Pith. sign in

REVIEW 3 major objections 5 minor 91 references

Neural Dynamic Modes: Computational Imaging of Dynamical Systems from Sparse Observations

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Sparse, noisy observations of an unknown dynamical system can recover its full space-time field and support short-term forecasting, if the field is modeled as a few neural spatial modes evolving under a linear operator.

desk verdict Honest proof-of-concept that nails the EHT-style case; the weather leg needs real metrics and a stronger baseline before the 'general tool' claim holds. read the letter →

arxiv 2507.03094 v1 pith:ZFCJXE3L submitted 2025-07-03 cs.CV astro-ph.IMcs.LGphysics.ao-ph

Ali SaraerToosi , Renbo Tu , Kamyar Azizzadenesheli , Aviad Levis This is my paper · ORCID
classification cs.CVastro-ph.IMcs.LGphysics.ao-ph
keywords neuralimplicitrepresentationsdynamicmodedecompositionsparseobservationsdataassimilationblackholeimagingweathernowcastingKoopmanoperatorspatio-temporalreconstruction
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 introduces NeuralDMD, a method for reconstructing continuous spatio-temporal fields from sparse, noisy measurements when the governing equations are unknown. The key move is to represent the field as a small sum of spatial modes, each evolving exponentially in time through growth, decay, or oscillation, with the modes, spectra, and initial coefficients parameterized by neural networks and fit directly to the observations. The authors argue this combines the expressiveness of neural implicit representations with the interpretability and predictive structure of Dynamic Mode Decomposition. They demonstrate on two realistic problems — wind-speed fields over North America from 10% or 1% of grid points, and plasma dynamics around Sgr A* from sparse interferometric visibilities — that NeuralDMD reconstructs the dynamics more accurately than established baselines and can extrapolate short-term forecasts. If correct, the method offers a general tool for imaging and forecasting unknown dynamics in geoscience, astronomy, and beyond.

What carries the argument

The central object is the modal decomposition identity $I_{\theta}(x,y,t) = w_0(x,y) b_0 + 2 \mathrm{Re}\sum_{j=1}^{r/2} w_j(x,y) e^{\Omega_j t} b_j$, a Dynamic Mode Decomposition style linear-operator representation made continuous by neural fields. Three MLPs carry the representation: a coordinate network for spatial modes $w_j(x,y)$, a spectral network for complex eigenvalues $\Omega_j$ (with decay constrained to $[-2,0]$ to forbid explosive growth), and an initial-state network for $b_j$. The loss is the squared error between the model and sparse observations — pixels in the weather task or interferometric visibilities in the black-hole task, where the sampling operator reflects Earth-rotation synthesis. The whole set of parameters is optimized jointly, so the initial state is learned from all frames rather than from the first frame as in classical DMD.

What would settle it

Take a sparse-observation sequence of a system with a known strongly nonlinear behavior at the observed scale, such as colliding solitons or a Duffing oscillator sampled at a handful of pixels, and fit NeuralDMD; if extrapolated frames quickly diverge from ground truth while a nonlinear baseline tracks them, the linear-operator premise is falsified. The paper itself shows the sharpening version: on unblurred GRMHD frames the reconstruction error rises and nonlinear structures are missed, so real EHT data with comparable small-scale structure would be a direct test.

Watch

Extended reading notes

Core claim

NeuralDMD posits that an observed dynamical field $I(x,y,t)$ can be written as $I_{\theta}(x,y,t) = w_0(x,y) b_0 + 2 \mathrm{Re}\sum_{j=1}^{r/2} w_j(x,y) e^{\Omega_j t} b_j$, where $w_j$ are spatial modes, $\Omega_j = \alpha_j + i\omega_j$ are complex temporal spectra, and $b_j$ are initial-state coefficients. A coordinate MLP outputs the modes at arbitrary positions, a spectral network outputs the eigenvalues, and an initial-state network outputs $b$; the whole set is optimized jointly against either sparse pixel samples or sparse Fourier visibilities. Once trained, the recovered linear operator gives a low-dimensional, interpretable description of the dynamics and can be advanced in time for forecasting. The paper reports that on sparse ERA5 wind data and on synthetic ngEHT observations of general-relativistic magnetohydrodynamic (GRMHD) black-hole simulations, this scheme outperforms a 3D-VAR data-assimilation baseline, a plain spatio-temporal neural representation, and StarWarps, and remains stable under noise up to 12% visibility error and as few as 20 Fourier components per frame.

Load-bearing premise

The load-bearing premise is that the observed dynamics are well approximated by a time-invariant linear operator over the window of observation and extrapolation; if the real process has strong nonlinearities or time-varying behavior at the observed scale, the reconstructed modes and forecasts will be biased.

Editorial extensions

If this is right

  • Recovered modes and spectra give a physical description of the dynamics, such as a zeroth-order mode capturing the mean frame and higher modes encoding spiral structures in the accretion flow, with a stability constraint ensuring modes decay or oscillate rather than grow explosively.
  • The model can extrapolate beyond the observation window; in the black-hole experiments NeuralDMD keeps reconstructed frames close to ground truth over a 3.2-hour horizon while a plain neural representation diverges.
  • Because the representation is continuous in coordinates, it evaluates at arbitrary positions and does not scale memory with the number of pixels, unlike grid-based DMD.
  • The same framework handles both pixel-domain and Fourier-domain measurements, so it can apply to weather station networks and radio interferometers without changing the architecture.
  • Performance improves as observational coverage increases: with the expanded ngEHT+ array the average L2 error drops from 0.024 to 0.009, so the method's error tracks the information content of the measured visibilities.

