the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 30 Aug 2024
Improving the EnSRF in the Community Inversion Framework: a case study with ICON-ART 2024.01
Abstract. The Community Inversion Framework (CIF) brings together methods for estimating greenhouse gas fluxes from atmospheric observations. While the analytical and variational optimization methods implemented in CIF are operational and have proved to be accurate and efficient, the initial ensemble method was found to be incomplete and could hardly be compared to other ensemble methods employed in the inversion community, mainly owing to strong performance limitations and absence of localization methods. In this paper, we present and evaluate a more efficient implementation of the ensemble mode. As a first step, we chose to implement the serial and batch versions of the Ensemble Square Root Filter (EnSRF) algorithm because it is widely employed in the inversion community. We provide a comprehensive description of the technical implementation in CIF and the useful features it can provide to users. Finally, we demonstrate the capabilities of the CIF-EnSRF system using a large number of synthetic experiments over Europe, exploring the system’s sensitivity to multiple parameters that can be tuned by users. As expected, the results are sensitive to the ensemble size and localization parameters. Other tested parameters, such as the number of lags, the propagation factors, or the localization function can also have a substantial influence on the results. We also introduce and provide a way of interpreting a set of metrics that are automatically computed by CIF and that can help assessing the success of inversions and comparing them. This work complements previous efforts focused on other inversion methods within CIF. With the integration of these new ensemble algorithms, any chemical transport model (CTM), including models without existing adjoint, can now perform inversions using CIF, leveraging its robust capabilities.
-
Notice on discussion status
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
-
Preprint
(17459 KB)
-
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(17459 KB) - Metadata XML
- BibTeX
- EndNote
- Final revised paper
Journal article(s) based on this preprint
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2024-2197', Arne Babenhauserheide, 15 Oct 2024
General Comments
The manuscript describes the addition of an Ensemble Kalman Smoother
(EnSRF) to the CIF inversion model and tests them with the ICON-ART
atmospheric tracer model.
The manuscript is well written and structured and overall in very good
shape. Its descriptions of Kalman methods are great and the topic is
relevant. The experimental setup and benchmark are also
appropriate for a new implementation. The evaluation of the results
could be a bit more extensive, though.
Figures and tables are correct and informative, though some of the
figures would benefit from added data in tables to be able to compare
numerical values.
The ease of comparing different methods with exactly the same setup
except for the inverse method is important:
https://doi.org/10.5194/acp-15-9747-2015 (own work) showed that
implementation details can have a bigger impact on the fluxes than the
method itself and using a unified assimilation system should allow to
avoid that.
I recommend publishing with minor changes and am willing to review
again.
Specific comments
Abstract:
- L 5: "mere efficient implementation": Table D1 shows a benchmark of
the new code. Is there a benchmark of the old code to support this
description?
- L 15: what does ICON-ART bring as advantage? Why does this use
ICON-ART? It is in the title, so it should be in the abstract.
Introduction:
- L 39: very nice description! (nothing to change, just a note)
- L 74: please add a short overview of numerical weather prediction
(NWP), tracer models, ICON, and ART. The reports on ICON may be a
good resource; either to cite relevant reports, or to add a link in
a footnote:
https://www.dwd.de/DE/leistungen/reports_on_icon/archiv/archiv_reports_on_icon.html
Maybe note that more details are provided in 4.1.1.
2.2
- L 178: what happens in case of nonlinear effects like OH-destruction
of CH4? This is not a problem, but should be noted.
2.3.2:
- L 218: “R matrix is diagonal”: for which observations is this true?
What’s the scale of weather patterns?
3.1.1:
- L 293: nlag 2: please note below this the reason for using nlag 2.
Also maybe reference the results from Table 3. Maybe also the
benchmark in Table D1.
- L 311: “eigenvalues and eigenvectors are already computed”: how well
does this scale? Is there a practical limit to the model resolution
due to this inversion? Also maybe reference Table D1 (it shows that
CIF performance is not the bottleneck).
- L 338: it is not clear whether the ensemble in the runs in the
publication is generated from the B matrix. Please note explicitly
whether it is or why it is not.
3.1.2:
- L 356 to 358: this part is hard to understand. Suggestion: “for
window 2 of the full period, this is the second time it is optimized
from the observations.”
