Pith. sign in

REVIEW 3 major objections 5 minor 64 references

A foundation model fine-tuned for air quality can forecast skillfully while lacking the chemical constraints that keep process-based models physically consistent.

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 · deepseek-v4-flash

2026-08-01 09:24 UTC pith:CY7IEZC2

load-bearing objection First serious mechanistic-interpretability pass at a chemistry fine-tuned weather FM, with a reusable evaluation template; the headline chemistry claim is under-supported by one missing control run. the 3 major comments →

arxiv 2607.20778 v1 pith:CY7IEZC2 submitted 2026-07-22 cs.LG physics.ao-ph

Toward Mechanistic Interpretability of an AI Foundation Model Fine-Tuned for Atmospheric Chemistry

Jason Y. Hu , Ivan Higuera-Mendieta , Patrick Obin Sturm , Makoto M. Kelp This is my paper
classification cs.LG physics.ao-ph
keywords mechanistic interpretabilityfoundation modelsatmospheric chemistryair quality forecastingozone-NOx couplingsparse autoencoderschemical transport modelsreanalysis evaluation
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.

This paper asks whether a foundation model fine-tuned for air-quality forecasting has learned real atmospheric chemistry or only statistical patterns in reanalysis data. Studying the Aurora model, the authors find that it reproduces a first-order ozone-suppression response to added reactive nitrogen, yet also predicts negative concentrations, produces NO–NO2–O3 combinations that violate a simple photostationary relationship, and relaxes wildfire plumes toward background. Internally, its latent representations stay organized around the meteorology inherited from pretraining, and sparse-autoencoder features that causally steer the forecast do not map onto individual chemical processes. The authors conclude that benchmark skill on historical reanalysis is not enough to trust composition forecasts; models should also be judged on internal chemical consistency.

Core claim

On the paper's own terms, the discovery is that a skillful end-to-end AI composition forecaster can carry useful short-range chemical sensitivities while lacking the constraints that make a process-based model physically consistent. Aurora Air Pollution responds to a uniform +5 ppb NOx perturbation by suppressing surface ozone in 89% of grid cells at 12 h (median −6.1 ppb), consistent with NO titration; but the same forecasts contain negative ozone, the NOx partitioning and Leighton-derived photolysis diagnostics are noisy and extensively masked by nonpositive concentrations, and fire-borne CO plumes are smoothed toward background within 24–48 h. Internally, the dominant PCA modes of encoder

What carries the argument

The argument rests on three probes applied to Aurora Air Pollution: (1) controlled NOx perturbation rollouts (uniform +0.5, +5, and +20 ppb additions to NO and NO2 in the initial conditions, compared against baseline over five days) to test for learned NOx–ozone coupling; (2) photochemical consistency diagnostics — the NO2 fraction fNO2 = [NO2]/([NO]+[NO2]) and a Leighton-derived photolysis proxy J1 ≈ k3[NO][O3]/[NO2] — computed from predicted fields and used to detect chemically inconsistent species combinations; and (3) sparse autoencoders (AuroraScope) trained on operator residual-stream activations, with zero-ablation steering to determine whether individual learned features causally alt

Load-bearing premise

The NOx perturbation experiments assume that uniformly adding +5 ppb to NO and NO2 is a chemically meaningful probe, and that the resulting ozone decline reflects learned NOx–ozone chemistry rather than the model's generic reaction to any large, out-of-distribution input.

What would settle it

Perturb an inert or weakly coupled species (e.g., CO added uniformly by +5 ppb, matching the NOx experiment). If surface ozone falls by a comparable amount and negative concentrations appear at similar rates, the ozone-suppression result is an artifact of perturbation magnitude, not evidence of learned chemistry.

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

If this is right

  • Short-lead RMSE skill against reanalysis does not establish that an AI forecaster has learned chemistry; a model can look accurate while producing impossible states.
  • Models trained without positivity or cross-species constraints can output negative ozone and other chemically inconsistent combinations even in operationally plausible forecasts.
  • Autoregressive rollout damps localized emission features such as wildfire plumes, so exposure-relevant pollution maxima may be under-forecast even when broad patterns match.
  • Composition forecasts intended for policy should be evaluated with mechanistic and consistency diagnostics in addition to aggregate error metrics.
  • Causally steering single sparse features shifts forecasts in spatially structured but species-entangled ways, so feature-level interventions do not by themselves yield process-level interpretation.

Where Pith is reading between the lines

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

  • Adding an inert-species control perturbation (e.g., CO or SO2 at the same magnitudes) would test whether the ozone-suppression result is specific to NOx or merely a generic response to out-of-distribution inputs; the paper does not include such a control.
  • Because fine-tuning uses low-rank updates to a pretrained weather model, the persistence of meteorological organization may be a general property of fine-tuned Earth-system foundation models; applying this test battery across models and checkpoints would reveal how much chemistry structure fine-tuning can actually add.
  • If the tropical ozone increases under +0.5 ppb NOx are driven by learned geographic correlations rather than VOC-sensitive chemistry, the model may mispredict under changing land use and emission patterns—an implication the paper leaves open.
  • A practical extension: use the same sparse-autoencoder steering pipeline on forecasts initialized from independent reanalyses to test whether discovered features are robust or initialization-specific.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper analyzes Microsoft's Aurora Air Pollution, a fine-tuned Earth system foundation model, using a combination of external chemical diagnostics, perturbation experiments, latent-space visualization, and sparse-autoencoder (SAE) feature steering. The authors report that Aurora reproduces some broad composition patterns and a first-order ozone response to NOx perturbations, but also produces negative concentrations, chemically inconsistent species combinations, patch-scale artifacts, and weak wildfire plume persistence. They conclude that the model has learned useful short-range chemical sensitivities without acquiring the physical constraints of a process-based model, and that its internal representations remain largely meteorological. They release the trained SAEs as AuroraScope.