Reading between the lines

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

  • Editorial inference: the linear-operator premise implies the method will be most reliable on dynamics whose dominant features are already smoothed at the observed scale, as the paper's own blur experiment suggests; applying it to unblurred, strongly nonlinear flows would likely require more modes or a non-exponential temporal basis.
  • Editorial inference: a natural extension is to make the spectrum time-dependent or to couple NeuralDMD with a learned Koopman lifting, which could cover systems whose linearity only appears in a higher-dimensional embedding.
  • Editorial inference: because the initial-state coefficients are learned from the full sequence rather than the first frame, the method should be tested on streaming or partially observed data, where it may be able to re-estimate the coefficients online as new frames arrive.
  • Editorial inference: the visibility-domain loss is currently limited to Gaussian thermal noise; adding closure phases and closure amplitudes would be the decisive test for real Event Horizon Telescope data.
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 / 5 minor

Summary. The paper introduces NeuralDMD, a method that reconstructs a continuous spatio-temporal field I(x,y,t) from sparse noisy measurements by representing the field as a low-rank linear dynamical system: a coordinate-based MLP outputs spatial modes w_j(x,y), while learned spectra Omega_j and initial-state coefficients b_j govern temporal evolution (Eqs. 4-5). The framework is instantiated both for sparse pixel observations (weather) and for sparse Fourier visibilities (interferometry, Eq. 7). The authors validate it on two simulated scientific imaging tasks: assimilation of ERA5 10 m wind-speed fields from 10% or 1% of grid points with a simplified 3D-VAR baseline, and ngEHT/ngEHT+ interferometric imaging of GRMHD black-hole movies with neural-representation and StarWarps baselines. The claimed contribution is a model-free, interpretable, continuous imaging framework that outperforms established baselines in both domains and supports short-term forecasting.

Significance. If the central claims are established, this would be a useful contribution: it combines the interpretability of DMD with the flexibility of neural implicit fields, handles extremely sparse and noisy observations, and produces continuous reconstructions with predictive extrapolation. The paper is strengthened by clearly stated optimization objectives (Eqs. 5 and 7), realistic synthetic observation generation with eht-imaging and ngehtsim, an extensive robustness study in the supplement, an analytic orbiting-hot-spot validation, and an unusually candid discussion of limitations (Sec. 5, Supp. S3). The main weakness is that the strongest advertised claim---outperforming established baselines on weather and black-hole imaging---is not yet supported: the weather experiment lacks quantitative evaluation, and the load-bearing linearity assumption is validated only for blurred black-hole images, not for the forced nonlinear weather system. The framework is a promising proof-of-concept, but the general claim needs to be either substantiated with additional experiments or substantially narrowed.

major comments (3)
  1. [Sec. 4.1, Fig. 3] The weather experiment presents no quantitative accuracy comparison. The text asserts that 3D-VAR degrades rapidly while NeuralDMD maintains high reconstruction quality, but Figure 3 is visual only: no RMSE, MAE, SSIM, or similar metrics are reported for the 10% or 1% coverage cases, and no numerical evaluation is given for the April 8 held-out forecast. Because the abstract explicitly claims that NeuralDMD outperforms established baselines in the weather domain, this evidence is insufficient. Please report held-out metrics with error bars over multiple random station subsets, and compare against a stronger and more standard baseline (e.g., operational 3D-VAR, kriging interpolation, or a recent learning-based data-assimilation method) rather than the self-described 'minimalist' implementation that is 'not on par with a fully operational 3D-VAR'.
  2. [Sec. 2.1, Eq. (1); Supp. Fig. S2] The method is committed to a time-invariant linear operator with no external forcing over both the assimilation and forecast windows. This assumption is explicitly acknowledged in Sec. 2.1, but it is validated only for the black-hole imaging case, where Supp. Fig. S2 shows that performance improves with increasing blur and that unblurred images contain nonlinear structures NeuralDMD misses. For the ERA5 wind-field experiment, the dynamics are externally forced by the atmosphere, and no analogous linearity diagnostic is provided. As a result, the weather experiment does not test the method's core assumption. Please add a linearity diagnostic for the weather data (e.g., the fraction of variance explained by the learned r modes on dense training data, or the fit residual of the linear model across the assimilation window) and, ideally, a nonlinear baseline (such as a ConvLSTM or neural ODE) to determine whether the linear inductive bias is responsible for the reported behavior.
  3. [Abstract; Sec. 3.1] The term 'model-free' is misleading as stated. Equations (1) and (5) impose a specific time-invariant, finite-rank linear dynamical model, and the method requires choosing several hyperparameters (number of modes r, positional-encoding degree L, spectral bounds for Omega, the Omega-t scaling constant, and the mode-pruning threshold). DMD is equation-free, but it is not model-free in the usual sense. Please qualify the claim (e.g., 'governing-equation-free' or 'not requiring known physics') or justify why 'model-free' is appropriate despite the strong linear inductive bias.
minor comments (5)
  1. [Abstract; Sec. 4.2] The abstract says NeuralDMD is used for 'recovering the evolution of plasma near the Galactic-center black hole,' but all black-hole results are obtained from simulated GRMHD movies with synthetic ngEHT observations, not from real EHT data. Please make clear in the abstract that this validation is on simulated observations.
  2. [Sec. 3.1, Eq. (5)] The notation reuses Theta for both the optimal parameters and the optimization variable; Eq. (4) uses Theta^star, so Eq. (5) should also write Theta^star = arg min_Theta for consistency and clarity.
  3. [Sec. 4.1, Fig. 3] The caption of Figure 3 does not define the color scale or normalization of the wind-speed maps, and the sparse sample locations are not shown; please add this information so the visual comparison is interpretable.
  4. [Supp. S6] The exact values of the number of modes r, positional-encoding degree L, batch size, and any experiment-specific hyperparameters are not listed; please provide a table with these settings for the weather, face-on GRMHD, edge-on GRMHD, hot-spot, and robustness experiments to support reproducibility.
  5. [Throughout] There are several typographical and formatting errors, including 'to to a broad range' in Sec. 1, 'with a continuousparameterization' in Sec. 3.1, and 'anugular velocity' in Supp. S5; in addition, References [46] and [47] are duplicate entries of the same book by Kutz et al.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: NeuralDMD fits a linear-modal model to training observations and extrapolates to held-out frames; no prediction reduces to its inputs by construction.

