Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Physics-informed sensor coverage through structure preserving machine learning

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read This paper argues that a machine-learning surrogate which enforces conservation laws exactly, through conditional neural Whitney forms, turns sparse sensor measurements into accurate source maps, and that using those maps to guide mobile se

desk verdict A useful architecture paper with an honestly conditional theory and a confounded headline comparison; the adaptive Lloyd loop and conditional source head are the real contributions. read the letter →

arxiv 2509.10363 v1 pith:6USNZUTR submitted 2025-09-12 cs.LG cs.NAmath.NA

classification cs.LGcs.NAmath.NA MSC 35Q9368T0735R30
keywords optimalcoveragedigitaltwinsoperatorlearningadaptivesamplinginverseproblemsstructurepreservationphysics-informedmachineconditionalneuralWhitneyforms
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that when a neural network learns an inverse map from sparse sensor data to an unknown pollution-like source, hard-wiring the conservation structure of the transport PDE into the network is not just a constraint but a performance advantage. It constructs a digital twin—a reduced-order finite element model in which the field, flux, and source are represented as learnable Whitney forms conditioned on sensor readings—so that every prediction satisfies discrete conservation by construction. Around this model it wraps a two-step closed loop: estimate the source from current measurements, then move sensors with a geodesic Lloyd algorithm using the predicted source as an importance density. If the claims hold, the concrete payoff is that mobile robotic sensors can find and converge to a hidden source with fewer, noisier measurements than black-box learned surrogates would allow.

What carries the argument

Conditional Neural Whitney Forms (CNWF): a data-driven reduced-order model in which the conserved scalar is a 0-form, the flux a 1-form, and the source a 0-form, all expanded in a learned partition-of-unity basis constructed from fine-scale Whitney forms. A transformer encoder takes the unordered set of sensor measurements and conditions three heads—basis mixing weights, a nonlinear learnable flux, and a nonnegative source—so the discrete conservation law εδᵀM₁δu + δᵀM₁Nθ(u,z) = M₀fθ(z) holds exactly. The second mechanism is a geodesic Lloyd algorithm that treats the normalized predicted source as an importance density and moves sensors along geodesics toward projection-restricted centroids,

What would settle it

Compute the empirical ratio ∥ρ(xk)−ρtrue∥∞/Jρtrue(xk) along an adaptive-sampling run on a real or experimental transport setup (not synthetic FEM). If no positive constant CΦ exists, or if the ratio grows as sensors approach the predicted source, Theorem 5.3's error-bound decrease fails. More directly, a deployment where sensors converge to a wrong location predicted by a smooth but inaccurate importance map would falsify the claim that regularity alone suffices for localization.

Watch

Extended reading notes

Core claim

The central claim is that structure preservation provides an effective inductive bias for source identification: enforcing the discrete conservation law ∇·F = f on a learned reduced Whitney-form basis makes the sensor-to-source map regular enough to be both invertible in practice and useful as a coverage objective. The paper reports that CNWF recovers source distributions with lower Wasserstein error than physics-agnostic MLP and transformer baselines across circular, Gulf-of-Mexico, and maze geometries, produces source fields that, when fed back into the original PDE, give more faithful scalar-field reconstructions, and that using the predicted source as an importance function inside a geod

Load-bearing premise

The convergence and localization guarantees hold only if the learned importance density remains close enough to the true source distribution in a way that depends on model training and cannot be generally guaranteed (as the paper's Remark 3 concedes), and the experiments assume the synthetic FEM advection–diffusion settings with known velocity fields are representative of real deployments.