- Figure 1 is beautiful and clear!
- L 361: why “nlag + 1”? The diagram only seems to show nlag.
- L 362 to 363: this note may be better placed at the examples, 3.1.1
at line 294.
3.2
- L 405: “a wealth of” this is marketing language. Please give a short
abstract of the features instead.
3.2.1:
- L 417: is it possible to quantify how much was possible before and
how big simulations could be used before and how big they can be
now? Maybe rough, like “Factor X”?
3.3.3:
- eq 39 (line 479): if J(xa) is 0, this is 1. Are higher values better?
4.1.1:
- line 569: are higher resolution possible with this implementation?
For example 6.5km? This might significantly improve the possible
gains from high frequency column measurements. Are there boundary
conditions for the Europe area?
4.2.1:
- Figure 3: please note that this is the best result from the LEVEL 1 comparison.
- Figure 3: please add more explanation to these results. Why are the
neighboring cells so different in the error reduction (c)? Also
please provide a table with the values of d (the clearest result) so
readers can quantitatively compare the error reduction per country.
4.2.2:
- L 641: reference the results of this experiment in 3.1.1 to motivate
why nlag 2 is used in the example.
- Table 2: why is the maximum lag 3? Would lag 5 be possible? Which
lags are used in current CTDAS runs? Reference Table D1.
4.3:
- Figure 5, d: why does this show France, when the maps in Figure 4
show little change in MUR in France? It could be good to split out
MUR from Figure 5 and add a set of plots for MUR in different
countries. And give possible explanations of the differences.
Portugal has no direct observations, but changes. MUR in UK also
seems to change much more than in France. There are also countries
in eastern Europe with bigger changes. Are these effects from being
at the border of the simulated area?
- Line 693: as above, see the maps in Figure 4. Other countries seem
to show bigger changes than France.
- line 751: the note about the heaviside function seems to be evading
a result. Maybe compare with the other functions at larger
correlation length?
- line 760: “to the best of our knowledge” sounds like this sentence
should not be stated. Remove?
4.5:
- line 777: “consistent with established literature”: please reference
appendix B.
Appendix D:
- Table D1: please reference this table from table 2, page 26. Please
give some additional information, how this benchmark data was gathered.
Technical corrections
- L 18: add a comma: (N2O), or
- L 231: the note duplicates parts of line 218. Please disentangle the
two sentences to avoid the duplication.
- L 249: wrong hyphenation: anal-ysis — might need explicit hyphenation via \hyphenation{a-na-ly-sis}
- L 283: which format is used? Where is it described? Citation, or a
footnote with a link?
- L 314 to 122: a long discussion for a simple point. Suggestion: “the
default behavior of CIF is to erase the scaling factors after
conversion. But some models (including ICON-ART) need scaling
factors as import, so they are preserved here.” Feel free to shorten
further.
- L 333: “and a scaling” this seems to be missing the word “factor”.
- line 635: please add a note that the results are discussed in section 4.3.
Appendix B:
- Figure B1: please label the colorbars (scaling factor / unitless)
- Figure B2: please label the colorbars with the variable shown (uncertainty reduction / unitless)
Appendix C:
- Figure C1: please show the lines with different line styles to be recognizable in grayscale.Citation: https://doi.org/10.5194/egusphere-2024-2197-RC1 -
RC2: 'Comment on egusphere-2024-2197', Anonymous Referee #2, 11 Nov 2024
In this study, the authors improved and upgraded the initial version of the EnSRF method in CIF, including enhancements to the batch optimization and localization methods, as well as improvements in technical details to increase assimilation speed, post-processing speed, and overall performance.
The manuscript is well-written and clear in structure. Especially, the description of the theoretical process is very comprehensive and detailed. Additionally, it includes the technical implementation and the characteristics of both the initial and current versions. The synthetic experiments demonstrate the performance of the CIF-EnSRF method and test the impact of different parameter schemes. An additional comparison with CTDAS is provided to prove that the current version of CIF-EnSRF is almost equivalent in effect to CTDAS. I recommend that this manuscript be accepted after the following clarifications and revisions.
