Detection of critical PM<sub>2.5</sub> emission sources and their contributions to a heavy haze episode in Beijing, China, using an adjoint model

Zhai, Shixian; An, Xingqin; Zhao, Tianliang; Sun, Zhaobin; Wang, Wei; Hou, Qing; Guo, Zengyuan; Wang, Chao

doi:https://doi.org/10.5194/acp-18-6241-2018

Articles | Volume 18, issue 9

https://doi.org/10.5194/acp-18-6241-2018

© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 3.0 License.

Special issue:

Regional transport and transformation of air pollution in...

https://doi.org/10.5194/acp-18-6241-2018

© Author(s) 2018. This work is distributed under
the Creative Commons Attribution 3.0 License.

Articles | Volume 18, issue 9

Research article

|

03 May 2018

Research article |

| 03 May 2018

Detection of critical PM_2.5 emission sources and their contributions to a heavy haze episode in Beijing, China, using an adjoint model

Shixian Zhai, Xingqin An, Tianliang Zhao, Zhaobin Sun, Wei Wang, Qing Hou, Zengyuan Guo, and Chao Wang

Download

Final revised paper (published on 03 May 2018)
Supplement to the final revised paper
Preprint (discussion started on 25 Nov 2016)
Supplement to the preprint

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

RC1: 'Review comments on “Detecting critical PM2.5 emission sources and their contributions to a heavy haze episode in Beijing, China by using an adjoint model” by Zhai et al.', Anonymous Referee #2, 13 Jan 2017
- AC1: 'Response to referee #2', Shixian Zhai, 14 Feb 2017
SC1: 'short comments', Guiqian Tang, 13 Feb 2017
RC2: 'review', Anonymous Referee #3, 21 Feb 2017

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Shixian Zhai on behalf of the Authors (26 Apr 2017) Author's response Manuscript

ED: Referee Nomination & Report Request started (30 Apr 2017) by Renyi Zhang

RR by Anonymous Referee #3 (15 May 2017)

Suggestions for revision or reasons for rejection

The revised manuscript by Zhai et al. has addressed many of the previous reviewer concerns, particularly with regards to expanding their work to better demonstrate the value of their adjoint modeling. They also include additional comparison to observations, and expanded comparison to previous work. Still, I found the description of the treatment of secondary aerosol species such as sulfate and nitrate very confusing (not clear if these are even modeled as secondary species), the model evaluation showed significant model overestimation, and the presentation of adjoint sensitivities for the purpose of source attribution is not considered or presented carefully. I believe addressing these issues constitutes another round of major revisions, after which this article will be more suitable for publication in ACP.

Major comments:

A large fraction of the PM2.5 concentrations in the pollution episodes studied here consist of secondary inorganic aerosol. However, there are several issues related to the treatment of such species that need to be clarified:

1. It is not clear if ISORROPIA and it’s adjoint were used in this study. The authors state that the GRAPES-CUASE aerosol adjoint includes CAM, citing Gong et al., 2003, but that version of CAM did not use ISORROPIA, which wasn’t added until Zhou et al. (2012). This needs to be clarified. Developing the adjoint of ISORROPIA is an extremely challenging task, the subject of an entire manuscript by developers of ISORROPIA (Capps et al., 2012). If the authors did include ISORROPIA, did they develop their own version, or use the one from Capps 2012 (ANISORROPIA)?

2. This issue is further confused by the fact that (Section 3.3) the authors state the model includes emissions of sulfate and nitrate. But sulfate and nitrate are species that are formed secondarily in the atmosphere, and the emitted species should be SO2 and NOx (and NH3). If instead the model somehow approximates inorganic PM2.5 as being a primary, rather than secondary, aerosol, the validity and accuracy of this assumption needs to be discussed, as it would be in conflict with most regional air quality models. It might though explain the very high bias of the GRAPES-CUASE model compared to the observations.

