Pith. sign in

REVIEW 3 major objections 4 minor 150 references

City-scale pollution attribution is identifiable exactly when the background-projected, lagged transport response matrix has full column rank, and its smallest singular value sets the noise-robust ceiling.

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

2026-08-04 01:29 UTC pith:CN35ZOYJ

load-bearing objection A sound framework with careful controlled experiments and honest limitation statements, but the real-data New Delhi demonstration skips the paper's own effective-rank criterion and shows no uncertainty intervals—fixable, but it needs to be fixed. the 3 major comments →

arxiv 2608.00050 v1 pith:CN35ZOYJ submitted 2026-07-25 eess.SP cs.CEcs.LG

Identifiability-Aware Source Apportionment in City-Scale Advection-Diffusion Systems

Ankit Bhardwaj , Lakshminarayanan Subramanian This is my paper
classification eess.SP cs.CEcs.LG
keywords source apportionmentidentifiabilityprojected lagged response matrixsingular value analysissparse sensor networksemission inventoriesurban air pollutionnonnegative estimation
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 tries to establish that, in inventory-based source apportionment from sparse urban sensors, the question of which declared source groups produced the observed pollution has a precise, computable answer. Attribution is identifiable exactly when the projected lagged response matrix — the sensor-time fingerprint of every source–basis coefficient after wind transport, lag, sparse sampling, and background removal — has full column rank, and it degrades gracefully only when that matrix's smallest singular value is large. Predictive fit is not evidence of identifiability, so the paper builds a procedure, IASA, that fits nonnegative source–basis coefficients and reports identifiability diagnostics, weak-coefficient flags, and conservative report groups side by side. If this is right, an apportionment study should report the finest attribution the sensing geometry supports, not the most detailed vector the inventory permits.

Core claim

Conditional on a declared inventory, a prechosen nonnegative temporal activity basis, a wind-driven transport operator, a background basis, and a lag window, the paper's discovery is that coefficient identifiability reduces to the projected lagged response matrix \(\tilde H_\Phi = P_Q^\perp H^{\rm lag}_\Phi\). Each column of \(H^{\rm lag}_\Phi\) is the sensor-time signature that one unit of a source–basis coefficient would leave after transport and sparse observation. After projecting out the background space, exact recovery of all nonnegative coefficients holds if and only if \(\mathrm{rank}(\tilde H_\Phi) = J\), and noise-robust recovery obeys \(\|\hat c - c\|_2 \le 2\|E\|_2 / \sigma_J(\ti

What carries the argument

The projected lagged response matrix \(\tilde H_\Phi\) — the array of sensor-time fingerprints produced by one unit of each source–basis coefficient after wind transport, time lag, sparse sensing, and projection off the background space — is the load-bearing object. Its full column rank is the exact-identifiability criterion; its smallest singular value \(\sigma_J\) is the noise-amplification constant; and the derived diagnostics (coefficient visibility, background absorption, pairwise coherence, ray distance, effective rank) all read off the same matrix. Because this object is computable before any fitting, it cleanly separates what is fitted from what is diagnosed.

Load-bearing premise

Everything rests on the declared source maps, the predeclared temporal activity basis, and the transport operator being the true generating structure for the observed field; an omitted source whose sensor signature lies inside the span of the fitted response and background is absorbed and stays invisible to residual checks.

What would settle it

Run a controlled two-source experiment with known coefficients under a steady single-direction wind and a sensor layout that makes \(\tilde H_\Phi\) rank-deficient; if any fit returns a unique, unflagged split, the rank criterion is falsified. Conversely, if a full-rank \(\tilde H_\Phi\) with very small \(\sigma_J\) produces recovery far better than the \(2\|E\|_2/\sigma_J\) bound, the robustness claim fails.

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

If this is right

  • A model that fits sensor readings well can still be unable to tell source groups apart; attribution claims without identifiability diagnostics are incomplete.
  • The finest defensible source resolution can be computed before collecting new data by evaluating \(\tilde H_\Phi\) over historical or simulated wind windows.
  • A too-flexible background model can absorb source-driven signal, driving the smallest singular value to zero while fitted residuals stay essentially unchanged.
  • When two source fingerprints are near-proportional, the honest output is a merged report group or a weak-visibility flag, not a confident per-source split.
  • Sensor placement and wind diversity should be judged by how they shape fingerprint conditioning and coherence, not by coverage statistics alone.