Editorial extensions

If this is right

  • Structure-preserving inverse surrogates generalize to unseen sensor counts and configurations without retraining, because the transformer encoder is permutation-invariant and conditions the same PDE operator.
  • Adaptive sampling improves source prediction for the structure-preserving model more than for black-box baselines, so physical consistency and informative sensing compound.
  • The learned reduced basis is interpretable: one partition often localizes the source, others encode the downstream plume, giving operators a visual diagnostic.
  • The closed-loop convergence theorems imply that, under regularity and accuracy conditions, sensors converge exponentially to the true source location under continuous Lloyd dynamics.
  • Enforcing conservation by construction removes the need for handcrafted regularization and keeps predictions PDE-consistent even in data-sparse or out-of-distribution regimes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the theorems reduce closed-loop performance to a single regularity/accuracy relation between the learned importance density and the true source, one testable extension is to monitor that relation online and throttle sensor movement when it is violated—a safety mechanism the paper does not propose.
  • The same conditional-Whitney-form construction could be applied to inverse problems beyond advection–diffusion, such as electromagnetic or elastic source localization, wherever a de Rham complex supplies the conserved quantities.
  • The claim that regularity is a sufficient condition for localization suggests a direct comparison against an explicitly regularized (e.g., Lipschitz-constrained) black-box inverse model, to isolate whether it is the conservation structure or merely the smoothness of predictions that drives the gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a structure-preserving machine learning framework for inverse source identification and adaptive sensor placement in advection-diffusion systems. The core model, Conditional Neural Whitney Forms (CNWF), couples a transformer encoder with a learned reduced Whitney-form basis, a learned nonlinear flux, and a learned source term, all solved under a discrete conservation law. The predicted source density is then used as an importance function in a geodesic Lloyd algorithm for sensor repositioning. The authors provide conditional convergence theorems (Theorems 5.1-5.4) and report experiments on circular, Gulf-of-Mexico, and maze geometries, comparing CNWF against MLP and transformer baselines on Wasserstein distance, source RMSE, and forward-consistency metrics. The main advertised claim is that hard conservation structure provides an effective inductive bias for source identification and that the adaptive loop further improves localization.

Significance. If the empirical claims hold, the paper offers a useful demonstration that exact discrete conservation plus a reduced basis can regularize an ill-posed sensor-to-source inverse map. The strengths include experiments on three geometrically distinct synthetic benchmarks, consistent improvement of the adaptive Lloyd loop across all cases, and a forward-consistency metric that quantifies compatibility with the original PDE. The analysis is honest in stating several key assumptions as conditional. However, the central empirical claim is not isolated by the experimental design, and the convergence theory rests on assumptions that are explicitly acknowledged as not generally guaranteed. The result is a promising but not yet fully supported contribution.

major comments (3)
  1. [§6.2, §7.1, Tables 1-2] The central claim that structure preservation provides an effective inductive bias is confounded by the training and inference protocol. In (5.10), CNWF is trained with both a scalar-field reconstruction term ||uθ - utrue||^2 and the Wasserstein source loss, while the MLP and transformer baselines are described as being trained to predict the source field directly. CNWF also solves the learned discrete PDE at inference, whereas the baselines map sensor data directly to ρ. Thus the reported gains (e.g., W2 1.16e-3 vs 7.37e-3 in Table 1) may be due to extra field supervision or forward-model regularization rather than the hard FEEC/conservation structure. An ablation is needed, e.g., a transformer trained with the same field-matching loss and a soft advection-diffusion penalty, or a CNWF variant with the conservation constraint relaxed. Without such a control, H1 is not fully supported. In
  2. [§5.7, Theorem 5.2 proof] The proof of Theorem 5.2 contains a sign inconsistency. After defining the Lloyd energy change ∆J^Lloyd_k < 0, the proof states that the descent requires "2CΩ||ρk − ρtrue||∞ < ∆J^Lloyd_k", which is impossible when the right-hand side is negative. The intended condition should be 2CΩ||ρk − ρtrue||∞ < −∆J^Lloyd_k, which is consistent with the theorem statement. As written, the proof does not establish the claimed result. This is a local but load-bearing error in the convergence analysis and should be corrected.
  3. [§5.7, Theorem 5.3 and Remark 3; §5.8, Theorem 5.4, (5.17)] The theoretical support for the positive feedback loop is explicitly conditional. Theorem 5.3 assumes ||ρ(xk) − ρtrue||∞ ≤ CΦ Jρtrue(xk), and Remark 3 states that this "cannot generally be guaranteed" and is "dependent on model training and construction." The paper does not estimate CΦ or verify the condition empirically; the correlation in Figure 9 is not a verification of the bound. Similarly, Theorem 5.4 assumes the contractive Lipschitz condition (5.17) on centroids of learned importance fields, but no evidence is provided that the CNWF density satisfies it; the appendix constructs a class of bump functions satisfying this condition, which is not shown to describe the learned model. These limitations should be stated prominently, and the abstract should not imply unconditional convergence.
minor comments (4)
  1. [§1, Eq. (1.5)] The text says "F is a prescribed optimal coverage functional," but the displayed equation uses G(f). Since F already denotes the flux, this notation is confusing and should be harmonized.
  2. [§4.4, §5.4, §7.1, Appendix A] Typos: "architechtures" (§4.4), "nonegativity" (§5.4), "We hypothesis" (§7.1), "amendeable" (Appendix A, likely intended "amenable"), and inconsistent use of "P´eclet" formatting.
  3. [Figures 8, 9, 11-13] The experimental figures report averages but no variance or number of trials for some curves. Adding shaded error bands or error bars would help assess the reliability of the reported improvements.
  4. [Table 3] The hyperparameter table lists "Data cache reset tolerance 25 ×" without specifying units or interpretation. Please clarify.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the source-identification claim is a supervised generalization result and the convergence theorems are explicitly conditional, not hidden restatements of fitted inputs.

