arxiv: 2605.04535 · v1 · submitted 2026-05-06 · 💻 cs.LG · cs.NA· math.NA· physics.comp-ph· stat.AP· stat.ML

Recognition: unknown

From Video-to-PDE: Data-Driven Discovery of Nonlinear Dye Plume Dynamics

Cesar Acosta-Minoli, Sayantan Sarkar

Pith reviewed 2026-05-08 16:55 UTC · model grok-4.3

classification 💻 cs.LG cs.NAmath.NAphysics.comp-phstat.APstat.ML

keywords data-driven PDE discoveryvideo to PDEdye plume dynamicsweak-form sparse regressionCole-Hopf transformationphysics-informed neural networkscontinuum model identification

0 comments

The pith

Uncalibrated video of an ink plume yields a compact nonlinear PDE that predicts its evolution and reduces to a linear advection-diffusion equation via Cole-Hopf transformation.

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

The paper establishes a staged pipeline that turns raw grayscale video into a scalar field, extracts a time-dependent bulk velocity from the intensity centroid, and uses weak-form sparse regression on a restricted gradient library to identify an effective transport law. After coefficient refinement with an inverse physics-informed network and uncertainty quantification via chronological bootstrap, the resulting model outperforms standard advection-diffusion baselines on held-out frames while retaining a positive diffusion coefficient. This shows that visual intensity data alone can support structurally interpretable continuum models when discovery, calibration, and validation are kept separate.

Core claim

The selected reduced model u_t + v(t)·∇u = 9.005 |∇u|^2 + 0.666 Δu is recovered from the video data; it outperforms advection-diffusion baselines on held-out frames, retains a positive Laplacian coefficient, and admits a Cole-Hopf reduction to a linear advection-diffusion equation.

What carries the argument

The video-to-PDE pipeline that converts grayscale frames to a normalized scalar field, isolates bulk drift via intensity-weighted centroid, performs weak-form sparse regression on compact gradient libraries, and refines coefficients with an inverse physics-informed network followed by forward-rollout recalibration.

If this is right

The identified nonlinear term allows the plume model to be transformed into a linear advection-diffusion equation, simplifying analytic or numerical treatment.
Positive Laplacian coefficient indicates that the effective diffusion remains physically consistent even after data-driven selection.
The separation of discovery, calibration, and uncertainty stages produces models that remain predictive on unseen frames without overfitting to noise in differentiation.
Compact gradient-based libraries suffice once collinearity diagnostics eliminate overcomplete candidates.

Where Pith is reading between the lines

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

The same staged pipeline could be applied to other scalar visual fields such as temperature or concentration maps from environmental imagery.
The Cole-Hopf reduction raises the possibility that the discovered nonlinearity corresponds to a known transformation in the underlying physics rather than an artifact of the imaging process.
Chronological block bootstrap uncertainty estimates could be used to decide when additional video data would meaningfully tighten coefficient bounds.
If the intensity-to-concentration mapping is monotonic but nonlinear, the recovered coefficients would require rescaling before quantitative physical interpretation.

Load-bearing premise

The normalized grayscale image intensity can be treated as a faithful scalar concentration field whose evolution is governed by a compact gradient-based transport law that is discoverable from the video without missing essential physical terms.

What would settle it

A direct comparison of the model's forward predictions against additional held-out video frames from the same plume, or an independent laboratory experiment with known dye concentration under controlled advection, that shows systematic deviation beyond the reported bootstrap uncertainty bands.

Figures

Figures reproduced from arXiv: 2605.04535 by Cesar Acosta-Minoli, Sayantan Sarkar.

**Figure 1.** Figure 1: Experimental setup. A top-view camera records the spreading of an ink droplet over the spatial domain Ω; grayscale frames are converted into the image-derived scalar field 𝑢(𝑥, 𝑦, 𝑡) view at source ↗

**Figure 2.** Figure 2: Preprocessing diagnostic. Each row shows an uncropped grayscale frame with the retained crop region indicated, the corresponding cropped region, and the final field 𝑢(𝑥, 𝑦, 𝑡). Colour represents the normalised intensity, not physical dye colour. 𝐿𝑥 = 𝐿𝑦 = 200 image-coordinate units, with 𝑇 = 34 s. Spatial coordinates are image units, not calibrated lengths view at source ↗