full rationale

I walked the derivation chain from Eq. (1) through Eqs. (5) and (7). The model class is a linear combination of spatial modes with exponential time dependence; the parameters are optimized against sparse pixel or visibility observations, and forecasting is obtained by advancing the fitted linear dynamics beyond the training window. This is a fit-then-extrapolate procedure, not a circular derivation: the reported prediction targets (April 8 weather; the grey-band EHT frames in Fig. 5) are held out from the loss. No fitted parameter is renamed as a prediction, and no quantity used in training is claimed as an independent forecast. Self-citations (e.g., refs. [48]–[50], [89]) appear as background or methodological motivation and are not used as load-bearing evidence for the central claims. No uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via citation: the linear-modal decomposition is stated directly in Sec. 2.1 and optimized explicitly in Sec. 3. The limitations are candidly acknowledged in Sec. 5 and Supp. S3, including the reliance on linearity at EHT-blurred resolutions and the omission of phase errors; these affect external validity and correctness risk but do not make the derivation circular. In both experiments the error metrics are computed against genuinely held-out frames, so the central empirical claim is self-contained rather than forced by construction.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The method introduces no new physical entities. Its free parameters are the mode count, the positional encoding degree, the spectral bounds, a time-scaling constant, and a pruning threshold. The central assumptions are linearity of the observed dynamics, stationarity, the Fourier visibility model without phase errors, and smoothness of the neural modes. These are all stated in the paper, though the linearity assumption is the most fragile.

free parameters (5)
  • number of modes r = not stated (set a priori; higher-order decaying modes discarded post hoc, Supp. S9)
    The model capacity is fixed by choosing r; the paper says 'We a priori set the number of modes and discard high-order fast decaying modes in a post-processing step' (Supp. S9). Results depend on this choice.
  • positional encoding degree L = L = 4
    Chosen by hand in Sec. 3.1 as a regularization trade-off for smooth fields; affects the spatial frequency content of the modes.
  • spectral bounds for Omega = alpha in [-2,0], omega in [0,160]
    Imposed in the spectral network (Sec. 3.1) to enforce decay and bounded oscillation; ad hoc physical constraints that shape the optimization.
  • Omega-t scaling constant = order 10^2
    Supp. S6: 'We multiply Omega t with a constant value of order 10^2 to bring the Omega values to a physical range...'.
  • mode pruning threshold = alpha < -0.05 discarded
    Supp. S8/S9: high-order modes with alpha < -0.05 are treated as not impacting reconstruction and are discarded in post-processing; this choice affects interpretability but not the fit.
assumptions (6)
  • domain assumption Observed dynamics are approximately governed by a time-invariant linear operator over the observation and extrapolation window.
    Stated in Sec. 2.1 and used to justify the DMD decomposition (Eq. 1). The paper supports it for EHT via Fig. S2 (error decreases as images are blurred), but it is load-bearing: strong nonlinearities would break the model.
  • domain assumption No external inputs and no time-varying dynamics (Koopman/linear approximation is sufficient).
    Sec. 2.1 assumes '(i) no external inputs affect the system... and (ii) the system's dynamics are approximately constant over time'.
  • standard math Real-valued field implies complex-conjugate mode pairs and a static zeroth mode (Omega_0=0).
    Sec. 3.1: 'since we are modeling real-valued dynamical systems, where modes appear as complex conjugate pairs... we set Omega 0 = 0.' This is mathematically standard for real signals.
  • domain assumption Interferometric visibility is a noiseless linear Fourier transform of the sky brightness plus Gaussian thermal noise; phase errors are negligible.
    Sec. 4.2.2: 'we consider only Gaussian thermal noise and amplitude errors and neglect phase errors, a significant source of uncertainty'. The forward model (Eq. 7) relies on this.
  • domain assumption The neural mode fields w_j(x,y) are smooth enough to be captured by a low-frequency MLP with L=4 positional encoding.
    Sec. 3.1: 'we rely on a low positional encoding degree, L=4, as a way to regularize the recovery of smoothly varying fields'. This limits the spatial complexity of recoverable modes.
  • domain assumption For weather, a fixed uniform random subset of grid points approximates a real station network.
    Sec. 4.1: 'sample only a small, fixed subset of grid points mimicking sparse station or satellite observations by uniform random sampling.' The paper acknowledges that real networks are uneven with dropouts (Discussion).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Dynamic Modes: Computational Imaging of Dynamical Systems from Sparse Observations." pith.science (2026). https://pith.science/paper/ZFCJXE3L

@misc{pith2026250703094,
  author       = {Pith},
  title        = {Pith review of: Neural Dynamic Modes: Computational Imaging of Dynamical Systems from Sparse Observations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZFCJXE3L}},
  note         = {Machine review of arXiv:2507.03094}
}
read the original abstract

Dynamical systems are ubiquitous within science and engineering, from turbulent flow across aircraft wings to structural variability of proteins. Although some systems are well understood and simulated, scientific imaging often confronts never-before-seen dynamics observed through indirect, noisy, and highly sparse measurements. We present NeuralDMD, a model-free framework that combines neural implicit representations with Dynamic Mode Decomposition (DMD) to reconstruct continuous spatio-temporal dynamics from such measurements. The expressiveness of neural representations enables capturing complex spatial structures, while the linear dynamical modes of DMD introduce an inductive bias that guides training and supports stable, low-dimensional representations and forecasting. We validate NeuralDMD on two real-world problems: reconstructing near-surface wind-speed fields over North America from sparse station observations, and recovering the evolution of plasma near the Galactic-center black hole, Sgr A*. In both cases, NeuralDMD outperforms established baselines, demonstrating its potential as a general tool for imaging dynamical systems across geoscience, astronomy, and beyond.