Significance. If the central claims hold, this is a valuable and timely contribution: it is the first mechanistic interpretability study of an atmospheric-chemistry foundation model, it introduces a reusable SAE suite (AuroraScope), and its diagnostics are external to the model, involving no fitted parameters. The paper also makes a falsifiable prediction — that a skillful reanalysis-trained FM can lack robust chemical constraints — which is important for the deployment of such models in policy-relevant air quality forecasting. The public release of code and SAEs strengthens reproducibility. However, as described below, the main evidence for the positive chemical-sensitivity claim is currently incomplete, and several conclusions are stated more strongly than the presented experiments support.

major comments (3)
  1. [§2.2.2, §3.2, Text S1] The central evidence that Aurora has 'learned useful short-range chemical sensitivities' is the global NOx perturbation experiment. A uniform +5 ppb offset to NO and NO2 across all grid cells, pressure levels, and input timesteps is far outside the training distribution, especially in the stratosphere and clean remote regions. The observed response — median surface ozone −6.1 ppb at 12 h, negative in 89% of cells, including nonphysical negative ozone — could be produced by a generic autoregressive reaction to any large anomalous input, such as regression to the mean, output clipping, or learned spatial covariance between polluted states and low ozone. No control perturbation with a chemically inert or weakly coupled species (e.g., CO or SO2) is presented. The manuscript itself acknowledges in §3.2 that 'it is not possible to further interpret its response to an initial condition perturba
  2. [§3.4] The claim that the latent representations remain 'largely organized around meteorology inherited during pretraining' rests on visual inspection of RGB maps of the first three PCA components. No quantitative comparison is made between the principal components and known meteorological fields (temperature, winds, geopotential, etc.) or chemical fields, and no null model (e.g., PCA on randomized or shuffled latents) is provided. The paper itself repeatedly notes that 'PCA alone cannot determine' the physical meaning of the modes. Because this is one of the paper's central conclusions, it should be supported by a quantitative test, such as correlation of PC loadings with candidate fields, or explicitly labeled as a qualitative hypothesis rather than a demonstrated result.
  3. [§3.2] The headline statistics (median −6.1 ppb, 89% of grid cells, sign reversal at 72 h) are reported without any measure of spread across the year-long set of initializations. Since the model is autoregressive and initialized from different CAMS states, it would be straightforward to report the distribution or bootstrap confidence intervals over initializations. As written, it is unclear whether these numbers are stable and robust or dominated by a few seasons or weather regimes. This is needed to support the quantitative strength of the claimed ozone response.
minor comments (5)
  1. [Eq. (3)] The quantity v_k(t) is described as the 'fraction of token reconstruction variance' attributable to feature k, but the formula uses ||x(t)-b_dec||^2, which is a squared norm, not a centered variance. Consider renaming it 'fraction of squared reconstruction magnitude' or explicitly defining the centering.
  2. [Figure 3/S1] The distribution of surface ozone under the +0.5 ppb perturbation is visually almost indistinguishable from baseline. Plotting differences from baseline, or using separate panels for each perturbation, would make the response easier to read.
  3. [§2.2.2] The perturbation is applied to both input timesteps, so it is not a pure initial-condition perturbation in the usual sense; it also modifies the second input frame. This choice is legitimate but should be stated explicitly as a design decision, since it may affect the interpretation of the 12-h response.
  4. [§3.1] The phrase 'ship track structures' appears without explanation. If these are known artifacts in CAMS or in the model, a brief description or reference would help the reader.
  5. [General] Some notation is used inconsistently: the model is called 'Aurora Air Pollution' in some places and 'Aurora' in others. For clarity, define the checkpoint name at first use and maintain it throughout.

Circularity Check

0 steps flagged

No significant circularity: the paper's central diagnostics use external photochemical relationships and CAMS comparisons, and its self-citations are background only.

full rationale

The paper's central claims about Aurora's learned chemistry are tested against external, independently established relationships: the Leighton photostationary relationship for NO2 photolysis, the known NOx-O3 titration/regime behavior, and comparisons with CAMS forecasts and FINN fire detections. These diagnostics are not fitted to Aurora's outputs, and no parameter is extracted from Aurora and then renamed as a prediction. The NOx perturbation experiments lack a chemically inert control perturbation, which is a real external-validity concern (the response could be a generic out-of-distribution artifact), but this is not a circularity: the paper itself flags the ambiguity in §3.2, stating 'It is not clear if Aurora ... is extending an emissions perturbation of NOx or rolling out a response consistent with atmospheric chemical relationships.' The SAE steering experiments select features by variance contribution and then measure the effect of ablation; the causal effect is measured independently and is not defined in terms of the selection criterion, so this is not a reduction by construction. Self-citations (Kelp et al., Sturm & Wexler) appear only as background motivation or as prior technical context, not as load-bearing justification for the paper's conclusions, and no uniqueness theorem or ansatz is imported from the authors' prior work. The internal-representation analysis is descriptive and explicitly hedged. Overall, the derivation chain is self-contained, with no circular step that equates an input to an output by definition or fits a parameter and then calls it a prediction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical entities are postulated. The analysis relies on standard atmospheric-chemistry assumptions (Leighton ratio, CAMS as reference) and on hand-chosen experimental thresholds (perturbation sizes, PCA components, steering selection). The central claims do not depend on fitted parameters in the traditional sense; the paper is an evaluation of an existing model.