**Figure 3.** Figure 3: Snapshots of 𝑢(𝑥, 𝑦, 𝑡) (top of each panel) and the middle horizontal cross-section (bottom). The plume undergoes spreading, deformation, and weak drift over the observation window. 2.3. Preprocessing Pipeline Each raw frame is converted to grayscale, cropped to the rectangular region (𝑥0 , 𝑦0 , 𝑤, ℎ) = (450, 0, 1080, 1080) in pixel units, the eight-pixel boundary is removed, intensity is inverted and norm… view at source ↗

**Figure 4.** Figure 4: Intensity-weighted centroid trajectory of the processed plume; the vertical image axis follows the image-coordinate convention. 3. Methodology 3.1. Centre-of-Mass Drift Estimation The processed plume exhibits a weak bulk drift in addition to its intrinsic spreading. To decouple the two, we estimate a time-dependent drift velocity from the image-derived field. With 𝑢(𝑥, 𝑦, 𝑡) the processed observable, the t… view at source ↗

**Figure 5.** Figure 5: Column-correlation heatmaps of the weak feature matrix Θ. The full library exhibits strong correlations among constant, 𝑢, 𝑢 2 , and nonlinear gradient terms, explaining its large condition number; the reduced libraries show clearer feature separation. 18% of runs and is therefore not robust. The full library and C-both select many terms with high frequency, but this reflects an overcomplete feature set ra… view at source ↗

**Figure 6.** Figure 6: Active-term counts under STLSQ as the threshold varies. C and C-alt retain meaningful sparse structures over a broad threshold range; the full library has a more complex support path. (a) A (b) B (c) C (d) C-alt (e) C-both (f) Full view at source ↗

**Figure 7.** Figure 7: Coefficient paths under the STLSQ threshold sweep. Reduced libraries show a single dominant nonlinear-gradient coefficient; the full library shows competing correlated features. to whether the advection coefficients are learned or fixed at unity. Centroid accuracy, however, is not: the C and Calt models have Mode B centroid RMSE below 0.05, against values near 1.1 under Mode A. The learned-advection mode … view at source ↗

**Figure 8.** Figure 8: Selection frequencies over 100 random weak-centre samples. (a) A (b) B (c) C (d) C-alt (e) C-both (f) Full view at source ↗

**Figure 9.** Figure 9: Coefficient stability over 100 random weak-centre samples; markers show the mean among selected runs and bars show one standard deviation. 4.4. iPINN-Refined Coefficients The iPINN refinement ( view at source ↗

**Figure 10.** Figure 10: Validation rollout rRMSE under Mode A (learned advection) and Mode B (measured advection, coefficient 1.0). Mode B substantially improves the advection–diffusion baselines and preserves centroid trajectories for the nonlineargradient models. rollout-based calibration, with both initialisations remaining near 11% validation rRMSE, indicating that the weightedgradient structure has reduced expressive free… view at source ↗

**Figure 11.** Figure 11: iPINN coefficient and loss trajectories for the C and C-alt models view at source ↗

**Figure 12.** Figure 12: Direct validation comparison between weak-SINDy/STLSQ and iPINN-refined coefficients. The iPINN-refined C model gives the lowest pixel-wise rRMSE and front-radius RMSE; weak-SINDy/STLSQ coefficients preserve the centroid more accurately view at source ↗

**Figure 13.** Figure 13: Gaussian smoothing sensitivity for a representative frame. Top row: smoothed frames at increasing 𝜎. Middle row: difference 𝑢𝜎 − 𝑢0 from the unsmoothed frame. The selected value 𝜎 = 1.0 preserves the front profile while reducing pixel-scale noise view at source ↗