- P6, L160: The dimensions [n×N] should follow Z, not B, as B has dimensions[n×n].
- P24, Figures 3a and 3b show the map of the true and posterior scaling factors. However, I have a question: why not include a comparison of the maps for the prior, posterior, and true fluxes, as flux maps could provide a more direct view of their spatial distribution? Additionally, are the values of the prior scaling factors equal to 1?
- Section 4.2.2 and Table 2: In LEVEL2 experiment (F), the authors investigated the impact of adjusting the mean and variance of the ensemble on the inversion results. Please provide the values for the mean and variance of the ensemble in both the adjusted and unadjusted experiments, i.e., what were the mean and variance values used in Table 2 for EXP_F_f1* and EXP_F_t1*? Only by providing the specific data parameters and comparing the extent of their changes can the importance of this parameter be better demonstrated.
- In Section 4, a series of synthetic experiments for CO2 flux inversion were conducted to demonstrate the performance of the CIF-EnSRF method. In Appendix B, the authors conduct a further comparison with CTDAS, but this time they selected the synthetic inversion of CH4. There are some notable differences between CH4 and CO2 flux inversions; for example, the assimilation module for CH4 flux inversion involves generating separate perturbations for natural and anthropogenic emission categories from the B matrix, which is not the case for CO2. Could you clarify why CO2 flux inversion experiments were used earlier, but in Appendix B, CH4 flux inversion experiments were chosen for comparison instead of continuing with CO2 flux inversion?
- P37, L844-L858: What are the differences between the CTDAS configuration and the CIF-EnSRF configuration here? It would be helpful to provide a comparison of their configurations and explain the similarities and differences.
- While the ICON-ART model is highlighted in the manuscript title, it is not mentioned in the abstract or introduction. It might be helpful to include at least a brief mention of ICON-ART in these sections.
Citation: https://doi.org/10.5194/egusphere-2024-2197-RC2 -
AC1: 'Comment on egusphere-2024-2197', Joel Thanwerdas, 20 Dec 2024
We thank the two referees for their invaluable insights, which have greatly enhanced the quality of the paper. We provide here a comprehensive response to the comments received. Referee#1's comments are in red and Referee#2's comments are in blue. For each comment, an answer is provided in normal text and the modifications from the new version of the manuscript are provided in bold and small text.
Interactive discussion
Status: closed
-
RC1: 'Comment on egusphere-2024-2197', Arne Babenhauserheide, 15 Oct 2024
General Comments
The manuscript describes the addition of an Ensemble Kalman Smoother
(EnSRF) to the CIF inversion model and tests them with the ICON-ART
atmospheric tracer model.
The manuscript is well written and structured and overall in very good
shape. Its descriptions of Kalman methods are great and the topic is
relevant. The experimental setup and benchmark are also
appropriate for a new implementation. The evaluation of the results
could be a bit more extensive, though.
Figures and tables are correct and informative, though some of the
figures would benefit from added data in tables to be able to compare
numerical values.
The ease of comparing different methods with exactly the same setup
except for the inverse method is important:
https://doi.org/10.5194/acp-15-9747-2015 (own work) showed that
implementation details can have a bigger impact on the fluxes than the
method itself and using a unified assimilation system should allow to
avoid that.
I recommend publishing with minor changes and am willing to review
again.
Specific comments
Abstract:
- L 5: "mere efficient implementation": Table D1 shows a benchmark of
the new code. Is there a benchmark of the old code to support this
description?
- L 15: what does ICON-ART bring as advantage? Why does this use
ICON-ART? It is in the title, so it should be in the abstract.
Introduction:
- L 39: very nice description! (nothing to change, just a note)
- L 74: please add a short overview of numerical weather prediction
(NWP), tracer models, ICON, and ART. The reports on ICON may be a
good resource; either to cite relevant reports, or to add a link in
a footnote:
https://www.dwd.de/DE/leistungen/reports_on_icon/archiv/archiv_reports_on_icon.html
Maybe note that more details are provided in 4.1.1.
2.2
- L 178: what happens in case of nonlinear effects like OH-destruction
of CH4? This is not a problem, but should be noted.
2.3.2:
- L 218: “R matrix is diagonal”: for which observations is this true?