3. Even if the authors did include secondary inorganic aerosols in their adjoint modeling, the accuracy of their adjoint model for estimating the contribution of emissions perturbations to PM2.5 concentrations is undermined by two issues: adjoint code accuracy and nonlinearity of PM2.5 formation (even within a perfect adjoint code).
3a Adjoint accuracy: While trying to investigate this in the present paper and previous manuscripts, I came to realize that the GRAPES-CAUSE model has not demonstrated the accuracy of the emissions sensitivity coefficients, which is particularly important if the authors are calculating the sensitivity of secondary aerosol species (e.g. nitrate) with respect to precursor emissions (NOx). Maybe they don’t include secondary aerosols in their work (not clear, see previous question), but if they do they need to provide detailed evaluation of adjoint sensitivities to demonstrate it is working. In An et al. (2016) they present results for the passing the Lagrange condition for a particular model configuration, but the details of that configuration were not clear (how long was the simulation? what species were included in xrow?), and results specifically relating to aerosol thermodynamics were not reported. They only showed tests with respect to concentrations, not emissions. Given the difficulties identified in Capps et al. (2012) with evaluating the extremely nonlinear, discontinuous ISORROPIA behavior using finite difference approaches, I’m skeptical as to the accuracy of the aerosol adjoint sensitivities for SO2, NOx and NH3. It would helpful if they could show some finite difference tests relating estimates of the change in PM2.5 owing to perturbations of emissions of SO2, NOx and NH3 in particular grid cells, for perturbations of various magnitudes. Not only would these tests evaluate the accuracy of the model, they would also help determine the range of perturbations over which the adjoint sensitivities can be interpreted as source contributions, as discussed below.

3b Nonlinearity: For nonlinear models of PM2.5 formation, adjoint sensitivities can not be directly equated to source contributions, even if they are correctly coded. In my original response, I requested the authors clarify this issue in their description of the use of adjoint sensitivities, but their modifications in this regard fell short. The abstract presents adjoint sensitivities as contributions, with no consideration for this issue. The first paragraph of the introduction (and several other locations_ highlights the benefits of adjoint sensitivity analysis without recognizing the limitation of this approach for evaluating complete source contributions. The authors need to think more carefully about how their results can be considered. Over what range of emissions perturbations, to which species, are their sensitivities accurately representative? Did they test this? Or there are other studies in the literature to reference here? Does this impact their comparison to the forward modeling study of Zhai et al (2016)?

4. I appreciate the additional content on model evaluation. Still, the performance here seems to be presented in a somewhat overly optimistic light. The authors did not deeply consider reasons why their model seems to overestimate PM2.5 peaks by >200 ug/m3. They mention the possibility of model resolution error, but the biases seem large and typically in the opposite direction of model resolution bias, although if the authors know that the measurement sites are located in locations that are more pristine than the surrounding areas, this should be stated. It seems that other sources of error, such as uncertainties in emissions, or treatment of all PM2.5 as primary species (see other comments with regards to sulfate, nitrate, etc.) may be to blame. Lastly, and even more importantly, the authors should figure out how these types of model performance deficiencies impact their interpretation of adjoint sensitivities. Can these be used to estimate uncertainties in their sensitivity analysis as well? The authors also did not address my previous specific questions about the ability to simulate such peaks without inclusion of heterogenous chemical reactions recently identified (Wang et al., PNAS, 2016, doi:10.1073/pnas.1616540113; Cheng et al., Science Advances, 2016, doi:10.1126/sciadv.1601530) — I would have thought their model would be low biased, not high biased. Please explain.

5: Most of the sensitivity results appear to be integrated across all sectors and species, and reflect the influence of all PM2.5 emissions. It would be useful, however, to know the extent to which different species and sector contribute as well, as policies generally target particular species and sectors. It would also show another benefit of the adjoint modeling approach, which provides information across species and sectors without any additional model runs.

Minor comments:

- Abstract: missing a period.

- ISORROPIA is mis-spelled.

- I was curious about their use of the word “unequilibrated”, so I checked Henze et al. (2007). In the latter, this word was used specifically with regards to the input variables for the aerosol thermodynamic equilibrium calculation. It does not make sense to use in the current manuscript as written — I suggest “…integrated to save the model state variables (concentrations) in checkpoint files…”.

Abstract, last sentence: This is a key point, but it needs rewording. Suggest: “…controlling air pollutant sources from regions identified using adjoint sensitive analysis would lead to greater PM2.5 reductions per source control than from regions with the greatest emission intensities.”

Introduction: The phrase “adjoint sensitive zone” needs to reworded, or at least defined, as “zones with maximum adjoint sensitivities” or similar.

Fig 5: Why is there also not a black line connecting the triangles showing the observation points? This would make it easier to compare the patterns of the observations as compared to the model.

Section 4.3: Although now defined earlier, as requested, it might be good to restate here what is defined as “local”.