full rationale

The core empirical claim (CNWF outperforms MLP/transformer baselines) is a held-out comparison: the model is trained under the PDE-constrained objective (5.10)-(5.11) with supervision on both the scalar field and true source, and then evaluated on fresh sensor configurations and velocity fields. Evaluation of a trained model on held-out data is not a circular prediction; it is generalization. The convergence results in Sections 5.6-5.8 are explicitly sufficient-condition theorems: Theorem 5.1 assumes a bound on ||ρ_{k+1}-ρ_k||, Theorem 5.2 assumes ΔJ_Lloyd < -2CΩ||ρ-ρ_true||, Theorem 5.3 assumes ||ρ-ρ_true|| ≤ CΦJ_true, and Theorem 5.4 assumes the contractivity condition (5.17). The paper repeatedly flags that these conditions 'cannot generally be guaranteed' and are 'dependent on model training' (Remark 3; Section 5.7). A theorem that derives a decrease in an assumed upper bound from a decrease in coverage is a valid conditional, not a claim that the bound was independently established. The CNWF architecture is inherited from the authors' prior work [36], but the present paper's headline result is an empirical benchmark against physics-agnostic baselines, so the self-citation is not load-bearing for the central claim; no uniqueness theorem is imported to forbid alternatives. The conservation-law ansatz is introduced explicitly in (1.3)/(5.1), not smuggled in via citation. The transformer baseline comparison is confounded by extra PDE supervision and forward-solve regularization, but that is an experimental-control concern, not a circularity. Overall the derivation chain is self-contained or explicitly conditional, so no circular step is established.

Assumptions & free parameters 8 free parameters · 7 assumptions · 0 invented entities

The method introduces no new physical entity. It relies on learned reduced bases, flux networks, and source networks, which are parameterized quantities fitted to synthetic data rather than newly postulated objects. The analysis is heavily conditional: the convergence theorems assume contractivity and error-to-coverage bounds that the authors explicitly mark as not guaranteed.

free parameters (8)
  • Neural network parameters theta = not reported
    All trainable weights in the transformer encoder, source head, flux head, and basis head are fit to synthetic FEM data through the constrained optimization in Eq. (5.10). They determine the predicted source, flux, and reduced basis.
  • Diffusive stabilization coefficient epsilon = not reported
    Trainable coefficient in the flux model (5.2) that stabilizes the reduced problem; its fitted value is not stated.
  • Nonlinear flux gain alpha = not reported
    Hand-tuned gain in the residual flux network (Appendix A) balancing nonlinearity and numerical stability; value not stated.
  • Lloyd relaxation step alpha = not reported
    Relaxation parameter in Eq. (5.13) controlling sensor update length; value not stated for experiments.
  • Inner Lloyd iterations m = not reported
    Number of inner Lloyd steps in Algorithm 5.2; chosen per experiment but not stated.
  • Number of partition-of-unity control volumes = 5
    Reduced basis size from Table 3; determines the dimension of the reduced conservation system and the expressiveness of the source representation.
  • Peclet number = 1000
    Set to 1e3 in Appendix B; controls the advection-diffusion balance of every test problem.
  • Source bump radius r = 0.07
    Radius of the compact bump source in Eq. (6.1); fixes the scale of the localization problem and is common across experiments.
assumptions (7)
  • standard math The learned reduced-order model inherits existence, stability, and consistency from Lax-Milgram theory for the data-driven Whitney-form discretization, cited to [36, Thm 2.1] and [63, Thms 3.1-5].
    The paper does not reprove these results; the well-posedness of Eq. (5.7) and of the PDE-constrained training problem (5.10) rests on them.
  • domain assumption The true field obeys steady-state advection-diffusion (3.1) with a divergence-free velocity, fixed boundary conditions, and a compactly supported source.
    This is the generative forward model for all synthetic data and the assumed physical setting for source identification.
  • domain assumption Sensors observe scalar field, velocity field, and Peclet number with additive noise as in Eq. (3.2).
    The conditioning input z includes v(xi) and Pe; if velocity is unavailable in a deployment, the method's inputs change.
  • ad hoc to paper Prediction error is bounded by true coverage energy: the inequality in Theorem 5.3, with the constant C_Phi.
    The authors state in Remark 3 that this condition cannot generally be guaranteed and depends on model training; the theorem's conclusion is a direct consequence of this assumption.
  • ad hoc to paper Contractivity and local Lipschitz condition (5.17) on centroids of learned importance fields in Theorem 5.4.
    This is the central assumption that makes exponential convergence to the source location follow. The appendix shows a constructed class of bump functions satisfies it, but it is not verified for the trained CNWF models.
  • standard math Lloyd's algorithm provides strict energy descent for a fixed importance field in convex Euclidean settings.
    Classical result from [14] used to define Delta-J-Lloyd as strictly negative.
  • standard math Geodesic centroids exist and are unique via Hadamard manifold theory [2] for Eq. (5.15).
    Used to justify centroids in the geodesic Lloyd update; the experiments approximate centroids by projection in Eq. (5.16).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-informed sensor coverage through structure preserving machine learning." pith.science (2026). https://pith.science/paper/6USNZUTR

