the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 31 Aug 2026
Constraining the Effects of Anthropogenic Emissions and Meteorological Variabilities on summertime PM2.5 and surface ozone changes over eastern China during 2015–2024
Abstract. Severe fine particulate matter (PM2.5) and ozone pollution pose major air quality concerns in China. Chemical transport models (CTMs) are widely used to attribute their long-term trends to anthropogenic emissions and meteorological variabilities, but inherent uncertainties often yield inconsistent results across models. Here, we develop a bias correction framework that integrates a multi-layer perceptron (MLP) neural network to jointly constrain PM2.5 and ozone attributions simulated by two independent CTMs, CMAQ and GEOS-Chem, over eastern China during 2015–2024. This framework improves simulated pollutant concentrations, spatial distributions, and interannual variabilities, providing a robust foundation for attribution analysis. During 2015–2024, anthropogenic emission changes accounted for 69–83 % of the observed PM2.5 decline and 56–63 % of ozone changes across both corrected models, with meteorological variabilities contributing the remainder. Emission changes dominate the PM2.5 decreases in 2015–2019, but the benefits weaken during 2019–2024 despite continued declines in major PM2.5 precursors. For ozone, anthropogenic emissions drive increases during 2015–2019, but become effective for mitigating ozone thereafter. Meteorological variabilities exacerbate ozone increase in 2015–2019 but exert smaller impacts in later years. Crucially, the MLP correction reduces the inter-model attribution discrepancies from 136 % to 14 % for PM2.5 and from 19 % to 8 % for ozone. These findings confirm the dominant role of emission changes in decadal (2015–2024) trends in summertime PM2.5 and ozone in China, and underscore that sustained, coordinated emission reductions remain essential for mitigating both pollutants. Furthermore, they highlight the value of coupling state-of-the-art, mechanism-based CTMs with machine learning to reliably attribute long-term air quality trends.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Atmospheric Chemistry and Physics.
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.- Preprint
(3059 KB) - Metadata XML
-
Supplement
(1651 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-4864', Anonymous Referee #1, 23 Sep 2026
-
RC2: 'Comment on egusphere-2026-4864', Anonymous Referee #2, 23 Sep 2026
Zeng et al. present a comprehensive source attribution of recent trends in major air pollutants in China. They correct biases in current CTMs using a machine learning method, which seems promising in terms of better predicting concentrations, but several points should be clarified to deliver a more robust message regarding source attribution.
- Why does the analysis only focus on summertime? Winter PM2.5 is an issue of larger concern.
- How do the authors account for the spatial representativeness when comparing grid vs observations? For example, there could be 10 observations within one grid cell where ground sites are densely located, but where the sites are sparsely located, only one station data may be used for comparison with the grid-averaged simulated value.
-
Figure 2: What does the shading represent in panel a?
- Table 1: What is the rationale for selecting these input features (particularly the chemical species)?
- Equation 3: How do the authors account for background levels? Ozone has a significant regional background (which may have a temporal trend) over East Asia.
- Line 520: Please describe in more detail the relationship between PM2.5 decline and photolysis of ozone.
- Section 4.3: Please mention that the meteorological variables are from GEOS-FP and WRF in the beginning of the section. Do meteorological observations support the trends shown in the models?
- Section 4.4: The large difference in WRF-Chem and GEOS-Chem before performing MLP correction is surprising, but it is difficult to agree that reduced inter-model difference (from MLP correction) provides more robust attribution results. Can the authors provide a better reason why the bias-corrected model results give more accurate predictions on source attribution?
