pith. sign in

arxiv: 2311.08754 · v3 · submitted 2023-11-15 · ⚛️ physics.flu-dyn

Machine-learning based flow field estimation using floating sensor locations

Tomoya Oura , Reno Miura , Koji Fukagata This is my paper

Pith reviewed 2026-05-24 06:21 UTC · model grok-4.3

classification ⚛️ physics.flu-dyn
keywords flow field estimationmachine learningfloating sensorsfluid dynamicssensor trajectoriesvelocity reconstructiontwo-dimensional flows
0
0 comments X

The pith

A machine learning model can estimate flow fields from floating sensor locations alone without using any fluid equations.

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

The paper introduces a machine learning technique for estimating fluid velocity fields using only the recorded positions of floating sensors over time. Instead of relying on known velocity data or the equations that govern fluid motion, the model learns to produce velocity fields that explain the observed sensor movements when the sensors are imagined to drift with the flow. Tests on flow past a cylinder, forced turbulence, and ocean currents show that the estimated fields capture key features like wakes and coherent structures, and the accuracy rivals methods that do incorporate the governing equations. This matters because it suggests a way to infer flows from limited, real-world sensor data in settings where full physics knowledge is missing or hard to apply.

Core claim

The central claim is that a neural network can be trained solely on sensor location data to output velocity fields whose integration produces sensor trajectories matching the observations, and that these fields accurately capture the underlying flow structures in two-dimensional examples without any reference to the Navier-Stokes equations or similar governing laws.

What carries the argument

The machine learning model that generates velocity fields so the time variation of sensor motion matches the given location data.

Load-bearing premise

That a velocity field reproducing the observed sensor paths must be the true underlying flow rather than some other field consistent with those paths.

What would settle it

Run the method on a known exact flow with recorded sensor paths and check whether the output velocity field matches the true field inside the regions traversed by the sensors.

Figures

Figures reproduced from arXiv: 2311.08754 by Koji Fukagata, Reno Miura, Tomoya Oura.

Figure 1
Figure 1. Figure 1: The schematic drawing of the present method. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The statistics of the forced HIT: (a) 𝑢 2 rms, (b) the Taylor Reynolds number, and (c) the energy spectrum. The warm-coloured range is adopted for the validation. 0 π 2π 0 π 2π (a) uf , Ground truth 0 π 2π (b) ˜uf , Estimated 0 π 2π (c) vf , Ground truth 0 π 2π (d) ˜vf , Estimated −1 0 1 u f , v f x y [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The ground truth and estimated velocity fields of the forced HIT at [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The dependence on the number of sensors 𝑁𝑠 in the case of the forced HIT. (a–h) The ground-truth and estimated velocity fields at 𝑡 = 40. The grey dots in (b–f) represent the sensor locations. In (g) and (h), the sensor locations are not printed for better readability. (i) The error norm. (j) The energy spectrum, where the black solid line shows the ground truth. with 𝑁𝑠 = 512, the error norm is 0.14 ± 0.0… view at source ↗
Figure 5
Figure 5. Figure 5: The estimated velocity fields of the forced HIT at [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (a–f) The ground truth and estimated velocity fields of the ocean currents for the number of sensors 𝑁𝑠 = 8 and 512. The grey dots in (c) and (d) represent the sensor locations. In (e) and (f), the sensor locations are not printed for better readability. (g) The error norm depending on 𝑁𝑠. The flow and sensor motion is best viewed in the Supplemental Movie 2. fluid flows or ground-truth velocity fields are… view at source ↗
read the original abstract

Based on machine learning techniques, we propose a novel method to estimate flow fields using only floating sensor locations. This method does not require either ground-truth velocity fields or governing equations for fluid flows, which is attractive for practical applications. The machine learning model is supposed to generate accurate velocity fields so that the time variation of sensor motion is consistent with the given data of sensor locations. To validate the method, the estimation accuracy, the dependence on the number of sensors, the time intervals for the sensor location data, and the robustness to noise are investigated using three examples of two-dimensional flows: the flow around a circular cylinder, the forced homogeneous isotropic turbulence, and the ocean currents. These investigations demonstrate the performance and practicality of this method, revealing that the accuracy can be comparable to the state-of-the-art physics-informed neural networks (PINNs)-based method even without any assumption of governing equations. Moreover, we observe that the present method can estimate the major structures, such as periodic wakes behind a cylinder, coherent structures in the forced turbulence, and stable ocean currents, with only a few sensors. We believe the present method can provide effective utilization of floating sensor observations in various fields.