**Figure 14.** Figure 14: Middle horizontal cross-section under different smoothing levels view at source ↗

read the original abstract

Inferring continuum models directly from video is hampered by two facts: the recorded field is uncalibrated image intensity rather than a physical state, and direct numerical differentiation of noisy frames is unstable. We develop a video-to-PDE pipeline that converts grayscale recordings of an ink plume into a normalised scalar field $u(x,y,t)$, isolates a bulk drift $\mathbf{v}(t)$ from intrinsic spreading via the intensity-weighted centroid, and identifies an effective transport law by weak-form sparse regression. Conditioning, threshold-sweep and random-centre diagnostics show that overcomplete libraries are strongly collinear; the search is therefore restricted to compact gradient-based libraries. Coefficients are refined by an inverse physics-informed network and recalibrated against forward rollouts, with a chronological block bootstrap quantifying uncertainty. The selected reduced model $u_t+\mathbf v(t)\!\cdot\!\nabla u = 9.005\,|\nabla u|^{2}+0.666\,\Delta u$ outperforms advection--diffusion baselines on held-out frames, retains a positive Laplacian coefficient, and admits a Cole--Hopf reduction to a linear advection--diffusion equation. The framework demonstrates that uncalibrated visual data can yield compact, predictive and structurally interpretable continuum models when discovery, calibration and uncertainty are treated as distinct stages.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This paper builds a video-to-PDE pipeline for a dye plume that produces a nonlinear model outperforming baselines on held-out frames, though the physical meaning depends on an uncalibrated intensity mapping.

read the letter

The main takeaway is a practical pipeline that converts raw video of an ink plume into a compact PDE. They normalize the grayscale frames to a scalar field u, extract a time-dependent bulk drift v(t) from the intensity-weighted centroid, run weak-form sparse regression on a library restricted after collinearity diagnostics, refine the coefficients with an inverse PINN, and recalibrate against forward rollouts while using chronological bootstrap for uncertainty. The resulting model u_t + v·∇u = 9.005 |∇u|^2 + 0.666 Δu beats plain advection-diffusion on held-out frames, keeps a positive diffusion coefficient, and reduces via Cole-Hopf to linear advection-diffusion. Treating discovery, calibration, and uncertainty as separate stages is a sensible way to work with uncalibrated visual data, and the threshold-sweep and random-centre checks add concrete robustness steps. The collinearity restriction is a useful practical move that avoids overfit libraries. The soft spot is the core assumption that normalized intensity directly stands in for the concentration field u. Standard dye transport is linear advection-diffusion in physical concentration, yet the large nonlinear coefficient is consistent with a nonlinear camera response or the normalization step itself. Without an independent calibration that maps intensity to actual dye amount, the outperformance on video data shows a good fit to the observations but does not yet confirm that the equation describes the underlying fluid dynamics rather than the imaging process. It is a single demonstration case, so broader applicability remains open. This is worth a reading group for people working on data-driven continuum models from observational video in fluids or environmental settings. The methods are structured enough and the validation steps are explicit enough that a serious editor should send it for peer review, with the main request being more direct evidence on the intensity-to-concentration relation.

Referee Report

2 major / 2 minor

Summary. The paper proposes a video-to-PDE pipeline that converts grayscale recordings of an ink plume into a normalized scalar field u(x,y,t), extracts a bulk drift v(t) via the intensity-weighted centroid, identifies an effective transport law through weak-form sparse regression on a restricted compact gradient-based library (after collinearity diagnostics), refines coefficients with an inverse physics-informed neural network and forward-rollout recalibration, and quantifies uncertainty via chronological block bootstrap. The selected model u_t + v(t)·∇u = 9.005 |∇u|^2 + 0.666 Δu is reported to outperform advection-diffusion baselines on held-out frames, retain a positive Laplacian coefficient, and admit a Cole-Hopf reduction to a linear advection-diffusion equation.

Significance. If the mapping from image intensity to the state variable holds, the work demonstrates a practical, multi-stage framework for extracting compact, predictive, and structurally interpretable continuum models from uncalibrated visual data. Strengths include explicit separation of discovery, calibration, and uncertainty stages; diagnostics for library collinearity (threshold sweeps, random-centre tests); bootstrap uncertainty; and the Cole-Hopf interpretability that links the nonlinear term to a possible logarithmic transformation. This approach addresses noisy differentiation and overcomplete libraries in a reproducible manner and could advance data-driven modeling in fluid dynamics and scientific machine learning.

major comments (2)