free parameters (4)
  • NOx perturbation offsets = +0.5, +5, +20 ppb
    Hand-chosen perturbation magnitudes. The central 'ozone suppression' result is reported at +5 ppb, and the qualitative response could depend on this range.
  • Steering feature selection count = top 10 by max per-token variance
    Hand-chosen threshold for selecting features to ablate; two case studies are then presented from this set, so the selection is post hoc.
  • SAE expansion factor = 8×
    Chosen SAE hyperparameter; sets the dictionary size (475,136 features) and affects the granularity and interpretability of discovered features.
  • Principal component count = 3
    PCA visualizations use the first three PCs mapped to RGB; the qualitative PCA conclusions depend on this choice.
axioms (4)
  • domain assumption Leighton photostationary relationship J1 ≈ k3[NO][O3]/[NO2]
    Used as a consistency diagnostic for fast NOx-O3 cycling (Section 2.2.1). The paper acknowledges it is idealized and masks nonpositive cells, but still interprets the noisiness of the field as evidence of inconsistency.
  • domain assumption CAMS fields are adequate initial conditions and reference for evaluating Aurora
    The entire evaluation compares Aurora to CAMS at matched initialization times (Sections 2.1, 2.2). Any biases in CAMS would propagate into the assessment of Aurora's chemical behavior.
  • domain assumption Zero-ablation of an SAE feature is a valid causal intervention
    Steering experiments remove a feature's decoder contribution and interpret forecast changes as causal involvement. Nonlinear interactions between features could confound this interpretation (Section 2.4).
  • ad hoc to paper PCA projection and visual inspection can distinguish meteorological from chemical organization in latent spaces
    The conclusion that latent representations are 'largely organized around meteorology' rests on visual assessment of RGB-mapped principal components. The authors note PCA alone cannot assign physical meaning, but the claim is still drawn from such inspection (Section 3.4).

pith-pipeline@v1.3.0-alltime-deepseek · 18681 in / 9651 out tokens · 89341 ms · 2026-08-01T09:24:13.491727+00:00 · methodology

0 comments
read the original abstract

Weather forecasting foundation models (FMs) are increasingly fine-tuned to predict air quality, offering fast global pollution forecasts at lower computational cost than conventional chemical transport models. These FMs are typically trained on reanalysis data and generate forecasts through autoregressive rollout. They do not explicitly represent governing physical or chemical processes. Therefore, high forecast skill does not reveal whether a model has learned physical mechanisms or exploits statistical regularities in its training data. Here, we present the first study of what a FM fine-tuned for atmospheric chemistry has learned by examining Microsoft's Aurora model. We impose controlled chemical perturbations on its forecasts and test them against known photochemical relationships. We then examine the internal representations that generate these forecasts. We find that Aurora captures a first-order ozone response to reactive nitrogen but does not enforce the chemical constraints that a process-based model encodes. It generates chemically inconsistent combinations of related species and relaxes localized emission features such as wildfire plumes toward background. Internally, its representations remain largely organized around the meteorology inherited during pretraining, with little structure specific to chemistry. Using sparse autoencoders, we identify internal components that causally control the chemical forecast but do not map cleanly onto individual atmospheric processes. This work provides a framework for testing whether AI forecasting systems learn atmospheric chemistry from reanalysis data. As these models are increasingly positioned to inform environmental policy decisions, we argue that composition forecasts should also be judged by their internal mechanisms rather than by benchmark skill alone.

Figures

Figures reproduced from arXiv: 2607.20778 by Ivan Higuera-Mendieta, Jason Y. Hu, Makoto M. Kelp, Patrick Obin Sturm.

Figure 1
Figure 1. Figure 1: Mechanistic interpretability of the Aurora Earth system foundation model. [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Calculated NOx ratio and NO2 photolysis rate in CAMS and Aurora. Both simulations are on September 1 2022 at 12:00 UTC. The NOx ratio is calculated as [NO2] [NO]+[NO2] . The NO2 photolysis rate is calculated as k3[NO][O3] [NO2] , which assumes photostationarity based on the Leighton relationship. Depicted is the first rollout step (+12h) for both models at surface level (1000 hPa). White areas on the map i… view at source ↗
Figure 3
Figure 3. Figure 3: Surface ozone response to global NOx perturbations in Aurora. Panel A shows the difference in surface ozone concentrations compared to a baseline where global NOx is uniformly increased by 0.5 and 5 ppb for every grid cell at every atmospheric pressure level over both initialization timesteps. Panel B shows the probability distribution of surface ozone concentrations under the different baseline and elevat… view at source ↗
Figure 4
Figure 4. Figure 4: Wildfire emissions and transport over Spain in CAMS and Aurora. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: PCA uncovers coherent structures in Aurora’s latent levels. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example features learned by Aurorascope SAEs. A range of features are rep￾resented, including regional transport (upper left), whole-globe activation (upper right), diurnal cycles (lower left), and a spurious feature for a latitude band (lower right). All features are ex￾tracted from the SAEs on the first step of a rollout initialized at September 01, 2022 +00:00 UTC. Resolution differences are reflective … view at source ↗
Figure 7
Figure 7. Figure 7: Steering Aurora forecasts by zero-ablating [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

64 extracted references · 29 canonical work pages · 5 internal anchors

  1. [1]

    , Henze, D K

    anenbergEstimatesGlobalBurden2018a APACrefauthors Anenberg, S C. , Henze, D K. , Tinney, V. , Kinney, P L. , Raich, W. , Fann, N. Kuylenstierna, J C. APACrefauthors \ 2018 . Estimates of the Global Burden of Ambient PM2 .5, Ozone , and NO2 on Asthma Incidence and Emergency Room Visits Estimates of the Global Burden of Ambient PM2 .5, Ozone , and NO2 on As...

    work page doi:10.1289/ehp3766 2018

  2. [2]

    \ Gavves, E

    bereskaMechanisticInterpretabilityAI2024 APACrefauthors Bereska, L. \ Gavves, E. APACrefauthors \ 2024 . Mechanistic Interpretability for AI Safety -- A Review Mechanistic Interpretability for AI Safety -- A Review . arXiv . APACrefDOI doi:10.48550/arXiv.2404.14082 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2404.14082 2024

  3. [3]

    , Xie, L

    biAccurateMediumrangeGlobal2023 APACrefauthors Bi, K. , Xie, L. , Zhang, H. , Chen, X. , Gu, X. \ Tian, Q. APACrefauthors \ 2023 . Accurate Medium-Range Global Weather Forecasting with 3D Neural Networks Accurate medium-range global weather forecasting with 3D neural networks . Nature 619 7970 533--538 . APACrefDOI doi:10.1038/s41586-023-06185-3 APACrefDOI

    doi:10.1038/s41586-023-06185-3 2023

  4. [4]