Citation: https://doi.org/10.5194/egusphere-2026-4864-RC2 -
RC3: 'Comment on egusphere-2026-4864', Anonymous Referee #3, 14 Oct 2026
The authors train a two-output MLP on hourly model biases (sim − obs) of summertime PM2.5 and MDA8 ozone from CMAQ and GEOS-Chem over 139 grid cells (NCP+YRD, 398 sites) for 2015-2019, apply it to fixed-emission (2015) sensitivity runs for 2016-2024, define ΔMet from the corrected fixed-emission differences, and assign ΔEmis = ΔObs − ΔMet as a residual. The data assembly is careful and many internal decompositions close arithmetically, and the "diminishing marginal PM2.5 benefits after 2019" finding is genuinely interesting. However, the framework's core assumption is untested, and nothing carries an uncertainty estimate.(1) Eq. (3) defines ΔEmis ≡ ΔObs − ΔMet; the framework never independently estimates the emission effect. "Both models agree emissions dominate (69-83%)" therefore reduces to "both corrected ΔMet trends are small." Sect. 4.3 half-concedes this; the Abstract/Conclusions do not. State the closure-by-construction logic wherever percentages are quoted.(2) The MLP learns bias as a mapping from base-scenario features during 2015-2019 and is then applied to fixed-emission runs across a different emissions world and later years, that bias structure transfers across scenarios is the load-bearing assumption and it is never validated. Cheap direct test available to the authors: score the corrected fixed-emission simulations against observations for the 2020-2024 years the MLP never saw, per region/pollutant.(3) Run single-target MLPs with identical features vs. the joint two-target network, on skill and downstream attribution. Without it, "coupled responses / shared sensitivities" is decoration. Relatedly: activation function never named, optimizer only "Adam," no framework/seed policy, code "available on request", not acceptable for a method-centric paper.(4) The 2015 is both trend year-zero and the fixed-emission anchor; 2019 is an anomalously ozone-favorable hot-drought year (the paper itself explains its +10.5/12.3 ppb anomaly meteorologically). Endpoint trends anchored on extremes mechanically flatten the 2019-2024 slope. Show period-mean differences (2015-2017 vs 2022-2024) as a robustness check before "diminishing marginal benefits" stays a headline claim. The text also places the policy break "after 2020," yet both sub-periods touch 2019.(5) Free-running WRF (from 1° FNL) vs. assimilated GEOS-FP can alone produce the NCP precipitation disagreement. Add a driver-only diagnostic (both met fields through identical metric definitions, no CTM) so readers can separate reanalysis choice from chemistry-scheme structure. Also please double check the description of Fig. 10 (WRF vs GEOS), which might be wrong.(6) No fold-to-fold spread of CV metrics, no per-region skill table for corrected vs uncorrected fields, no significance tests for the per-cell trend fields in Figs. 7/9, no propagated CIs on the attribution ranges, no sensitivity to the site-selection rule.(7) FNR regime classification needs full methodology before the "52%/33% VOC-limited" percentages carry interpretive weight.Citation: https://doi.org/
10.5194/egusphere-2026-4864-RC3
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 376 | 196 | 67 | 639 | 107 | 78 | 71 |
- HTML: 376
- PDF: 196
- XML: 67
- Total: 639
- Supplement: 107
- BibTeX: 78
- EndNote: 71
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Ganquan Zeng
Xiang Weng
Haofan Wang
Haolin Wang
Guowen He
We develop a novel, dual-CTM bias correction framework to attribute summertime PM2.5 and ozone changes over eastern China during 2015–2024. The framework substantially reduces the CTM biases and reconciles the inter-model discrepancies in the attribution. Emission reductions dominate both the PM2.5 decline and ozone increase, but there is a marked transition of their role after 2019. Persistent unfavorable meteorological conditions contribute to ozone increase especially before 2019.
We develop a novel, dual-CTM bias correction framework to attribute summertime PM2.5 and ozone...
This manuscript develops a dual-CTM, dual-pollutant machine-learning bias-correction framework to attribute summertime PM₂.₅ and MDA8 ozone changes over eastern China during 2015–2024 to anthropogenic emission changes and meteorological variability. The authors apply a multi-output multilayer perceptron (MLP) to correct PM₂.₅ and ozone biases in CMAQ and GEOS-Chem simulations. The trained MLP models are subsequently applied to simulations in which anthropogenic emissions are fixed at the 2015 level. Changes in the corrected fixed-emission simulations are interpreted as meteorological contributions, while the residual between the observed change and the meteorological contribution is attributed to anthropogenic emission changes. The topic is important and timely. The attempt to constrain two structurally different CTMs using a consistent observational and machine-learning framework is potentially useful. I think the paper should be published after addressing the followings.