Where Pith is reading between the lines

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

  • The natural reframing is from point estimates to resolution certificates: the deliverable of an apportionment study becomes the coarsest-to-finest scale at which the sensing system can defend a split, and anything finer is labeled non-identifiable.
  • The same geometry transfers to other sparse linear inverse problems with known spatial maps transported to fixed receivers — satellite-column emissions tracing, indoor source localization, groundwater plume attribution — whenever the forward response can be simulated.
  • A testable extension is to make \(\sigma_J\) an explicit optimization objective for sensor placement or wind-window selection, maximizing the minimum singular value; the paper's prospective adequacy analysis points in this direction without optimizing it.
  • Policy users should read real-world results as conditional on inventory completeness: the residual test is one-sided, and an omitted source whose signature resembles the fitted ones is absorbed and invisible.

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 / 4 minor

Summary. The paper proposes IASA, an identifiability-aware source-apportionment framework for sparse urban air-quality sensor networks with inventory-based source models. It represents source activity as nonnegative combinations of a predeclared temporal basis, transports the resulting source maps through a wind-conditioned lagged Gaussian-puff operator, projects out a low-rank background basis, and identifies the projected lagged response matrix \widetilde H_\Phi as the object governing coefficient identifiability. Proposition 4.2 gives the rank condition for exact identifiability; Proposition 4.3 bounds noise sensitivity by 1/σ_J(\widetilde H_\Phi). The paper defines a diagnostic panel (numerical/effective rank, singular values, visibility, background absorption, pairwise coherence, ray distance), a source-ambiguity graph, conservative report groups, and a fit/diagnosis separation enforced by predeclared thresholds. Controlled experiments on a New Delhi platform test conditioning, wind/layout diversity, background stress, transport error, lag selection, temporal-basis recovery, inventory robustness, and baseline comparisons against NNLS, CMB, and PMF. Observed New Delhi weeks 1–4 are reported as full-rank with four singleton report groups and a population-to-brick-kiln apportionment swing.

Significance. If the methodology holds up, the paper makes a useful contribution: it converts the well-known but often informal concern that receptor-model attributions can be non-identifiable into a concrete, computable set of diagnostics, and it refuses to report finer attributions than the projected response matrix supports. The controlled experiments are carefully designed, with predeclared thresholds, honest separation of fit from diagnosis, explicit structural-mismatch testing, and unusually detailed reproducibility and sanity-gate appendices. The propositions in Appendix A are correct, and the paper is candid about its limitations, including the residual-invisibility of an omitted source whose signature lies in span[H_lag_Phi, Q] and the uncalibrated status of the observed-data adequacy test. However, the observed New Delhi study, which supports the abstract's claim that IASA 'reports the finest attribution,' does not itself apply the paper's own noise-dependent effective-rank criterion and contains no reported uncertainty intervals. These gaps are load-bearing for the real-data demonstration and require revision.

