Pith. sign in

REVIEW 3 major objections 4 minor 16 references

Data-driven pressure field prediction for ships in regular sea states

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A convolutional-autoencoder/regression surrogate trained on time-averaged CFD pressure fields predicts the full wave-pressure field around a free-floating container ship from three sea-state parameters, yielding a held-out mean absolute…

desk verdict Solid engineering study of a CAE/NN surrogate for ship wave pressures, but the routing claim is unsupported by the ship-fixed-only validation and the time-averaging penalty in inertial-frame drag. read the letter →

arxiv 2505.06014 v1 pith:FRYAGJLR submitted 2025-05-09 physics.flu-dyn physics.data-an

Malte Loft , Henning Schwarz , Thomas Rung This is my paper · ORCID
classification physics.flu-dynphysics.data-an
keywords surrogatemodelingmultiphaseflowspressurefieldsseawayshiproutingmachinelearningconvolutionalautoencoderaddedresistance
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 aims to show that a data-driven surrogate can replace slow CFD when a routing system needs the added resistance a ship feels in waves. Instead of predicting a single resistance number, it reconstructs the entire three-dimensional, time-averaged pressure field around a free-floating container ship from just three sea-state parameters: wave height, wave length and encounter angle. A convolutional autoencoder compresses those fields into a ten-dimensional latent space and a small neural network maps the three parameters into that space, so a full pressure field can be generated almost instantly. On held-out sea states the surrogate reproduces the ship-fixed drag coefficient with a mean absolute error of 1.77e-5, about twice the error of the interpolation and integration step alone. The practical point is that if the method generalizes, existing CFD data sets become reusable fuel-saving tools for onboard routing.

What carries the argument

The load-bearing mechanism is the two-block CAE/NN surrogate. The convolutional autoencoder performs nonlinear order reduction: it squeezes a 128x64x96 pressure field down to ten latent numbers and reconstructs the field from them, capturing spatial dependencies with relatively few parameters. The fully connected regression network (200-200-10) maps the three sea-state control parameters into that same latent space, closing the loop from weather data to pressure field. Training uses a simultaneous loss that combines reconstruction error and prediction error on the full field, with the latent-matching term set to zero; interior ship cells are filled by iterated boundary means (Int.-Mean) so the convolutional filters near the hull are not corrupted by zeros.

What would settle it

Run the surrogate in the inertial frame by combining its predicted ship-fixed pressure field with the CFD pitch-angle history (or a separately predicted pitch angle), and compare the global-coordinate mean absolute drag error against the 0.94e-5 benchmark for instantaneous integration; if the error stays at the time-averaged level of 3.42e-5 rather than dropping, the surrogate does not deliver routing-grade resistance in earth-fixed coordinates.

Watch

Extended reading notes

Core claim

The central claim is that time-averaged, ship-fixed pressure deviations from calm water lie on a low-dimensional manifold parameterized by wave height, wave length and encounter angle, and that a CAE/NN pair can learn that manifold from 285 CFD-derived fields (150 simulated, 135 mirrored). The encoder reduces each field to ten latent variables, the regression network predicts the latent code from the three sea-state inputs, and the decoder outputs the full pressure field, which is then projected onto the hull and integrated to give added drag. The paper reports a held-out mean absolute error of 1.77e-5 for the ship-fixed drag coefficient with the finer meta-grid and iterative interior averaging, with head-sea cases reproduced well and the beam-sea short-wave case e1-3m as a notable outlier attributed to sparse data coverage. The authors evaluate only in the ship-fixed coordinate system, because for the inertial frame their own transfer analysis shows time-averaging introduces a larger error (3.42e-5) than integrating instantaneous fields (0.94e-5).

Load-bearing premise

The load-bearing premise is that a time-averaged pressure field written in the ship's own moving frame contains enough information to predict the added resistance a routing system would act on; the paper only tests that premise in the ship-fixed frame, and its own data-transfer check shows that averaging raises the global-coordinate drag error from 0.94e-5 to 3.42e-5.

