Deep learning reveals a stronger fossil fuel influence than biomass burning in shaping remote tropospheric ozone

Chaoqun Ma; Hang Su; Yafang Cheng

REVIEW 2 major objections 1 minor 2 references

Fossil fuel emissions contribute over three times more ozone to the remote troposphere than biomass burning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.3

2026-06-27 14:03 UTC pith:XZFWEESI

load-bearing objection DL source attribution for remote ozone may simply reproduce CTM biases rather than resolve the tracer discrepancy. the 2 major comments →

arxiv 2606.09793 v1 pith:XZFWEESI submitted 2026-06-08 physics.ao-ph

Deep learning reveals a stronger fossil fuel influence than biomass burning in shaping remote tropospheric ozone

Chaoqun Ma , Hang Su , Yafang Cheng This is my paper

classification physics.ao-ph

keywords tropospheric ozonefossil fuel emissionsbiomass burningsource attributiondeep learningremote atmosphereatmospheric chemistrychemical transport model

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Tracer studies had indicated biomass burning supplies far more remote tropospheric ozone than fossil fuels, creating a sharp conflict with global chemical transport models. The paper traces the mismatch to the high sensitivity of tracer ratios to differences in tracer lifetimes after long-range transport. A deep learning framework trained to combine observations with model simulations is used to separate the source contributions. The resulting attribution shows fossil fuel sources supply more than three times as much ozone as biomass burning in remote regions. This identification of the dominant source directly affects which emission controls would most effectively lower remote ozone levels.

Core claim

The discrepancy between observation-based tracer analyses and state-of-the-art models arises primarily from the strong sensitivity of tracer methods to differences in tracer lifetimes, especially after extended transport to remote regions. A deep learning framework that synthesizes global observations and chemical transport model simulations accurately infers source contributions and shows that fossil fuel emissions contribute over three times more O3 to the remote troposphere than biomass burning.

What carries the argument

Deep learning framework that synthesizes global observations and chemical transport model simulations to infer ozone source contributions.

Load-bearing premise

The deep learning model trained on chemical transport simulations can separate fossil fuel and biomass burning contributions without inheriting biases from the simulations' own chemistry and transport assumptions.

What would settle it

Independent remote-site measurements or an alternative attribution technique that yields a biomass-burning contribution larger than or comparable to the fossil-fuel contribution would falsify the central claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Phasing out fossil fuels is the most effective single action for lowering remote tropospheric ozone.
Tracer-based estimates systematically overstate biomass burning contributions once lifetime differences are accounted for.
Global chemical transport models are consistent with the deep learning attribution once the tracer bias is removed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same deep learning separation approach could be applied to other long-lived atmospheric species whose source attributions currently rely on lifetime-sensitive tracers.
Policy models that rely on tracer-derived biomass burning fractions may need downward revision of those fractions when estimating ozone impacts.
If the deep learning result holds, the net effect of biomass burning on remote oxidation capacity is smaller than tracer studies alone would imply.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

DL source attribution for remote ozone may simply reproduce CTM biases rather than resolve the tracer discrepancy.

read the letter

The key point here is that the paper claims fossil fuel emissions drive over three times as much remote tropospheric ozone as biomass burning, using a deep learning model to reconcile observations with chemical transport models. This is presented as resolving a long-standing discrepancy where tracers suggest the opposite.

The work does something useful by pointing to tracer lifetime sensitivity as the source of the mismatch and then building a DL synthesizer. That idea has merit if the method can be shown to work independently.

However, the central weakness is that the DL is trained on CTM simulations, which already favor fossil fuel dominance in their ozone chemistry and transport. The abstract mentions the discrepancy arises from lifetime differences, but there's no indication of a test where the network is checked against perturbed lifetimes or different physics to confirm it recovers the right attributions. Without that, the result could just reflect the training distribution rather than new insight from observations.

The abstract supplies no specifics on the network architecture, how it was trained, validation against held-out data, or error bars on the factor of three. This makes the quantitative claim difficult to evaluate.