@misc{pith2026250910363,
  author       = {Pith},
  title        = {Pith review of: Physics-informed sensor coverage through structure preserving machine learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6USNZUTR}},
  note         = {Machine review of arXiv:2509.10363}
}
read the original abstract

We present a machine learning framework for adaptive source localization in which agents use a structure-preserving digital twin of a coupled hydrodynamic-transport system for real-time trajectory planning and data assimilation. The twin is constructed with conditional neural Whitney forms (CNWF), coupling the numerical guarantees of finite element exterior calculus (FEEC) with transformer-based operator learning. The resulting model preserves discrete conservation, and adapts in real time to streaming sensor data. It employs a conditional attention mechanism to identify: a reduced Whitney-form basis; reduced integral balance equations; and a source field, each compatible with given sensor measurements. The induced reduced-order environmental model retains the stability and consistency of standard finite-element simulation, yielding a physically realizable, regular mapping from sensor data to the source field. We propose a staggered scheme that alternates between evaluating the digital twin and applying Lloyd's algorithm to guide sensor placement, with analysis providing conditions for monotone improvement of a coverage functional. Using the predicted source field as an importance function within an optimal-recovery scheme, we demonstrate recovery of point sources under continuity assumptions, highlighting the role of regularity as a sufficient condition for localization. Experimental comparisons with physics-agnostic transformer architectures show improved accuracy in complex geometries when physical constraints are enforced, indicating that structure preservation provides an effective inductive bias for source identification.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Structure-Preserving Learning Improves Geometry Generalization in Neural PDEs

    cs.LG 2026-02 conditional novelty 6.0 of 10

    A geometry-conditioned Whitney-form neural network that solves a learned discrete conservation law improves out-of-distribution geometry generalization for steady-state PDEs compared with regression-based neural operators.

Reference graph

Works this paper leans on

73 extracted references · 9 linked inside Pith · cited by 1 Pith paper

  1. [36]

    Structure-preserving digital twins via conditional neural whitney forms

    Brooks Kinch, Benjamin Shaffer, Elizabeth Armstrong, Michael Meehan, John Hewson, and Nathaniel Trask. Structure-preserving digital twins via conditional neural whitney forms. arXiv preprint arXiv:2508.06981 , 2025

  2. [1]

    Data-driven whitney forms for structure-preserving control volume analysis

    Jonas A Actor, Xiaozhe Hu, Andy Huang, Scott A Roberts, and Nathaniel Trask. Data-driven whitney forms for structure-preserving control volume analysis. Journal of Computational Physics, 496:112520, 2024

  3. [2]

    Riemannian lp center of mass: existence, uniqueness, and convexity

    Bijan Afsari. Riemannian lp center of mass: existence, uniqueness, and convexity. Proceedings of the American Mathematical Society , 139(2):655–673, 2011

  4. [3]

    Model-based solution techniques for the source localization problem

    Mehmet E Alpay and Molly H Shor. Model-based solution techniques for the source localization problem. IEEE transactions on control systems technology , 8(6):895–904, 2000