    , Bruinsma, W P

    bodnarFoundationModelEarth2025 APACrefauthors Bodnar, C. , Bruinsma, W P. , Lucic, A. , Stanley, M. , Allen, A. , Brandstetter, J. Perdikaris, P. APACrefauthors \ 2025 . A Foundation Model for the Earth System A foundation model for the Earth system . Nature 641 8065 1180--1187 . APACrefDOI doi:10.1038/s41586-025-09005-y APACrefDOI

    doi:10.1038/s41586-025-09005-y 2025

  5. [5]

    o m, A. , Bareeva, D. \ H \

    bommerFindingRightXAI2024 APACrefauthors Bommer, P L. , Kretschmer, M. , Hedstr \"o m, A. , Bareeva, D. \ H \"o hne, M M C. APACrefauthors \ 2024 . Finding the Right XAI Method --- A Guide for the Evaluation and Ranking of Explainable AI Methods in Climate Science Finding the Right XAI Method --- A Guide for the Evaluation and Ranking of Explainable AI Me...

    doi:10.1175/aies-d-23-0074.1 2024

  6. [6]

    \ Jacob, D J

    brasseurModelingAtmosphericChemistry2017 APACrefauthors Brasseur, G P. \ Jacob, D J. APACrefauthors \ 2017 . Modeling of Atmospheric Chemistry Modeling of Atmospheric Chemistry . Cambridge Cambridge University Press . APACrefDOI doi:10.1017/9781316544754 APACrefDOI

    work page doi:10.1017/9781316544754 2017

  7. [7]

    , Beucler, T

    brenowitzInterpretingStabilizingMachineLearning2020 APACrefauthors Brenowitz, N D. , Beucler, T. , Pritchard, M. \ Bretherton, C S. APACrefauthors \ 2020 . Interpreting and Stabilizing Machine-Learning Parametrizations of Convection Interpreting and Stabilizing Machine-Learning Parametrizations of Convection . Journal of the Atmospheric Sciences 77 12 435...

    doi:10.1175/jas-d-20-0082.1 2020

  8. [8]

    \ Stockwell, W R

    calvertAcidGenerationTroposphere1983 APACrefauthors Calvert, J G. \ Stockwell, W R. APACrefauthors \ 1983 . Acid Generation in the Troposphere by Gas-Phase Chemistry Acid generation in the troposphere by gas-phase chemistry . Environmental Science & Technology 17 9 428A-443A . APACrefDOI doi:10.1021/es00115a727 APACrefDOI

    work page doi:10.1021/es00115a727 1983

  9. [9]

    , Fern \'a ndez-Torres , M \'A

    camps-vallsArtificialIntelligenceModeling2025 APACrefauthors Camps-Valls , G. , Fern \'a ndez-Torres , M \'A . , Cohrs, K H. , H \"o hl, A. , Castelletti, A. , Pacal, A. Williams, T. APACrefauthors \ 2025 . Artificial Intelligence for Modeling and Understanding Extreme Weather and Climate Events Artificial intelligence for modeling and understanding extre...

    doi:10.1038/s41467-025-56573-8 2025

  10. [10]

    , Siu, M

    chalnevImprovingSteeringVectors2024 APACrefauthors Chalnev, S. , Siu, M. \ Conmy, A. APACrefauthors \ 2024 . Improving Steering Vectors by Targeting Sparse Autoencoder Features Improving Steering Vectors by Targeting Sparse Autoencoder Features . arXiv . APACrefDOI doi:10.48550/arXiv.2411.02193 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.02193 2024

  11. [11]

    , Liu, G

    chouCausalLanguageControl2025 APACrefauthors Chou, C T. , Liu, G. , Sun, J. , Blondin, C. , Zhu, K. , Sharma, V. \ O'Brien, S. APACrefauthors \ 2025 . Causal Language Control in Multilingual Transformers via Sparse Feature Steering Causal Language Control in Multilingual Transformers via Sparse Feature Steering . arXiv . APACrefDOI doi:10.48550/arXiv.2507...

    doi:10.48550/arxiv.2507.13410 2025

  12. [12]

    No One Knows the State of the Art in Geospatial Foundation Models

    corleyNoOneKnows2026 APACrefauthors Corley, I. , Lehmann, N. , Robinson, C. , Tseng, G. , Fuller, A. , Alemohammad, H. Kerner, H. APACrefauthors \ 2026 . No One Knows the State of the Art in Geospatial Foundation Models No One Knows the State of the Art in Geospatial Foundation Models . arXiv . APACrefDOI doi:10.48550/arXiv.2605.12678 APACrefDOI

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.12678 2026

  13. [13]

    , Ewart, A

    cunninghamSparseAutoencodersFind2023 APACrefauthors Cunningham, H. , Ewart, A. , Riggs, L. , Huben, R. \ Sharkey, L. APACrefauthors \ 2023 . Sparse Autoencoders Find Highly Interpretable Features in Language Models Sparse Autoencoders Find Highly Interpretable Features in Language Models . arXiv . APACrefDOI doi:10.48550/arXiv.2309.08600 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.08600 2023

  14. [14]

    , Amini, H

    diEnsemblebasedModelPM252019 APACrefauthors Di, Q. , Amini, H. , Shi, L. , Kloog, I. , Silvern, R. , Kelly, J. Schwartz, J. APACrefauthors \ 2019 . An Ensemble-Based Model of PM2 .5 Concentration across the Contiguous United States with High Spatiotemporal Resolution An ensemble-based model of PM2 .5 concentration across the contiguous United States with ...

    arXiv 2019

  15. [15]

    , Long, M S

    easthamGEOSChemHighPerformance2018 APACrefauthors Eastham, S D. , Long, M S. , Keller, C A. , Lundgren, E. , Yantosca, R M. , Zhuang, J. Jacob, D J. APACrefauthors \ 2018 . GEOS-Chem High Performance ( GCHP V11-02c): A next-Generation Implementation of the GEOS-Chem Chemical Transport Model for Massively Parallel Applications GEOS-Chem High Performance ( ...