This paper is aimed at specialists in atmospheric chemistry who care about ozone sources and emission controls. A reader in that group could find the approach interesting as a proof of concept, but the missing details mean it would need major work to be convincing. It is worth sending to peer review so that referees can check the methods section and any supplementary validation.

I would recommend engaging with it in review rather than desk rejecting, because the topic is relevant and the proposed method could be valuable if the circularity issue is addressed.

Referee Report

2 major / 1 minor

Summary. The manuscript claims that the long-standing discrepancy between tracer-based analyses (suggesting biomass burning contributes 2-10 times more remote tropospheric O3 than fossil fuels) and chemical transport models is an artifact of tracer lifetime sensitivity after long-range transport. A deep learning framework trained on CTM simulations and global observations is used to infer source contributions, yielding the result that fossil fuel emissions contribute over three times more O3 to the remote troposphere than biomass burning.

Significance. If the DL attribution proves robust and independent of training-data biases, the result would reconcile an important observational-modeling tension in atmospheric chemistry and strengthen the case for fossil-fuel emission controls as the dominant lever for remote O3 mitigation. The synthesis of observations with CTM output via DL is a potentially generalizable technique, but its validity rests on validation steps not evident from the provided description.

major comments (2)

[Abstract] Abstract: no architecture details, training procedure, validation metrics, or error analysis are supplied, so the central quantitative claim (fossil fuels >3 imes biomass burning) cannot be checked for derivation gaps or post-hoc choices.
[DL framework description] DL framework (implicit in the synthesis step): the network is trained on CTM output whose own ozone chemistry and transport assumptions are already known to produce lower biomass-burning O3; without an explicit test that the network recovers known source ratios when tracer lifetimes are artificially varied or when CTM physics are perturbed, the reported inference reduces to quantities defined by the models.

minor comments (1)

[Results] Provide uncertainty ranges or sensitivity tests on the factor of 'over three times' rather than a point estimate.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight important aspects of transparency and robustness for the deep learning framework. We address each major comment below and will revise the manuscript accordingly to improve clarity and address potential concerns about dependence on CTM assumptions.

read point-by-point responses

Referee: [Abstract] Abstract: no architecture details, training procedure, validation metrics, or error analysis are supplied, so the central quantitative claim (fossil fuels >3 times biomass burning) cannot be checked for derivation gaps or post-hoc choices.

Authors: We agree that the abstract would benefit from additional details to support assessment of the quantitative results. In the revised manuscript, we will expand the abstract to concisely include the DL architecture (e.g., convolutional or recurrent layers used), training procedure (dataset split, loss function, and optimizer), validation metrics (e.g., mean absolute error on held-out simulations and observations), and error analysis (e.g., uncertainty quantification via ensemble methods). These additions will allow readers to better evaluate the central claim without altering the abstract's length substantially. revision: yes
Referee: [DL framework description] DL framework (implicit in the synthesis step): the network is trained on CTM output whose own ozone chemistry and transport assumptions are already known to produce lower biomass-burning O3; without an explicit test that the network recovers known source ratios when tracer lifetimes are artificially varied or when CTM physics are perturbed, the reported inference reduces to quantities defined by the models.

Authors: We appreciate this concern regarding possible inheritance of CTM biases. The DL framework is trained on CTM-derived source labels but uses real global observations as primary inputs to learn mappings that are constrained by data; validation occurs on independent observational subsets. This design aims to leverage observations to mitigate CTM-specific assumptions, consistent with the manuscript's emphasis on reconciling tracer-model discrepancies via lifetime sensitivity. To directly address the referee's request for explicit tests, we will add experiments in a revised methods or supplementary section that artificially vary tracer lifetimes in the training CTM output and confirm the network recovers the corresponding source ratio adjustments, demonstrating robustness. revision: partial

