the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
EMEPai 1.0.0: A fast regional data-driven air-quality forecast model for Europe
Abstract. Air-quality forecasting over Europe relies largely on chemical transport models, whose computational cost limits forecast resolution and frequency. This study presents EMEPai, a regional data-driven forecasting system designed to emulate atmospheric composition fields produced by the EMEP/MSC-W modeling and data-assimilation framework. The model uses a graph-based encoder–processor–decoder architecture to generate three-hourly forecasts for NO2, O3, PM10, PM2.5, and SO2 over Europe at 0.1° resolution, with lead times up to 96 hours.
Against surface observations, its performance exceeds that of the corresponding EMEP hindcast while approaching the skill of the assimilated reanalysis. Reduced skill is found for SO2 and during high-concentration events. In an operational forecast setting, EMEPai performs comparably to or better than the CAMS regional ensemble across the evaluated pollutants and forecast lead times. A 96-hour forecast requires only seconds on a single GPU, with the full forecast workflow approximately 230–460 times faster than the corresponding EMEP/MSC-W worfklow.
- Preprint
(5284 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 25 Dec 2026)
- RC1: 'Comment on egusphere-2026-5855', Anonymous Referee #1, 17 Nov 2026 reply
-
RC2: 'Comment on egusphere-2026-5855', Anonymous Referee #2, 19 Nov 2026
reply
The manuscript presents a regional graph-based model for forecasting five surface air pollutants over Europe at 0.1° resolution. The combination of a Europe-wide application, training partly on assimilated concentration fields, and evaluation against operational CAMS forecasts makes this a potentially useful contribution. The reported inference speed is also encouraging. However, the current manuscript does not describe several essential choices sufficiently clear and relies too heavily on aggregate statistics and qualitative descriptions of the results. Substantial revision is needed before accptance. My main concerns are detailed below.
1. Clarify the training data, assimilation procedure, and independence of the evaluation (Sections 2.1 and 2.4; lines 92–100 and 121–130).
The distinction between the validated reanalysis and the hindcast is central to the paper, but neither product is adequately introduced. Please explain how each was generated, including the meteorological inputs, observation sources, assimilated species, assimilation method and frequency, and any relevant differences in model configuration. State whether the authors generated the assimilated fields specifically for this study or used an existing offline reanalysis product. It should also be clear whether assimilation produces temporally discontinuous concentration updates and how the model is trained across these updates.
Explain how the hindcast and reanalysis periods are combined during training, whether they receive different sampling weights, and whether the network is informed of the data source. The stated “22 years of hindcasts (1990–2012)” is inconsistent with the inclusive date range, which spans 23 years; please reconcile the dates and sample counts.
The withholding of 30% of stations is useful, but further information is needed: are the withheld stations fixed across years, selected separately for each pollutant, or changed annually? This would not necessarily invalidate temporal testing, but the distinction between independence from test-year assimilation and generalization to previously unseen stations should be explicit. Please document the station screen method, observation quality control, completeness criteria, and model-to-station interpolation.
2. Make the model and training procedure reproducible in the main text (Sections 2.1–2.2, Table 2, and Appendix E).
The architecture is the core methodological component and should not be described mainly through generic references and in an appendix. Move Figure E1 and the essential architectural description into the main text. Report the actual mesh levels, regional graph construction, node and edge counts, encoder connection radius, number of transformer blocks, latent dimension, attention heads, and total trainable parameters. Explain how the regional domain is represented on the spherical mesh and how information is handled at its edges.
Please provide a mathematical definition of the prediction update and loss, including area weighting, species weighting, and averaging across rollout steps. Clarify whether mean–standard-deviation normalization is performed separately for each variable, over which spatial and temporal samples, and using training data only (Table 2 input scaling). Explain whether the loss is calculated in standardized or physical units and how concentration increments are scaled. These details are particularly relevant to the interpretation of the SO₂ weight.
Table 2 indicates a two-step, six-hour rollout during fine-tuning; state this directly in Section 2.2. Specify the longer rollout lengths tested and support the claim that they did not improve validation performance with a compact sensitivity table. Explain the learning-rate schedule in each stage, including whether warm-up and optimizer state are reset at fine-tuning.
3. Clarify the forcing variables and operational forecast timeline (Table 1, Sections 2.3–2.5, and Appendix D).
Although three-hourly sampling is stated, the preprocessing of individual forcings remains unclear. Specify whether emissions are surface fluxes or vertically integrated quantities, their units, how sector totals are combined, and how hourly emissions are converted to three-hourly inputs. Likewise, distinguish instantaneous meteorological variables from accumulated or interval-averaged fluxes and precipitation. An explicit update equation should show which forcing times are supplied when predicting the next concentration state.
Move Figure D1 and its essential explanation into the main text. The description of a five-day operational run, a 24-hour spin-up, and a 96-hour forecast needs to be reconciled with the forecast-time axes and the first-day statistics. State the exact valid times of the two initialization fields, the meteorological forecast cycle used during each interval, and whether lead time is measured from the model initialization or from the start of the published forecast. Specify the initialization source and verification window for Figures 1–2 and Table 3 as well. This is necessary to determine whether the comparisons use equivalent forecast ages and information availability.