Editorial extensions

If this is right

  • A routing system could query the surrogate for many wave scenarios in the time it takes one CFD run to advance a few time steps, making added-resistance-aware route optimization feasible in practice.
  • Because the machine-learning stage adds only about twice the error of the pure interpolation/integration pipeline, the accuracy ceiling is currently set by data transfer to the structured meta-grid rather than by the autoencoder itself.
  • Head-sea conditions (encounter angles near 180 degrees) are predicted with the lowest errors, so initial deployment is most credible for head-sea routing.
  • The time-averaged surrogate is validated only for ship-fixed drag; a global-coordinate routing target would need a pitch-angle estimate or an instantaneous-field variant.
  • A latent dimension of ten already sits at the accuracy/cost sweet spot compared with dimensions five and twenty, so the approach does not need a large latent space to capture the wave-response manifold.

Reading between the lines

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

  • The strongest unstated consequence is negative: if routing requires resistance in the earth-fixed frame, the surrogate in its current form is incomplete, and the missing piece is quantified by the paper's own data (3.42e-5 for time-averaged versus 0.94e-5 for instantaneous integration).
  • A natural test of the method's ceiling is whether adding training cases near beam seas or explicitly encoding the mirror symmetry between beta and 360 degrees minus beta shrinks the outlier error of case e1-3m; the paper's data-coverage argument predicts it would.
  • The same CAE/NN pipeline should transfer to other hulls and to other field quantities such as air concentration or local wave elevation, provided a similarly smooth latent manifold exists; the paper's broader claim about complex multiphase flow fields is extrapolation rather than demonstration.
  • For an operational routing system, coupling this pressure-field surrogate with a cheap pitch-angle regression could recover most of the global-frame accuracy gap, since pitch enters the inertial-coordinate force integration in a nearly linear way.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes an offline/online surrogate-modeling pipeline for predicting wave-induced added resistance of a free-floating KRISO container ship. A convolutional autoencoder reduces time-averaged, ship-fixed CFD pressure fields to a low-dimensional latent space, and a fully connected neural network maps the three sea-state parameters (wave height, encounter angle, wave length) to that latent space; the decoded pressure fields are then projected onto the hull and integrated to obtain added-resistance coefficients. The manuscript evaluates the pipeline for 150 CFD-simulated regular sea states (augmented to 285 cases by mirroring) on two structured meta-grids, reports data-transfer errors separately from machine-learning errors, and presents per-case drag and pressure-field errors on 18 held-out test cases. The stated headline result is a mean ship-fixed added-drag error of 1.77e-5 for the best configuration on meta-grid B with interior cells filled by iterative averaging (Section 5.2, Table 5).

Significance. If the claimed accuracy carried over to the routing-relevant quantity, the work would be a useful engineering contribution: it demonstrates a complete chain from CFD data through projection, dimension reduction, regression, and force integration, with per-case error tables, a careful separation of data-transfer versus machine-learning errors, openly available training data, and honest reporting of outlier cases such as the beam-sea case e1-3m. The data-transfer verification in Section 5.1 is a particular strength because it isolates the interpolation error before any ML is introduced. However, the central routing claim in the abstract and introduction requires added resistance in the inertial/global frame, and the manuscript deliberately evaluates the surrogate only in the ship-fixed frame; Section 5.1 shows that time-averaging degrades the global-frame drag error substantially. That gap, together with unresolved issues around test-set selection and mirrored-case leakage, means the paper currently supports a narrower claim than the one it advertises.