major comments (3)
  1. [Section 5.2, Table 14; Prop. 4.3 and Appendix B] The observed New Delhi identifiability conclusion omits the paper's own effective-rank criterion. Section 5.2 states that every week has 'full numerical rank (7 free coefficients...), finite-conditioned, low max coherence, empty weak set, no cross-source ambiguous pairs, and four singleton report components,' and Table 14 lists σ1, σJ, numerical rank, and max coherence. But no observation-noise scale σe, no τσ = σe√N, and no reff(τσ) are given. Proposition 4.3 and the effective-rank definition in Appendix B make clear that numerical full rank alone does not establish noise-robust identifiability: if reff(τσ) < 7, some coefficients should be flagged weak or grouped, and the 'four singleton report components' conclusion would not be supported at the actual noise level. Please report σe (from sensor calibration or a bootstrap residual estimate), the resulting reff(τσ) per week, or explicitl
  2. [Section 5.2, Eq. (19), Appendix G, Table 15] The subsection titled 'New Delhi Proxy Apportionment and Uncertainty' reports no uncertainty. Table 15 gives only point estimates of weekly proxy shares. The protocol in Eq. (19) and Appendix G specifies active-set/ridge covariance intervals and, when transport ensembles are supplied, empirical quantiles across ensemble refits. None of these appear in Section 5.2 or Table 15. Consequently, the substantial week-to-week swing from population-dominated (weeks 1–3) to brick-kiln-dominated (week 4) is not quantified, and statements such as 'the week-4 brick-kiln dominance ... fits the end-of-season firing push' go beyond what the reported data support. Please add coefficient and group-share intervals (or explicitly state that no calibrated noise model and no transport ensemble are available, so uncertainty is uncalibrated), and soften the causal narrative accordingly.
  3. [Section 5.2 and Figure 2(a)] The observed New Delhi resolution is also conditional on an unvalidated transport operator and wind field. The controlled recovery experiments use the same Gaussian-puff operator as generator and inference, so they validate within-family recovery, not the adequacy of the puff approximation for real Delhi transport; the one structural-mismatch experiment (Table 9) tests the adequacy test's power but does not validate the operator on the observed window. The paper honestly states that dense wind truth is unavailable and that the learned wind imputer did not beat a city-mean baseline, yet the abstract's 'finest attribution' claim inherits the unvalidated operator. Please add a sensitivity analysis of the New Delhi report groups under transport-ensemble or wind-imputation perturbations, or explicitly restrict the real-data conclusion to the declared operator and imputed wind field with no tr
minor comments (4)
  1. [Table 3 and Table 17] Table 3 says τρ is 'predeclared, near 1' but does not give the actual value used. Table 17 reports maximum coherence up to 0.958 with no merge; this is only interpretable if τρ is specified. Please state the exact threshold used in the reported experiments.
  2. [Figure 2(a)] The caption says 'bars σ1/σJ (left, log)', which is ambiguous: the text describes σ1 and σJ as separate quantities. Clarify whether the bars show both singular values or their ratio.
  3. [Section 5.2, first paragraph] The finding that the learned wind model did not beat a non-spatial city-mean baseline is important for interpreting the imputed wind field. It appears only as a brief statement; please give the held-out error comparison so readers can assess the magnitude of the imputation problem.
  4. [Section 4, 'Identifiable Source Resolution'] The transitive over-merging behavior of the source-ambiguity graph (A–B and B–C merging A–C) is described, which is good. It would be helpful to state explicitly that the final singleton-vs-merged report for the observed weeks is the deterministic component output of this graph and not a direct claim about pairwise distinguishability of every source pair.

Circularity Check

0 steps flagged

No significant circularity: identifiability claims are derived from the independently constructed projected response matrix; self-citations are peripheral.

full rationale

The paper's central chain is self-contained. The projected response matrix eH_Phi is built from declared source maps, temporal basis, wind-transport response, and a metadata-only background basis (Eqs. 1�C5), and exact identifiability is characterized by rank(eH_Phi)=J with a direct rank-nullity proof (Proposition 4.2). Noise robustness is derived from the pseudoinverse bound and sigma_J (Proposition 4.3), not from fitted coefficients. Diagnostics and thresholds are predeclared and explicitly separated from the fit (Algorithm 1, Table 3). The controlled recovery experiments reuse the same puff operator as generator and inference, but this is a matched self-consistency check, and the auxiliary advection-diffusion simulator provides an independent structural-mismatch test (Section 5.1, Appendix H). The paper's self-citations (Bhardwaj et al. 2025; Bhardwaj, Balashankar, and Subramanian 2025) appear only in related-work positioning and are not load-bearing. The observed New Delhi 'four singleton' conclusion omits the paper's own noise-dependent effective-rank criterion (tau_sigma = sigma_e sqrt(N)) and reports no sigma_e, so that claim is under-supported; the paper also marks residual adequacy as uncalibrated and concedes that in-span omitted sources are residual-invisible. These are limitations or correctness gaps, not circular reductions: no equation in the observed-data section is equal to its input by construction.

Axiom & Free-Parameter Ledger

8 free parameters · 7 axioms · 0 invented entities

The paper introduces no new physical entity, force, particle, or conserved quantity. Its novel objects (projected lagged response matrix, source ambiguity graph, report groups, per-sensor footprints) are all mathematical constructions derived from declared inputs, so the invented-entity ledger is empty. The most important unpaid inputs are the inventory maps, the temporal dictionary, the transport operator, and the background basis, rather than any new physical postulate.