  5. [4]

    Finite element exterior calculus

    Douglas N Arnold. Finite element exterior calculus . SIAM, 2018

  6. [5]

    Finite element exterior calculus, 25 Fig

    Douglas N Arnold, Richard S Falk, and Ragnar Winther. Finite element exterior calculus, 25 Fig. 13. Change in coverage costs from Figure 12. homological techniques, and applications. Acta numerica, 15:1–155, 2006

  7. [6]

    Solving inverse problems using data-driven models

    Simon Arridge, Peter Maass, Ozan ¨Oktem, and Carola-Bibiane Sch¨ onlieb. Solving inverse problems using data-driven models. Acta Numerica, 28:1–174, 2019

  8. [7]

    Iterative methods for approximate solution of inverse problems , volume 577

    Anatolii Borisovich Bakushinsky and M Yu Kokurin. Iterative methods for approximate solution of inverse problems , volume 577. Springer Science & Business Media, 2005

Show all 73 references
  1. [8]

    Inverse source problems in transport equations

    Guillaume Bal and Alexandru Tamasan. Inverse source problems in transport equations. SIAM journal on mathematical analysis , 39(1):57–76, 2007

  2. [9]

    Clifford neural layers for pde modeling

    Johannes Brandstetter, Rianne van den Berg, Max Welling, and Jayesh K Gupta. Clifford neural layers for pde modeling. arXiv preprint arXiv:2209.04934 , 2022

  3. [10]

    Physics-informed neural networks (pinns) for fluid mechanics: A review

    Shengze Cai, Zhiping Mao, Zhicheng Wang, Minglang Yin, and George Em Karniadakis. Physics-informed neural networks (pinns) for fluid mechanics: A review. Acta Mechan- ica Sinica, 37(12):1727–1738, 2021

  4. [11]

    Nonlinear least squares for inverse problems: theoretical foundations and step- by-step guide for applications

    Guy Chavent. Nonlinear least squares for inverse problems: theoretical foundations and step- by-step guide for applications . Springer Science & Business Media, 2010

  5. [12]

    Neural symplectic form: Learning hamiltonian equations on general coordinate systems

    Yuhan Chen, Takashi Matsubara, and Takaharu Yaguchi. Neural symplectic form: Learning hamiltonian equations on general coordinate systems. Advances in Neural Information Processing Systems, 34:16659–16670, 2021

  6. [13]

    Group equivariant convolutional networks

    Taco Cohen and Max Welling. Group equivariant convolutional networks. In International conference on machine learning , pages 2990–2999. PMLR, 2016

  7. [14]

    Coverage control for mobile sensing networks

    Jorge Cortes, Sonia Martinez, Timur Karatas, and Francesco Bullo. Coverage control for mobile sensing networks. IEEE Transactions on robotics and Automation , 20(2):243–255, 2004

  8. [15]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems , 26, 2013

  9. [16]

    Deep learning architectures for nonlinear operator functions and nonlinear inverse problems

    Maarten V de Hoop, Matti Lassas, and Christopher A Wong. Deep learning architectures for nonlinear operator functions and nonlinear inverse problems. Mathematical Statistics and Learning, 4(1):1–86, 2022

  10. [17]

    Discrete exterior calculus

    Mathieu Desbrun, Anil N Hirani, Melvin Leok, and Jerrold E Marsden. Discrete exterior calculus. arXiv preprint math/0508341 , 2005

  11. [18]

    Bacterium-inspired robots for environmental monitoring

    Amit Dhariwal, Gaurav S Sukhatme, and Aristides AG Requicha. Bacterium-inspired robots for environmental monitoring. In IEEE International Conference on Robotics and Au- tomation, 2004. Proceedings. ICRA’04. 2004 , volume 2, pages 1436–1443. IEEE, 2004

  12. [19]

    Centroidal voronoi tessellations: Applications and algorithms

    Qiang Du, Vance Faber, and Max Gunzburger. Centroidal voronoi tessellations: Applications and algorithms. SIAM review, 41(4):637–676, 1999

  13. [20]

    Regularization of inverse problems

    Heinz W Engl and Ronny Ramlau. Regularization of inverse problems. In Encyclopedia of 26 applied and computational mathematics , pages 1233–1241. Springer, 2015

  14. [21]

    Gmsh: A 3-d finite element mesh generator with built-in pre-and post-processing facilities