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.

Referee Report

3 major / 2 minor

Summary. The paper proposes a machine-learning method to estimate 2D flow velocity fields solely from time series of floating sensor locations. A neural network parameterizes the velocity field v(x,t); its parameters are optimized so that numerical integration of the sensor ODEs dx_i/dt = v(x_i,t) reproduces the observed sensor trajectories. No ground-truth velocities or governing equations are used. Validation on cylinder wake, forced isotropic turbulence, and ocean currents shows that major flow structures can be recovered with few sensors and that accuracy is comparable to PINNs-based methods.

Significance. If the recovered fields are shown to be the true underlying flows rather than arbitrary fields consistent with the finite set of trajectories, the method would be practically useful for Lagrangian-only data scenarios in oceanography and experiments where governing equations are unavailable. The absence of equation assumptions is a potential strength, but the underdetermined inverse problem must be resolved for the claim to hold.

major comments (3)
  1. [Method] Method section (around the loss function and optimization): the training objective minimizes only the integrated sensor-position mismatch. With a finite number of sensors the map from velocity field to trajectories is many-to-one; nothing in the formulation (no physics constraints, no explicit regularization beyond network architecture) selects the physical solution over other consistent fields. This directly undermines the central claim that the estimated field is the true flow rather than merely trajectory-consistent.
  2. [Results] Results, sensor-count dependence (figures showing 4–20 sensors): the reported recovery of “major structures” with few sensors is consistent with the underdetermination concern. Without a quantitative demonstration that additional sensors systematically reduce error toward an independently known ground-truth field (rather than simply adding more constraints), the claim that accuracy is comparable to PINNs remains unproven.
  3. [Numerical examples] Cylinder and turbulence examples: while qualitative agreement with expected wakes and coherent structures is shown, the manuscript must supply point-wise velocity or vorticity error norms against ground truth (or against the PINN baseline) to establish that the learned field is not an arbitrary interpolant of the paths.
minor comments (2)
  1. [Method] Clarify the precise neural-network architecture, activation functions, and any implicit regularization (e.g., weight decay) used for the velocity-field representation.
  2. [Method] Specify the numerical integrator and time-stepping scheme employed for the sensor ODEs, and report sensitivity of results to integrator choice.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the scope and limitations of our method. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation of quantitative results and to better articulate the method's assumptions.

read point-by-point responses

  1. Referee: [Method] Method section (around the loss function and optimization): the training objective minimizes only the integrated sensor-position mismatch. With a finite number of sensors the map from velocity field to trajectories is many-to-one; nothing in the formulation (no physics constraints, no explicit regularization beyond network architecture) selects the physical solution over other consistent fields. This directly undermines the central claim that the estimated field is the true flow rather than merely trajectory-consistent.

    Authors: We acknowledge that the inverse problem is underdetermined with finite sensors and that the loss function alone does not enforce uniqueness. Our formulation deliberately omits governing equations to enable application where such equations are unavailable. The neural network architecture provides implicit regularization, and in the tested cases the optimized fields recover the dominant structures observed in the ground-truth data. We do not claim the recovered field is the unique true flow, only that it is consistent with the trajectories and reproduces major features. We will revise the manuscript to explicitly state this distinction, add a limitations paragraph on non-uniqueness, and discuss how the network capacity influences the selected solution. revision: partial

  2. Referee: [Results] Results, sensor-count dependence (figures showing 4–20 sensors): the reported recovery of “major structures” with few sensors is consistent with the underdetermination concern. Without a quantitative demonstration that additional sensors systematically reduce error toward an independently known ground-truth field (rather than simply adding more constraints), the claim that accuracy is comparable to PINNs remains unproven.

    Authors: Ground-truth velocity fields are available for the cylinder and turbulence examples. We will add quantitative error analysis (L2 norms of velocity) versus sensor count, showing systematic error reduction, and include direct numerical comparison to the PINN baseline errors. This will be presented in revised figures and a new table. revision: yes

  3. Referee: [Numerical examples] Cylinder and turbulence examples: while qualitative agreement with expected wakes and coherent structures is shown, the manuscript must supply point-wise velocity or vorticity error norms against ground truth (or against the PINN baseline) to establish that the learned field is not an arbitrary interpolant of the paths.