    2018

  16. [16]

    Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs

    fengBenchmarkingImprovingMonitors2026 APACrefauthors Feng, D. , Srivastava, P. , Dragan, A. \ Laidlaw, C. APACrefauthors \ 2026 . Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs . arXiv . APACrefDOI doi:10.48550/arXiv.2605.21602 APACrefDOI

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.21602 2026

  17. [17]

    , Estrada, J C

    gomez-pedreroIncrementalPCAAlgorithm2022 APACrefauthors G \'o mez-Pedrero , J A. , Estrada, J C. , Alonso, J. , Quiroga, J A. \ Vargas, J. APACrefauthors \ 2022 . Incremental PCA Algorithm for Fringe Pattern Demodulation Incremental PCA algorithm for fringe pattern demodulation . Optics Express 30 8 12278--12293 . APACrefDOI doi:10.1364/OE.452463 APACrefDOI

    work page doi:10.1364/oe.452463 2022

  18. [18]

    , Hossaini, R

    gouldsbroughMachineLearningApproach2024a APACrefauthors Gouldsbrough, L. , Hossaini, R. , Eastoe, E. , Young, P J. \ Vieno, M. APACrefauthors \ 2024 . A Machine Learning Approach to Downscale EMEP4UK : Analysis of UK Ozone Variability and Trends A machine learning approach to downscale EMEP4UK : Analysis of UK ozone variability and trends . Atmospheric Ch...

    work page doi:10.5194/acp-24-3163-2024 2024

  19. [19]

    , Kelp, M M

    hickmanApplicationsMachineLearning2025a APACrefauthors Hickman, S H M. , Kelp, M M. , Griffiths, P T. , Doerksen, K. , Miyazaki, K. , Pennington, E A. Liu, Z. APACrefauthors \ 2025 . Applications of Machine Learning and Artificial Intelligence in Tropospheric Ozone Research Applications of Machine Learning and Artificial Intelligence in Tropospheric Ozone...

    work page doi:10.5194/gmd-18-8777-2025 2025

  20. [20]

    , Shen, Y

    huLoRALowRankAdaptation2021 APACrefauthors Hu, E J. , Shen, Y. , Wallis, P. , Allen-Zhu , Z. , Li, Y. , Wang, S. Chen, W. APACrefauthors \ 2021 . LoRA : Low-Rank Adaptation of Large Language Models LoRA : Low-Rank Adaptation of Large Language Models . arXiv . APACrefDOI doi:10.48550/arXiv.2106.09685 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.09685 2021

  21. [21]

    \ Evans, M J

    ivattImprovingPredictionAtmospheric2020b APACrefauthors Ivatt, P D. \ Evans, M J. APACrefauthors \ 2020 . Improving the Prediction of an Atmospheric Chemistry Transport Model Using Gradient-Boosted Regression Trees Improving the prediction of an atmospheric chemistry transport model using gradient-boosted regression trees . Atmospheric Chemistry and Physi...

    work page doi:10.5194/acp-20-8063-2020 2020

  22. [22]

    PhysMetrics.Weather: An Evaluation Framework for Physical Consistency in ML Weather Models

    kasteleynPhysMetricsWeatherEvaluationFramework2026 APACrefauthors Kasteleyn, E. , Maier, T. , Lauer, A. , Eyring, V. , Gentine, P. \ Lucic, A. APACrefauthors \ 2026 . PhysMetrics . Weather : An Evaluation Framework for Physical Consistency in ML Weather Models PhysMetrics . Weather : An Evaluation Framework for Physical Consistency in ML Weather Models . ...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.10642 2026

  23. [23]

    \ Evans, M J

    kellerApplicationRandomForest2019 APACrefauthors Keller, C A. \ Evans, M J. APACrefauthors \ 2019 . Application of Random Forest Regression to the Calculation of Gas-Phase Chemistry within the GEOS-Chem Chemistry Model V10 Application of random forest regression to the calculation of gas-phase chemistry within the GEOS-Chem chemistry model v10 . Geoscient...

    work page doi:10.5194/gmd-12-1209-2019 2019

  24. [24]

    , Knowland, K E

    kellerDescriptionNASAGEOS2021 APACrefauthors Keller, C A. , Knowland, K E. , Duncan, B N. , Liu, J. , Anderson, D C. , Das, S. Pawson, S. APACrefauthors \ 2021 . Description of the NASA GEOS Composition Forecast Modeling System GEOS-CF v1.0 Description of the NASA GEOS Composition Forecast Modeling System GEOS-CF v1.0 . Journal of Advances in Modeling Ear...

    work page doi:10.1029/2020ms002413 2021

  25. [25]

    , Hickman, S

    kelpAI4O3FoundationalData2025a APACrefauthors Kelp, M. , Hickman, S. , Miyazaki, K. , Chang, K L. , Griffiths, P. , Zhu, Q. Schultz, M G. APACrefauthors \ 2025 . AI4O3 : A Foundational Data Collection for Artificial Intelligence in Tropospheric Ozone Research AI4O3 : A Foundational Data Collection for Artificial Intelligence in Tropospheric Ozone Research...

    2025

  26. [26]

    , Jacob, D J

    kelpStableGeneralMachineLearned2020 APACrefauthors Kelp, M. , Jacob, D J. , Kutz, J N. , Marshall, J D. \ Tessum, C W. APACrefauthors \ 2020 . Toward Stable , General Machine-Learned Models of the Atmospheric Chemical System Toward Stable , General Machine-Learned Models of the Atmospheric Chemical System . Journal of Geophysical Research: Atmospheres 125...

    work page doi:10.1029/2020jd032759 2020

  27. [27]

    , Jacob, D J

    kelpOnlineLearnedNeuralNetwork2022 APACrefauthors Kelp, M. , Jacob, D J. , Lin, H. \ Sulprizio, M P. APACrefauthors \ 2022 . An Online-Learned Neural Network Chemical Solver for Stable Long-Term Global Simulations of Atmospheric Chemistry An Online-Learned Neural Network Chemical Solver for Stable Long-Term Global Simulations of Atmospheric Chemistry . Jo...