Circularity Check

0 steps flagged

No significant circularity; DL inference combines CTM training with independent observations

full rationale

The paper's derivation chain develops a DL framework trained on CTM simulations to infer source attributions from global observations, then uses that to attribute remote O3 contributions. No quoted equations or steps in the abstract or described text reduce the final attribution (FF > BB by factor >3) to a fitted parameter or self-citation by construction. The lifetime-sensitivity argument for tracer discrepancy is presented as an independent analysis, and the DL result is framed as synthesizing external data rather than renaming CTM outputs. The derivation remains self-contained against the provided benchmarks without load-bearing self-citation or ansatz smuggling.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review limits visibility into specific parameters or entities; the central claim rests on the assumption that chemical transport model simulations provide unbiased training targets for source attribution.

axioms (1)

domain assumption Chemical transport models provide sufficiently accurate representations of ozone production, loss, and transport for use as DL training targets
Invoked when the DL framework is described as synthesizing observations and model simulations.

pith-pipeline@v0.9.1-grok · 5682 in / 1104 out tokens · 27539 ms · 2026-06-27T14:03:28.736348+00:00 · methodology

0 comments

read the original abstract

Tropospheric ozone (O3) is a key greenhouse gas and atmospheric oxidant, yet its sources in the remote troposphere remain strongly debated. Observation-based tracer analyses suggest that O3 attributed to biomass burning is much greater than that from fossil fuel sources (by a factor of ~2-10), contradicting state-of-the-art global models. Here we show that this discrepancy primarily arises from the strong sensitivity of tracer methods to differences in tracer lifetimes, especially after extended transport to the remote regions. To resolve this discrepancy, we develop a deep learning (DL) framework that synthesizes global observations and chemical transport model simulations. The DL approach accurately infers source contributions and reveals that fossil fuel emissions contribute over three times more O3 to the remote troposphere than biomass burning. Our findings underscore that phasing out fossil fuels remains the most powerful lever for mitigating remote tropospheric ozone.

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages · 1 internal anchor

[1]

OT method

anthropogenic emissions of reactive gases and aerosols from the Community Emissions Data System (CEDS). Geosci. Model Dev. 11 , 369 - 408, doi:10.5194/gmd - 11 - 369 - 2018 (2018). 41 Stettler, M. E. J., Boies, A. M., Petzold, A. & Barrett, S. R. H. Global Civil Aviation Black Carbon Emissions. Environ. Sci. Technol. 47 , 10397 - 10404, doi:10.1021/es4013...

work page doi:10.5194/gmd 2018
[2]

Consistent Individualized Feature Attribution for Tree Ensembles

BB10abovePBL is the same as Base in Fig. 1B except BB emissions are scaled by a factor of 10 and released entirely above the planetary boundary layer (PBL) . FF2 is the same as Base except a nthropogenic emissions were doubled . Despite these changes, both scenarios retain the same spatial distribution of BB emissions and wind fields as in Fig. 1A, and th...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/j.atmosenv.2012.05.002 2017

[1] [1]

OT method

anthropogenic emissions of reactive gases and aerosols from the Community Emissions Data System (CEDS). Geosci. Model Dev. 11 , 369 - 408, doi:10.5194/gmd - 11 - 369 - 2018 (2018). 41 Stettler, M. E. J., Boies, A. M., Petzold, A. & Barrett, S. R. H. Global Civil Aviation Black Carbon Emissions. Environ. Sci. Technol. 47 , 10397 - 10404, doi:10.1021/es4013...

work page doi:10.5194/gmd 2018

[2] [2]

Consistent Individualized Feature Attribution for Tree Ensembles

BB10abovePBL is the same as Base in Fig. 1B except BB emissions are scaled by a factor of 10 and released entirely above the planetary boundary layer (PBL) . FF2 is the same as Base except a nthropogenic emissions were doubled . Despite these changes, both scenarios retain the same spatial distribution of BB emissions and wind fields as in Fig. 1A, and th...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/j.atmosenv.2012.05.002 2017