major comments (3)
  1. [§5.2, Table 5 vs. §5.1] The headline accuracy of E'_D = 1.77e-5 (Table 5) is computed in the ship-fixed coordinate system, while the routing motivation requires the inertial-frame added resistance, i.e., the x1-component in Eq. (1). Section 5.1 reports that transferring time-averaged pressure fields to meta-grid B and integrating with an averaged pitch angle yields a mean global-frame drag error of 3.42e-5, compared with 0.94e-5 when instantaneous fields are integrated; the paper explicitly states that only the ship-fixed system will be used to judge the surrogate. Because the ML surrogate is trained on time-averaged fields and no model of the instantaneous pitch angle is described, the route-relevant global-frame error is at least the time-averaging error of 3.42e-5 unless the ML contribution compensates, for which no evidence is provided. The authors should either evaluate the full surrogate in the inertial frame or clearly restrict the claims to ship-fixed added resistance.
  2. [§4.1, Dataset split and mirrored test cases] The test set in Table 5 contains nine mirrored cases marked 'm' (e.g., b2-3m, d1-2m, e1-3m), which are exact reflections of simulated cases at the complementary encounter angle. The paper does not state whether the non-mirrored counterpart of each mirrored test case was excluded from the training and validation sets. If any such counterpart appears in training, the corresponding test errors are not independent because the mirrored pressure field is a deterministic transformation of the original. The authors should describe the random split explicitly and either exclude counterpart cases or report the test errors separately for mirrored and non-mirrored cases.
  3. [§5.2, Table 5, configuration selection] The reported mean error of 1.77e-5 corresponds to the configuration (meta-grid B, Int.-Mean, simultaneous training) that performs best on the same 18-case test table used for evaluation. Selecting the best of four configurations on the test set introduces selection bias, so the reported accuracy is optimistic relative to what would be obtained on a truly unseen configuration choice. The authors should either use a separate hold-out set for configuration selection or present the selection rule and its uncertainty transparently.
minor comments (4)
  1. [§1, Introduction] The sentence 'Thanks to the the now widely available hardware and software' contains a duplicated article; it should read 'Thanks to the now widely available hardware and software'.
  2. [§5.2, Table 5 caption and text] The symbol E'_D is introduced in the Table 5 caption but the text in Section 5.2 defines E_D and E_p; the notation should be unified so the ship-fixed drag error has a single name.
  3. [Appendix A, Table 6] Some numeric entries contain stray characters, e.g., '7 .765' in the c1-2 row, '4 .160' in the f3-2 row, and '2 .467' in the j4-1 row; these should be corrected.
  4. [References] The citation in Section 4.3 contains a stray question mark ('(?Wu et al. [2021], Schwarz et al. [2025b])'), and the reference to Muzaferija et al. contains a typo ('Syposium' instead of 'Symposium').

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported surrogate accuracy is evaluated on held-out sea states, not on fitted values or equations that reduce to their own inputs.

full rationale

The derivation chain is self-contained. The CAE/NN surrogate is trained on time-averaged pressure fields from 249 CFD cases and evaluated on 18 held-out parameter combinations (Section 4.1, Table 5); the reported mean drag error of 1.77e-5 is an out-of-sample error rather than a training residual or a fitted quantity renamed as a prediction. The pressure-to-force integration path is verified separately against CFD in Section 5.1, and the paper explicitly quantifies the global-frame penalty of time averaging (3.42e-5 versus 0.94e-5 relative to instantaneous integration). The later decision to evaluate the full surrogate only in ship-fixed coordinates is an acknowledged limitation of the claimed routing applicability, not a circular reduction. Self-citations to the authors' own CFD solver, motion model, Hilbert-window selection, and CAE/NN architecture choices provide methods and prior design context, but none of these citations defines the measured test metric or logically forces the reported outcome. No equation is defined in terms of its own target, and no fitted parameter is relabeled as a prediction. The malformed citation marker '(?Wu et al. [2021], Schwarz et al. [2025b])' in Section 4.3 is noted; it is a formatting/reference issue and does not carry any circularity load.