free parameters (8)
  • Gaussian puff dispersion parameters sigma_parallel, sigma_perp, t_min
    Covariance model in Eq. 13 controls the shape and width of each transported fingerprint; no numerical values are given in the submission, but they directly affect eH_Phi and all identifiability diagnostics.
  • Lag convergence threshold tau_L = 1e-3 (default)
    Eq. 3 selects the primary lag L=16 for New Delhi from this threshold; a different threshold changes the response matrix and the conditioning reported.
  • Ambiguity coherence threshold tau_rho
    Set 'predeclared, near 1' in Table 3; it controls which fingerprint pairs enter the ambiguous-pair set and therefore the conservative report groups.
  • Visibility floor tau_v
    Set as SNR_min * sigma_e in Table 3; it determines the weak coefficient set W and which pairs are eligible for coherence-based merging.
  • Effective-rank noise floor tau_sigma = sigma_e * sqrt(N) (default)
    Table 3; this threshold determines reff and therefore whether the network supports all J coefficients at the stated observation-noise level.
  • Background basis effective-rank cap / rank-four Delhi basis = capped at 8; rank 4 for observed Delhi
    Chosen before fitting from metadata only; the rank and exact columns determine background absorption and the projection that defines eH_Phi.
  • Regularizer lambda in Eq. 6/22
    The ridge/inventory-prior strength is not given a numerical value; it affects the coefficient fit and the active-set covariance uncertainty.
  • Wind imputer Gaussian kernel bandwidth and grid query parameters
    The kernel coordinate-query imputer produces the gridded transport field, but its bandwidth and grid settings are not specified, so the response matrix cannot be independently reconstructed.
axioms (7)
  • domain assumption The declared inventory maps S (Guttikunda-Calori industry/brick, GPWv4 population, Google road traffic) approximate the true Delhi source spatial structure.
    Used in Section 5, Table 4, and Appendix H to build the four source groups; if these proxies are wrong, the identifiability conclusions apply to the wrong source model.
  • domain assumption The open-boundary Gaussian puff operator with the anisotropic covariance model (Eqs. 12-13) is a sufficient approximation of real advection-diffusion transport for inference.
    This operator builds H^lag_Phi for both controlled generation and inference; transport error is bounded in Proposition E.1 but the real-world adequacy is only tested against one auxiliary simulator.
  • domain assumption Each source's activity is a nonnegative combination of the predeclared temporal dictionary Phi; the dictionary itself is not learned from the sparse observations.
    Section 3 explicitly assumes this to avoid bilinear non-identifiability; if real source dynamics are not in span(Phi), the reconstructed activity trajectories are structurally biased.
  • domain assumption The background basis Q, built only from timestamps, day labels, sensor identity, and coordinates, captures regional background and smooth trends without absorbing source-driven signal.
    Q defines the projection P_Q^perp that produces eH_Phi; the source-like background stress test shows the method detects absorption, but the normal rank-four Delhi basis is itself an assumption.
  • domain assumption CPCB hourly PM2.5 and wind records are accurate, and missingness in PM2.5 is handled by row masking without imputation.
    The observed platform uses government records and the valid-value mask; any systematic sensor error propagates into observations and fitted coefficients.
  • domain assumption The kriged initial-condition baseline estimated from two Pusa monitors and propagated through the puff operator approximates the true initial pollution field.
    Appendix H declares and subtracts this zero-source transported baseline before fitting source activity; the paper notes this itself is an ill-posed spatial estimate.
  • standard math Uniform identifiability over the nonnegative orthant follows from rank-nullity together with nonnegative feasibility of the kernel counterexample.
    This is the proof of Proposition 4.2 in Appendix A; it is correct textbook linear algebra.

pith-pipeline@v1.3.0-alltime-deepseek · 27417 in / 11214 out tokens · 131354 ms · 2026-08-04T01:29:15.376230+00:00 · methodology

0 comments
read the original abstract

Source apportionment from sparse urban air-quality sensors is an inverse problem limited by sensor placement, wind-driven transport, background variation, and noise. Known or proxy emission inventories make attribution meaningful by restricting the unknown source field to a finite set of candidate groups, but do not guarantee those groups are distinguishable from the observations. We represent time-varying source activity with a low-dimensional nonnegative temporal basis and formulate inventory-based apportionment as a wind-conditioned lagged inverse problem in which each source--basis coefficient produces a sensor-time fingerprint. After projecting out a separate low-dimensional background space, the relevant object is the projected lagged response matrix $\widetilde H_\Phi$: exact identifiability at the chosen basis resolution requires its full column rank, while noise-robust attribution is controlled by its singular values, coefficient visibility, background absorption, pairwise coherence, and ray distance. We propose an identifiability-aware apportionment (IASA) framework that estimates nonnegative source--basis coefficients, reconstructs activity trajectories, and reports uncertainty and conservative grouping recommendations for indistinguishable sources. We instantiate it on a New Delhi platform built from government PM$_{2.5}$ and wind records, regulatory sensor locations, and four proxy source groups, and define controlled and observed evaluations of recovery, ambiguity, wind diversity, background stress, transport error, inventory robustness, and residual adequacy. IASA reports the attribution resolution defensible under the declared inventories, transport, background, lag, and noise rather than the most detailed possible vector.