    Authors: We agree that qualitative agreement alone is insufficient. We will add point-wise error norms (velocity and vorticity L2 and point-wise errors) against ground truth for both examples, together with the corresponding PINN errors, to quantify the accuracy and support the comparability statement. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method is external-data inverse fit

full rationale

The paper trains an ML model (NN-parameterized velocity field) to minimize mismatch between integrated sensor trajectories dx/dt = v(x,t) and observed location time series. This is a standard data-fitting inverse problem whose output is the fitted v; the paper does not claim any first-principles derivation or prediction that reduces to the inputs by construction. Validation uses external benchmark flows (cylinder wake, forced turbulence, ocean currents) with reported comparisons to PINNs and ground-truth structures. No self-citations, self-definitional steps, or fitted-input-renamed-as-prediction appear in the abstract or described method. The central claim (comparable accuracy without governing equations) rests on empirical performance against independent data, not on internal reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the trainability of an ML model on sensor location time series alone and the assumption that matching sensor motions implies accurate flow reconstruction. No explicit free parameters or invented entities are identifiable from the abstract.

free parameters (1)
  • machine learning model parameters
    The neural network or ML model weights are optimized during training to satisfy the consistency condition with sensor data.
axioms (1)
  • domain assumption Sensor motion is determined by the fluid velocity at its location
    This is the fundamental link allowing inference of the flow field from sensor trajectories.

pith-pipeline@v0.9.0 · 5737 in / 1325 out tokens · 46894 ms · 2026-05-24T06:21:37.468721+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

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

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence ...

    work page

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

    work page

  3. [3]

    Callaham, J. L. , Maeda, K. & Brunton, S. L. 2019 Robust flow reconstruction from limited measurements via sparse representation . Phys. Rev. Fluids 4 (10), 103907

    work page 2019

  4. [4]

    Charney, J. G. 1971 Geostrophic turbulence . J. Atmos. Sci. 28 (6), 1087--1095

    work page 1971

  5. [5]

    Erichson, N. B. , Mathelin, L. , Yao, Z. , Brunton, S. L. , Mahoney, M. W. & Kutz, J. N. 2020 Shallow neural networks for fluid flow reconstruction with limited sensors . Proc. R. Soc. A 476 (2238), 20200097

    work page 2020

  6. [6]

    , Maulik, R

    Fukami, K. , Maulik, R. , Ramachandra, N. , Fukagata, K. & Taira, K. 2021 Global field reconstruction from sparse sensors with Voronoi tessellation-assisted deep learning . Nat. Mach. Intell. 3 (11), 945--951

    work page 2021

  7. [7]

    Hansen, D. V. & Poulain, P-M. 1996 Quality control and interpolations of WOCE - TOGA drifter data . J. Atmos. Ocean. Technol. 13 (4), 900--909

    work page 1996

  8. [8]

    Kingma, D. P. & Ba, J. 2017 Adam: A method for stochastic optimization, arXiv:arXiv: 1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  9. [9]

    , Bottou, L

    LeCun , Y. , Bottou, L. , Bengio, Y. & Haffner, P. 1998 Gradient-based learning applied to document recognition . Proc. IEEE 86 (11), 2278--2324

    work page 1998

  10. [10]

    , Santangelo, P

    Legras, B. , Santangelo, P. & Benzi, R. 1988 High-resolution numerical experiments for forced two-dimensional turbulence . Europhys. Lett. 5 (1), 37--42

    work page 1988

  11. [11]

    Leith, C. E. 1968 Diffusion approximation for two‐dimensional turbulence . Phys. Fluids 11 (3), 671--672

    work page 1968

  12. [12]

    Lundgren, T. S. 2003 Linearly forced isotropic turbulence . Tech. Rep.\/ . Center for Turbulence Research, Stanford

    work page 2003

  13. [13]

    Maltrud, M. E. & Vallis, G. K. 1991 Energy spectra and coherent structures in forced two-dimensional and beta-plane turbulence . J. Fluid Mech. 228 , 321--342

    work page 1991

  14. [14]