Figures

Figures reproduced from arXiv: 2507.03094 by the authors.

Figure 1
Figure 1. We introduce a general and interpretable framework for [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Block diagram of the NEURALDMD architecture. The model consists of three interacting MLPs, together define the spatio￾temporal dynamics: (1) Θw, which maps positionally encoded spatial coordinates (x, y) to mode values wj (x, y); (2) ΘΩ, which outputs the complex spectral components – growth/decay rate α and oscillations ω – from a learnable latent vector; and (3) Θb, which maps another learnable latent vector to th… view at source ↗
Figure 3
Figure 3. A comparison of NEURALDMD with the data assimilation baseline 3D-VAR. The results are shown on sparsely sampled ERA5 datasets of wind velocity magnitudes. [First row]: Results for 5 hours from the initial frame [Second row] 60 hours from the initial frame with 10% of data given as sparse observations, [Third Row] 60 hours from the initial frame with 1% of the data given. [Fourth row] demonstrates NEURALDMD ’s abilit… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Reconstruction of a GRMHD sequence from ngEHT -synthetic interferometric data. Rows (top to bottom): ground truth, NEURALDMD, neural representation, and StarWarps. The left two frames are data-constrained reconstructions (“data-fitting”), the right two are model extrap…
Figure 5
Figure 5. Figure 5: Reconstruction error for NEURALDMD (orange) and the neural representation (blue). Solid curves give total error, dashed curves the dynamics-only error (mean image removed), and reconstructed frames are shown at selected points. Data-fitting relies on observations from …
Figure 6
Figure 6. Figure 6: This figure highlights the robustness of N [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

91 extracted references · 76 canonical work pages

  1. [1]

    Nathan Kutz

    Travis Askham and J. Nathan Kutz. Variable projection methods for an optimized dynamic mode decomposition,

  2. [2]

    Broderick, Ra ´ul Carballo-Rubio, Vitor Cardoso, and et al

    Dimitry Ayzenberg, Lindy Blackburn, Richard Brito, Silke Britzen, Avery E. Broderick, Ra ´ul Carballo-Rubio, Vitor Cardoso, and et al. Andrew Chael. Fundamental physics op- portunities with the next-generation event horizon telescope,

  3. [3]

    Neural operators for accelerating scientific simulations and design

    Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Miguel Liu-Schiaffini, Jean Kossaifi, and Anima Anandku- mar. Neural operators for accelerating scientific simulations and design. Nature Reviews Physics, 6(5):320–328, 2024. 2

    2024

  4. [4]

    McVicar, Lorenzo Minola, Juan I

    Cesar Azorin-Molina, Jes ´us As´ın, Tim R. McVicar, Lorenzo Minola, Juan I. L ´opez-Moreno, Sergio M. Vicente-Serrano, and Deliang Chen. Evaluating anemometer drift: A statisti- cal approach to correct biases in wind speed measurement. Atmospheric Research, 203:175–188, 2018. 8

    2018

  5. [5]

    R. N. Bannister. A review of operational methods of varia- tional and ensemble-variational data assimilation. Quarterly Journal of the Royal Meteorological Society, 143(703):607– 633, 2017. 2

    2017

  6. [6]

    Bouman, Michael D

    Katherine L. Bouman, Michael D. Johnson, Adrian V . Dalca, Andrew A. Chael, Freek Roelofs, Sheperd S. Doeleman, and William T. Freeman. Reconstructing video of time- varying sources from radio interferometric measurements. IEEE Transactions on Computational Imaging , 4(4):512– 527, 2018. 1, 2

    2018

  7. [7]

    Bouman, Michael D

    Katherine L. Bouman, Michael D. Johnson, Adrian V . Dalca, Andrew A. Chael, Freek Roelofs, Sheperd S. Doeleman, and William T. Freeman. Reconstructing video from interfero- metric measurements of time-varying sources, 2018. 2, 7

    2018

  8. [8]

    Broderick and Abraham Loeb

    Avery E. Broderick and Abraham Loeb. Imaging bright- spots in the accretion flow near the black hole horizon of sgr a. Monthly Notices of the Royal Astronomical Society, 2005. 7

    2005

Show all 91 references
  1. [9]

    Hexplane: A fast representa- tion for dynamic scenes, 2023

    Ang Cao and Justin Johnson. Hexplane: A fast representa- tion for dynamic scenes, 2023. 2

    2023

  2. [10]

    Neural space–time model for dy- namic multi-shot imaging

    Ruiming Cao, Nikita Divekar, James Nu ˜nez, Srigokul Upad- hyayula, and Laura Waller. Neural space–time model for dy- namic multi-shot imaging. Nature Methods, 21:2336–2341,

  3. [11]

    eht-imaging, 2022

    Andrew Chael. eht-imaging, 2022. 5

    2022

  4. [12]

    The role of electron heating physics in images and variability of the galactic centre black hole sagittarius a*

    Andrew Chael, Michael Rowan, Ramesh Narayan, Michael Johnson, and Lorenzo Sironi. The role of electron heating physics in images and variability of the galactic centre black hole sagittarius a*. Monthly Notices of the Royal Astronom- ical Society, 478(4):5209–5229, 2018

    2018

  5. [13]

    Chael, Michael D

    Andrew A. Chael, Michael D. Johnson, Ramesh Narayan, Sheperd S. Doeleman, John F. C. Wardle, and Katherine L. Bouman. High-resolution linear polarimetric imaging for the event horizon telescope. The Astrophysical Journal, 829(1): 11, 2016. 5