Figures

Figures reproduced from arXiv: 2608.00050 by Ankit Bhardwaj, Lakshminarayanan Subramanian.

Figure 1
Figure 1. Figure 1: Controlled experiments. (a) IASA vs. three identifiability-blind baselines (plain NNLS, chemical mass balance, PMF/NMF) on the wind/geometry collapse and source-like background stress; bars are apportionment share error (log) and the row below marks whether each method raised an identifiability flag (IASA flags both; no baseline does). (b) Conditioning, not noise, sets the coefficient-recovery ceiling (clo… view at source ↗
Figure 2
Figure 2. Figure 2: Observed New Delhi. (a) Weeks 1–4: every week is full rank; bars σ1/σJ (left, log) and max eligible coherence (red, right axis 0–1) ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.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

150 extracted references · 2 canonical work pages

  1. [1]

    T.; and Kunisch, K

    Banks, H. T.; and Kunisch, K. 2012. Estimation techniques for distributed parameter systems. Springer Science & Business Media

    2012

  2. [2]

    Beck, A.; and Teboulle, M. 2009. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM Journal on Imaging Sciences, 2(1): 183--202

    2009

  3. [3]

    Belis, C.; Favez, O.; Mircea, M.; Diapouli, E.; Manousakas, M.; Vratolis, S.; Gilardoni, S.; Paglione, M.; Decesari, S.; Mocnik, G.; et al. 2019. European guide on air pollution source apportionment with receptor models—Revised version 2019. Publications Office, LU

    2019

  4. [4]

    R.; and Hopke, P

    Belis, C.; Karagulian, F.; Larsen, B. R.; and Hopke, P. 2013. Critical review and meta-analysis of ambient particulate matter source apportionment using receptor models in Europe. Atmospheric Environment, 69: 94--108

    2013

  5. [5]

    Bhardwaj, A.; Balashankar, A.; Iyer, S.; Soans, N.; Sudarshan, A.; Pande, R.; and Subramanian, L. 2025. Comprehensive monitoring of air pollution hotspots using sparse sensor networks. ACM Journal on Computing and Sustainable Societies, 3(4): 1--36

    2025

  6. [7]

    Carrassi, A.; Bocquet, M.; Bertino, L.; and Evensen, G. 2018. Data assimilation in the geosciences: An overview of methods, issues, and perspectives. Wiley Interdisciplinary Reviews: Climate Change, 9(5): e535

    2018

  7. [8]

    Center for International Earth Science Information Network (CIESIN), Columbia University . 2018. Gridded Population of the World, Version 4 (GPWv4): Population Density, Revision 11

    2018

  8. [9]

    Chen, C.-T. 1984. Linear system theory and design, volume 301. Holt, Rinehart and Winston New York

    1984

  9. [10]

    Chen, L.; Xu, J.; Wu, B.; and Huang, J. 2023. Group-aware graph neural network for nationwide city air quality forecasting. ACM Transactions on Knowledge Discovery from Data, 18(3): 1--20

    2023

  10. [11]

    L.; Ewing, R

    Colton, D. L.; Ewing, R. E.; Rundell, W.; et al. 1990. Inverse problems in partial differential equations, volume 42. Siam

    1990

  11. [12]

    CPCB. 2023. Air Quality Data. https://cpcb.nic.in/

    2023

  12. [13]

    CPCB. 2025. CPCB Data Portal. https://app.cpcbccr.com/ccr/#/caaqm-dashboard-all/caaqm-landing/caaqm-comparison-data

    2025

  13. [14]

    W.; Hanke, M.; and Neubauer, A

    Engl, H. W.; Hanke, M.; and Neubauer, A. 1996. Regularization of inverse problems, volume 375. Springer Science & Business Media

    1996

  14. [15]

    Google . 2024. Google Maps. https://maps.google.com. Accessed: 2026-03-30

    2024

  15. [16]

    K.; and Calori, G

    Guttikunda, S. K.; and Calori, G. 2013. A GIS based emissions inventory at 1 km × 1 km spatial resolution for air pollution analysis in Delhi, India. Atmospheric Environment, 67: 101--111

    2013

  16. [17]

    Hopke, P. K. 2016. Review of receptor modeling methods for source apportionment. Journal of the Air & Waste Management Association, 66(3): 237--259

    2016

  17. [18]

    K.; Dai, Q.; Li, L.; and Feng, Y

    Hopke, P. K.; Dai, Q.; Li, L.; and Feng, Y. 2020. Global review of recent source apportionments for airborne particulate matter. Science of The Total Environment, 740: 140091