    Masumoto, Y. et al. 2004 A fifty-year eddy-resolving simulation of the world ocean: Preliminary outcomes of OFES ( OGCM for the earth simulator) . J. Earth Simulator 1 , 35--56

    work page 2004

  15. [15]

    , Rempfer, D

    Mokhasi, P. , Rempfer, D. & Kandala, S. 2009 Predictive flow-field estimation . Phys. D: Nonlinear Phenom. 238 (3), 290--308

    work page 2009

  16. [16]

    , Nguimatsia, S

    Podvin, B. , Nguimatsia, S. , Foucaut, J-M. , Cuvier, C. & Fraigneau, Y. 2018 On combining linear stochastic estimation and proper orthogonal decomposition for flow reconstruction . Exp. Fluids 59 (3), 58

    work page 2018

  17. [17]

    , Perdikarism, P

    Raissi, M. , Perdikarism, P. & Karniadakis, G. E. 2019 Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations . J. Comput. Phys. 378 , 686--707

    work page 2019

  18. [18]

    Rumelhart, D. E. , Hinton, G. E. & Williams, R. J. 1986 Learning representations by back-propagating errors . Nature 323 (6088), 533--536

    work page 1986

  19. [19]

    & Staples, A

    San, O. & Staples, A. E. 2013 Stationary two-dimensional turbulence statistics using a Markovian forcing scheme . Comput. Fluids 71 , 1--18

    work page 2013

  20. [20]

    , Nonaka, M

    Sasaki, H. , Nonaka, M. , Masumoto, Y. , Sasai, Y. , Uehara, H. & Sakuma, H. 2008 An Eddy-Resolving Hindcast Simulation of the Quasiglobal Ocean from 1950 to 2003 on the Earth Simulator\/ , pp. 157--185 . Springer New York

    work page 2008

  21. [21]

    , Sasai, Y

    Sasaki, H. , Sasai, Y. , Kawahara, S. , Furuichi, M. , Araki, F. , Ishida, A. , Yamanaka, Y. , Masumoto, Y. & Sakuma, H. 2004 a\/ A series of eddy-resolving ocean simulations in the world ocean - OFES ( OGCM for the earth simulator) project. In OCEANS `04\/ , pp. 1535--1541

    work page 2004

  22. [22]

    , Ishida, A

    Sasaki, Y. , Ishida, A. , Yamanaka, Y. & Sasaki, H. 2004 b\/ Chlorofluorocarbons in a global ocean eddy-resolving OGCM : Pathway and formation of antarctic bottom water . Geophysical Research Letters 31 (12)

    work page 2004

  23. [23]

    Schorghofer, N 2000 Energy spectra of steady two-dimensional turbulent flows . Phys. Rev. E 61 (6), 6572--6577

    work page 2000

  24. [24]

    , Nair, A

    Taira, K. , Nair, A. G. & Brunton, S. L. 2016 Network structure of two-dimensional decaying isotropic turbulence . J. Fluid Mech. 795 , R2

    work page 2016

  25. [25]

    , Rafiee, M

    Tinka, A. , Rafiee, M. & Bayen, A. M. 2013 Floating sensor networks for river studies . IEEE Syst. J. 7 (1), 36--49

    work page 2013

  26. [26]

    , Percelay, J

    Tossavainen, O-P. , Percelay, J. , Tinka, A. , Wu, A. & Bayen, A. M. 2008 Ensemble Kalman filter based state estimation in 2D shallow water equations using Lagrangian sensing and state augmentation. In 2008 47th IEEE Conference on Decision and Control\/ , pp. 1783--1790

    work page 2008

  27. [27]

    Wong, A. P. S. et al. 2020 Argo data 1999^^e2^^80^^932019: Two million temperature-salinity profiles and subsurface velocity observations from a global array of profiling floats . Front. Mar. Sci. 7 , 700

    work page 2020

  28. [28]

    & Heimbach, P

    Wunsch, C. & Heimbach, P. 2007 Practical global oceanic state estimation . Phys. D: Nonlinear Phenom. 230 (1), 197--208

    work page 2007

  29. [29]

    , Shats, M

    Xia, H. , Shats, M. & Falkovich, G. 2009 Spectrally condensed turbulence in thin layers . Phys. Fluids 21 (12), 125101

    work page 2009