    2016

  6. [14]

    Chael, Michael D

    Andrew A. Chael, Michael D. Johnson, Katherine L. Bouman, Lindy L. Blackburn, Kazunori Akiyama, and Ramesh Narayan. Interferometric imaging directly with clo- sure phases and closure amplitudes. Astrophys. J., 857:23,

  7. [15]

    Layered dynamic tex- tures

    Antoni Chan and Nuno Vasconcelos. Layered dynamic tex- tures. In Advances in Neural Information Processing Sys- tems. MIT Press, 2005. 2

    2005

  8. [16]

    Variational layered dy- namic textures

    Antoni Chan and Nuno Vasconcelos. Variational layered dy- namic textures. pages 1062–1069, 2009. 2

    2009

  9. [17]

    Chiaramonte, Peter Yichen Chen, and Eitan Grinspun

    Yue Chang, Otman Benchekroun, Maurizio M. Chiaramonte, Peter Yichen Chen, and Eitan Grinspun. Neural representa- tion of shape-dependent laplacian eigenfunctions, 2024. 3

    2024

  10. [18]

    Automated dis- covery of fundamental variables hidden in experimental data

    Boyuan Chen, Kuang Huang, Sunand Raghupathi, Ishaan Chandratreya, Qiang Du, and Hod Lipson. Automated dis- covery of fundamental variables hidden in experimental data. Nature Computational Science, 2(7):433–442, 2022. 2

    2022

  11. [19]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations,

  12. [20]

    Ef- ficient deep data assimilation with sparse observations and time-varying sensors

    Sibo Cheng, Che Liu, Yike Guo, and Rossella Arcucci. Ef- ficient deep data assimilation with sparse observations and time-varying sensors. Journal of Computational Physics , 496:112581, 2024. 1

    2024

  13. [21]

    Conroy, George N

    Nicholas S. Conroy, George N. Wong, Dominic W. Pesce, Daniel C. M. Palumbo, and Avery E. Broderick. Rota- tion in event horizon telescope movies. arXiv preprint arXiv:2304.03826, 2023. 4

    2023 arXiv

  14. [22]

    Courtier, E

    P. Courtier, E. Andersson, W. A. Heckley, J. Pailleux, D. Vasiljevic, M. Hamrud, A. Hollingsworth, F. Rabier, and M. Fisher. The ecmwf implementation of three-dimensional variational assimilation (3d-var): Part i: formulation. Quar- terly Journal of the Royal Meteorological So...

    1998

  15. [23]

    Dynamic mode decomposition and koopman theory, 2022

    Sourya Dey. Dynamic mode decomposition and koopman theory, 2022. 3

    2022

  16. [24]

    Vedant Dhruv, Ben Prather, George Wong, and Charles F. Gammie. A survey of general relativistic magnetohydrody- namic models for black hole accretion systems, 2025. 5

    2025

  17. [25]

    Dynamic textures

    Gianfranco Doretto, Alessandro Chiuso, Yingnian Wu, and Stefano Soatto. Dynamic textures. International Journal of Computer Vision, 51:91–109, 2003. 2

    2003

  18. [26]

    First M87 event horizon telescope results

    Event Horizon Telescope Collaboration. First M87 event horizon telescope results. III. Data processing and calibra- tion. Astrophys. J. Lett., 875:L4, 2019. 8

    2019

  19. [27]

    First M87 Event Horizon Telescope Results

    Event Horizon Telescope Collaboration et al. First M87 Event Horizon Telescope Results. II. Array and Instrumenta- tion. The Astrophysical Journal Letters, 875(L2), 2019. 2

    2019

  20. [28]

    First Sagittar- ius A* Event Horizon Telescope Results

    Event Horizon Telescope Collaboration et al. First Sagittar- ius A* Event Horizon Telescope Results. III: Imaging of the Galactic Center Supermassive Black Hole. The Astrophysi- cal Journal Letters, 930(L14), 2023. 2

    2023

  21. [29]

    Multiwavelength light curves of two remarkable sagittarius a* flares

    GG Fazio, JL Hora, G Witzel, SP Willner, MLN Ashby, F Baganoff, E Becklin, S Carey, D Haggard, C Gammie, et al. Multiwavelength light curves of two remarkable sagittarius a* flares. The Astrophysical Journal, 864(1):58, 2018. 7

    2018

  22. [30]

    Fish, Michael D

    Vincent L. Fish, Michael D. Johnson, Sheperd S. Doeleman, Avery E. Broderick, Dimitrios Psaltis, and et al. Persis- tent asymmetric structure of sagittarius a* on event horizon scales. Astrophys. J., 820:90, 2016. 8

    2016

  23. [31]

    K-planes: Explicit radiance fields in space, time, and appearance, 2023

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance, 2023. 2

    2023

  24. [32]

    Gammie, Jonathan C

    Charles F. Gammie, Jonathan C. McKinney, and Gabor Toth. Harm: A numerical scheme for general relativistic mag- netohydrodynamics. The Astrophysical Journal , 589(1): 444–457, 2003. 5

    2003

  25. [33]

    Gaussianflow: Splatting gaussian dynamics for 4d content creation, 2024

    Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wen- chao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splatting gaussian dynamics for 4d content creation, 2024. 2

    2024

  26. [34]

    Learning physics from video: Unsuper- vised physical parameter estimation for continuous dynami- cal systems, 2025

    Alejandro Casta ˜neda Garcia, Jan van Gemert, Daan Brinks, and Nergis T¨omen. Learning physics from video: Unsuper- vised physical parameter estimation for continuous dynami- cal systems, 2025. 2

    2025

  27. [35]

    Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, Andr´as Hor ´anyi, Joaqu ´ın Mu ˜noz-Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, Adrian Sim- mons, Cornel Soci, Saleh Abdalla, Xavier Abellan, Gian- paolo Balsamo, Peter Bechtold, Gionata Biav...

