Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Quantifying data needs in surrogate modeling for flow fields in two-dimensional stirred tanks with physics-informed neural networks

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Six velocity labels train a stirred-tank flow surrogate to ~3% error

desk verdict A careful, honestly reported benchmark on data needs for PINNs in a 2D stirred-tank setup, whose headline '3% across Re 50-5000' overstates what the paper's own Figure 15 shows at Re=50. read the letter →

arxiv 2507.11640 v2 pith:XZ4XKKLT submitted 2025-07-15 cs.CE

classification cs.CE
keywords physics-informedneuralnetworkssurrogatemodelingstirredtankNavier-StokesequationsdatarequirementsReynoldsnumberinterpolationlabel-efficientlearning
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

This paper measures how much labeled data a vanilla physics-informed neural network actually needs to become a usable surrogate for steady laminar flow in a two-dimensional stirred tank. A vanilla PINN is a neural network trained by minimizing the residual of the governing Navier–Stokes equations along with any available data, without extra tricks such as domain decomposition or hard boundary enforcement. The central result is that six velocity labels suffice: three points placed in the high-error region just beyond the impeller tips at each of the two boundary Reynolds numbers, Re=50 and Re=5000, give roughly 3% normalized ℓ1 errors for velocity and pressure at unseen interior Reynolds numbers such as Re=1000. Replacing those labels with a simple polynomial approximation of the velocity profile works nearly as well, giving about 2.5% errors across most of the range. The reason to care is practical: a classical supervised network needed thousands of labels to approach the same accuracy, so physics constraints can cut simulation or measurement effort by orders of magnitude.

What carries the argument

The load-bearing mechanism is the composite loss of a physics-informed neural network: it combines the steady incompressible Navier–Stokes PDE residual evaluated at collocation points, the Dirichlet boundary-condition residuals at the tank wall and stirrer blades, and a data residual against the labeled points. Because the PDE residual constrains every output to be a solution of the physics, the network can interpolate across Reynolds number almost without data; the small data term only has to fix the parametric endpoints and the localized high-error zone. The targeted-label recipe consists of three ingredients: placing velocity labels in the annular region 0.04 m ≤ r ≤ 0.07 m just after the impeller tips, where vanilla PINN errors concentrate; reducing the domain to a symmetric quadrant with an extra symmetry boundary residual; and, in the approximate-data variant, using a piecewise polynomial for the velocity magnitude, scaled by the tip speed, as a stand-in for real velocity labels.

What would settle it

Evaluate the same six-label, two-endpoint PINN at finely resolved intermediate Reynolds numbers (for example Re=250, 750, and 1500) on a converged mesh; if the normalized ℓ1 velocity or pressure error at any of those points exceeds the roughly 6–7% seen at Re=50 rather than staying near 3%, the boundary-only labeling strategy is not working in the interior. The claim would also be tested by applying the recipe to a parameter range known to contain a flow transition, where endpoint labels alone should fail.

Watch

Extended reading notes

Core claim

The authors claim that for the 2D stirred-tank benchmark, the solution family over Reynolds numbers 50–5000 is well captured by a vanilla PINN whose data loss is fed only by labels at the extremes of the parameter range. With three targeted velocity labels per boundary Re value, placed where r is between 0.04 m and 0.07 m—the annulus just outside the impeller where base PINN errors peak—and using the problem's four-fold symmetry to train on one quadrant, the model achieves mean normalized errors of 3.15% (velocity) and 2.93% (pressure) at Re=1000. Without labels the same architecture gives 11.00% velocity and 12.35% pressure error at that Reynolds number. If the labels are replaced by 256 points sampled from the paper's polynomial approximation of the velocity magnitude profile, errors settle around 2.5% for both fields for Re between 100 and 6000, including extrapolation beyond the training range. The paper also reports that a supervised neural network reaches comparable accuracy only with on the order of 10,000 labels per Reynolds number, and that a boundary-informed network omitting the PDE residual needs labels from at least four Reynolds numbers.

Load-bearing premise

The six-datapoint claim assumes the solution changes smoothly enough with Reynolds number that physics constraints can fill in the whole interior range from endpoints Re=50 and Re=5000 alone; nothing in the paper proves that smoothness or bounds the interpolation error.

Editorial extensions

If this is right

  • With one high-fidelity CFD solution at each end of the Reynolds range, a vanilla PINN can produce a surrogate accurate to about 3% for interior Reynolds numbers, at least for smooth steady laminar regimes.
  • Approximate, physics-derived labels are nearly as good as expensive simulation labels, so users can avoid high-fidelity data generation when a rough velocity profile is known.
  • The comparison quantifies a practical guideline: choose a classical NN or BINN only when thousands of labels are available; otherwise the PDE residual earns its extra training cost.
  • As few as eight labels per Reynolds number already reduce vanilla PINN errors from about 11–12% to about 6–7%, and 64 labels per Reynolds number reach the point of diminishing returns.