Assumptions & free parameters 9 free parameters · 8 assumptions · 0 invented entities

The central claim rests on the CFD data being representative and on several modeling choices (4DoF motion, virtual spring, regular waves, linear wave generation, mirroring symmetry, ship-fixed averaging). None of these is externally validated in the paper; the ML hyperparameters are additional free parameters that were chosen by the authors and affect the reported accuracy.

free parameters (9)
  • Latent space dimension = 10
    Default used for the final model; Section 5.3 compares 5, 10, and 20 and reports 10 as the best compromise.
  • Loss weights nu_reconst, nu_predict, nu_latent = 1, 1, 0
    Chosen manually in Section 4.3; the authors state that nu_latent values of 0.1 or higher cause training problems and nu_latent=1e-3 gives similar results to 0.
  • Mini-batch size = 2 for CAE and simultaneous training, 128 for regression
    Section 4.3 states larger mini-batches lead to poorer generalization.
  • Learning rate = 1e-3 (Adam)
    Set in Section 4.3 for all training runs.
  • Meta-grid B resolution = 128x64x96 cells
    Choice of interpolation grid affects pressure field accuracy; the finer grid B gives the best drag errors in Table 5.
  • Inverse-distance interpolation neighbor count = 5
    Section 3.1 defines the pressure transfer using the 5 closest CFD cells.
  • Early stopping patience = 50 epochs
    Used for CAE and regression training in Section 4.3.
  • LeakyReLU slope = 0.01
    Stated in Section 4.2 for all hidden layers.
  • Virtual spring stiffness per simulation
    Adjusted for each run to keep the spring-mass natural frequency below the wave frequency (Section 3). Not quantified; exceptions occur when wave speed matches ship speed, affecting the motion-pressure data.
assumptions (8)
  • domain assumption RANS k-omega MSST and VoF accurately represent wave-induced pressure fields
    Solver described in Section 3; no experimental validation of added resistance is presented in this paper.
  • domain assumption Added resistance is pressure-dominated and friction changes are negligible
    Assumption stated in Section 2.1 and checked in Table 4, which reports mean |Delta Cf/Cf,0| = 0.7 percent.
  • ad hoc to paper 4DoF motion with roll and yaw suppressed, plus a virtual spring, gives representative vessel response
    Section 3 states roll and yaw are suppressed and a virtual spring keeps the ship on course; exceptions occur when wave speed matches ship speed.
  • domain assumption Time-averaged pressure fields in ship-fixed coordinates are sufficient for surrogate modeling and routing
    Section 2.3 and 5.1-5.2 discard the instantaneous strategy, yet global-frame accuracy for the averaged strategy is worse (3.42e-5 versus 0.94e-5).
  • ad hoc to paper Mirroring pressure fields across the x'2 axis exactly represents beta+180 scenarios
    Section 2.2 flips fields for 0-160 degrees to obtain 200-360 degrees; requires hull symmetry and assumes no roll-induced asymmetry.
  • domain assumption Linear wave theory generates the incident waves
    Section 3 states arbitrary sea states are generated with an implicit forcing approach based on linear wave theory.
  • domain assumption Interpolation to the structured meta-grid preserves the pressure information needed for drag integration
    Section 3.1 and 5.1 describe inverse-distance interpolation; transfer errors are small on average but localized high-pressure peaks are smoothed.
  • ad hoc to paper The random train/validation/test split does not place mirrored twins in different splits in a way that leaks information
    Not stated in the paper; test cases in Table 5 include mirrored entries and if their originals are in training, the hull symmetry makes them near-duplicates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data-driven pressure field prediction for ships in regular sea states." pith.science (2026). https://pith.science/paper/FRYAGJLR

@misc{pith2026250506014,
  author       = {Pith},
  title        = {Pith review of: Data-driven pressure field prediction for ships in regular sea states},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FRYAGJLR}},
  note         = {Machine review of arXiv:2505.06014}
}
read the original abstract