What’s the scale of weather patterns?
3.1.1:
- L 293: nlag 2: please note below this the reason for using nlag 2.
Also maybe reference the results from Table 3. Maybe also the
benchmark in Table D1.
- L 311: “eigenvalues and eigenvectors are already computed”: how well
does this scale? Is there a practical limit to the model resolution
due to this inversion? Also maybe reference Table D1 (it shows that
CIF performance is not the bottleneck).
- L 338: it is not clear whether the ensemble in the runs in the
publication is generated from the B matrix. Please note explicitly
whether it is or why it is not.
3.1.2:
- L 356 to 358: this part is hard to understand. Suggestion: “for
window 2 of the full period, this is the second time it is optimized
from the observations.”
- Figure 1 is beautiful and clear!
- L 361: why “nlag + 1”? The diagram only seems to show nlag.
- L 362 to 363: this note may be better placed at the examples, 3.1.1
at line 294.
3.2
- L 405: “a wealth of” this is marketing language. Please give a short
abstract of the features instead.
3.2.1:
- L 417: is it possible to quantify how much was possible before and
how big simulations could be used before and how big they can be
now? Maybe rough, like “Factor X”?
3.3.3:
- eq 39 (line 479): if J(xa) is 0, this is 1. Are higher values better?
4.1.1:
- line 569: are higher resolution possible with this implementation?
For example 6.5km? This might significantly improve the possible
gains from high frequency column measurements. Are there boundary
conditions for the Europe area?
4.2.1:
- Figure 3: please note that this is the best result from the LEVEL 1 comparison.
- Figure 3: please add more explanation to these results. Why are the
neighboring cells so different in the error reduction (c)? Also
please provide a table with the values of d (the clearest result) so
readers can quantitatively compare the error reduction per country.
4.2.2:
- L 641: reference the results of this experiment in 3.1.1 to motivate
why nlag 2 is used in the example.
- Table 2: why is the maximum lag 3? Would lag 5 be possible? Which
lags are used in current CTDAS runs? Reference Table D1.
4.3:
- Figure 5, d: why does this show France, when the maps in Figure 4
show little change in MUR in France? It could be good to split out
MUR from Figure 5 and add a set of plots for MUR in different
countries. And give possible explanations of the differences.
Portugal has no direct observations, but changes. MUR in UK also
seems to change much more than in France. There are also countries
in eastern Europe with bigger changes. Are these effects from being
at the border of the simulated area?
- Line 693: as above, see the maps in Figure 4. Other countries seem
to show bigger changes than France.
- line 751: the note about the heaviside function seems to be evading
a result. Maybe compare with the other functions at larger
correlation length?
- line 760: “to the best of our knowledge” sounds like this sentence
should not be stated. Remove?
4.5:
- line 777: “consistent with established literature”: please reference
appendix B.
Appendix D:
- Table D1: please reference this table from table 2, page 26. Please
give some additional information, how this benchmark data was gathered.
Technical corrections
- L 18: add a comma: (N2O), or
- L 231: the note duplicates parts of line 218. Please disentangle the
two sentences to avoid the duplication.
- L 249: wrong hyphenation: anal-ysis — might need explicit hyphenation via \hyphenation{a-na-ly-sis}
- L 283: which format is used? Where is it described? Citation, or a
footnote with a link?
- L 314 to 122: a long discussion for a simple point. Suggestion: “the
default behavior of CIF is to erase the scaling factors after
conversion. But some models (including ICON-ART) need scaling
factors as import, so they are preserved here.” Feel free to shorten
further.
- L 333: “and a scaling” this seems to be missing the word “factor”.
- line 635: please add a note that the results are discussed in section 4.3.
Appendix B:
- Figure B1: please label the colorbars (scaling factor / unitless)
- Figure B2: please label the colorbars with the variable shown (uncertainty reduction / unitless)
Appendix C:
- Figure C1: please show the lines with different line styles to be recognizable in grayscale.Citation: https://doi.org/10.5194/egusphere-2024-2197-RC1 -
RC2: 'Comment on egusphere-2024-2197', Anonymous Referee #2, 11 Nov 2024
In this study, the authors improved and upgraded the initial version of the EnSRF method in CIF, including enhancements to the batch optimization and localization methods, as well as improvements in technical details to increase assimilation speed, post-processing speed, and overall performance.
