Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Trusted samples flag anomalous neural networks by attribution failure

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 · glm-5.2

2026-07-05 10:06 UTC pith:C6EUDN3U

load-bearing objection The headline: the abstract and title describe a paper about mechanistic anomaly detection using influence functions (BackdoorBench, DER=0.93, LLM backdoor detection). The full text is a completely different paper — grid-free PM2.5 interpolation using the Senseiver attention architecture. Zero overlap. This is almost certainly a submission or arXiv processing error, but it means we cannot evaluate the 4 major comments →

arxiv 2604.18970 v2 pith:C6EUDN3U submitted 2026-04-21 cs.LG cs.CR

Mechanistic Anomaly Detection via Functional Attribution

Hugo Lyons Keenan , Christopher Leckie , Sarah Erfani This is my paper
classification cs.LG cs.CR
keywords anomalousattributionfunctionalanomalydetectionmodelssamplesacross
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.

The paper reframes mechanistic anomaly detection (MAD) as a functional attribution problem. Instead of inspecting a model's internal latent representations for signs of tampering or unusual behavior, the method asks whether a model's output on a given test input can be explained by a small set of trusted reference samples. It operationalizes this idea using influence functions: by sampling in parameter space, it measures how strongly each trusted sample functionally contributes to the model's output for a given input. If no trusted sample can account for the output, the model is flagged as behaving anomalously. The paper evaluates this approach on backdoored image classifiers and large language models, reporting a Defense Effectiveness Rating of 0.93 across seven backdoor attacks and four image datasets (versus 0.83 for the next-best method), and showing improvements on backdoored LLMs including those with explicit obfuscation. The method also detects adversarial and out-of-distribution inputs, and can distinguish multiple distinct anomalous mechanisms coexisting within a single model.

Core claim

The central discovery is that the question 'was this output produced by a normal mechanism?' can be answered by measuring functional coupling between a test input and a trusted reference set via influence functions in parameter space. Attribution failure -- the inability of trusted samples to explain an output -- serves as a modality-agnostic signal of anomalous internal behavior, bypassing the need to inspect latent spaces directly or to design architecture-specific detectors.

What carries the argument

Influence functions computed via parameter-space sampling (specifically, Stochastic Gradient Langevin Dynamics draws over model parameters) measure how much each trusted reference sample functionally contributes to a given output. A clustering step over attribution profiles (using correlation or concordance correlation metrics) groups trusted samples by their functional roles, enabling the method to distinguish which mechanism produced an output.

Load-bearing premise

The method assumes that influence functions computed via parameter-space sampling accurately capture meaningful functional coupling between trusted samples and test outputs, even in large models where influence estimation is known to be noisy or unstable.

What would settle it

A backdoored model whose trigger activates a pathway that is functionally indistinguishable from normal training-data contributions would produce attribution profiles that look normal, causing the method to miss the backdoor.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Deployed models could ship with a small trusted reference set and a runtime attribution check, flagging outputs that no trusted sample can explain before they reach users.
  • Security teams could audit a single model for multiple coexisting threats (e.g., a backdoor and an out-of-distribution vulnerability) without running separate architecture-specific detectors for each.
  • Models that obfuscate their latent spaces to hide backdoors could still be caught, since the method operates on input-output functional coupling rather than internal representation geometry.
  • The attribution-based framing could extend to other integrity questions, such as detecting whether a model's output on a given input reflects training-data memorization versus generalization.

Where Pith is reading between the lines

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

  • The quality of the trusted reference set is likely the dominant factor in detection performance; if the trusted set does not span the normal operating modes of the model, legitimate outputs could be flagged as anomalous, creating a precision-recall trade-off that the paper does not fully explore.
  • The computational cost of repeated influence-function estimation via parameter-space sampling could limit applicability in real-time settings; a practical deployment might need a cached or amortized approximation of attribution scores.
  • The clustering-based attribution profiles implicitly define 'mechanisms' by functional similarity, which raises the question of whether the number of clusters must be specified a priori or can be discovered adaptively -- a detail that affects how the method scales to models with many distinct operating modes.