Hide

RR by Anonymous Referee #2 (18 May 2017)

Suggestions for revision or reasons for rejection

This study attempts to quantitatively identify the aerosol contributions from different regions to a typical Beijing haze event by using aerosol adjoint model. Since the quantifying of relative contributions of local emission and regional transport is a critical issue for air pollution management and is still in debate, this study could be interest of readership of ACP journal given the relatively comprehensive analysis and solid discussions performed in this study. But several concepts below still need a little more clarification before going to publication.
1. In several places (e.g., the titles of section 4.2, Figs. 2 and 7-10), it seems this study focuses on primary emission sources. But the modeling study by GRAPES–CUACE adopted in this study did consider the chemistry of secondary aerosol formation and thus the secondary source of PM2.5, according to the description in section 3.2. Why the authors did not report the results related to secondary aerosol source? Or what the exact definition of “primary emission sources” here? To avoid confusion, please also clarify this in abstract. That is, denote that this study is working on primary emission sources only.
2. In section 4.3, what the exact definition of “local” and “surrounding emission”? For example, how far geographically away from the receipt site in Beijing could be treated as the local emission? For example, the emission in Beijing municipality is local or surrounding ones? In addition, if the gas precursors were transported from a location of the long distance to the receipt site in Beijing and formed secondary aerosols, would this portion of locally formed aerosols be attributed to local or surrounding emission in this modeling study?
3. The content in Figure 1 (particularly Fig. 1e-h) is not readable. Please provide clearer one. In the caption of Figure 11, both red and blue shaded in the figure denote sensitive Huabei?

Hide

ED: Reconsider after major revisions (12 Jul 2017) by Renyi Zhang

AR by Shixian Zhai on behalf of the Authors (27 Jul 2017) Author's response Manuscript

ED: Referee Nomination & Report Request started (14 Aug 2017) by Renyi Zhang

RR by Anonymous Referee #2 (17 Aug 2017)

RR by Anonymous Referee #1 (28 Aug 2017)

Suggestions for revision or reasons for rejection

Review comments on “Detecting critical PM2.5 emission sources and their contributions to a
heavy haze episode in Beijing, China by using an adjoint model ” by Shixian Zhai et al.

General comments

This manuscript practices the application of an adjoint model of GRAPES-CUACE developed by An et al. (2016) in tracking PM2.5 influential emissions during a typical haze episode in Beijing, China’s massive capital. As a kind alternative source apportionment technique, the mode using the gradient of the objective function to emissions, is able to quantitatively present the emission contribution to PM2.5 concentrations. The results are impressive, and the study is helpful to understand how the regional emissions impact local PM2.5 levels, especially in Beijing. The subject provides an alternative way to conduct sensitivity analysis of emissions to concentrations. It is meaningful to air quality agencies to develop emission control measures and emission reduction plan. However, the basic assumption of “the primary particulate emission sources and PM2.5 concentrations have an approximately linear relationship” works only under the condition of calm weather. This assumption, from my point of view, may be very questionable if the enhanced inflow and high background particulate levels due partly to strong transport.
The quality of the manuscript including all figures has been significantly improved as the result of revised version. I recommend it be accepted to publish in Atmospheric Chemistry and Physics via minor revisions.

Major comments

The case study is based on single mechanism (CAM+RADM2) for aerosol and gas phase. I am wondering if different mechanisms affect PM2.5 concentrations. The conclusion may become stronger if the authors could provide the level of uncertainties of the results caused by different mechanisms.

For comparing adjoint model results with Models-3/CMAQ simulations, the authors should use Source Apportionment (SA) tool attached to CMAQ rather than different sensitivity analysis CMAQ runs.

Minor comments

Abstract, Line 32 - 34 on Page 1: “This work also reflects …” might be “The results imply that sensitive regional emissions reduction be more efficient than individual peak emission control for improving regional PM2.5 air quality”.

Line 43 on Page 2: ‘’Therefore, in concentration source sensitivity analysis problems,” should be “Therefore, in concentration source sensitivity analysis, the adjoint method is more computationally efficient than others such as traditional finite difference method ”

Line 57 on Page 2: “Research using approaches…have revealed that ...” should be “Researches using approaches …have revealed that...”.

Line 79 -80 on Page 3: “...during which time two PM2.5 concentration peaks occur and are set as the objective functions.” might be “during the episode two PM2.5 concentration peaks ...”