[Abstract and §2] Abstract and §2 (field normalization and library selection): The central claim that the pipeline yields models of 'nonlinear dye plume dynamics' rests on treating normalized grayscale intensity as a faithful scalar concentration field u(x,y,t) governed by the discovered gradient-based law. The large nonlinear coefficient (9.005) is consistent with u ∝ log(physical concentration), yet no calibration experiment, independent intensity-to-concentration mapping, or test against known linear advection-diffusion behavior in physical concentration is provided. This assumption is load-bearing for the asserted physical interpretability and structural reduction.
[§4] §4 (held-out validation and performance): The claim that the reduced model outperforms advection-diffusion baselines on chronologically held-out frames is central to the predictive utility, but the manuscript provides no quantitative metrics (e.g., L2 prediction errors, relative norms, or R² values) comparing the discovered PDE against the baselines. Without these numbers, the magnitude and robustness of the reported outperformance cannot be assessed.

minor comments (2)

[Notation] Notation for the time-dependent velocity should be uniformly boldface (v(t)) in all equations and text for consistency.
[Figures] Figure captions for validation plots should explicitly indicate the frame ranges or time blocks used for training versus held-out testing.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the scope of our video-to-PDE pipeline and the need for clearer validation metrics. We address each major comment below, indicating where revisions will be made to improve clarity without overstating the results.

read point-by-point responses

Referee: [Abstract and §2] Abstract and §2 (field normalization and library selection): The central claim that the pipeline yields models of 'nonlinear dye plume dynamics' rests on treating normalized grayscale intensity as a faithful scalar concentration field u(x,y,t) governed by the discovered gradient-based law. The large nonlinear coefficient (9.005) is consistent with u ∝ log(physical concentration), yet no calibration experiment, independent intensity-to-concentration mapping, or test against known linear advection-diffusion behavior in physical concentration is provided. This assumption is load-bearing for the asserted physical interpretability and structural reduction.

Authors: We agree that the mapping from normalized grayscale intensity to physical dye concentration is not calibrated experimentally in this work. The pipeline is designed to operate directly on the extracted scalar field u(x,y,t) derived from video intensity, and the resulting PDE is therefore an effective model for the observed intensity dynamics rather than a direct description of physical concentration. The Cole-Hopf reduction is presented as a mathematical property of the discovered equation that offers structural insight and is consistent with a possible logarithmic relation to concentration, but we do not claim a calibrated physical mapping. In the revision we will update the abstract and Section 2 to state explicitly that the model governs normalized intensity and to qualify the physical interpretation as suggestive, thereby scoping the load-bearing assumption appropriately. revision: partial
Referee: [§4] §4 (held-out validation and performance): The claim that the reduced model outperforms advection-diffusion baselines on chronologically held-out frames is central to the predictive utility, but the manuscript provides no quantitative metrics (e.g., L2 prediction errors, relative norms, or R² values) comparing the discovered PDE against the baselines. Without these numbers, the magnitude and robustness of the reported outperformance cannot be assessed.

Authors: We accept that the outperformance statement in Section 4 is currently qualitative. Although the manuscript reports that the discovered model performs better than advection-diffusion baselines on held-out frames, explicit numerical comparisons are not tabulated. We will add quantitative metrics—L2 prediction errors, relative norms, and R² values—for the discovered PDE and the baselines on the chronologically held-out data. These will appear in a revised table or figure panel in Section 4 of the resubmission. revision: yes

Circularity Check

0 steps flagged

No significant circularity; model discovery and validation are data-driven with held-out evaluation

full rationale

The paper's core pipeline—normalizing video intensity to u(x,y,t), extracting v(t) via centroids, applying weak-form sparse regression on a restricted gradient-based library, refining coefficients via inverse PINN and forward rollouts, and evaluating predictive performance on chronologically held-out frames with bootstrap—operates directly on observed data without presupposing the final PDE form or coefficients. The selected equation is not equivalent to its inputs by construction; outperformance on unseen frames and the positive Laplacian term are empirical outcomes. The Cole-Hopf reduction is a mathematical property noted after discovery, not a load-bearing assumption or self-referential step. No self-citations, fitted-input-as-prediction, or ansatz smuggling are evident in the derivation chain.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on two fitted numerical coefficients and two key domain assumptions about data representation and library completeness; no new physical entities are postulated.