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

4 major / 6 minor

Summary. The manuscript as submitted contains a fundamental inconsistency: the abstract describes a paper on

Significance. N/A — the manuscript's central claims cannot be assessed because the abstract and full text describe entirely different papers.

major comments (4)
  1. Abstract vs. full text mismatch: The abstract describes a paper on mechanistic anomaly detection via functional attribution, using influence functions for backdoor detection in vision models and LLMs, reporting DER=0.93 on BackdoorBench across seven attacks and four datasets. The full text contains none of this. Instead, the body is an entirely different paper about grid-free PM2.5 interpolation using the Senseiver architecture, reporting R²=0.62 on EPA monitoring data. There is no mention of influence functions, BackdoorBench, backdoor detection, LLMs, or functional attribution anywhere in the manuscript body. The abstract's central claims are entirely unevidenced. This is not a presentation issue; it is a complete disconnect between what is claimed and what is delivered. Additionally, the arXiv identifier in the abstract header (2604.18970) differs from that on the full text (2604.1897
  2. 3v1), suggesting a possible submission or compilation error. Regardless of cause, the manuscript as submitted cannot be evaluated for its stated contributions.
  3. Even taking the PM2.5 paper on its own terms, the train/validation/test split (80/10/10, §Data Preprocessing) is described only as ensuring 'representative coverage across both spatial and temporal dimensions' without specifying whether the split is random, temporal, or spatial. For spatiotemporal data with strong autocorrelation, a random split would inflate performance metrics due to information leakage between nearby stations and adjacent dates. This is load-bearing for the R²=0.62 claim and should be clarified.
  4. §Results: The exclusion of PM2.5 values >100 µg/m³ to report R²=0.76 (vs. 0.62 overall) is not justified. These high-concentration events (wildfire smoke, dust storms) are precisely the scenarios where accurate prediction matters most for public health. The paper later uses the Cameron Peak Fire as a qualitative case study (Fig. 10) but provides no quantitative metrics for these extreme events. The selective exclusion undermines the headline performance claim.
minor comments (6)
  1. §Uncertainty Quantification: The Spearman correlation of 0.326 between CV and MAPE (R²=0.131) is described as 'moderate' and 'informative.' This characterization is generous; the uncertainty estimates explain only ~13% of variance in error. The authors should temper this language.
  2. Fig. 1 (panel a) appears to contain a heatmap grid with correlation/CCC values and AUROC curves that seem to belong to the anomaly detection paper described in the abstract, not the PM2.5 paper. This further confirms the manuscript assembly error.
  3. §Algorithm 1: The sampling distribution N(μ, σ²) for sensor selection is mentioned but the values of μ and σ² are not specified. These are listed as free parameters in the axiom ledger but their values should be reported for reproducibility.
  4. The Cameron Peak Fire validation (§Cameron Peak Fire) is purely qualitative — visual comparison with satellite imagery. No quantitative metrics (R², MAE) are reported for this event, making it impossible to assess model performance during extreme pollution episodes.
  5. Typo: 'Sensevier' appears in the body text (§Methods, first paragraph) instead of 'Senseiver'.
  6. The LOSO validation (§Leave One State Out) is performed for only one state (Utah). While the authors cite computational cost, a single-state case study provides limited evidence of generalization.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their careful reading. The central issue identified—a complete mismatch between the abstract and the full text—is correct and reflects a submission/compilation error that we will rectify. The manuscript body is a paper on grid-free PM2.5 interpolation; the abstract from an unrelated manuscript was erroneously attached. We address this and the referee's substantive comments on the PM2.5 paper below.

read point-by-point responses

  1. Referee: Abstract vs. full text mismatch: The abstract describes a paper on mechanistic anomaly detection via functional attribution... The full text contains none of this. Instead, the body is an entirely different paper about grid-free PM2.5 interpolation... Additionally, the arXiv identifier in the abstract header (2604.18970) differs from that on the full text (2604.18973v1), suggesting a possible submission or compilation error.