    work page doi:10.1029/2021ms002926 2022

  28. [28]

    Orders-of-magnitude speedup in atmospheric chemistry modeling through neural network-based emulation

    kelpOrdersofmagnitudeSpeedupAtmospheric2018 APACrefauthors Kelp, M. , Tessum, C W. \ Marshall, J D. APACrefauthors \ 2018 . Orders-of-Magnitude Speedup in Atmospheric Chemistry Modeling through Neural Network-Based Emulation Orders-of-magnitude speedup in atmospheric chemistry modeling through neural network-based emulation . arXiv . APACrefDOI doi:10.485...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1808.03874 2018

  29. [29]

    kingmaAdamMethodStochastic2017 APACrefauthors Kingma, D P. \ Ba, J. APACrefauthors \ 2017 . Adam: A Method for Stochastic Optimization Adam: A Method for Stochastic Optimization . arXiv . APACrefDOI doi:10.48550/arXiv.1412.6980 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2017

  30. [30]

    , Raghunathan, A

    kumarFineTuningCanDistort2022a APACrefauthors Kumar, A. , Raghunathan, A. , Jones, R. , Ma, T. \ Liang, P. APACrefauthors \ 2022 . Fine- Tuning Can Distort Pretrained Features and Underperform Out-of-Distribution Fine- Tuning can Distort Pretrained Features and Underperform Out-of-Distribution . arXiv . APACrefDOI doi:10.48550/arXiv.2202.10054 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2202.10054 2022

  31. [31]

    Machine learning for climate physics and simulations

    laiMachineLearningClimate2024 APACrefauthors Lai, C Y. , Hassanzadeh, P. , Sheshadri, A. , Sonnewald, M. , Ferrari, R. \ Balaji, V. APACrefauthors \ 2024 . Machine Learning for Climate Physics and Simulations Machine learning for climate physics and simulations . arXiv . APACrefDOI doi:10.48550/arXiv.2404.13227 APACrefDOI

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2404.13227 2024

  32. [32]

    , Sanchez-Gonzalez , A

    lamLearningSkillfulMediumrange2023 APACrefauthors Lam, R. , Sanchez-Gonzalez , A. , Willson, M. , Wirnsberger, P. , Fortunato, M. , Alet, F. Battaglia, P. APACrefauthors \ 2023 . Learning Skillful Medium-Range Global Weather Forecasting Learning skillful medium-range global weather forecasting . Science 382 6677 1416--1421 . APACrefDOI doi:10.1126/science...

    doi:10.1126/science.adi2336 2023

  33. [33]

    , Alexe, M

    langAIFSECMWFsDatadriven2024a APACrefauthors Lang, S. , Alexe, M. , Chantry, M. , Dramsch, J. , Pinault, F. , Raoult, B. Rabier, F. APACrefauthors \ 2024 . AIFS -- ECMWF 's Data-Driven Forecasting System AIFS -- ECMWF 's data-driven forecasting system . arXiv . APACrefDOI doi:10.48550/arXiv.2406.01465 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.01465 2024

  34. [34]

    , Chen, A S

    leeSurgicalFineTuningImproves2023 APACrefauthors Lee, Y. , Chen, A S. , Tajwar, F. , Kumar, A. , Yao, H. , Liang, P. \ Finn, C. APACrefauthors \ 2023 . Surgical Fine-Tuning Improves Adaptation to Distribution Shifts Surgical Fine-Tuning Improves Adaptation to Distribution Shifts . arXiv . APACrefDOI doi:10.48550/arXiv.2210.11466 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2210.11466 2023

  35. [35]

    APACrefauthors \ 1961

    leightonPhotochemistryAirPollution2012 APACrefauthors Leighton, P. APACrefauthors \ 1961 . Photochemistry of Air Pollution Photochemistry of Air Pollution . Elsevier

    1961

  36. [36]

    , Chaudhari, P

    liRethinkingHyperparametersFinetuning2020 APACrefauthors Li, H. , Chaudhari, P. , Yang, H. , Lam, M. , Ravichandran, A. , Bhotika, R. \ Soatto, S. APACrefauthors \ 2020 . Rethinking the Hyperparameters for Fine-tuning Rethinking the Hyperparameters for Fine-tuning . arXiv . APACrefDOI doi:10.48550/arXiv.2002.11770 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2002.11770 2020

  37. [37]

    , Jin, X

    liGrowingImpactsFire2026 APACrefauthors Li, Y. , Jin, X. , Kelp, M. , Sun, H Z. \ Qiu, M. APACrefauthors \ 2026 . Growing Impacts of Fire Smoke on Ozone Pollution and Associated Mortality Burden in the United States Growing impacts of fire smoke on ozone pollution and associated mortality burden in the United States . Science Advances 12 18 eaec2903 . APA...

    work page doi:10.1126/sciadv.aec2903 2026

  38. [38]

    , Doherty, R M

    liuCorrectingOzoneBiases2022a APACrefauthors Liu, Z. , Doherty, R M. , Wild, O. , O'Connor, F M. \ Turnock, S T. APACrefauthors \ 2022 . Correcting Ozone Biases in a Global Chemistry--Climate Model: Implications for Future Ozone Correcting ozone biases in a global chemistry--climate model: Implications for future ozone . Atmospheric Chemistry and Physics ...

    work page doi:10.5194/acp-22-12543-2022 2022

  39. [39]

    \ Lee, S I

    lundbergUnifiedApproachInterpreting2017 APACrefauthors Lundberg, S. \ Lee, S I. APACrefauthors \ 2017 . A Unified Approach to Interpreting Model Predictions A Unified Approach to Interpreting Model Predictions . arXiv . APACrefDOI doi:10.48550/arXiv.1705.07874 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.1705.07874 2017

  40. [40]

    \ Ouellette, N T

    macmillanMechanisticUnderstandingDatadriven2025a APACrefauthors MacMillan, T. \ Ouellette, N T. APACrefauthors \ 2025 . Towards Mechanistic Understanding in a Data-Driven Weather Model: Internal Activations Reveal Interpretable Physical Features Towards mechanistic understanding in a data-driven weather model: Internal activations reveal interpretable phy...