4. Assess the consequences of omitting explicit boundary conditions (Section 2.3; line 120).
The training CTM uses climatological boundary conditions, whereas no boundary conditions are explicitly imposed during EMEPai inference. Please explain what the network is expected to learn about inflow and background concentrations under this setup. Evaluate errors as a function of distance from the domain boundary, forecast lead time, and inflow conditions, and include representative transport episodes. A sensitivity test with an imposed boundary treatment or a changed domain extent would strengthen the assessment. At minimum, provide quantitative diagnostics establishing where the present treatment is adequate and acknowledge limitations for externally driven pollution events.
5. Define the evaluation metrics and support operational comparisons with uncertainty estimates (Tables 3–4, Figures 3 and C1).
The "median station-level temporal correlation" requires a precise definition (line 178). State whether a separate correlation is calculated across forecast dates for every station and lead time before taking the station median, and explain how missing data, low-variance series, and station eligibility are handled. Clarify whether the correlations in Tables 3–4 are pooled correlations or summaries of individual-station correlations. These statistics measure different aspects of skill and should not be compared without explanation.
Correlation alone is insufficient for the full 96-hour evaluation. Add lead-dependent RMSE and bias, together with the distribution of station-level performance differences. Use paired uncertainty estimates that account for temporal dependence and, where applicable, spatial dependence. This is especially necessary for small differences such as the PM2.5 RMSE of 6.19 versus 6.28 in Table 4.
The operational evaluation covers approximately November–June, so its representativeness for a full annual cycle, especially summer ozone and episodic pollution, remains uncertain. Extend the assessment to a full year if available, or clearly restrict the claims to the evaluated period.
6. Strengthen the analysis of spatial errors, high concentrations, and SO2 performance (Section 3.1 and Discussion).
A representative spatial snapshot and domain-averaged time series are insufficient to characterize a model whose intended application includes local pollution episodes. Add a difference-map column to Figure 1, spatial maps of evaluation-period errors, and summaries by region and season. Quantify the reported smoothing and attenuation of peaks using concentration distributions.
SO2 deserves a more diagnostic analysis. Its daily correlation in Table 3 is only 0.20, and the fixed-lead correlations in Figure 3 are lower still. Explain what these different aggregations reveal rather than attributing weak performance mainly to noisy observations near detection limits. Provide the experiments supporting the claim that a loss weight of 0.2 improves the other pollutants without degrading SO2, including equal weighting, the selected weighting, and the stated zero-weight case. Report all species so that the trade-off is visible.
Appendix F already contains useful single-variable and grouped ablations. Bring its main findings into the results and clarify the verification target, nRMSE definition, evaluation year, bootstrap confidence level, and resampling procedure.
7. Align the claims with the reported results and define the scope of the speed comparison (Abstract, Section 3.2, Conclusions, and Appendix B).
The statement that EMEPai matches or outperforms the ensemble for all variables is too broad. In Table 4, PM10 RMSE is higher for EMEPai than for the ensemble (9.97 versus 9.82), and its particulate-matter biases are larger: 19% versus 3% for PM25 and 15% versus −10% for PM10. Identify the metric, pollutant, and lead time for each improvement. Likewise, Table 3 shows appreciable gaps from EMEP-vra for particulate matter; for example, PM25 RMSE increases from 3.66 to 4.19. Quantify these differences instead of relying on phrases such as "approaches the skill."
Appendix B appropriately acknowledges differences in model scope and spin-up. These qualifications should also accompany the main computational claims. The operational system still depends on externally generated meteorology and EMEP analysis initial conditions, so distinguish neural-network execution time, the timed inference workflow, and total operational latency. Summarize the hardware, forecast duration, spin-up in a comparison table.
Specific comments:
1. Figures 2 and C2: Separate daily and monthly series, or use distinct panels, to reduce overlapping curves. State the spatial/station averaging method, forecast lead or day, and treatment of missing observations. Support the textual descriptions of seasonal and episodic agreement with numerical summaries. In Appendix C, the description of Figure C2 refers to EMEP-vra and EMEP-hi, whereas the figure shows operational ENSEMBLE, EMEP, and EMEPai; correct this mismatch.
2. Tables 3–4: Correct the concentration and RMSE units from µg⁻³ to µg m⁻³. Define normalized bias mathematically, report sample counts, and use consistent numerical precision. Clarify the particulate-bias comparison in lines 176–177 and how the quoted differences of 2.5 and 3.7 µg m⁻³ were calculated.
3. Line 174: “June 31 2026” is not a valid date. Supply the actual final evaluation date and use it consistently throughout the text and captions.
4. Loss and physical consistency: Explain the area-weighted MSE in the methods, rather than introducing its interpretation mainly in the conclusions. State whether non-negative concentrations and PM25 ≤ PM10 are enforced or checked.
5. Discuss the limitations of predicting only surface species with the listed precursor emissions and meteorological forcings.