    2020

  18. [19]

    R.; Balashankar, A.; Aeberhard, W

    Iyer, S. R.; Balashankar, A.; Aeberhard, W. H.; Bhattacharyya, S.; Rusconi, G.; Jose, L.; Soans, N.; Sudarshan, A.; Pande, R.; and Subramanian, L. 2022. Modeling fine-grained spatio-temporal pollution maps with low-cost sensors. npj Climate and Atmospheric Science, 5(1): 76

    2022

  19. [20]

    u ss-Ust \

    Karagulian, F.; Belis, C. A.; Dora, C. F. C.; Pr \"u ss-Ust \"u n, A. M.; Bonjour, S.; Adair-Rohani, H.; and Amann, M. 2015. Contributions to cities' ambient particulate matter (PM): A systematic review of local source contributions at global level. Atmospheric environment, 120: 475--483

    2015

  20. [23]

    Li, Z.; Zheng, H.; Kovachki, N.; Jin, D.; Chen, H.; Liu, B.; Azizzadenesheli, K.; and Anandkumar, A. 2024. Physics-informed neural operator for learning partial differential equations. ACM/IMS Journal of Data Science, 1(3): 1--27

    2024

  21. [24]

    Ljung, L. 1987. System Identification: Theory for the User. Prentice Hall

    1987

  22. [25]

    Mircea, M.; Calori, G.; Pirovano, G.; Belis, C.; et al. 2020. European guide on air pollution source apportionment for particulate matter with source oriented models and their combined use with receptor models. Publications Office of the European Union, LU

    2020

  23. [26]

    Paatero, P.; and Tapper, U. 1994. Positive matrix factorization: A non-negative factor model with optimal utilization of error estimates of data values. Environmetrics, 5(2): 111--126

    1994

  24. [27]

    Pant, P.; and Harrison, R. M. 2012. Critical review of receptor modelling for particulate matter: a case study of India. Atmospheric Environment, 49: 1--12

    2012

  25. [28]

    Raissi, M.; Perdikaris, P.; and Karniadakis, G. E. 2019. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378: 686--707

    2019

  26. [29]

    Seber, G. A. F.; and Lee, A. J. 2003. Linear Regression Analysis. Hoboken, NJ: Wiley, 2nd edition

    2003

  27. [30]

    Tarantola, A. 2005. Inverse problem theory and methods for model parameter estimation. SIAM

    2005

  28. [31]

    A.; Querol, X.; Alastuey, A.; Harrison, R

    Viana, M.; Kuhlbusch, T. A.; Querol, X.; Alastuey, A.; Harrison, R. M.; Hopke, P. K.; Winiwarter, W.; Vallius, M.; Szidat, S.; Pr \'e v \^o t, A. S.; et al. 2008. Source apportionment of particulate matter in Europe: a review of methods and results. Journal of aerosol science, 39(10): 827--849

    2008

  29. [32]

    Wang, B.; Sun, Z.; Jiang, X.; Zeng, J.; and Liu, R. 2023. Kalman filter and its application in data assimilation. Atmosphere, 14(8): 1319

    2023

  30. [33]

    G.; Zhu, T.; Chow, J

    Watson, J. G.; Zhu, T.; Chow, J. C.; Engelbrecht, J.; Fujita, E. M.; and Wilson, W. E. 2002. Receptor modeling application framework for particle source apportionment. Chemosphere, 49(9): 1093--1136

    2002

  31. [36]

    Wiley Interdisciplinary Reviews: Climate Change , volume=

    Data assimilation in the geosciences: An overview of methods, issues, and perspectives , author=. Wiley Interdisciplinary Reviews: Climate Change , volume=. 2018 , publisher=

    2018

  32. [37]

    Atmosphere , volume=

    Kalman filter and its application in data assimilation , author=. Atmosphere , volume=. 2023 , publisher=

    2023

  33. [38]

    arXiv preprint arXiv:1711.10561 , year=

    Physics informed deep learning (part i): Data-driven solutions of nonlinear partial differential equations , author=. arXiv preprint arXiv:1711.10561 , year=

    Pith/arXiv arXiv

  34. [39]

    Journal of Computational physics , volume=

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , author=. Journal of Computational physics , volume=. 2019 , publisher=

    2019

  35. [40]

    ACM/IMS Journal of Data Science , volume=

    Physics-informed neural operator for learning partial differential equations , author=. ACM/IMS Journal of Data Science , volume=. 2024 , publisher=