    Christophe Geuzaine and Jean-Fran¸ cois Remacle. Gmsh: A 3-d finite element mesh generator with built-in pre-and post-processing facilities. International journal for numerical methods in engineering, 79(11):1309–1331, 2009

  15. [22]

    Hamiltonian neural networks

    Samuel Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks. Ad- vances in neural information processing systems , 32, 2019

  16. [23]

    scikit-fem: A python package for finite element assembly

    Tom Gustafsson and Geordie Drummond Mcbain. scikit-fem: A python package for finite element assembly. Journal of Open Source Software , 5(52):2369, 2020

  17. [24]

    Multi-agent search for source localization in a turbulent medium

    Hadi Hajieghrary, M Ani Hsieh, and Ira B Schwartz. Multi-agent search for source localization in a turbulent medium. Physics Letters A , 380(20):1698–1705, 2016

  18. [25]

    Information theoretic source seeking strate- gies for multiagent plume tracking in turbulent fields

    Hadi Hajieghrary, Daniel Mox, and M Ani Hsieh. Information theoretic source seeking strate- gies for multiagent plume tracking in turbulent fields. Journal of Marine Science and Engineering, 5(1):3, 2017

  19. [26]

    Structure-preserving neural networks

    Quercus Hern´ andez, Alberto Bad ´ ıas, David Gonz´ alez, Francisco Chinesta, and El ´ ıas Cueto. Structure-preserving neural networks. Journal of Computational Physics , 426:109950, 2021

  20. [27]

    Hycom surface velocity fields for the gulf of mexico and the florida straits at 1km resolution for january 2014 and july 2014, 2019

    Pat Hogan. Hycom surface velocity fields for the gulf of mexico and the florida straits at 1km resolution for january 2014 and july 2014, 2019

  21. [28]

    Inverse source problems

    Victor Isakov. Inverse source problems. Number 34. American Mathematical Soc., 1990

  22. [29]

    Physics- informed neural networks for inverse problems in supersonic flows

    Ameya D Jagtap, Zhiping Mao, Nikolaus Adams, and George Em Karniadakis. Physics- informed neural networks for inverse problems in supersonic flows. Journal of Compu- tational Physics , 466:111402, 2022

  23. [30]

    A structure-preserving domain decomposition method for data-driven modeling

    Shuai Jiang, Jonas Actor, Scott Roberts, and Nathaniel Trask. A structure-preserving domain decomposition method for data-driven modeling. arXiv preprint arXiv:2406.05571 , 2024

  24. [31]

    Deep learning methods for inverse problems

    Shima Kamyab, Zohreh Azimifar, Rasool Sabzi, and Paul Fieguth. Deep learning methods for inverse problems. PeerJ Computer Science , 8:e951, 2022

  25. [32]

    Physics-informed machine learning

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics , 3(6):422–440, 2021

  26. [33]

    Generalized coverage control for time- varying density functions

    James Kennedy, Airlie Chapman, and Peter M Dower. Generalized coverage control for time- varying density functions. In 2019 18th European Control Conference (ECC), pages 71–76. IEEE, 2019

  27. [34]

    Model-based active source identification in complex environments

    Reza Khodayi-mehr, Wilkins Aquino, and Michael M Zavlanos. Model-based active source identification in complex environments. IEEE Transactions on Robotics , 35(3):633–652, 2019

  28. [35]

    Computing geodesic paths on manifolds

    Ron Kimmel and James A Sethian. Computing geodesic paths on manifolds. Proceedings of the national academy of Sciences , 95(15):8431–8435, 1998

  29. [37]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

  30. [38]

    Machine learning for groundwater pollution source identification and monitoring network optimization

    Yiannis N Kontos, Theodosios Kassandros, Konstantinos Perifanos, Marios Karampasis, Kon- stantinos L Katsifarakis, and Kostas Karatzas. Machine learning for groundwater pollution source identification and monitoring network optimization. Neural Computing and Appli- cations, 34...