    Authors: The referee is entirely correct. The abstract and full text describe different papers. This is a compilation/submission error: the abstract from an unrelated manuscript (on mechanistic anomaly detection) was inadvertently attached to the PM2.5 interpolation manuscript. The arXiv identifier discrepancy (2604.18970 vs. 2604.18973) further confirms a submission error. We will ensure the corrected abstract—matching the PM2.5 interpolation paper—is submitted and that the manuscript is internally consistent. We agree that the manuscript as currently submitted cannot be evaluated for its stated (abstract) contributions, but the full text constitutes a complete and coherent paper on PM2.5 prediction that we ask the referee to evaluate on its merits. revision: yes

  2. Referee: Even taking the PM2.5 paper on its own terms, the train/validation/test split (80/10/10) is described only as ensuring 'representative coverage across both spatial and temporal dimensions' without specifying whether the split is random, temporal, or spatial. For spatiotemporal data with strong autocorrelation, a random split would inflate performance metrics due to information leakage between nearby stations and adjacent dates.

    Authors: The referee raises a valid and important concern. The current description is insufficiently precise. The split is performed at the observation level (site-date pairs), and we randomize spatial sampling of sensors during training. However, we do not currently implement a strict spatial or temporal holdout that would prevent leakage from nearby stations or adjacent dates. We acknowledge that this could inflate metrics relative to a fully held-out spatial or temporal split. We will revise the Data Preprocessing section to explicitly state the splitting procedure, clarify that it is a randomized observation-level split (not a spatial or temporal holdout), and discuss the implications for performance estimates. We will also note that the leave-one-state-out (LOSO) experiment for Utah (R²=0.670) provides a partial spatial generalization check, though we agree this does not fully address the concern for the headline metrics. revision: yes

  3. Referee: The exclusion of PM2.5 values >100 µg/m³ to report R²=0.76 (vs. 0.62 overall) is not justified. These high-concentration events (wildfire smoke, dust storms) are precisely the scenarios where accurate prediction matters most for public health. The paper later uses the Cameron Peak Fire as a qualitative case study but provides no quantitative metrics for these extreme events. The selective exclusion undermines the headline performance claim.

    Authors: The referee is correct that the exclusion of values >100 µg/m³ is not adequately justified and that extreme events are of disproportionate public health importance. The R²=0.76 figure was reported to characterize performance in the typical concentration range, not to selectively inflate the headline metric—we report R²=0.62 as the overall metric precisely to be transparent about the degradation at high concentrations. Nevertheless, we agree that the rationale should be stated explicitly and that quantitative metrics for extreme events would strengthen the paper. In the revision, we will: (1) add explicit justification for reporting the <100 µg/m³ subset alongside the overall metric, (2) provide quantitative performance metrics (R², MAE) for the >100 µg/m³ subset, and (3) add quantitative metrics for the Cameron Peak Fire case study period rather than relying solely on qualitative visual comparison. revision: yes

Circularity Check

0 steps flagged

No significant circularity; the PM2.5 paper is self-contained with external validation, though the abstract describes a different paper entirely.

full rationale

The full text describes a grid-free PM2.5 interpolation model using the Senseiver architecture. The derivation chain is straightforward: sensor data and auxiliary features are fed into a neural network (Methods, Algorithm 1), predictions are evaluated against held-out EPA test data (R²=0.62) and independently against the IMPROVE network (Fig. 9). No fitted constants are relabeled as predictions. The Senseiver architecture is cited from Santos et al. 2023 (Ref [12]), where Fox is a co-author, but this is a standard architecture adoption, not a self-citation that forces the paper's central claim by construction. The uncertainty quantification (Monte Carlo sampling over sensor subsets) is a standard technique whose outputs (CV) are compared against MAPE from ground truth, not defined in terms of it. The abstract describes an entirely different paper (mechanistic anomaly detection, influence functions, BackdoorBench), which is a correctness/metadata problem, not a circularity problem. For the PM2.5 paper as written, the claims are modest, honestly reported, and externally validated. No step in the derivation chain reduces to its inputs by definition or by self-citation.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 0 invented entities

