A Geographically Weighted Gaussian Process Regression Emulator of the GCHP 13.0.0 Global Air Quality Model

Wong, Anthony Y. H.; Eastham, Sebastian D.; Monier, Erwan; Selin, Noelle E.

doi:https://doi.org/10.5194/egusphere-2025-2663

Preprints

https://doi.org/10.5194/egusphere-2025-2663

Preprints

14 Jul 2025

| 14 Jul 2025

Status: this preprint is open for discussion and under review for Geoscientific Model Development (GMD).

A Geographically Weighted Gaussian Process Regression Emulator of the GCHP 13.0.0 Global Air Quality Model

Anthony Y. H. Wong, Sebastian D. Eastham, Erwan Monier, and Noelle E. Selin

Abstract. Air quality modelling has been an essential tool to study the impacts of socio-economic changes and policies on air quality and associated social costs due to human health impacts. However, high computational and human resource demands limit the use of state-of-the-art air quality models outside of the atmospheric science community. We address this limitation by training Geographically Weighted Gaussian Process Regressors (GW-GPR) on the outputs of a series of perturbation experiments from the high-fidelity GEOS-Chem High Performance global chemical transport model (GCHP 13.0.0). The Gaussian Process Regressor relates changes in annual mean surface anthropogenic PM_2.5 to changes in short-lived air pollutant emissions and atmospheric CH₄ and CO₂ levels for each GCHP model grid cell. In comparison to existing widely adopted linearized and regionalized approaches, our method can account for sub-regional changes in air pollutant emission patterns and incorporates the non-linear response of secondary air pollutants to precursor and greenhouse gas emissions. We evaluate and demonstrate the utility of our model by predicting the global distribution of PM_2.5 in 2050 (relative to 2014) under 4 sets of climate and air pollution control policy scenarios. The emulator reproduces grid cell-level changes in anthropogenic PM_2.5 (R² = 0.94 – 0.99 over the 4 scenarios tested), and associated global changes in premature mortalities at 95 % confidence level, while requiring < 10 seconds of CPU time (vs. ~3000 CPU hours for GCHP) for each scenario. The emulator is also able to capture projected global trends of population-weighted PM_2.5 from the AerChemMIP ensemble within the ensemble range. To our knowledge, the GW-GPR emulator is the first global-scale emulator operating at grid cell level with explicit consideration of non-linearities in atmospheric chemistry, climate change, and uncertainties resulting from both chemistry and climate variability. The accuracy, speed and simplicity of the emulator also show the capability of machine learning algorithms in emulating global air quality models, and make air quality modelling accessible for global climate/air pollution scenario analysis and integrated assessment.

Received: 05 Jun 2025 – Discussion started: 14 Jul 2025

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Anthony Y. H. Wong, Sebastian D. Eastham, Erwan Monier, and Noelle E. Selin

Status: open (until 24 Sep 2025)

Post a comment Subscribe to comment alert

CEC1:
'Comment on egusphere-2025-2663 - No compliance with the policy of the journal', Juan Antonio Añel, 28 Jul 2025 reply

Dear authors,
Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".

https://www.geoscientific-model-development.net/policies/code_and_data_policy.html
First, you have archived part of your code on GitHub. However, GitHub is not a suitable repository for scientific publication. GitHub itself instructs authors to use other long-term archival and publishing alternatives, such as Zenodo.
Also, there are problems with your data. For example, for the GAINS data you cite a paper previously published in our journal, but the mentioned paper does not contain a permanent repository for it. Therefore, please, share the data in one of the acceptable repositories listed in our policy. Regarding the unprocessed output data you do not share it, and ask to contact the authors. We can not accept this, unless you provide a good reason that prevents you of sharing the data, such as for example the size of the files. We expect that you share the output variables from the simulations that you use in your work.
Therefore, the current situation with your manuscript is irregular. Please, publish your code and data in one of the appropriate repositories and reply to this comment with the relevant information (link and a permanent identifier for it (e.g. DOI)) as soon as possible, as we can not accept manuscripts in Discussions that do not comply with our policy. Also, you must include a modified 'Code and Data Availability' section in a potentially reviewed manuscript, containing the information of the new repositories.
I must note that if you do not fix this problem, we cannot continue with the peer-review process or accept your manuscript for publication in our journal.
Juan A. Añel