  31. [39]

    Multirobot control using time- varying density functions

    Sung G Lee, Yancy Diaz-Mercado, and Magnus Egerstedt. Multirobot control using time- varying density functions. IEEE Transactions on robotics, 31(2):489–493, 2015

  32. [40]

    Controlled coverage using time-varying density functions

    Sung G Lee and Magnus Egerstedt. Controlled coverage using time-varying density functions. IF AC Proceedings Volumes, 46(27):220–226, 2013

  33. [41]

    Fourier neural operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895 , 2020

  34. [42]

    Explainable ai: A review of machine learning interpretability methods

    Pantelis Linardatos, Vasilis Papastefanopoulos, and Sotiris Kotsiantis. Explainable ai: A review of machine learning interpretability methods. Entropy, 23(1):18, 2020

  35. [43]

    Least squares quantization in pcm

    Stuart Lloyd. Least squares quantization in pcm. IEEE transactions on information theory , 28(2):129–137, 1982

  36. [44]

    Whitney forms and their extensions

    Jonni Lohi and Lauri Kettunen. Whitney forms and their extensions. Journal of Computational and Applied Mathematics , 393:113520, 2021

  37. [45]

    Learn- ing nonlinear operators via deeponet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learn- ing nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence , 3(3):218–229, 2021

  38. [46]

    Distributed environmental 27 modeling and adaptive sampling for multi-robot sensor coverage

    Wenhao Luo, Changjoo Nam, George Kantor, and Katia Sycara. Distributed environmental 27 modeling and adaptive sampling for multi-robot sensor coverage. In Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems , pages 1488–1496, 2019

  39. [47]

    Olfaction-based mobile robot navi- gation

    Lino Marques, Urbano Nunes, and An ´ ıbal T de Almeida. Olfaction-based mobile robot navi- gation. Thin solid films , 418(1):51–58, 2002

  40. [48]

    Source localization by spatially distributed electronic noses for advection and diffusion

    J¨ org Matthes, Lutz Groll, and Hubert B Keller. Source localization by spatially distributed electronic noses for advection and diffusion. IEEE Transactions on Signal Processing , 53(5):1711–1719, 2005

  41. [49]

    Estimates on the generalization error of physics- informed neural networks for approximating a class of inverse problems for pdes

    Siddhartha Mishra and Roberto Molinaro. Estimates on the generalization error of physics- informed neural networks for approximating a class of inverse problems for pdes. IMA Journal of Numerical Analysis , 42(2):981–1022, 2022

  42. [50]

    The discrete geodesic problem

    Joseph SB Mitchell, David M Mount, and Christos H Papadimitriou. The discrete geodesic problem. SIAM Journal on Computing , 16(4):647–668, 1987

  43. [51]

    Neural inverse oper- ators for solving pde inverse problems

    Roberto Molinaro, Yunan Yang, Bj¨ orn Engquist, and Siddhartha Mishra. Neural inverse oper- ators for solving pde inverse problems. arXiv preprint arXiv:2301.11167 , 2023

  44. [52]

    Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chil- amkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, An- dreas K¨ opf, Edward Z. Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chil- amkurthy, Benoit Steine...

  45. [53]

    Thermodynamically consistent physics-informed neural networks for hyperbolic systems

    Ravi G Patel, Indu Manickam, Nathaniel A Trask, Mitchell A Wood, Myoungkyu Lee, Ignacio Tomas, and Eric C Cyr. Thermodynamically consistent physics-informed neural networks for hyperbolic systems. Journal of Computational Physics , 449:110754, 2022

  46. [54]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. 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

  47. [55]

    A machine learning approach to identifying point source locations in photoacoustic data

    Austin Reiter and Muyinatu A Lediju Bell. A machine learning approach to identifying point source locations in photoacoustic data. In Photons Plus Ultrasound: Imaging and Sensing 2017, volume 10064, pages 504–509. SPIE, 2017

  48. [56]

    Corbijn van Willenswaard

    Nicolas Renaud, Stef Smeets, and Lars J. Corbijn van Willenswaard. nanomesh

  49. [57]

    On the definition and importance of interpretability in scientific machine learning

    Conor Rowan and Alireza Doostan. On the definition and importance of interpretability in scientific machine learning. arXiv preprint arXiv:2505.13510 , 2025

  50. [58]

    A compar- ison of reactive robot chemotaxis algorithms

    R Andrew Russell, Alireza Bab-Hadiashar, Rod L Shepherd, and Gordon G Wallace. A compar- ison of reactive robot chemotaxis algorithms. Robotics and Autonomous Systems, 45(2):83– 97, 2003

  51. [59]

    Decentralized minimum-energy coverage control for time-varying density functions

    Maria Santos, Siddharth Mayya, Gennaro Notomista, and Magnus Egerstedt. Decentralized minimum-energy coverage control for time-varying density functions. In2019 international symposium on multi-robot and multi-agent systems (MRS) , pages 155–161. IEEE, 2019

  52. [60]

    A distributed formation-based odor source localization algorithm-design, implementation, and wind tun- nel evaluation

    Jorge M Soares, A Pedro Aguiar, Ant´ onio M Pascoal, and Alcherio Martinoli. A distributed formation-based odor source localization algorithm-design, implementation, and wind tun- nel evaluation. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 18...