    doi:10.48550/arxiv.2512.24440 2025

  41. [41]

    , Sharps, K

    millsOzonePollutionWill2018 APACrefauthors Mills, G. , Sharps, K. , Simpson, D. , Pleijel, H. , Broberg, M. , Uddling, J. Van Dingenen, R. APACrefauthors \ 2018 . Ozone Pollution Will Compromise Efforts to Increase Global Wheat Production Ozone pollution will compromise efforts to increase global wheat production . Global Change Biology 24 8 3560--3574 . ...

    work page doi:10.1111/gcb.14157 2018

  42. [42]

    , Chan, L

    nandaProgressMeasuresGrokking2023 APACrefauthors Nanda, N. , Chan, L. , Lieberum, T. , Smith, J. \ Steinhardt, J. APACrefauthors \ 2023 . Progress Measures for Grokking via Mechanistic Interpretability Progress measures for grokking via mechanistic interpretability . arXiv . APACrefDOI doi:10.48550/arXiv.2301.05217 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.05217 2023

  43. [43]

    APACrefauthors \ 2012

    nationalresearchcouncilNationalStrategyAdvancing2012 APACrefauthors National Research Council . APACrefauthors \ 2012 . A National Strategy for Advancing Climate Modeling A National Strategy for Advancing Climate Modeling \ . The National Academies Press . APACrefDOI doi:10.17226/13430 APACrefDOI

    work page doi:10.17226/13430 2012

  44. [44]

    , Brandstetter, J

    nguyenClimaXFoundationModel2023 APACrefauthors Nguyen, T. , Brandstetter, J. , Kapoor, A. , Gupta, J K. \ Grover, A. APACrefauthors \ 2023 . ClimaX : A Foundation Model for Weather and Climate ClimaX : A foundation model for weather and climate . arXiv . APACrefDOI doi:10.48550/arXiv.2301.10343 APACrefDOI

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.10343 2023

  45. [45]

    , Jacob, D J

    oakBiascorrectedGEMSGeostationary2024 APACrefauthors Oak, Y J. , Jacob, D J. , Balasus, N. , Yang, L H. , Chong, H. , Park, J. Kim, J. APACrefauthors \ 2024 . A Bias-Corrected GEMS Geostationary Satellite Product for Nitrogen Dioxide Using Machine Learning to Enforce Consistency with the TROPOMI Satellite Instrument A bias-corrected GEMS geostationary sat...

    2024

  46. [46]

    , Cammarata, N

    olahZoomIntroductionCircuits2020 APACrefauthors Olah, C. , Cammarata, N. , Schubert, L. , Goh, G. , Petrov, M. \ Carter, S. APACrefauthors \ 2020 . Zoom In : An Introduction to Circuits Zoom In : An Introduction to Circuits . Distill 5 3 . APACrefDOI doi:10.23915/distill.00024.001 APACrefDOI

    doi:10.23915/distill.00024.001 2020

  47. [47]

    qiuWildfireSmokeExposure2025b APACrefauthors Qiu, M. , Li, J. , Gould, C F. , Jing, R. , Kelp, M. , Childs, M L. Burke, M. APACrefauthors \ 2025 . Wildfire Smoke Exposure and Mortality Burden in the USA under Climate Change Wildfire smoke exposure and mortality burden in the USA under climate change . Nature 647 8091 935--943 . APACrefDOI doi:10.1038/s415...

    work page doi:10.1038/s41586-025-09611-w 2025

  48. [48]

    APACrefauthors \ 2020

    raspCoupledOnlineLearning2020a APACrefauthors Rasp, S. APACrefauthors \ 2020 . Coupled Online Learning as a Way to Tackle Instabilities and Biases in Neural Network Parameterizations: General Algorithms and Lorenz 96 Case Study (v1.0) Coupled online learning as a way to tackle instabilities and biases in neural network parameterizations: General algorithm...

    doi:10.5194/gmd-13-2185-2020 2020

  49. [49]

    , Wang, J

    sadeghitabasGFSPoweredMachineLearning2025 APACrefauthors Sadeghi Tabas, S. , Wang, J. , Lei, W. , Row, M. , Zhang, Z. , Zhu, L. Carley, J R. APACrefauthors \ 2025 . GFS-Powered Machine Learning Weather Prediction : A Comparative Study on Training GraphCast with NOAA 's GDAS Data for Global Weather Forecasts GFS-Powered Machine Learning Weather Prediction ...

    work page doi:10.25923/xd3y-wy31 2025

  50. [50]

    \ Pandis, S N

    seinfeldAtmosphericChemistryPhysics2016 APACrefauthors Seinfeld, J H. \ Pandis, S N. APACrefauthors \ 2006 . Atmospheric Chemistry and Physics : From Air Pollution to Climate Change Atmospheric Chemistry and Physics : From Air Pollution to Climate Change . John Wiley & Sons

    2006

  51. [51]

    , Jacob, D J

    shenAdaptiveMethodSpeeding2020 APACrefauthors Shen, L. , Jacob, D J. , Santillana, M. , Wang, X. \ Chen, W. APACrefauthors \ 2020 . An Adaptive Method for Speeding up the Numerical Integration of Chemical Mechanisms in Atmospheric Chemistry Models: Application to GEOS-Chem Version 12.0.0 An adaptive method for speeding up the numerical integration of chem...