Geosci. Model Dev. Executive Editor

Reply

Citation: https://doi.org/10.5194/egusphere-2025-2663-CEC1
- AC1: 'Reply on CEC1', Anthony Y. H. Wong, 28 Jul 2025 reply
  
  Dear Editor,
  Thanks for pointing out the irregularities. We are working on addressing them, and provide the updated version of manuscript/repositories once we finish our work.
  On behalf of all co-authors,
  Anthony Y. H. Wong
  
  Reply
  
  Citation: https://doi.org/10.5194/egusphere-2025-2663-AC1
- AC2:
  'Reply on CEC1', Anthony Y. H. Wong, 05 Aug 2025 reply
  
  Dear Editor,
  We have uploaded a new version of our data and code for this manuscript (doi:10.5281/zenodo.16741084) , which includes a frozen and self-contained version of the emulator training and execution code. Nothing in the Github repository is required to reproduce the results of the paper. However, since we expect active maintenance and development of our emulator, we still hope to provide the GitHub link for the readers of our manuscript.
  In addition, the input (GAINS, IGSM and FAOSTAT) and output (pollutant emissions for each IGSM-GAINS-TAPS scenarios, which are used to drive the GCHP) of TAPS model are included.
  We also provide more detailed description about the content of the Zenodo repository. In particular, we explain how due to data size, we cannot share the whole raw hourly output archive from our GCHP simulations, and how the processed GCHP output (annual mean PM2.5) are included as part of the training and testing sets in the Zenodo repository.
  We will revise our data availability statement accordingly in the next version of the manuscript to reflect these changes. Particularly, we will emphasize that the Zenodo repository is self-contained to reproduce the result the emulator, without requiring any additional code from the GitHub repository.
  We hope these changes will make our manuscript compliant with the policy of GMD. Many thanks.
  On behalf of all authors,
  Anthony Y. H. Wong
  
  Reply
  
  Citation: https://doi.org/10.5194/egusphere-2025-2663-AC2
  - CEC2: 'Reply on AC2', Juan Antonio Añel, 05 Aug 2025 reply
    
    Dear authors,
    Many thanks for addressing the outstanding issues in your manuscript. We can consider the current version of your manuscript in compliance with the code and data policy of the journal.
    Juan A. Añel
    Geosci. Model Dev. Executive Editor
    
    Reply
    
    Citation: https://doi.org/10.5194/egusphere-2025-2663-CEC2
RC1: 'Comment on egusphere-2025-2663', Anonymous Referee #1, 08 Aug 2025 reply

Air quality is a global issue with significant impacts on human health. Understanding the likely long-term trend for air quality and the potential utility of different policy interventions is an important use of our scientific understanding as manifested in a range of chemical transport models. As the authors argue, the computational burden of such models make the development and use of reduced complexity models useful. In this paper the authors present a methodology of generating a reduced complexity model for surface PM2.5 concentrations based on the GCHP model driven by a somewhat complex series of climate and emissions models.
In general, I don’t think there is anything significantly wrong with the methodology outlined here. However, it is hard to make that evaluation as I found the methodology difficult to follow and understand. Both the modelling methodology to generate the training data and then Gaussian process emulation are not explained clearly and given the information provided it would be difficult to replicate. The paper needs to be significantly re-written to make it easier to better explain what has been done. I have not indicated every problem that I see in the paper. I’ve highlighted the major problems as I see it and have indicated some others as I went through the text.
I would suggest that the editor consider either “major corrections,” or “reject and re-submit,” and the authors substantially re-write the paper, and it is then re-reviewed by a different set of reviewers.
Major problems.
Modelling tools
One major concern is the clarity on how the base modelling is described. Ultimately the output comes from GCHP but the meteorology for this comes from CAM which is provided with climate information from the MIT IGSM. There are emissions from GAINS via TAPs etc. It is unclear what the inputs and outputs are from each of these models and the configurations that are being used. This ultimately raises several questions revolving around the consistency between the emissions assumptions used for the driving climate / meteorology and the air pollution? This section of the methods needs to be significantly clarified.
Some of this is laid out in Figure 2 which should come earlier in the methodology section but there needs to be a significant effort to re-write the methodology is a more coherent and straight forwards way. I would start with the MIT IGSM, explain the configuration for that and the time-period of the model runs. Then what are the outputs from that and how are they used by the subsequent modelling tools. Then move onto the next model and explain its inputs and outputs.
Reduced form modelling
There is also a lack of clarity about how the gaussian process emulation is being undertaken. What is being predicted? Is it the PM2.5 or the individual tracers needed to calculate PM2.5? Is it all the GCHP tracers? In some places it suggests its PM2.5 but then there are discussions of the anthropogenic PM2.5 which would suggest that individual components have been predicted and then in other places its seem that the individual components are being predicted. What exactly are the input / output variables to make the prediction? A table of some sort would be useful here. It would also be useful to know more explicitly what the training data consisted of. Data from what period to what period. Again, this isn’t clear. How is this a “geographically weighted” approach. This isn’t clear in the description. What is meant by this?
Improvement of the methodology
There are many ways a reduced form model can be produced. This paper describes one way. However, there is only real utility in the methods if it is “better” than other approaches. Table 3 includes columns describing the results from a Multilinear regression emulator approach. It would be useful if this could be described in the methodology more and highlighted in more detail. How does Figure 6 look like with the MLR approach? Is the additional burden of the Gaussian Process system “worth it” compared to the simpler MLR approach?