Line 111 - 112 on Page 4: “Knowledge of the impacts of emission sources on pollutant concentrations can help enact effective air pollution control strategies.” can be re-written into “Understanding of the impacts of emissions on pollutant concentrations is helpful to develop effective air pollution control strategies.”

Line 133 - 134 on Page 5:”In Fig. S3, we can see that the adjoint sensitivity results are close to the finite difference results,...”, How? I could not see it.

Line 454 - 455 on Page 15: “Contributions to PM2.5 concentration peaks from local Beijing and its surrounding provinces were compared.” may be replaced by “The peak PM2.5 contributions from Beijing local emissions and neighbor provinces sources were well compared” if you like.

Hide

RR by Anonymous Referee #3 (19 Sep 2017)

Suggestions for revision or reasons for rejection

The revised article by Zhai et al. includes several clarifications and updates. First, they clarify that their adjoint modeling does not include secondary aerosol formation. Sulfate and nitrate are instead treated as primarily emitted species. However, they do provide additional verification of the adjoint modeling results for primary species; these are linear and thus their interpretation of these sensitivities as being representative of large emissions perturbations is consistent with the forward model. These don’t happen to be consistent with reality though, as sulfate and nitrate are not primary species. The potential impacts of these assumptions on their forward model evaluation is not considered.

Overall, I request that the authors remove sulfate and nitrate from their analysis and focus on their adjoint sensitivity analysis and source apportionment of primary species (BC, OC, dust, sea salt) until the point that their model is mature enough to include secondary species. Treating sulfate and nitrate as primary species, without any evaluation of the errors in this assumption, is not a good approach. Removing this problematic aspect would alleviate much of the concerns from my review, and the other reviewer, regarding this paper.

Detailed comments in response to their revisions are provided below. Addressing these may involve another round of major revisions, to remove sulfate and nitrate from this study.

Specific comments:

- Because the authors do not include the adjoint of thermodynamic partitioning, they can not properly include secondary aerosols in their simulation. Their work-around for this is to treat sulfate and nitrate as primary species. This is very non-physical - sulfate and nitrate are secondary species, not emitted species, and any air quality model from the past two decades treats them as such. If they authors intend to introduce this major change in modeling, they need to quantify how much of an impact it makes on their forward model performance. This could be done, for example, by comparing complete forward model simulations (that include aerosol thermodynamics and secondary PM formation) with their primary-only version. Otherwise, their reference to earlier studies that have evaluated the performance of the forward model (e.g., Zhou et al., 2012; Want et al., 2015a, 2015b, Jiang et al., 2015) have no baring here.

- The statement that “adjoint sensitivity analysis are not significantly affected by overestimation of PM2.5” is not correct. The accuracy of the adjoint model is directly related to accuracy of the version of the forward model for which the adjoint is created. This appears to be a version of the forward model that does not include secondary aerosol formation and instead treats sulfate and nitrate as primary emissions. This is not a forward model that is recommended, nor one that has been previously evaluated. While it is true that relative sensitivities from a linear model may be correct regardless of the absolute estimate of PM from the forward model, the authors would have to keep their study limited to the linear components of PM2.5 (i.e., species that are actually primary in the atmosphere: BC, OC, dust and sea salt), for this statement to apply. It is also correct that the sensitivities from an adjoint model may match the same sensitivities in the forward model, as the authors show in Fig S3 — but this does not mean these sensitivities match real-world behavior.

- The authors’ response to my question about the possible sources of uncertainty leading to overestimation of PM2.5 concentrations by 200 ug/m3 was to provide a generic, obvious, list of factors that contribute to forward model performance: emissions, atmospheric processing, and meteorology. This provides zero insight into the nature of the uncertainty in their model. They do not specifically quantify or estimate how treatment of all PM2.5 as primary may impact their results. They do not specifically quantify, estimate, or even mention the lack of heterogenous chemistry in even the complete forward GRAPES-CUACE model that has been implicated in several recent papers (see my previous review) as playing a key role in leading to peak PM2.5 episodes.

- The fact that this modeling study only includes primary aerosols needs to be stated in the abstract and introduction, not just in the detailed description of the model (3.2).

- Section 3.2 needs to explicitly state that the authors do not include an adjoint of aerosol thermodynamic partitioning — this still is not explicitly stated.