6. Correct typographical and grammatical errors, including "worfklow" (line 10) "have provided effective" (line 38) and the incomplete sentence beginning "A separate version trained..." (line 241).
Citation: https://doi.org/10.5194/egusphere-2026-5855-RC2
Data sets
EMEPai v1.0.0: Archive of source code for training and inference, and checkpoint and input data for inference Halvard Sutterud et al. https://doi.org/10.5281/zenodo.22955196
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 328 | 132 | 20 | 480 | 27 | 24 |
- HTML: 328
- PDF: 132
- XML: 20
- Total: 480
- BibTeX: 27
- EndNote: 24
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
A. General comments
The authors present EMEPai, a data-driven surrogate of the EMEP/MSC-W modelling and data-assimilation system. It produces three-hourly forecasts of NO2, O3, PM10, PM2.5 and SO2 over Europe at 0.1° resolution. As the authors claim, it is the first purely data-driven, multi-pollutant system at the same scale as the CAMS regional ensemble. It follows the global data-driven systems Aurora and AIFS-COMPO, and the recent regional AI weather models (e.g. BRIS, Varda-single). This development was therefore foreseeable, but it is a significant proof of concept given the added complexity of atmospheric composition at a regional scale (importance of emissions, small scale meteorology, non-linear behaviour of chemistry, etc).
The results are promising. Against withheld stations in the 2023 test year, EMEPai outperforms the EMEP hindcast and approaches the assimilated reanalysis. In the operational evaluation (November 2025–June 2026) it is on par with or better than the CAMS regional ensemble median.
The manuscript is well written, clear and fairly comprehensive, and the code and checkpoints are archived. Besides, the authors are open about the remaining weaknesses: SO2, smoothing of fine-scale structure, reduced skill for high concentrations, and a positive PM bias in operational mode.
My main requests are:
B. Specific comments
B1. Evaluation and validation
Pollution episodes. The Conclusions acknowledge that smoothing reduces the representation of high-concentration episodes and exceedances, but this is not quantified. Please add threshold-exceedance scores, and/or short case studies of notable episodes, such as the late-June ozone heat wave.
Operational period. The period covers winter and spring and ends in June, so it includes very little of the ozone season. I suggest extending the evaluation over the whole summer, which is especially important for ozone.
Choice of reference. Please justify the use of EMEP forecasts and reanalyses, rather than the CAMS ensemble median, as the training reference and as the basis of the test-year validation. Are there known EMEP biases that could be inherited by EMEPai? For the 2023 test year, why not also compare with the CAMS validated-reanalysis ensemble median?
Emission mismatch. Please justify training on EMEP emissions while operational runs use CAMS emissions. Section 4 attributes the positive PM bias to this mismatch, but this is not tested. Annual totals of both inventories (see l. 204) would help. Ideally, run EMEPai with EMEP emissions over the operational period to isolate the effect. Table 4 also suggests the explanation is incomplete. For PM2.5 the EMEP forecast is itself biased high (+22%), but for PM10 EMEPai (+15%) is much more biased than EMEP (+3%).
Daily means vs. maxima. Tables 3 and 4 use daily means. Please move to the standard regulatory metrics (daily maximum for NO2 and for O3) or explain why this is not done (e.g. three-hourly output).
Station independence and uncertainty. The 2023 test uses stations withheld from the assimilation, but the operational evaluation uses all EEA-UTD stations. Some of these were assimilated in the 2013–2021 reanalysis targets used for training. Please also report operational scores on the withheld subset, or discuss the fact that the model has in principle learnt site-specific corrections where the validation is computed.
B2. Model design and training
Surface-only prognostic variables. Learning surface concentrations only is a strong choice, opposite to chemistry-transport models, and the results are surprisingly good. It should nevertheless be discussed, as it limits the explainability of the fields. The model has no explicit representation of processes in the lower atmosphere (transport above the surface layer, vertical mixing, deposition, wet scavenging, etc). These are only implicitly captured through surface metorological forcings.
Orography. Please state whether orography is prescribed as a forcing and, if not, why.
Species and emission forcings. Why are CO and NO not considered, and are they part of future improvements?
Natural emissions. Only anthropogenic emissions are supplied (Table 1), and Appendix A states that biogenic emissions are not included. Please add a sentence on desert dust, sea salt (and wildfire emissions) and the likely impact on PM, particularly PM10, and on ozone performance. These sources are represented in the CAMS ensemble.
Boundary conditions and long-range transport. The training fields use climatological boundary conditions (l. 92–93) and none are imposed at inference (l. 119–120). Please discuss the consequences for long-range transport, particularly for ozone and PM (e.g. Saharan dust, intercontinental ozone, smoke plumes). The CAMS members use real-time global boundary conditions. Maps of grid-point temporal correlation with the CAMS reanalysis would help show the magnitude of possible discrepancies near the domain edges.
Time step. The reasons for the three-hourly step are explained (l. 89–91). For pollutants with a strong diurnal cycle, however, hourly output is preferable. Please discuss the feasibility of an hourly step in future versions.
C. Technical corrections