    1999

  28. [36]

    Hersbach, B

    H. Hersbach, B. Bell, P. Berrisford, G. Biavati, A. Hor´anyi, J. Mu˜noz Sabater, J. Nicolas, C. Peubey, R. Radu, I. Rozum, D. Schepers, A. Simmons, C. Soci, D. Dee, and J.-N. Th ´epaut. ERA5 hourly data on single levels from 1940 to present. Copernicus Climate Change Service (...

    1940

  29. [37]

    Neural implicit representations for physical param- eter inference from a single video

    Florian Hofherr, Lukas Koestler, Florian Bernard, and Daniel Cremers. Neural implicit representations for physical param- eter inference from a single video. In 2023 IEEE/CVF Win- ter Conference on Applications of Computer Vision (WACV), pages 2092–2102, 2023. 2

    2023

  30. [38]

    Learning to Decompose and Disentan- gle Representations for Video Prediction

    Jun-Ting Hsieh, Bingbin Liu, De-An Huang, Li Fei-Fei, and Juan Carlos Niebles. Learning to Decompose and Disentan- gle Representations for Video Prediction. arXiv e-prints, art. arXiv:1806.04166, 2018. 2

    2018 arXiv

  31. [39]

    Issaoun, M

    S. Issaoun, M. D. Johnson, L. Blackburn, A. Broderick, P. Tiede, M. Wielgus, S. S. Doeleman, H. Falcke, K. Akiyama, G. C. Bower, C. D. Brinkerink, A. Chael, I. Cho, J. L. G´omez, A. Hern ´andez-G´omez, D. Hughes, M. Kino, T. P. Krichbaum, E. Liuzzo, L. Loinard, S. Markoff, D. ...

    2021

  32. [40]

    Physics-as-inverse-graphics: Unsupervised physical param- eter estimation from video, 2020

    Miguel Jaques, Michael Burke, and Timothy Hospedales. Physics-as-inverse-graphics: Unsupervised physical param- eter estimation from video, 2020. 2

    2020

  33. [41]

    The black hole explorer: motivation and vision

    Michael Johnson, Kazunori Akiyama, Rebecca Baturin, Bryan Bilyeu, and Blackburn et al. The black hole explorer: motivation and vision. InSpace Telescopes and Instrumenta- tion 2024: Optical, Infrared, and Millimeter Wave, page 90. SPIE, 2024. 2

    2024

  34. [42]

    Johnson, Ramesh Narayan, Dimitrios Psaltis, Lindy Blackburn, Yuri Y

    Michael D. Johnson, Ramesh Narayan, Dimitrios Psaltis, Lindy Blackburn, Yuri Y . Kovalev, Carl R. Gwinn, Guang- Yao Zhao, Geoffrey C. Bower, James M. Moran, Motoki Kino, Michael Kramer, Kazunori Akiyama, Jason Dexter, Avery E. Broderick, and Lorenzo Sironi. The scattering and ...

    2018

  35. [43]

    Re- solving horizon-scale dynamics of sagittarius a*, 2023

    Jakob Knollm ¨uller, Philipp Arras, and Torsten Enßlin. Re- solving horizon-scale dynamics of sagittarius a*, 2023. 1, 2

    2023

  36. [44]

    Neural operator: learning maps between function spaces with applications to pdes

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Aziz- zadenesheli, Kaushik Bhattacharya, Andrew Stuart, and An- ima Anandkumar. Neural operator: learning maps between function spaces with applications to pdes. J. Mach. Learn. Res., 24(1), 2023. 2

    2023

  37. [45]

    Nathan Kutz, Steven L

    J. Nathan Kutz, Steven L. Brunton, Bingni W. Brunton, and Joshua L. Proctor. Dynamic Mode Decomposition: Data- Driven Modeling of Complex Systems. SIAM, 2016. 2, 3

    2016

  38. [46]

    Nathan Kutz, Steven L

    J. Nathan Kutz, Steven L. Brunton, Bingni W. Brunton, and Joshua L. Proctor. Dynamic Mode Decomposition: Data- Driven Modeling of Complex Systems. Society for Industrial and Applied Mathematics, Philadelphia, PA, 2016. 3

    2016

  39. [47]

    Nathan Kutz, Steven L

    J. Nathan Kutz, Steven L. Brunton, Bingni W. Brunton, and Joshua L. Proctor. Dynamic Mode Decomposition. Society for Industrial and Applied Mathematics, Philadelphia, PA,

  40. [48]

    Tropp, Charles F

    Aviad Levis, Daeyoung Lee, Joel A. Tropp, Charles F. Gam- mie, and Katherine L. Bouman. Inference of black hole fluid-dynamics from sparse interferometric measurements. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2340–2349, 2021. 2

    2021

  41. [49]

    Gravitationally lensed black hole emission tomography

    Aviad Levis, Pratul P Srinivasan, Andrew A Chael, Ren Ng, and Katherine L Bouman. Gravitationally lensed black hole emission tomography. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , pages 19841–19850, 2022. 1

    2022

  42. [50]

    Orbital polarimetric tomography of a flare near the sagittarius a* supermassive black hole

    Aviad Levis, Andrew A Chael, Katherine L Bouman, Ma- ciek Wielgus, and Pratul P Srinivasan. Orbital polarimetric tomography of a flare near the sagittarius a* supermassive black hole. Nature Astronomy, 8(6):765–773, 2024. 1, 7

    2024

  43. [51]

    Dynibar: Neural dynamic image-based rendering, 2023

    Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering, 2023. 2

    2023

  44. [52]

