow" and/or "long-lived", they are not "removed" from the mechanism per se, rather they are solved using the analytical approach instead of the 4-th order Rosenbrock solver. Given this, I think you need to make clear here that when you say "coupled system", you mean all of the species and reactions which are solved using the Rosenbrock solver. All of the species that are "removed" or "excluded" from the coupled system, as you say later, are instead solved using the simple analytical approach.Line 99. Please can you define "distance" qualitively here before the quantitative definition.Line 112. The term "distance" is overloaded in the text to mean multiple different things. You have defined a new term "Euclidian distance" |Di-Dj|, which is a scaler, rather than the previous "distance" Di which is a vector. This Euclidian distance is then modified to cluster species together (I would call this something like "clustered distance"). Please make clear that it is this "clustered distance" that is stored in the matrix and used to calculate the cost function Z (eq 4).Line 113. When applying the 50% factor, is this applied iteratively? i.e. if two species are each others closest pair, will their Euclidian distance be reduced by a factor 0.5*0.5=0.25, as iterating through each species, or is the 50% factor only applied once? Please be clear which approach you are using.The 50% factor and 5 closest neighbours both seem quite arbitrary, but I think I can see how this approach will cause clustering of species into close families. What was the reason for the use of these two values, and did you test other values?Line 135. Would benefit from Figure S2 being in the main text here. Looking at Figure S2, I think you can make a fair cost-benefit argument that M=20, N=13 is well optimised as it is on the bottom-left of the 30% contour. It is clear from the contour lines that you get diminishing returns of the fraction of species if you were to increase M and N, hence selecting on the 30% contour seems reasonable. By selecting the bottom-left part of the 30% contour, you are minising both N and M.Basically, you can be more rigorous in the justification for why you used M=20, N=13 than simply saying you "choose" them.Line 137-139. Please move this text to the top of section 2.5, as this describes the training set used to derive the submechanisms and blocks.Line 196. Please again change "slow reactions" to "unimportant reactions". Please clarrify - are the unimportant reactions removed from each of the submechanisms using the original training data, or are they removed on the fly depending on the concentrations of species in each grid cell at each timestep?Removing the reactions from each of the submechanisms in advance seems like the more efficient approach to me. However, there is a risk that the approach becomes inconsistant if used in different time periods with different chemical conditions to the training data. For example, there will be risk in using submechanisms derived with present day training data in preindustrial conditions. This is relevant for application in Earth System models.line 229. The full mechanism is by definition not a submechanism. Say that it is the 21st "chemical regime".Line 236. slow reactions -> unimportant reactionsFigure 1. The line in panel a shows the fraction of species removed from the coupled mechanism. Call the slow reactions unimportant reactions.Figure 3. There are 21 chemical regimes, made up of M=20 submechanisms plus the whole mechanism.Figure 4. Number of chemical regimes is 21 (M+1), not 20.Supplementary materialline 12. think Z2 should be called f or f(M,N) instead. Use a different term to D for the gridcells, because D is already used for the distance vectors.line 18-20. Unclear how this algorithm works, don't know what is meant by "temperature" here.figure S5. presumably chemical regimes should go up to 21, not 20? Also, its the 21st regime that has 100% of the mechanism, regime 20 has 90% according to Figure 3.Figure S6. Please include panels showing biases for the key species. If any have consistant/growing biases, that should be commented on (especially concerning for ozone). What value of δ did you use here?
-
AC2: 'Reply on RC2', Lu Shen, 23 Aug 2021
The comment was uploaded in the form of a supplement: https://gmd.copernicus.org/preprints/gmd-2020-425/gmd-2020-425-AC2-supplement.pdf
-
AC2: 'Reply on RC2', Lu Shen, 23 Aug 2021
Lu Shen et al.
Data sets
Replication Data for: A machine learning-guided and accurate algorithm to halve the computational cost of atmospheric chemistry in Earth System models Shen et al. https://doi.org/10.7910/DVN/KASQOC
Model code and software
GEOS-Chem code GEOS-Chem Team https://doi.org/10.5281/zenodo.1343547
Lu Shen et al.
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 448 | 173 | 8 | 629 | 41 | 1 | 3 |
- HTML: 448
- PDF: 173
- XML: 8
- Total: 629
- Supplement: 41
- BibTeX: 1
- EndNote: 3
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1