    2024

  36. [41]

    Journal of computational science , volume=

    Combining data assimilation and machine learning to emulate a dynamical model from sparse and noisy observations: A case study with the Lorenz 96 model , author=. Journal of computational science , volume=. 2020 , publisher=

    2020

  37. [42]

    Physical Review Fluids , volume=

    Data assimilation empowered neural network parametrizations for subgrid processes in geophysical flows , author=. Physical Review Fluids , volume=. 2021 , publisher=

    2021

  38. [43]

    arXiv preprint arXiv:1910.00935 , year=

    Difftaichi: Differentiable programming for physical simulation , author=. arXiv preprint arXiv:1910.00935 , year=

    Pith/arXiv arXiv 1910

  39. [44]

    arXiv preprint arXiv:1707.01926 , year=

    Diffusion convolutional recurrent neural network: Data-driven traffic forecasting , author=. arXiv preprint arXiv:1707.01926 , year=

    Pith/arXiv arXiv

  40. [45]

    arXiv preprint arXiv:1709.04875 , year=

    Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting , author=. arXiv preprint arXiv:1709.04875 , year=

    Pith/arXiv arXiv

  41. [46]

    arXiv preprint arXiv:1906.00121 , year=

    Graph wavenet for deep spatial-temporal graph modeling , author=. arXiv preprint arXiv:1906.00121 , year=

    Pith/arXiv arXiv 1906

  42. [47]

    ACM Transactions on Knowledge Discovery from Data , volume=

    Group-aware graph neural network for nationwide city air quality forecasting , author=. ACM Transactions on Knowledge Discovery from Data , volume=. 2023 , publisher=

    2023

  43. [48]

    Mathematical geology , volume=

    The origins of kriging , author=. Mathematical geology , volume=. 1990 , publisher=

    1990

  44. [49]

    Scientific reports , volume=

    Recurrent neural networks for multivariate time series with missing values , author=. Scientific reports , volume=. 2018 , publisher=

    2018

  45. [50]

    Advances in neural information processing systems , volume=

    Brits: Bidirectional recurrent imputation for time series , author=. Advances in neural information processing systems , volume=

  46. [51]

    Expert Systems with Applications , volume=

    Saits: Self-attention-based imputation for time series , author=. Expert Systems with Applications , volume=. 2023 , publisher=

    2023

  47. [52]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    ImputeFormer: Low rankness-induced transformers for generalizable spatiotemporal imputation , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  48. [53]

    arXiv preprint arXiv:2001.02908 , year=

    Spatial-temporal transformer networks for traffic flow forecasting , author=. arXiv preprint arXiv:2001.02908 , year=

    Pith/arXiv arXiv 2001

  49. [54]

    Nature machine intelligence , volume=

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators , author=. Nature machine intelligence , volume=. 2021 , publisher=

    2021

  50. [55]

    Proceedings of the National Academy of Sciences , volume=

    Machine learning--accelerated computational fluid dynamics , author=. Proceedings of the National Academy of Sciences , volume=. 2021 , publisher=

    2021

  51. [56]

    Journal of Computational Science , volume=

    TransFlowNet: A physics-constrained Transformer framework for spatio-temporal super-resolution of flow simulations , author=. Journal of Computational Science , volume=. 2022 , publisher=

    2022

  52. [57]

    Machine Learning: Science and Technology , volume=

    Physics informed token transformer for solving partial differential equations , author=. Machine Learning: Science and Technology , volume=. 2024 , publisher=

    2024

  53. [58]

    ACM Transactions on Graphics (ToG) , volume=

    Physics informed neural fields for smoke reconstruction with sparse data , author=. ACM Transactions on Graphics (ToG) , volume=. 2022 , publisher=

    2022

  54. [59]

    Nature communications , volume=

    Deep learning for universal linear embeddings of nonlinear dynamics , author=. Nature communications , volume=. 2018 , publisher=

    2018

  55. [60]

    2019 , publisher=

    Spatio-temporal statistics with R , author=. 2019 , publisher=

    2019

  56. [61]

    The Econometrics Journal , volume=

    Interpolation of spatial data: Some theory for kriging , author=. The Econometrics Journal , volume=. 2016 , publisher=

    2016

  57. [62]

    2007 , publisher=

    Finite difference methods for ordinary and partial differential equations: steady-state and time-dependent problems , author=. 2007 , publisher=

    2007

  58. [63]

    2003 , publisher=

    The finite element method: linear static and dynamic finite element analysis , author=. 2003 , publisher=