Merchant shipping is responsible for more than 90% of the global trade and has a significant environmental impact, accounting for over 2% of global greenhouse gas emissions. Therefore, fuel-saving measures are becoming increasingly important in reducing the ecological footprint and increasing the fuel efficiency of maritime transport. Routing optimization systems, which require a rapid prediction of ambient-dependent fuel consumption, represent an essential pillar here, e.g. to reduce added resistances due to seaways and/or wind. The paper aims to predict the added resistance due to seaways. In contrast to conventional methods the goal is achieved by surrogate modeling of the entire pressure fields. To this end, an online/offline-procedure is applied to an exemplarily free-floating container vessel. The online approach to be trained consists of two building blocks, namely a convolutional autoencoder (CAE)-based order reduction step and a neural network-based (NN) regression step that links the reduced space of the autoencoder with three control parameters that describe the sea state (wave height/steepness, encounter angle and wave length). Training data is obtained from time-averaged values for simulating instantaneous ship motion and pressure fields. During the offline phase, the combined CAE/NN is trained to capture the time-averaged pressure fields for a variety of sea-state conditions. During ship operation (online phase), the surrogate model predicts the three-dimensional pressure fields in response to sea state conditions, projects the pressure fields onto the ship hull, and integrates the corresponding resistances to guide the route. The evaluation of the method shows promising results for the different building blocks and the concept could therefore represent an attractive approach for cost-effective surrogate modeling of complex multiphase flow fields.

Figures

Figures reproduced from arXiv: 2505.06014 by the authors.

Figure 1
Figure 1. Investigated KRISO container vessel (Λ = 31 [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the investigated scenarios and coordinate systems characterized by the global coordinates [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. CFD simulations are performed due to the according parameter spectrum (top). Three different [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Views of the unstructured CFD grid for the exemplary simulated KRISO container ship (KCS). [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Views of the employed two meta-grids. via an inverse distance approach, where the pressure values of the 5 closest CFD cells are used to determine the pressure in each cell of the meta-grid, cf. Figures 6 (a) and (b). Mind that locations inside the ship hull are marked…
Figure 6
Figure 6. Figure 6: Comparison of the simulated (a) and interpolated (b, meta-grid A) instantaneous pressure fields of an [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: For the separate training, the CAE is trained with the Adam optimizer (Kingma and Ba [2015]) and a [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 7
Figure 7. Figure 7: Offline and online phases of the neural network based surrogate model. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Analysis of the influences of the meta-grid data transfer by comparing the additional drag coefficients [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Analysis of the influences of the meta-grid data transfer by comparing the additional drag coefficients [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Comparison of pressure fields obtained from the surrogate model on meta-grid B with their [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Comparison of transferred CFD pressure fields on meta-grid A for case 8 with neighboring wavelength [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 6 canonical work pages

  1. [2]

    doi:10.1063/5.0020526. Y. Hou, H. Li, H. Chen, W. Wei, J. Wang, and Y. Huang. A novel deep U-Net-LSTM framework for time- sequenced hydrodynamics prediction of the SUBOFF AFF-8. Engineering Applications of Computational Fluid Mechanics, 16(1):630–645,

    doi:10.1063/5.0020526

  2. [4]

    doi:10.1007/s00158-022-03338-2. 14 Table 6: Overview of all local drag coefficients based on CFD simulations (∆ C′ D) as well as reconstructed values from time averaged (∆C′ D,a) and instantaneous flow data (∆C′ D,i) for two meta-grids A and B. The mean drag error E′ D :=||C′ D−C′ D,a,i||1 is given at the end of the table. Run ∆C′ D/10−5 ∆C′ D,a/10−5 ∆C′ ...