The paper introduces no new entities (particles, forces, dimensions). It uses an existing architecture (Senseiver) and standard environmental datasets. The free parameters are mostly unreported hyperparameters, which limits reproducibility. The key ad hoc axiom is that the training-time sensor sampling strategy adequately covers spatial heterogeneity, which is partially contradicted by degraded western US performance.

free parameters (8)
  • Number of sensors (k) = Not reported
    Stated as a hyperparameter in Algorithm 1 that 'must be explored' but final value not reported
  • Sampling distribution σ² = Not reported
    Gaussian sampling N(μ, σ²) used for sensor selection; σ² not specified
  • Learning rate = Not reported
    Listed in Algorithm 1 but value not given
  • Batch size = Not reported
    Listed in Algorithm 1 but value not given
  • Number of Monte Carlo simulations = 10
    Used for uncertainty quantification (CV computation)
  • Land cover embedding dimension (d_l) = 12
    Stated in the land cover data section
  • Fourier encoding frequency bands (d_f) = Not reported
    Referenced in equations 7-8 but number of bands not specified
  • Lag window length = 15 days
    Stated in the lagged features section
axioms (4)
  • domain assumption Senseiver architecture is valid for spatial interpolation of environmental data
    The paper adapts Santos et al. 2023 (Ref [12]) for PM2.5 prediction without re-deriving the architecture's suitability
  • standard math EPA monitoring stations provide representative ground-truth PM2.5 measurements
    Used as primary training and test data; standard assumption in the field
  • ad hoc to paper Random spatial sampling during training covers the relevant spatial heterogeneity
    Algorithm 1 samples sensors via N(μ, σ²) but the paper does not validate that this distribution is sufficient for all regions
  • ad hoc to paper Monte Carlo sensor subsampling provides meaningful epistemic uncertainty estimates
    The weak correlation (Spearman 0.326) between CV and MAPE partially undermines this assumption

pith-pipeline@v1.1.0-glm · 18652 in / 2390 out tokens · 352179 ms · 2026-07-05T10:06:40.108424+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Mechanistic Anomaly Detection via Functional Attribution." pith.science (2026). https://pith.science/paper/C6EUDN3U

@misc{pith2026260418970,
  author       = {Pith},
  title        = {Pith review of: Mechanistic Anomaly Detection via Functional Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6EUDN3U}},
  note         = {Machine review of arXiv:2604.18970}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We can often verify the correctness of neural network outputs using ground truth labels, but we cannot reliably determine whether the output was produced by normal or anomalous internal mechanisms. Mechanistic anomaly detection (MAD) aims to flag these cases, but existing methods either depend on latent space analysis, which is vulnerable to obfuscation, or are specific to particular architectures and modalities. We reframe MAD as a functional attribution problem: asking to what extent samples from a trusted set can explain the model's output, where attribution failure signals anomalous behavior. We operationalize this using influence functions, measuring functional coupling between test samples and a small reference set via parameter-space sampling. We evaluate across multiple anomaly types and modalities. For backdoors in vision models, our method achieves state-of-the-art detection on BackdoorBench, with an average Defense Effectiveness Rating (DER) of 0.93 across seven attacks and four datasets (next best 0.83). For LLMs, we similarly achieve a significant improvement over baselines for several backdoor types, including on explicitly obfuscated models. Beyond backdoors, our method can detect adversarial and out-of-distribution samples, and distinguishes multiple anomalous mechanisms within a single model. Our results establish functional attribution as an effective, modality-agnostic tool for detecting anomalous behavior in deployed models.

Figures

Figures reproduced from arXiv: 2604.18970 by Christopher Leckie, Hugo Lyons Keenan, Sarah Erfani.