Overall, it is difficult to evaluate the work here as I can’t really understand exactly what has been done. The paper is long and covers the development of the reduced form model and then some application. It would be worth thinking about whether these applications are useful. The health effects between the full model and the emulator are identical. This isn’t a surprise given Figure 6. Similarly, it’s not obvious to me that there is much utility in the work on the AerChemMIP comparison. The numbers calculated appear to be sensible, but do we learn much here? Given the performance of the emulator in Figure 17, this work appears to be just a statement about the performance of the GCHP/CAM/MIT IGCM system compared to the AerChemMIP models rather than the emulator. There is a lot of work done here but the paper feels long. The main conclusions get lost in this, both by the range of topics discussed and the way that they are explained.

General comment.
The word “level” does a lot of work in this paper. It is used to mean “concentration” (pollution level), the vertical coordinate of the model grid (vertical level), the “degree” of global warming (level of global warming), a spatial scale (global level). It would be useful if there could be some specificity in the different words used here.
The words “high fidelity” is used in several places. It’s not clear to me what this means. An alternative set of words should be used of more context given to what the authors mean. At a resolution of ~200km this is not a “high resolution” model.

I include specific comments here.
Title: The paper title suggests an emulation of the whole model, but I think only the PM2.5 concentrations have been emulated. GCHP at this spatial resolution (~200km by 250km) isn’t really an “Air quality” model, it’s an “atmospheric composition” model or something like that but most people would think an “air quality” model would have a substantially higher spatial resolution.
Abstract
Line 19 “Widely adopted.” I don’t think any of these techniques have been “widely” adopted. I would remove this comment.
Line 37 “Uncertainties resulting from both chemistry and climate variability” I understand what climate variability is. I’m not sure what chemistry variability. However, I’m not sure that the methodology used here addresses these issues. This should either be expanded to be clearer or removed.
Introduction
Line 36. Sustainable development goals. I don’t think air quality has been explicitly stated as part of the SDGs. There isn’t an AQ SDG which is surprising. The AQ goals are given as sub, sub, SDGs (3.9.1 and 11.6.2).
Line 37. This suggests that the only way to evaluate the future air quality is through offline models. However, there are online ESM approaches which are in general the more used for this kind of long-term projections.
Line 54. “Frequently applied in recent science and policy studies” References should be given.
Line 60. “Chemical regimes” What do the authors mean here? Ozone NOx-VOC limitations? Aerosol SO4-NO3-NH4 regimes?
Line 63. SOA is an important component of PM2.5
Line 64. What are the direct vs indirect impacts of climate change on air pollution? Changes in the meteorology? Increased temperatures? Can this be more specific.
Line 73. What is meany by “geographically weighted.” This is used a lot in the paper but there isn’t a definition of what this means.
Why was Gaussian Process Regression chosen over other methods? What is it and why is an appropriate tool to use for this problem?
Methods
As indicated earlier I found this difficult to understand. In the first paragraph the authors talk about GCHP but this study uses a chain of models to generate the PM2.5 concentrations under several climate and emissions scenarios. It is very hard to understand what they have done. This section should be re-written with an introduction to explain the system being used and then details of each model used given in turn. What information is being used by which models? How is the data transferred between these models. Figure 2 is a start for this. But the textual description should be clearer and more specific. What are the inputs into the MIT GCM? What are its configurations? What are the outputs? What are the inputs into CAM? What is the CAM configuration? And then then what are the outputs? What emissions is it using? etc
This whole section should be rewritten in a much more coherent way. Some more section headings to describe the MIT GCM, CAM, TAPS, GAINS, GCHP etc and the flow of information between them. Once the model framework has been outlined the experiments performed to develop the training data can be explained.
Line 215. The choice of L appears somewhat arbitrary. Presumably it has something to do with the lifetime of the compound and some mean wind-speed component. Given a ~250km gridbox and say a ~5ms-1 wind. The timescale for air to be blown out of the box is ~12 hours. Thus L=1 seems appropriate for NO and potentially for some of the shorter-lived VOCs. L=3 (36 hours) seems very short for CO. The authors should provide more of a chemical interpretation of this in their text or identify this as a weakness of their approach.
Section 2.3
What is a random variable? Is this a normally distributed variable? This is a bit confusing as it could be construed as a variable containing random numbers, but I don’t think this is what is meant. What is N in equation (1)?
Are the PM2.5 surface concentrations and the input variables normally distributed? My guess is that they are not and many of them are likely log normally distributed. Does this matter?
What bits of information are are being used here? Exactly what is being predicted and with what information?