- The phrase “rigorous mathematical verification” redundant and vague. Any type of mathematical verification is rigorous by definition, yet this statement still fails to convey the essence of what has been verified. The test they have performed is referred to in the adjoint modeling community as the Lagrange condition; that would be a more informative way to describe this evaluation.

- The additional finite difference tests are appreciated, and show that the adjoint sensitivities are rather consistent. The consistency of the sensitivities across different perturbation magnitudes is a consequence of the linearity of primary PM2.5. It does not match the known non-linear response of actual PM2.5 concentrations in China to changes in SO2 and NOx emissions (e.g., Wang et al., ES&T, 45, 9293-9300, 2011).

Hide

ED: Reconsider after major revisions (27 Sep 2017) by Renyi Zhang

AR by Shixian Zhai on behalf of the Authors (08 Nov 2017) Author's response Manuscript

ED: Referee Nomination & Report Request started (13 Nov 2017) by Renyi Zhang

RR by Anonymous Referee #2 (25 Nov 2017)

RR by Anonymous Referee #3 (27 Nov 2017)

Suggestions for revision or reasons for rejection

The authors have made it more clear that the forward model includes secondary aerosol formation.
As the primary results of this paper though are the source attribution results from the adjoint model, it still suffices to say they’re not really treating secondary PM2.5 in their present study — even if yes it is included in the forward model. So I appreciate that they have qualified their results as pertaining to “primary” species throughout the article.

However, I was expecting more. The approximation necessary for constructing the adjoint model using primary tracers for sulfate and nitrate are not explicitly stated, and some details and evaluation are missing. For example, how do the authors construct an emission inventory of aerosol nitrate? Do they assume a particular fraction of NOx becomes nitrate? Same question for sulfate and SO2? The authors need to be very clear about these sorts of details in order for the interpretation of their sensitivity results for these species to be meaningful; or they should leave out these species and focus on the primary species. They should still demonstrate, with their forward model alone, that sulfate and nitrate can be treated as primary species for these case studies. It’s not clear to me whether or not the full forward model including ISORPIA etc was used for the finite difference perturbation calculations included in figure S3 — the description of this critical verification test is quite limited — we aren’t told which species emissions are being reduced — all of them? It’s just very hard to imagine that in the full forward model a 90% change in NOx emissions would elicit a linear PM2.5 response. If that is true in their model, then they need to explain why their model differs from other previous studies examining the non-linear response of PM2.5 to NOx emissions changes.

- The first reviewer’s comment about the linearity of the adjoints only being correct for calm systems isn’t correct. That reviewer is mistakenly thinking of the adjoint of a weather model; in contrast here the active variables are only tracers. The transport of tracers is always linear, regardless of calm or dynamic systems. The non-linearities of concern are those related to the chemical environment.

- 165: A critical correction, the “GRAPES-CAUCE aerosol adjoint model is NOT capable of coupling major aerosol processes in CAM”…

- the manuscript requires grammatical proofreading throughout

Hide

ED: Reconsider after major revisions (19 Dec 2017) by Renyi Zhang

AR by Shixian Zhai on behalf of the Authors (28 Jan 2018) Author's response Manuscript

ED: Referee Nomination & Report Request started (19 Feb 2018) by Renyi Zhang

RR by Anonymous Referee #4 (19 Mar 2018)

ED: Publish subject to minor revisions (review by editor) (24 Mar 2018) by Renyi Zhang

AR by Shixian Zhai on behalf of the Authors (04 Apr 2018) Author's response Manuscript

ED: Publish as is (05 Apr 2018) by Renyi Zhang

AR by Shixian Zhai on behalf of the Authors (08 Apr 2018) Manuscript

Download

Article (4933 KB)
Full-text XML

Short summary

The GRAPES–CUACE aerosol adjoint model was developed and applied in detecting PM_2.5 sources for haze events in eastern China (EC). The response time of Beijing PM_2.5 pollution peaks to local and surrounding emissions is quantized for regional transport of air pollution over the EC. The adjoint results agreed well with the Models-3/CMAQ assessments. The adjoint method is powerful in simulating the receptor–source relationship and can be utilized in dynamic air quality control scheme design.

Detection of critical PM2.5 emission sources and their contributions to a heavy haze episode in Beijing, China, using an adjoint model

Download

Interactive discussion

Peer-review completion

Detection of critical PM_2.5 emission sources and their contributions to a heavy haze episode in Beijing, China, using an adjoint model