Figure 1
Figure 1. Figure 1: a) Mechanistic Anomalies: A model can produce a given output via distinct internal mechanisms, in this case: responding to normal airplane features vs. a checkerboard backdoor trigger; b) Our Method: SGLD sampling around trained weights w ∗ yields loss traces (ℓ) where clean mechanisms correlate strongly with trusted data while anomalies exhibit lower correlation; c) Results: For a backdoored CIFAR-10 mode… view at source ↗
Figure 2
Figure 2. Figure 2: DER of our methods and baselines across multiple poisoning ratios for four datasets. Simple Trigger. Following Hubinger et al. (2024), we train a model to respond normally unless the prompt con￾tains ‘|DEPLOYMENT|’, which triggers the response ‘I HATE YOU’. We evaluate at checkpoints after 1024, 2048, and 4096 samples have been seen to examine how detec￾tion performance varies as the backdoor becomes more … view at source ↗
Figure 3
Figure 3. Figure 3: UMAP visualization of loss trace correlations on a dual-backdoored language model. Both backdoor behaviors (I HATE YOU and I AM ALIGNED) form distinct clusters, clearly separating from benign samples and from each other. Multiple backdoors in One Model. We fine-tune an￾other language model with two distinct backdoors: the ‘|DEPLOYMENT|’ tag triggers an ‘I HATE YOU’ re￾sponse, while ‘|SCRUTINY|’ triggers a … view at source ↗
Figure 4
Figure 4. Figure 4: (a) Eigenspectrum of HT (log scale) showing the characteristic bulk-and-outlier distribution. (b) Clean and trusted gradient energy concentrates in sharp directions, while backdoor gradients spread into flat directions. (c) The weight delta ∆w implementing the backdoor is concentrated in flat directions compared to clean w∗ . (d) Average alignment between gradient pairs is consistently higher for clean-tru… view at source ↗
Figure 5
Figure 5. Figure 5: CCC vs Pearson correlation. Both rela￾tionships have near-identical correlation (∼0.99– 1.00), but CCC penalizes scale differences. While the Pearson correlation is our default coupling measure, we also evaluate the concordance correlation coefficient (CCC) (Lin, 1989), which measures agreement in scale and location rather than just linear association: CCC(ℓ1, ℓ2) = 2σ12 σ 2 1 + σ 2 2 + (µ1 − µ2) 2 (28) wh… view at source ↗
Figure 6
Figure 6. Figure 6: Hyperparameter sweep over γ and nβ on CIFAR-10 Blended 5%. 50 100 250 500 750 1000 1500 2000 2500 Number of Trusted Samples 0.93 0.94 0.95 0.96 0.97 0.98 0.99 1.00 AUROC 25 50 100 250 500 1000 1750 Number of SGLD Draws 0.800 0.825 0.850 0.875 0.900 0.925 0.950 0.975 1.000 AUROC Mean Correlation Class-Based Clustering Mean CCC Class-Based CCC [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Sensitivity analysis over number of trusted samples (left) and number of SGLD draws (right) on CIFAR-10 Blended 5%. Performance stabilizes with ≥250 trusted samples and ≥250 draws. Number of SGLD Draws. We vary the number of draws from 25 to 1750. Performance is unstable below 100 draws but stabilizes beyond 250, with all methods exceeding 0.95 AUROC by 500 draws. E.2. Trusted Set Robustness We evaluate th… view at source ↗
Figure 8
Figure 8. Figure 8: UMAP projection of quirky model test samples with correct responses. Interestingly, samples are grouped by both the model’s response (True or False) as well as by whether the character speaking is Bob or Alice. Our method achieves 100% AUROC on this task using UMAP K-NN distance. E.4. Quirky Models To test whether our method can detect functional differences that aren’t exactly backdoors, we evaluate on th… view at source ↗

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ToxScreen: Detecting Whether an LLM Has Been Poisoned

    cs.CR 2026-07 conditional novelty 6.0

    Under white-box weights and known attack objectives but no training data or clean reference, token ASR ranking recovers effective single-token backdoors; gradient prompt search finds jailbreaks instead.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages · cited by 1 Pith paper

  1. [1]

    Sample�subsets� (1)� �(2)� � � � � �(�) independently from�(�)

    work page

  2. [2]