    2020

  52. [52]

    shenEnhancingGlobalEstimation2024 APACrefauthors Shen, S. , Li, C. , van Donkelaar , A. , Jacobs, N. , Wang, C. \ Martin, R V. APACrefauthors \ 2024 . Enhancing Global Estimation of Fine Particulate Matter Concentrations by Including Geophysical a Priori Information in Deep Learning Enhancing Global Estimation of Fine Particulate Matter Concentrations by ...

    work page doi:10.1021/acsestair.3c00054 2024

  53. [53]

    \ Keller, C A

    silvaLimitationsXAIMethods2024 APACrefauthors Silva, S J. \ Keller, C A. APACrefauthors \ 2024 . Limitations of XAI Methods for Process-Level Understanding in the Atmospheric Sciences Limitations of XAI Methods for Process-Level Understanding in the Atmospheric Sciences . Artificial Intelligence for the Earth Systems 3 1 e230045 . APACrefDOI doi:10.1175/A...

    work page doi:10.1175/aies-d-23-0045.1 2024

  54. [54]

    , Manders, A

    sturmAdvectingSuperspeciesEfficiently2023 APACrefauthors Sturm, P O. , Manders, A. , Janssen, R. , Segers, A. , Wexler, A S. \ Lin, H X. APACrefauthors \ 2023 . Advecting Superspecies : Efficiently Modeling Transport of Organic Aerosol With a Mass-Conserving Dimensionality Reduction Method Advecting Superspecies : Efficiently Modeling Transport of Organic...

    2023

  55. [55]

    \ Wexler, A S

    sturmMassEnergyconservingFramework2020a APACrefauthors Sturm, P O. \ Wexler, A S. APACrefauthors \ 2020 . A Mass- and Energy-Conserving Framework for Using Machine Learning to Speed Computations: A Photochemistry Example A mass- and energy-conserving framework for using machine learning to speed computations: A photochemistry example . Geoscientific Model...

    work page doi:10.5194/gmd-13-4435-2020 2020

  56. [56]

    \ Wexler, A S

    sturmConservationLawsNeural2022a APACrefauthors Sturm, P O. \ Wexler, A S. APACrefauthors \ 2022 . Conservation Laws in a Neural Network Architecture: Enforcing the Atom Balance of a Julia-based Photochemical Model (v0.2.0) Conservation laws in a neural network architecture: Enforcing the atom balance of a Julia-based photochemical model (v0.2.0) . Geosci...

    work page doi:10.5194/gmd-15-3417-2022 2022

  57. [57]

    , Barnes, E A

    tomsPhysicallyInterpretableNeural2020 APACrefauthors Toms, B A. , Barnes, E A. \ Ebert-Uphoff , I. APACrefauthors \ 2020 . Physically Interpretable Neural Networks for the Geosciences : Applications to Earth System Variability Physically Interpretable Neural Networks for the Geosciences : Applications to Earth System Variability . Journal of Advances in M...

    doi:10.1029/2019ms002002 2020

  58. [58]

    , Archer-Nicholls , S

    weberChemistrydrivenChangesStrongly2022 APACrefauthors Weber, J. , Archer-Nicholls , S. , Abraham, N L. , Shin, Y M. , Griffiths, P. , Grosvenor, D P. Archibald, A T. APACrefauthors \ 2022 . Chemistry-Driven Changes Strongly Influence Climate Forcing from Vegetation Emissions Chemistry-driven changes strongly influence climate forcing from vegetation emis...

    work page doi:10.1038/s41467-022-34944-9 2022

  59. [59]

    , Kimura, Y

    gmd-16-3873-2023 APACrefauthors Wiedinmyer, C. , Kimura, Y. , McDonald-Buller, E C. , Emmons, L K. , Buchholz, R R. , Tang, W. Yokelson, R. APACrefauthors \ 2023 . The Fire Inventory from NCAR version 2.5: an updated global fire emissions model for climate and chemistry applications The fire inventory from ncar version 2.5: an updated global fire emission...

    2023

  60. [60]

    , Place, B K

    wiserAMOREIsopreneV10New2023 APACrefauthors Wiser, F. , Place, B K. , Sen, S. , Pye, H O T. , Yang, B. , Westervelt, D M. McNeill, V F. APACrefauthors \ 2023 . AMORE-Isoprene v1.0: A New Reduced Mechanism for Gas-Phase Isoprene Oxidation AMORE-Isoprene v1.0: A new reduced mechanism for gas-phase isoprene oxidation . Geoscientific Model Development 16 6 18...

    work page doi:10.5194/gmd-16-1801-2023 2023

  61. [61]

    , Zhao, C

    xiaAdvancingSophisticatedPhotochemistry2025 APACrefauthors Xia, Z. , Zhao, C. , Du, Q. , Yang, Z. , Zhang, M. \ Qiao, L. APACrefauthors \ 2025 . Advancing Sophisticated Photochemistry Simulation in Atmospheric Numerical Models With Artificial Intelligence PhotoChemistry ( AIPC ) Scheme Using the Feature-Mapping Subspace Self-Attention Algorithm Advancing ...

    2025

  62. [62]

    , Wang, Y

    xiaoGeneratingLongterm200320202022 APACrefauthors Xiao, Y. , Wang, Y. , Yuan, Q. , He, J. \ Zhang, L. APACrefauthors \ 2022 . Generating a Long-Term (2003-2020) Hourly 0.25 ^ Global PM2 .5 Dataset via Spatiotemporal Downscaling of CAMS with Deep Learning ( DeepCAMS ) Generating a long-term (2003-2020) hourly 0.25 ^ global PM2 .5 dataset via spatiotemporal...

    arXiv 2022

  63. [63]

    \ Liu, Q

    yuDeepLearningbasedDownscaling2021 APACrefauthors Yu, M. \ Liu, Q. APACrefauthors \ 2021 . Deep Learning-Based Downscaling of Tropospheric Nitrogen Dioxide Using Ground-Level and Satellite Observations Deep learning-based downscaling of tropospheric nitrogen dioxide using ground-level and satellite observations . Science of The Total Environment 773 14514...

    arXiv 2021

  64. [64]

    , Neumann, N

    zhuUncertainNaturalEmissions2026 APACrefauthors Zhu, Q. , Neumann, N. , Fiore, A M. , Pincus, R. , Guan, J. , Milly, G. Giani, P. APACrefauthors \ 2026 . Uncertain Natural Emissions Dampen the Increase in Tropospheric Hydroxyl Radical ( OH ) With Idealized Surface Warming Uncertain Natural Emissions Dampen the Increase in Tropospheric Hydroxyl Radical ( O...

    work page doi:10.1029/2025ms005248 2026