free parameters (2)

nonlinear coefficient = 9.005
The value 9.005 multiplying |∇u|^2 is obtained by inverse PINN refinement and forward recalibration.
diffusion coefficient = 0.666
The value 0.666 multiplying Δu is obtained by the same fitting procedure.

axioms (2)

domain assumption Normalized grayscale intensity represents the physical scalar concentration field u(x,y,t)
Invoked when converting raw video frames into the input field for regression.
ad hoc to paper The true dynamics lie within a compact gradient-based library
Adopted after observing strong collinearity in overcomplete libraries; restricts the search space.

pith-pipeline@v0.9.0 · 5542 in / 1723 out tokens · 112969 ms · 2026-05-08T16:55:02.592706+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 28 canonical work pages · 1 internal anchor

[1]

Ondata-drivenidentification:Isautomaticallydiscoveringequationsofmotionfromdataachimera? Nonlinear Dynamics 111, 6487–6498

Antonelli,G.,Chiaverini,S.,Lillo,P.D.,2022. Ondata-drivenidentification:Isautomaticallydiscoveringequationsofmotionfromdataachimera? Nonlinear Dynamics 111, 6487–6498. doi:10.1007/s11071-022-08192-x. Banerjee, C., Nguyen, K., Fookes, C., Karniadakis, G.,

work page doi:10.1007/s11071-022-08192-x 2022
[2]

Both, G.J., Choudhury, S., Sens, P., Kusters, R.,

doi:10.1145/3689037. Both, G.J., Choudhury, S., Sens, P., Kusters, R.,

work page doi:10.1145/3689037
[3]

Journal of Computational Physics 428, 109985

Deepmod: Deep learning for model discovery in noisy data. Journal of Computational Physics 428, 109985. doi:10.1016/j.jcp.2020.109985. Brunton, S.L., Proctor, J.L., Kutz, J.N.,

work page doi:10.1016/j.jcp.2020.109985 2020
[4]

Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou

Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences 113, 3932–3937. doi:10.1073/pnas.1517384113. Chu, M., Liu, L., Zheng, Q., Franz, A., Seidel, H.P., Theobalt, C., Zayer, R.,

work page doi:10.1073/pnas.1517384113
[5]

Cole, J.D.,

doi:10.1145/3528223.3530169. Cole, J.D.,

work page doi:10.1145/3528223.3530169
[6]

Scientific machine learning through physics-informed neural networks: where we are and what’s next.Journal of Scientific Computing, 92(3):88, 2022

doi:10.1007/s10915-022-01939-z. Dreisbach, M., Kiyani, E., Kriegseis, J., Karniadakis, G., Stroh, A.,

work page doi:10.1007/s10915-022-01939-z
[7]

Language Model Cascades: Token-Level Uncertainty and Beyond

Pinns4drops: Video-conditioned physics-informed neural networks for two-phase flow reconstruction. arXiv preprint arXiv:2411.15949 URL:https://arxiv.org/abs/2411.15949, doi:10.48550/arXiv. 2411.15949. Efron, B.,

work page internal anchor Pith review doi:10.48550/arxiv
[8]

doi: 10.1214/aos/ 1176342360

Bootstrap methods: Another look at the jackknife. The Annals of Statistics 7, 1–26. doi:https://doi.org/10.1214/aos/ 1176344552. Fasel, U., Kutz, J.N., Brunton, B.W., Brunton, S.L.,

work page doi:10.1214/aos/
[9]

Fasel , author J

Ensemble-sindy: Robust sparse model discovery in the low-data, high-noise limit, with active learning and control. Proceedings of the Royal Society A 478, 20210904. doi:10.1098/rspa.2021.0904. Fukami,K.,Murata,T.,Zhang,K.,Fukagata,K.,2021. Sparseidentificationofnonlineardynamicswithlow-dimensionalizedflowrepresentations. Journal of Fluid Mechanics 926, A1...

work page doi:10.1098/rspa.2021.0904 2021
[10]

Communications on Pure and Applied Mathematics 3, 201–230

The partial differential equation𝑢 𝑡 +𝑢𝑢 𝑥 =𝜇𝑢 𝑥𝑥. Communications on Pure and Applied Mathematics 3, 201–230. doi:10.1002/cpa.3160030302. Horn, B.K.P., Schunck, B.G.,

work page doi:10.1002/cpa.3160030302
[11]