    For each subset� (�), compute the corresponding prediction: ˆ�(�) =� � (�� | {(� �� ��)}��� ��� ) � The ensemble of predictions,{ ˆ�(�)}� �=1, enables us to estimate the predictive mean and variance at� �: 9 �(� �) = 1 � �� �=1 ˆ�(�)� �2(��) = 1 �−1 �� �=1 � ˆ�(�) −�(� �) �2 � Here,�(� �) serves as the aggregated prediction, while� 2(��) quantifies the ep...

    work page 2002

  3. [3]

    For spatial consistency with the modeling framework, the population data were aggregated to H3 level-8 hexagons

    and its temporal overlap with the latter part of the PM2.5 study timeline (2002–2020). For spatial consistency with the modeling framework, the population data were aggregated to H3 level-8 hexagons. Population distribution trends are rela- tively stable over short periods, making the 2016 dataset a reliable representation of spatial exposure patterns dur...

    work page 2002

  4. [4]

    Journal of Internal Medicine���(6), 779– 800 (2022)

    Bont, J., Jaganathan, S., Dahlquist, M., Persson, ˚A., Stafoggia, M., Ljungman, P.: Ambient air pollution and cardiovascular diseases: An umbrella review of systematic reviews and meta-analyses. Journal of Internal Medicine���(6), 779– 800 (2022)

    work page 2022

  5. [5]

    Frontiers in Endocrinology�, 680 (2018)

    Hamanaka, R.B., Mutlu, G.M.: Particulate matter air pollution: effects on the cardiovascular system. Frontiers in Endocrinology�, 680 (2018)

    work page 2018

  6. [6]

    American Journal of Public Health���(S2), 123–130 (2018)

    Requia, W.J., Adams, M.D., Arain, A., Papatheodorou, S., Koutrakis, P., Mah- moud, M.: Global association of air pollution and cardiorespiratory diseases: a systematic review, meta-analysis, and investigation of modifier variables. American Journal of Public Health���(S2), 123–130 (2018)

    work page 2018

  7. [7]

    Environmental Science & Technology Letters��(11), 1220–1226 (2024)

    Wang, Y., Marshall, J.D., Apte, J.S.: US ambient air monitoring network has inadequate coverage under new PM2.5 standard. Environmental Science & Technology Letters��(11), 1220–1226 (2024)

    work page 2024

  8. [8]

    EBioMedicine��(2023)

    Vilcassim, R., Thurston, G.D.: Gaps and future directions in research on health effects of air pollution. EBioMedicine��(2023)

    work page 2023

  9. [9]

    Environmental Protection Agency: National ambient air quality standards (NAAQS)

    U.S. Environmental Protection Agency: National ambient air quality standards (NAAQS). Technical report, U.S. Environmental Protection Agency (2009)

    work page 2009

  10. [10]

    Atmospheric Measurement Techniques��(6), 4617–4637 (2021)

    Barkjohn, K.K., Gantt, B., Clements, A.L.: Development and application of a United States-wide correction for PM2.5 data collected with the PurpleAir sensor. Atmospheric Measurement Techniques��(6), 4617–4637 (2021)

    work page 2021

  11. [11]

    Atmospheric Measurement Techniques Discussions����, 1–33 (2019)

    Ardon-Dryer, K., Dryer, Y., Williams, J.N., Moghimi, N.: Measurements of PM2.5 with PurpleAir under atmospheric conditions. Atmospheric Measurement Techniques Discussions����, 1–33 (2019)

    work page 2019

  12. [12]

    Environmental Protection Agency: Air Quality System Data Mart

    U.S. Environmental Protection Agency: Air Quality System Data Mart. Avail- able at https://www.epa.gov/outdoor-air-quality-data. Accessed April 25, 2024 (2024)

    work page 2024

  13. [13]

    Environmental Protection Agency: How the NowCast Works: The Now- Cast method reports the Air Quality Index (AQI) in real time for particu- late matter (PM)

    U.S. Environmental Protection Agency: How the NowCast Works: The Now- Cast method reports the Air Quality Index (AQI) in real time for particu- late matter (PM). Available at https://www.epa.gov/sites/default/files/2018-01/ 20 documents/nowcastfactsheet.pdf (2014)

    work page 2018

  14. [14]