  53. [61]

    Inverse problems: a bayesian perspective

    Andrew M Stuart. Inverse problems: a bayesian perspective. Acta numerica, 19:451–559, 2010

  54. [62]

    Inverse problem theory and methods for model parameter estimation

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

  55. [63]

    Enforcing exact physics in scientific machine learning: a data-driven exterior calculus on graphs

    Nathaniel Trask, Andy Huang, and Xiaozhe Hu. Enforcing exact physics in scientific machine learning: a data-driven exterior calculus on graphs. Journal of Computational Physics , 456:110969, 2022

  56. [64]

    Contaminant source identifica- tion using semi-supervised machine learning

    Velimir V Vesselinov, Boian S Alexandrov, and Daniel O’Malley. Contaminant source identifica- tion using semi-supervised machine learning. Journal of contaminant hydrology , 212:134– 142, 2018

  57. [65]

    Physics-informed neural network algorithm for solving forward and inverse problems of variable-order space-fractional advection–diffusion equations

    Shupeng Wang, Hui Zhang, and Xiaoyun Jiang. Physics-informed neural network algorithm for solving forward and inverse problems of variable-order space-fractional advection–diffusion equations. Neurocomputing, 535:64–82, 2023

  58. [66]

    Understanding and mitigating gradient flow pathologies in physics-informed neural networks

    Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing , 43(5):A3055–A3081, 2021

  59. [67]

    Bioinspired algorithm for autonomous sensor- driven guidance in turbulent chemical plumes

    DR Webster, KY Volyanskyy, and MJ Weissburg. Bioinspired algorithm for autonomous sensor- driven guidance in turbulent chemical plumes. Bioinspiration & biomimetics , 7(3):036023, 2012

  60. [68]

    Multiple source detection and localization in advection-diffusion processes using wireless sensor networks

    James Weimer, Bruno Sinopoli, and Bruce Krogh. Multiple source detection and localization in advection-diffusion processes using wireless sensor networks. In 2009 30th IEEE Real-Time 28 Systems Symposium, pages 333–342. IEEE, 2009

  61. [69]

    Foundational research gaps and future directions for digital twins

    Karen Willcox et al. Foundational research gaps and future directions for digital twins. Tech- nical report, National Academies of Sciences, Engineering, and Medicine, 2023. DOI: https://doi.org/10.17226/26894

  62. [70]

    Gradient-enhanced physics- informed neural networks for forward and inverse pde problems

    Jeremy Yu, Lu Lu, Xuhui Meng, and George Em Karniadakis. Gradient-enhanced physics- informed neural networks for forward and inverse pde problems. Computer Methods in Applied Mechanics and Engineering , 393:114823, 2022

  63. [71]

    Distributed robotics approach to chemical plume tracing

    Dimitri Zarzhitsky, Diana F Spears, and William M Spears. Distributed robotics approach to chemical plume tracing. In 2005 IEEE/RSJ International Conference on Intelligent Robots and Systems , pages 4034–4039. IEEE, 2005. Appendix A: Neural Network Parameterization Details. We...

  64. [72]

    We provide a proof for Theorem 5.4 Proof

    Appendix D: Proof for convergence to sensor location. We provide a proof for Theorem 5.4 Proof. In (5.17), if we let X1 = X be an initial configuration of sensors with a minimal distance m0 < mc, and X2 be a configuration with a sensor located at x∗, then the inequality become...

  65. [73]

    Using the simple bounds arctan( θ) ≤ θ and tan(θ) ≤ 2θ, it is straightforward to show (10.17) θ∗ 1 = arctan r′ 1 − r′ tan(θ1) ≤ 2r′ 1 − r′ θ1

    lying above the x1-axis. Using the simple bounds arctan( θ) ≤ θ and tan(θ) ≤ 2θ, it is straightforward to show (10.17) θ∗ 1 = arctan r′ 1 − r′ tan(θ1) ≤ 2r′ 1 − r′ θ1. A similar calculation follows for calculating angles below the x1-axis. Then θ∗/θ ≤ 2r′(1 − r′), and so I2/I1...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.