Horn and Brian G

Determining optical flow. Artificial Intelligence 17, 185–203. doi:10.1016/0004-3702(81)90024-2. Hu,M.K.,1962. Visualpatternrecognitionbymomentinvariants. IRETransactionsonInformationTheory8,179–187. doi:10.1109/TIT.1962. 1057692. Jaques,M.,Burke,M.,Hospedales,T.M.,2019. Physics-as-inverse-graphics:Jointunsupervisedlearningofobjectsandphysicsfromvideo. Co...

work page doi:10.1016/0004-3702(81)90024-2 1962
[12]

Physical Review Letters 129, 258001

Data-driven discovery of active nematic hydrodynamics. Physical Review Letters 129, 258001. doi:10.1103/PhysRevLett.129.258001. Minoli, Sarkar:Preprint submitted to ElsevierPage 20 of 22 Video-to-PDE Kaheman, K., Kutz, J.N., Brunton, S.L.,

work page doi:10.1103/physrevlett.129.258001
[13]

Proceedings of the Royal Society A 476, 20200279

Sindy-pi: A robust algorithm for parallel implicit sparse identification of nonlinear dynamics. Proceedings of the Royal Society A 476, 20200279. doi:10.1098/rspa.2020.0279. Kaptanoglu, A.A., de Silva, B.M., Fasel, U., Kaheman, K., Goldschmidt, A.J., Callaham, J.L., Delahunt, C.B., Nicolaou, Z.G., Champion, K., Loiseau, J.C., Kutz, J.N., Brunton, S.L.,

work page doi:10.1098/rspa.2020.0279 2020
[14]

Pysindy: A comprehensive python package for robust sparse system identification,

doi:10.21105/joss.03994. Kardar, M., Parisi, G., Zhang, Y.C.,

work page doi:10.21105/joss.03994
[15]

Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

Physics-informed machine learning. Nature Reviews Physics 3, 422–440. doi:https://doi.org/10.1038/s42254-021-00314-5. Krishnapriyan, A.S., Gholami, A., Zhe, S., Kirby, R.M., Mahoney, M.W.,

work page doi:10.1038/s42254-021-00314-5
[16]

26548–26560

Characterizing possible failure modes in physics-informed neural networks, in: Advances in Neural Information Processing Systems 34 (NeurIPS 2021), pp. 26548–26560. URL:https://proceedings. neurips.cc/paper/2021/file/df438e5206f31600e6ae4af72f2725f1-Paper.pdf. Künsch, H.R.,

2021
[17]

Schervish

The jackknife and the bootstrap for general stationary observations. The Annals of Statistics 17, 1217–1241. doi:https: //doi.org/10.1214/aos/1176347265. Larson, J., Menickelly, M., Wild, S.M.,

work page doi:10.1214/aos/1176347265
[18]

Measurement Science and Technology 33, 075406

Image convolution-based experimental technique for flame front detection and dimension estimation: a case study on laminar-to-transition jet diffusion flame height measurement. Measurement Science and Technology 33, 075406. doi:10.1088/1361-6501/ac65db. Liu,R.Y.,Singh,K.,1992. Movingblocksjackknifeandbootstrapcaptureweakdependence,in:LePage,R.,Billard,L.(...

work page doi:10.1088/1361-6501/ac65db 1992
[19]

Journal of Fluid Mechanics 844, 459–490

Sparse reduced-order modelling: sensor-based dynamics to full-state estimation. Journal of Fluid Mechanics 844, 459–490. doi:10.1017/jfm.2018.147. Long,Z.,Lu,Y.,Dong,B.,2019. PDE-Net2.0:LearningPDEsfromdatawithanumeric–symbolichybriddeepnetwork. JournalofComputational Physics 399, 108925. doi:10.1016/j.jcp.2019.108925. Mangan,N.M.,Kutz,J.N.,Brunton,S.L.,P...

work page doi:10.1017/jfm.2018.147 2018
[20]

Hidden physics models: Machine learning of nonlinear partial differential equations,