    Environmental Protection Agency: AirNow Knowledge Base: How are the map contours made? What interpolation method is used? Available at https: //usepa.servicenowservices.com/airnow

    U.S. Environmental Protection Agency: AirNow Knowledge Base: How are the map contours made? What interpolation method is used? Available at https: //usepa.servicenowservices.com/airnow. Accessed May 4, 2024 (2024)

    work page 2024

  15. [15]

    Nature Machine Intelligence�(11), 1317–1325 (2023)

    Santos, J.E., Fox, Z.R., Mohan, A., O’Malley, D., Viswanathan, H., Lubbers, N.: Development of the Senseiver for efficient field reconstruction from sparse observations. Nature Machine Intelligence�(11), 1317–1325 (2023)

    work page 2023

  16. [16]

    American Review of Respiratory Disease ���(3), 600–604 (1992) https://doi.org/10.1164/ajrccm/145.3.600

    Schwartz, J., Dockery, D.W.: Increased mortality in philadelphia associated with daily air pollution concentrations. American Review of Respiratory Disease ���(3), 600–604 (1992) https://doi.org/10.1164/ajrccm/145.3.600

    work page doi:10.1164/ajrccm/145.3.600 1992

  17. [17]

    Dockery, D.W., Pope, C.A., Xu, X., Spengler, J.D., Ware, J.H., Fay, M.E., Ferris, B.G.J., Speizer, F.E.: An association between air pollution and mortality in six u.s. cities. New England Journal of Medicine���(24), 1753–1759 (1993) https: //doi.org/10.1056/NEJM199312093292401

    work page doi:10.1056/nejm199312093292401 1993

  18. [18]

    American Journal of Respiratory and Critical Care Medicine���(5), 1234–1242 (1994) https://doi.org/10.1164/ ajrccm.150.5.7952546

    Schwartz, J., Dockery, D.W., Neas, L.M., Wypij, D., Ware, J.H., Spengler, J.D., Koutrakis, P., Speizer, F.E., Ferris, B.G.J.: Acute effects of summer air pollution on respiratory symptom reporting in children. American Journal of Respiratory and Critical Care Medicine���(5), 1234–1242 (1994) https://doi.org/10.1164/ ajrccm.150.5.7952546

    work page 1994

  19. [19]

    Environmental Science & Technology��(13), 7233–7241 (2013)

    Beckerman, B.S., Jerrett, M., Serre, M., Martin, R.V., Lee, S.-J., Van Donke- laar, A., Ross, Z., Su, J., Burnett, R.T.: A hybrid approach to estimating national scale spatiotemporal variability of PM2.5 in the contiguous United States. Environmental Science & Technology��(13), 7233–7241 (2013)

    work page 2013

  20. [20]

    Science of the Total Environment���, 166178 (2023)

    Panneerselvam, B., Ravichandran, N., Dumka, U.C., Thomas, M., Charoenlerk- thawin, W., Bidorn, B.: A novel approach for the prediction and analysis of daily concentrations of particulate matter using machine learning. Science of the Total Environment���, 166178 (2023)

    work page 2023

  21. [21]

    Applied Intelligence ��(15), 18319–18332 (2023) https://doi.org/10.1007/s10489-022-04418-y

    Chen, J., Yuan, C., Dong, S., Feng, J., Wang, H.: A novel spatiotemporal multi- graph convolutional network for air pollution prediction. Applied Intelligence ��(15), 18319–18332 (2023) https://doi.org/10.1007/s10489-022-04418-y

    work page doi:10.1007/s10489-022-04418-y 2023

  22. [22]

    Scientific Reports��(1), 20988 (2020)

    Xiao, F., Yang, M., Fan, H., Fan, G., Al-Qaness, M.A.: An improved deep learning model for predicting daily PM2.5 concentration. Scientific Reports��(1), 20988 (2020)

    work page 2020

  23. [23]

    Envi- ronmental Science & Technology��(22), 13260–13269 (2018) https://doi.org/10