    2003

  59. [64]

    Statistica Sinica , pages=

    Predictive spatio-temporal models for spatially sparse enviromental data , author=. Statistica Sinica , pages=. 2005 , publisher=

    2005

  60. [65]

    npj Climate and Atmospheric Science , volume=

    Modeling fine-grained spatio-temporal pollution maps with low-cost sensors , author=. npj Climate and Atmospheric Science , volume=. 2022 , publisher=

    2022

  61. [66]

    Advances in neural information processing systems , volume=

    Characterizing possible failure modes in physics-informed neural networks , author=. Advances in neural information processing systems , volume=

  62. [67]

    arXiv preprint arXiv:2010.08895 , year=

    Fourier neural operator for parametric partial differential equations , author=. arXiv preprint arXiv:2010.08895 , year=

    Pith/arXiv arXiv 2010

  63. [68]

    arXiv preprint arXiv:2003.03485 , year=

    Neural operator: Graph kernel network for partial differential equations , author=. arXiv preprint arXiv:2003.03485 , year=

    Pith/arXiv arXiv 2003

  64. [69]

    International Conference on Medical image computing and computer-assisted intervention , pages=

    U-net: Convolutional networks for biomedical image segmentation , author=. International Conference on Medical image computing and computer-assisted intervention , pages=. 2015 , organization=

    2015

  65. [70]

    arXiv preprint arXiv:2204.11127 , year=

    U-no: U-shaped neural operators , author=. arXiv preprint arXiv:2204.11127 , year=

    Pith/arXiv arXiv

  66. [71]

    NCEP FNL Operational Model Global Tropospheric Analyses, continuing from July 1999

    NCEP. NCEP FNL Operational Model Global Tropospheric Analyses, continuing from July 1999. 2000

    1999

  67. [72]

    Guttikunda and Giuseppe Calori , keywords =

    Sarath K. Guttikunda and Giuseppe Calori , keywords =. A GIS based emissions inventory at 1 km × 1 km spatial resolution for air pollution analysis in Delhi, India , journal =. 2013 , issn =. doi:https://doi.org/10.1016/j.atmosenv.2012.10.040 , url =

    work page doi:10.1016/j.atmosenv.2012.10.040 2013

  68. [73]

    doi:10.7927/H49C6VHW , url=

    Gridded Population of the World, Version 4 (GPWv4): Population Density, Revision 11 , year=. doi:10.7927/H49C6VHW , url=

    doi:10.7927/h49c6vhw

  69. [74]

    arXiv preprint arXiv:1309.6835 , year=

    Gaussian processes for big data , author=. arXiv preprint arXiv:1309.6835 , year=

    Pith/arXiv arXiv

  70. [75]

    Advances in neural information processing systems , volume=

    Fourier features let networks learn high frequency functions in low dimensional domains , author=. Advances in neural information processing systems , volume=

  71. [76]

    Advances in neural information processing systems , volume=

    Implicit neural representations with periodic activation functions , author=. Advances in neural information processing systems , volume=

  72. [77]

    Neural networks , volume=

    Approximation capabilities of multilayer feedforward networks , author=. Neural networks , volume=. 1991 , publisher=

    1991

  73. [78]

    arXiv preprint arXiv:1912.10077 , year=

    Are transformers universal approximators of sequence-to-sequence functions? , author=. arXiv preprint arXiv:1912.10077 , year=

    Pith/arXiv arXiv 1912

  74. [79]

    Communications of the ACM , volume=

    Multidimensional binary search trees used for associative searching , author=. Communications of the ACM , volume=. 1975 , publisher=

    1975

  75. [80]

    S. R. ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , title=. 2020 , volume=

    2020

  76. [81]

    Proceedings of the 34th International Conference on Machine Learning (ICML) , pages =

    Neural Message Passing for Quantum Chemistry , author =. Proceedings of the 34th International Conference on Machine Learning (ICML) , pages =. 2017 , volume =

    2017

  77. [82]

    International Conference on Learning Representations , year=

    Semi-Supervised Classification with Graph Convolutional Networks , author=. International Conference on Learning Representations , year=

  78. [83]

    International Conference on Learning Representations , year=

    Graph Attention Networks , author=. International Conference on Learning Representations , year=

  79. [84]

    2020 , eprint=

    Temporal Graph Networks for Deep Learning on Dynamic Graphs , author=. 2020 , eprint=

    2020

  80. [85]

    Air Pollution Monitoring 101 , howpublished =

Showing first 80 references.