Automatic block-length selection for the dependent bootstrap. Econometric Reviews 23, 53–70. doi:10.1081/ ETC-120028836. Raissi,M.,Karniadakis,G.E.,2018. Hiddenphysicsmodels:Machinelearningofnonlinearpartialdifferentialequations. JournalofComputational Physics 357, 125–141. doi:10.1016/j.jcp.2017.11.039. Raissi, M., Perdikaris, P., Karniadakis, G.,

work page doi:10.1016/j.jcp.2017.11.039 2018
[21]

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

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. URL:https://www. sciencedirect.com/science/article/pii/S0021999118307125, doi:https://doi.org/10.1016/j.jcp.2018.10.045. Raissi, M., Yazdani, A., Karniadakis, G.E.,

work page doi:10.1016/j.jcp.2018.10.045 2018
[22]

Science367(6481), 1026– 1030 (2020) https://doi.org/10.1126/science.aaw4741

Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations. Science 367, 1026–1030. doi:10.1126/science.aaw4741. Reinbold,P.A.K., Gurevich,D.R., Grigoriev,R.O.,2020. Usingnoisy orincomplete datatodiscover modelsofspatiotemporal dynamics. Physical Review E 101, 010203. doi:10.1103/PhysRevE.101.010203. Rudy,S.,Alla,A.,Brunton,S.L...

work page doi:10.1126/science.aaw4741 2020
[23]

Science Advances3(4), 1602614 (2017) https://doi

Data-driven discovery of partial differential equations. Science Advances 3, e1602614. doi:10.1126/sciadv.1602614. Savitzky,A.,Golay,M.J.E.,1964.Smoothinganddifferentiationofdatabysimplifiedleastsquaresprocedures.AnalyticalChemistry36,1627–1639. doi:10.1021/ac60214a047. Schaeffer, H., McCalla, S.G.,

work page doi:10.1126/sciadv.1602614 1964
[24]

Physical Review E 96, 023302

Sparse model selection via integral terms. Physical Review E 96, 023302. doi:10.1103/PhysRevE.96. 023302. Sharma, P., Chung, W.T., Akoush, B., Ihme, M.,

work page doi:10.1103/physreve.96
[25]

Sharma, W.T

doi:10.3390/en16052343. Shea,D.E.,Brunton,S.L.,Kutz,J.N.,2021. Sindy-bvp:Sparseidentificationofnonlineardynamicsforboundaryvalueproblems. PhysicalReview Research 3, 023255. doi:10.1103/PhysRevResearch.3.023255. Vinuesa, R., Brunton, S.L., McKeon, B.J.,

work page doi:10.3390/en16052343 2021
[26]

Nature Reviews Physics 5, 536–545

The transformative potential of machine learning for experiments in fluid mechanics. Nature Reviews Physics 5, 536–545. doi:10.1038/s42254-023-00622-y. Wentz, J., Doostan, A.,

work page doi:10.1038/s42254-023-00622-y
[27]

Computer Methods in Applied Mechanics and Engineering 413, 116096

Derivative-based sindy (dsindy): Addressing the challenge of discovering governing equations from noisy data. Computer Methods in Applied Mechanics and Engineering 413, 116096. doi:10.1016/j.cma.2023.116096. Minoli, Sarkar:Preprint submitted to ElsevierPage 21 of 22 Video-to-PDE Westerweel, J., Elsinga, G.E., Adrian, R.J.,

work page doi:10.1016/j.cma.2023.116096 2023
[28]

Annual Review of Fluid Mechanics 45, 409–436

Particle image velocimetry for complex and turbulent flows. Annual Review of Fluid Mechanics 45, 409–436. doi:10.1146/annurev-fluid-120710-101204. Yu, C., Bi, X., Fan, Y., 2023a. Deep learning for fluid velocity field estimation: A review. Ocean Engineering 271, 113693. doi:10.1016/j. oceaneng.2023.113693. Yu, H.X., Zheng, Y., Gao, Y., Deng, Y., Zhu, B., ...

work page doi:10.1146/annurev-fluid-120710-101204 2023
[29]

IEEE Access 7, 1404–1423

A unified framework for sparse relaxed regularized regression sr3. IEEE Access 7, 1404–1423. doi:10.1109/ACCESS.2018.2886528. Minoli, Sarkar:Preprint submitted to ElsevierPage 22 of 22

work page doi:10.1109/access.2018.2886528 2018