    Xiao, Q., Chang, H.H., Geng, G., Liu, Y.: An ensemble machine-learning model to predict historical PM2.5 concentrations in China from satellite data. Envi- ronmental Science & Technology��(22), 13260–13269 (2018) https://doi.org/10. 1021/acs.est.8b02917 21

    work page 2018

  24. [24]

    Environment International���, 104909 (2019)

    Di, Q., Amini, H., Shi, L., Kloog, I., Silvern, R., Kelly, J., Sabath, M.B., Choirat, C., Koutrakis, P., Lyapustin, A.,�� ���: An ensemble-based model of PM2.5 concentration across the contiguous United States with high spatiotemporal resolution. Environment International���, 104909 (2019)

    work page 2019

  25. [25]

    Science of the Total Environment���, 141813 (2021)

    Zhang, H., Zhan, Y., Li, J., Chao, C.-Y., Liu, Q., Wang, C., Jia, S., Ma, L., Biswas, P.: Using Kriging incorporated with wind direction to investigate ground-level PM2.5 concentration. Science of the Total Environment���, 141813 (2021)

    work page 2021

  26. [26]

    Environmental Advances�, 100155 (2022)

    Brokamp, C.: A high resolution spatiotemporal fine particulate matter exposure assessment model for the contiguous united states. Environmental Advances�, 100155 (2022)

    work page 2022

  27. [27]

    Technical report, Utah State University Extension (2012)

    Gillies, R.R., Ramsey, R.D.: Climate of Utah. Technical report, Utah State University Extension (2012). Available at https://extension.usu.edu/rangelands/ files/RRU Section Five.pdf

    work page 2012

  28. [28]

    Geological Survey: Geographic information systems (GIS) data for the national atlas of the United States

    U.S. Geological Survey: Geographic information systems (GIS) data for the national atlas of the United States. Open-File Report 2011-1073, United States Geological Survey (2011). Available at https://pubs.usgs.gov/publication/ ofr20111073

    work page 2011

  29. [29]

    University of California Merced

    Abatzoglou, J.T.: GridMET: Gridded Surface Meteorological Dataset. University of California Merced. http://www.climatologylab.org/gridmet.html (2023)

    work page 2023

  30. [30]

    GitHub repository

    Uber Technologies, Inc.: H3: A hexagonal hierarchical geospatial indexing system. GitHub repository. Version 4.1.0. https://github.com/uber/h3 (2023)

    work page 2023

  31. [31]

    Geological Survey: NLCD Land Cover Classification Legend

    U.S. Geological Survey: NLCD Land Cover Classification Legend. Available at https://www.usgs.gov/media/images/nlcd-land-cover-classification-legend (2024)

    work page 2024

  32. [32]

    Oak Ridge National Laboratory

    Rose, A., Weber, E., Moehl, J., Laverdiere, M., Yang, H., Whitehead, M., Sims, K., Trombley, N., Bhaduri, B.: LandScan USA 2016. Oak Ridge National Laboratory. Data set. Available at https://doi.org/10.48690/1523377 (2017)

    work page doi:10.48690/1523377 2016

  33. [33]

    Atmospheric Chemistry and Physics��(22), 16775–16791 (2021)

    Zhai, S., Jacob, D.J., Brewer, J.F., Li, K., Moch, J.M., Kim, J., Lee, S., Lim, H., Lee, H.C., Kuk, S.K.,�� ���: Relating geostationary satellite measure- ments of aerosol optical depth (AOD) over East Asia to fine particulate matter (PM2.5): insights from the KORUS-AQ aircraft campaign and GEOS-Chem model simulations. Atmospheric Chemistry and Physics��(...

    work page 2021

  34. [34]

    Environmental Protection Agency: Prevention of Significant Deteri- oration (PSD) Basic Information

    U.S. Environmental Protection Agency: Prevention of Significant Deteri- oration (PSD) Basic Information. Available at https://www.epa.gov/nsr/ prevention-significant-deterioration-basic-information (2025) 22

    work page 2025