    work page doi:10.1007/s00158-022-03338-2

  3. [5]

    doi:https://doi.org/10.1016/j.ast.2022.107629

    ISSN 1270-9638. doi:https://doi.org/10.1016/j.ast.2022.107629. URL https://www.sciencedirect.com/science/article/ 16 pii/S1270963822003030. S. Liu, A. Papanikolaou, and G. Zaraphonitis. Prediction of added resistance of ships in waves.Ocean Engineering, 38(4):641–650,

    work page arXiv 2022

  4. [14]

    doi:https://doi.org/10.1016/j.compfluid.2018.07.021

    ISSN 00457930. doi:https://doi.org/10.1016/j.compfluid.2018.07.021. S. V¨ olkner, J. Brunswig, and T. Rung. Analysis of Non-Conservative Interpolation Techniques in Overset Grid Finite-Volume Methods. Computers & Fluids , 148:39–55,

    doi:10.1016/j.compfluid.2018.07.021 2018

  5. [15]

    doi:10.1016/j.compfluid.2017.02.010. P. Wu, S. Gong, K. Pan, F. Qiu, W. Feng, and C. Pain. Reduced order model using convolutional auto-encoder with self-attention. Physics of Fluids , 33(7):077107,

    work page doi:10.1016/j.compfluid.2017.02.010 2017

  6. [1994]

    doi:10.2514/3.12149. B. Mi and W. Cheng. Intelligent aerodynamic modelling method for steady/unsteady flow fields of airfoils driven by flow field images based on modified u-net neural network. Engineering Applications of Computational Fluid Mechanics, 19(1):2440075,

    doi:10.2514/3.12149

  7. [2008]

    Accessed: 2023-09-14. A. Solera-Rico, C. Sanmiguel Vila, M. G´ omez-L´ opez, Y. Wang, A. Almashjary, S. Dawson, , and R. Vinuesa.β- variational autoencoders and transformers for reduced-order modelling of fluid flows. Nature Communications, 15, 02

    work page 2023

  8. [2009]

    Schwarz, P

    H. Schwarz, P. P. Lin, J.-P. M. Zemke, and T. Rung. Disentangled latent spaces for reduced order models using deterministic autoencoders. arXiv preprint arXiv:2502.14679 , 2025a. URL https://arxiv.org/abs/2502. 14679. H. Schwarz, M. ¨Uberr¨ uck, J.-P. M. Zemke, and T. Rung. Machine learning based prediction of ditching loads. AIAA Journal, 63(5):1835–1854...

    arXiv

Show all 16 references
  1. [2017]

    doi:10.1080/09377255.2017.1348654. X. Luo-Theilen and T. Rung. Numerical Analysis of the Installation Procedures of Offshore Structures. Ocean Engineering, 179:116–127,

    2017

  2. [2019]

    doi:10.1016/j.oceaneng.2019.03.004. A. L. Maas, A. Y. Hannun, and A. Y. Ng. Rectifier nonlinearities improve neural network acoustic models. In Proceedings of the 30th International Conference on Machine Learning , volume 28,

    2019 doi

  3. [2020]

    doi:10.1063/5.0012906

    ISSN 1070-6631. doi:10.1063/5.0012906. F. Chollet et al. Keras. https://keras.io,

    doi

  4. [2021]

    doi:10.1063/5.0051155. 17

    doi

  5. [2022]

    URL https://doi.org/10.1080/ 19942060.2022.2030802

    doi:10.1080/19942060.2022.2030802. URL https://doi.org/10.1080/ 19942060.2022.2030802. D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR),

    2022

  6. [2023]

    doi:10.1063/5.0156963. J. L. Lumley. The structure of inhomogeneous turbulent flows. Atmospheric turbulence and radio wave propagation, pages 166–178,

    doi

  7. [2024]

    doi:10.1038/s41467-024-45578-4. R. Swischuk, L. Mainini, B. Peherstorfer, and K. Willcox. Projection-based model reduction: Formu- lations for physics-based machine learning. Computers & Fluids , 179:704–717,

    doi

  8. [2025]

    URL https://doi.org/10.1080/ 19942060.2024.2440075

    doi:10.1080/19942060.2024.2440075. URL https://doi.org/10.1080/ 19942060.2024.2440075. M. Milano and P. Koumoutsakos. Neural network modeling for near wall turbulent flow.Journal of Computational Physics, 182(1):1–26,

    2024

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.