Section 2.4
I might move this description to be in the section 4 as it feels disjointed in the flow of the text.
Section 3.2
It would be useful there could be some description at the start of how the evaluation is going to take place, what is going to be contained in this section. As described early the metrics are only useful if they are compared to an alternative method of reduced model generation. It seems like this has been done but it would be useful if this could be the basis for the evaluation? Is the new approach better than the old, rather than providing metrics of the performance of the new model in isolation.
Line 293. This says that Figure 6 shows the comparison with the Delta Anthropogenic PM2.5 but the figure caption text says that it just Delta PM2.5. It’s not clear over what time period this is run for? What is the calculated delta between?
Line 300. Why does the standard deviation of the prediction and the MAE between the prediction and GCHP indicate that emulator SD is an appropriate measure of chemical and climate uncertainty. This should be explained in more detail.
Section 3.
Figure 9. This is very small and hard to read. I can’t see any dots in this figure, but they are described in the caption. The colour scale isn’t very useful as to my eye as everywhere seems to show a reduction other than a possibly over Bangladesh.
Table 3. The description of the multilinear approach is pages ahead in the document. It should be in the methods section and explained properly.
Is it clear why Section 3.3.2 is that rather than Section 3.4?
Discussion
I think the workflow has been applied to a number of applications outside of Engineering. There are a number of examples of similar approaches in atmospheric composition research.

Reply

Citation: https://doi.org/10.5194/egusphere-2025-2663-RC1

Anthony Y. H. Wong, Sebastian D. Eastham, Erwan Monier, and Noelle E. Selin

Data sets

Model for GW-GPR air quality emulator Anthony Y. H. Wong https://zenodo.org/records/15490120

Model code and software

Geographically weighted Gaussian Process Regression (GW-GPR) global air quality emulator Anthony Y. H. Wong https://github.com/ayhwong/GW-GPR

Models and Data for Reproducing the Result in "A Geographically Weighted Gaussian Process Regression Emulator of the GCHP 13.0.0 Global Air Quality Model Anthony Y. H. Wong https://zenodo.org/records/15484655

Anthony Y. H. Wong, Sebastian D. Eastham, Erwan Monier, and Noelle E. Selin

Viewed

Total article views: 1,054 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
991	45	18	1,054	14	16

HTML: 991
PDF: 45
XML: 18
Total: 1,054
BibTeX: 14
EndNote: 16

Views and downloads (calculated since 14 Jul 2025)

Month	HTML	PDF	XML	Total
Jul 2025	120	22	10	152
Aug 2025	413	14	8	435
Sep 2025	458	9	0	467

Cumulative views and downloads (calculated since 14 Jul 2025)

Month	HTML	PDF	XML	Total
Jul 2025	120	22	10	152
Aug 2025	413	14	8	435
Sep 2025	458	9	0	467

Viewed (geographical distribution)

Total article views: 1,052 (including HTML, PDF, and XML) Thereof 1,052 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 16 Sep 2025

Short summary

We developed a fast and accurate computer tool that predicts how air pollution levels will change around the world under different climate and policy choices. Using machine learning and real model data, our tool can estimate changes in harmful fine particulate pollution in seconds instead of thousands of hours. This makes it easier for researchers and policymakers to explore future air quality and health impacts under a wide range of scenarios.


Total:	0
HTML:	0
PDF:	0
XML:	0