    A. C. Lorenc. Analysis methods for numerical weather pre- diction. Quarterly Journal of the Royal Meteorological So- ciety, 112(474):1177–1194, 1986. 5

    1986

  45. [53]

    Donoho, Juan M

    Michael Lustig, David L. Donoho, Juan M. Santos, and John M. Pauly. Compressed sensing mri. IEEE Signal Pro- cessing Magazine, 25(2):72–82, 2008. 2

    2008

  46. [54]

    Gen- erative data assimilation of sparse weather station observa- tions at kilometer scales, 2025

    Peter Manshausen, Yair Cohen, Peter Harrington, Jaideep Pathak, Mike Pritchard, Piyush Garg, Morteza Mardani, Karthik Kashinath, Simon Byrne, and Noah Brenowitz. Gen- erative data assimilation of sparse weather station observa- tions at kilometer scales, 2025. 1

    2025

  47. [55]

    Lauer, and Fer- yal ¨Ozel

    Lia Medeiros, Dimitrios Psaltis, Tod R. Lauer, and Fer- yal ¨Ozel. Principal-component interferometric modeling (primo), an algorithm for eht data. i. reconstructing images from simulated eht observations. The Astrophysical Journal, 943(2):144, 2023. 2

    2023

  48. [56]

    Analysis of fluid flows via spectral properties of the koopman operator

    Igor Mezi ´c. Analysis of fluid flows via spectral properties of the koopman operator. Annual Review of Fluid Mechanics , 45:357–378, 2013. 2

    2013

  49. [57]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis, 2020. 2, 3

    2020

  50. [58]

    Monika Moscibrodzka and Charles F. Gammie. ipole - semi- analytic scheme for relativistic polarized radiative transport,

  51. [59]

    M ¨uller and A

    H. M ¨uller and A. P. Lobanov. Dog-hit: A novel vlbi mul- tiscale imaging approach. Astronomy &amp; Astrophysics, 666:A137, 2022. 2

    2022

  52. [60]

    Instant neural graphics primitives with a multires- olution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multires- olution hash encoding. ACM Transactions on Graphics, 41 (4):1–15, 2022. 3

    2022

  53. [61]

    A chandra/hetgs census of x-ray variability from sgr a* during 2012

    J Neilsen, MA Nowak, C Gammie, J Dexter, S Markoff, D Haggard, S Nayakshin, QD Wang, N Grosso, D Porquet, et al. A chandra/hetgs census of x-ray variability from sgr a* during 2012. The Astrophysical Journal, 774(1):42, 2013. 7

    2012

  54. [62]

    Brunton, and J

    Shaowu Pan, Steven L. Brunton, and J. Nathan Kutz. Neu- ral Implicit Flow: a mesh-agnostic dimensionality reduc- tion paradigm of spatio-temporal data. arXiv e-prints, art. arXiv:2204.03216, 2022. 2

    2022 arXiv

  55. [63]

    Pesce, Lindy Blackburn, Ryan Chaves, Shep- erd S

    Dominic W. Pesce, Lindy Blackburn, Ryan Chaves, Shep- erd S. Doeleman, Mark Freeman, Sara Issaoun, Michael D. Johnson, Greg Lindahl, Iniyan Natarajan, Scott N. Paine, Daniel C. M. Palumbo, Freek Roelofs, and Paul Tiede. nge- htsim, 2023. 5

    2023

  56. [64]

    Previtali, N

    D. Previtali, N. Valceschini, M. Mazzoleni, and F. Previdi. Identification of dynamic textures using dynamic mode de- composition. IFAC-PapersOnLine, 53(2):2423–2428, 2020. 21st IFAC World Congress. 2

    2020

  57. [65]

    Proctor, Steven L

    Joshua L. Proctor, Steven L. Brunton, and J. Nathan Kutz. Dynamic mode decomposition with control. SIAM Journal on Applied Dynamical Systems, 15(1):142–161, 2016. 3

    2016

  58. [66]

    Proctor, Steven L

    Joshua L. Proctor, Steven L. Brunton, and J. Nathan Kutz. Generalizing koopman theory to allow for inputs and control. SIAM Journal on Applied Dynamical Systems , 17(1):909– 930, 2018. 3

    2018

  59. [67]

    D-nerf: Neural radiance fields for dynamic scenes, 2020

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes, 2020. 2

    2020

  60. [68]

    Collection of historical weather data: Issues with missing values

    Fadoua Rafii and Tahar Kechadi. Collection of historical weather data: Issues with missing values. In Proceedings of the 4th International Conference on Smart City Applications (SCA ’19), Casablanca, Morocco, 2019. ACM. 5, 8

    2019

  61. [69]

    L4gm: Large 4d gaussian reconstruction model, 2024

    Jiawei Ren, Kevin Xie, Ashkan Mirzaei, Hanxue Liang, Xi- aohui Zeng, Karsten Kreis, Ziwei Liu, Antonio Torralba, Sanja Fidler, Seung Wook Kim, and Huan Ling. L4gm: Large 4d gaussian reconstruction model, 2024. 2

    2024

  62. [70]

    Ripperda, M

    B. Ripperda, M. Liska, K. Chatterjee, G. Musoke, A. A. Philippov, S. B. Markoff, A. Tchekhovskoy, and Z. Younsi. Black hole flares: Ejection of accreted magnetic flux through 3d plasmoid-mediated reconnection. The Astrophysical Journal Letters, 924(2):L32, 2022. 2

    2022

  63. [71]

    Risser and Michael F

    Mark D. Risser and Michael F. Wehner. The effect of geographic sampling on evaluation of extreme precipita- tion in high resolution climate models. arXiv preprint arXiv:1911.05103, 2019. 5, 8

    1911 arXiv

  64. [72]

    Rowley, Igor Mezi ´c, Saman Bagheri, Philipp Schlatter, and Dan S

    Clarence W. Rowley, Igor Mezi ´c, Saman Bagheri, Philipp Schlatter, and Dan S. Henningson. Spectral analysis of non- linear flows. Journal of Fluid Mechanics , 641:115–127,

  65. [73]

    Nathan Kutz

    Diya Sashidhar and J. Nathan Kutz. Bagging, optimized dynamic mode decomposition for robust, stable forecasting with spatial and temporal uncertainty quantification. Philo- sophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 380(2229)...