The manuscript is well-written and clear in structure. Especially, the description of the theoretical process is very comprehensive and detailed. Additionally, it includes the technical implementation and the characteristics of both the initial and current versions. The synthetic experiments demonstrate the performance of the CIF-EnSRF method and test the impact of different parameter schemes. An additional comparison with CTDAS is provided to prove that the current version of CIF-EnSRF is almost equivalent in effect to CTDAS. I recommend that this manuscript be accepted after the following clarifications and revisions.
- P6, L160: The dimensions [n×N] should follow Z, not B, as B has dimensions[n×n].
- P24, Figures 3a and 3b show the map of the true and posterior scaling factors. However, I have a question: why not include a comparison of the maps for the prior, posterior, and true fluxes, as flux maps could provide a more direct view of their spatial distribution? Additionally, are the values of the prior scaling factors equal to 1?
- Section 4.2.2 and Table 2: In LEVEL2 experiment (F), the authors investigated the impact of adjusting the mean and variance of the ensemble on the inversion results. Please provide the values for the mean and variance of the ensemble in both the adjusted and unadjusted experiments, i.e., what were the mean and variance values used in Table 2 for EXP_F_f1* and EXP_F_t1*? Only by providing the specific data parameters and comparing the extent of their changes can the importance of this parameter be better demonstrated.
- In Section 4, a series of synthetic experiments for CO2 flux inversion were conducted to demonstrate the performance of the CIF-EnSRF method. In Appendix B, the authors conduct a further comparison with CTDAS, but this time they selected the synthetic inversion of CH4. There are some notable differences between CH4 and CO2 flux inversions; for example, the assimilation module for CH4 flux inversion involves generating separate perturbations for natural and anthropogenic emission categories from the B matrix, which is not the case for CO2. Could you clarify why CO2 flux inversion experiments were used earlier, but in Appendix B, CH4 flux inversion experiments were chosen for comparison instead of continuing with CO2 flux inversion?
- P37, L844-L858: What are the differences between the CTDAS configuration and the CIF-EnSRF configuration here? It would be helpful to provide a comparison of their configurations and explain the similarities and differences.
- While the ICON-ART model is highlighted in the manuscript title, it is not mentioned in the abstract or introduction. It might be helpful to include at least a brief mention of ICON-ART in these sections.
Citation: https://doi.org/10.5194/egusphere-2024-2197-RC2 -
AC1: 'Comment on egusphere-2024-2197', Joel Thanwerdas, 20 Dec 2024
We thank the two referees for their invaluable insights, which have greatly enhanced the quality of the paper. We provide here a comprehensive response to the comments received. Referee#1's comments are in red and Referee#2's comments are in blue. For each comment, an answer is provided in normal text and the modifications from the new version of the manuscript are provided in bold and small text.
Peer review completion
Journal article(s) based on this preprint
Model code and software
The Community Inversion Framework: codes and documentation v1.2 Antoine Berchet, Espen Sollum, Isabelle Pison, Rona L. Thompson, Joël Thanwerdas, Audrey Fortems-Cheiney, Jacob C. A. van Peet, Elise Potier, Frédéric Chevallier, Grégoire Broquet, and Adrien Berchet https://doi.org/10.5281/zenodo.12742377
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 260 | 76 | 103 | 439 | 16 | 12 |
- HTML: 260
- PDF: 76
- XML: 103
- Total: 439
- BibTeX: 16
- EndNote: 12
Viewed (geographical distribution)
| Country | # | Views | % |
|---|---|---|---|
| United States of America | 1 | 96 | 22 |
| Germany | 2 | 68 | 16 |
| China | 3 | 32 | 7 |
| Finland | 4 | 21 | 4 |
| France | 5 | 19 | 4 |
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
- 96
Antoine Berchet
Lionel Constantin
Aki Tsuruta
Michael Steiner
Friedemann Reum
Stephan Henne
Dominik Brunner
The requested preprint has a corresponding peer-reviewed final revised paper. You are encouraged to refer to the final revised version.
- Preprint
(17459 KB) - Metadata XML