    2022

  66. [74]

    Peter J. Schmid. Dynamic mode decomposition of numerical and experimental data. Journal of Fluid Mechanics, 656:5– 28, 2010. 2

    2010

  67. [75]

    PETER J. SCHMID. Dynamic mode decomposition of nu- merical and experimental data. Journal of Fluid Mechanics, 656:5–28, 2010. 3

    2010

  68. [76]

    Convolutional lstm network: a machine learning approach for precipitation now- casting

    Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo. Convolutional lstm network: a machine learning approach for precipitation now- casting. In Proceedings of the 29th International Conference on Neural Information Processing Systems - Volume ...

    2015

  69. [77]

    Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ra- mamoorthi, Jonathan T

    Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ra- mamoorthi, Jonathan T. Barron, and Ren Ng. Fourier fea- tures let networks learn high frequency functions in low di- mensional domains, 2020. 4

    2020

  70. [78]

    First sagittar- ius a* event horizon telescope results

    The Event Horizon Telescope Collaboration. First sagittar- ius a* event horizon telescope results. i. the shadow of the supermassive black hole in the center of the milky way. The Astrophysical Journal Letters, 930(2):L12, 2022. 2

    2022

  71. [79]

    Richard Thompson, James M

    A. Richard Thompson, James M. Moran, and George W. Swenson. Very-Long-Baseline Interferometry, pages 391–

  72. [80]

    Broderick, Roman Gold, Mansour Karami, and Jorge A

    Paul Tiede, Hung-Yi Pu, Avery E. Broderick, Roman Gold, Mansour Karami, and Jorge A. Preciado-L ´opez. Spacetime tomography using the event horizon telescope. The Astro- physical Journal, 892(2):132, 2020. 7

    2020

  73. [81]

    Tu, Clarence W

    Jonathan H. Tu, Clarence W. Rowley, Dirk M. Luchtenburg, Steven L. Brunton, and J. Nathan Kutz. On dynamic mode decomposition: Theory and applications. Journal of Compu- tational Dynamics, 1(2):391–421, 2014. 2

    2014

  74. [82]

    Correspondence-free ma- terial reconstruction using sparse surface constraints

    Sebastian Weiss, Robert Maier, Daniel Cremers, R ¨udiger Westermann, and Nils Thuerey. Correspondence-free ma- terial reconstruction using sparse surface constraints. In 2020 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 4685–4694, 2020. 2

    2020

  75. [83]

    Keating, Venkatessh Ramakrishnan, Paul Tiede, Ed Fomalont, Sara Issaoun, Joey Neilsen, Michael A

    Maciek Wielgus, Nicola Marchili, Iv ´an Mart ´ı-Vidal, Gar- rett K. Keating, Venkatessh Ramakrishnan, Paul Tiede, Ed Fomalont, Sara Issaoun, Joey Neilsen, Michael A. Nowak, et al. Millimeter Light Curves of Sagittarius A* Observed during the 2017 Event Horizon Telescope Campai...

    2017

  76. [84]

    4d gaussian splatting for real-time dynamic scene rendering,

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering,

  77. [85]

    Yang, Matthias Kretzler, Sebastian Sudarski, Vikas Gulani, and Nicole Seiberlich

    Angela C.Y . Yang, Matthias Kretzler, Sebastian Sudarski, Vikas Gulani, and Nicole Seiberlich. Sparse reconstruction techniques in mri: Methods, applications, and challenges to clinical adoption. Investigative Radiology, 51(6):349–364,

  78. [86]

    Chenyu Zhang, Daniil Cherniavskii, Andrii Zadaianchuk, Antonios Tragoudaras, Antonios V ozikis, Thijmen Nijdam, Derck W. E. Prinzhorn, Mark Bodracska, Nicu Sebe, and Efstratios Gavves. Morpheus: Benchmarking physical rea- soning of video generative models with real physical ex...

    2025

  79. [87]

    Rowley, Eric A

    Hao Zhang, Clarence W. Rowley, Eric A. Deem, and Louis N. Cattafesta. Online dynamic mode decomposition for time-varying systems. SIAM Journal on Applied Dynam- ical Systems, 18(3):1586–1609, 2019. 3

    2019

  80. [88]

    Neural signed distance function inference through splatting 3d gaus- sians pulled on zero-level set, 2024

    Wenyuan Zhang, Yu-Shen Liu, and Zhizhong Han. Neural signed distance function inference through splatting 3d gaus- sians pulled on zero-level set, 2024. 2

    2024

  81. [89]

    Revealing the 3d cosmic web through gravitationally constrained neural fields

    Brandon Zhao, Aviad Levis, Liam Connor, Pratul P Srini- vasan, and Katherine Bouman. Revealing the 3d cosmic web through gravitationally constrained neural fields. In The Thirteenth International Conference on Learning Represen- tations. 3

  82. [90]

    hot-spots

    Ellen D. Zhong, Tristan Bepler, Bonnie Berger, and Joseph H. Davis. CryoDRGN: reconstruction of heteroge- neous cryo-EM structures using neural networks. Nature Methods, 18:176–185, 2021. 1 Neural Dynamic Modes: Computational Imaging of Dynamical Systems from Sparse Observatio...

    2021

  83. [483]

    Springer International Publishing, Cham, 2017. 2, 5

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.