Reading between the lines

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

  • Inference: the six-label recipe should transfer to other parametric fluid problems only when the solution manifold is smooth and non-bifurcating in the parameter; the paper's own robustness results show higher variance for models trained at just two Reynolds numbers, so a user should verify smoothness before relying on it.
  • Inference: the success of placing labels where the base model errs most suggests an active-learning loop—train without data, locate the largest residuals, label those few points—that could yield nearly the same accuracy with even fewer or better-chosen labels.
  • Inference: the approximate-profile result implies that an analytic or empirical prior about the flow can substitute for data labels, which may matter for three-dimensional stirred-tank surrogates, where high-fidelity labels are the main bottleneck.
  • Inference: because pressure labels are not needed, the method is compatible with experimental measurement campaigns that only resolve velocity.
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

4 major / 5 minor

Summary. The paper quantifies how many labeled data points are needed by a vanilla physics-informed neural network (PINN) to build a surrogate model for steady 2D stirred-tank flow over Reynolds numbers from 50 to 5000, comparing it with classical supervised neural networks and boundary-informed neural networks (BINNs). Using finite element reference solutions as ground truth, the authors report that adding a small number of labels greatly improves the vanilla PINN, that only boundary values of the Reynolds range are needed when PDE residuals are present, that six strategically placed velocity labels give about 3% error at Re=1000, and that approximate velocity labels give about 2.5% error over most of the range. The body of the paper explicitly acknowledges degraded accuracy at Re=50 and below, but this caveat is not reflected in the abstract and conclusion.

Significance. If the central quantitative claims survive revision, the paper is a useful, carefully controlled benchmark: it uses an independent high-resolution FEM reference, reports architecture, sampling, and loss-weight details in the appendix, includes ten-seed robustness experiments for random label placement, and compares PINN, BINN, and supervised NN under a common architecture. The main qualitative finding, that PINNs reduce label requirements for parametric surrogates, is well supported. The headline accuracy claim is currently too broad and needs qualification, and the six-datapoint result lacks a reported variance estimate.

major comments (4)
  1. [Abstract, Section 4.5, Figure 15] The claim of 'prediction errors around 3% across Reynolds numbers from 50 to 5000 using as few as six datapoints' is not supported by the paper's own evaluation. The targeted-label model achieves 3.15% velocity and 2.93% pressure error at Re_test=1000, but Figure 15 shows velocity error at Re_test=50 of roughly 7%, and Section 4.6 explicitly states that errors for Re_test=30 and Re_test=50 remain significantly larger than for other values because of the impeller-tip discontinuity. The abstract and conclusion should either quote the Re>=100 range or report the full-range error honestly.
  2. [Section 4.4, Section 4.5] The six-datapoint result is reported as a single training run, while the paper's own robustness analysis for the analogous two-Re configuration (64 labels per Re_train) shows standard deviations reaching 15-20% of the mean error for unseen Re values. Since strategic label placement is the basis of the headline data-efficiency claim, the authors should re-run the 3-labels-per-Re configuration over multiple seeds and report mean and spread. Without this, the reader cannot distinguish a robust finding from a favorable seed.
  3. [Section 4.6, Eq. (22)] The approximate-label experiment depends on the polynomial profile with R*=0.0875 inherited from prior work, but the paper offers only a visual comparison of the profile and no quantitative error of the approximate labels against the FEM reference in the label region. Because the approximate-label claim (around 2.5%) is a headline contribution, the authors should quantify how far the approximate labels deviate from the true profiles at the actual label locations and for the Re values sampled.
  4. [Section 4.3, Section 4.4] The conclusion that labels at only the two boundary Re values suffice to train an accurate PINN across the interior range rests on an interpolation assumption that the solution manifold is smooth in Re. The robustness standard deviations in Figure 12b are a partial proxy, but no uncertainty quantification over the Re axis is provided. The authors should either add a brief justification for the assumed smoothness or restrict the claim to the demonstrated range.
minor comments (5)
  1. [Section 4.6, first paragraph] The text contains a typo: 'sirred tank' should be 'stirred tank'.
  2. [Section 3.2, Eq. (16)] The phrase 'approaching approaching' contains a duplicated word and should be corrected.
  3. [Figure 18 caption] The caption describes panel (b) as showing 'velocity prediction error distributions', but the panel shows pressure errors; the caption should say 'pressure'.
  4. [Table 3 and Figure 4] The training-time comparison uses hyperparameters tuned for the PINN, which is disclosed in the text but should be restated in the table or figure caption so that the timing comparison is not misinterpreted as a hardware-independent measure.
  5. [Data availability statement] Making the FEM reference solutions and trained model outputs available in a repository, rather than only 'upon reasonable request', would strengthen reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: the six-datapoint and approximate-label results are empirical benchmarks against independent FEM solutions; at most a minor non-load-bearing self-citation for R*.

full rationale

The paper's central quantitative claims are not derived from their own inputs: the 3.15% velocity / 2.93% pressure errors with six targeted labels (Section 4.5, Fig. 13) and the ~2.5% errors with approximate labels (Section 4.6, Figs. 17-18) are measured against high-fidelity FEM reference solutions on a separate, much finer mesh (Section 3.3), so they are not equal to the training data by construction. The targeted-label experiment places labels in a region of known high error (r in [0.04,0.07]), but this is an active-learning choice, and accuracy is still evaluated on unseen test data. The approximate-label experiment uses an explicit polynomial ansatz (Eq. 22) with R*=0.0875 m, stated in the text and attributed to the authors' prior work [29]; the profile is an input to the PINN, not a rediscovery of the output, and the subsequent error evaluation is independent. The only self-referential element is the citation of [29] for the choice of R*, but the value is disclosed and is not a predicted quantity, so it does not make the claim circular. The abstract's unqualified 'across Reynolds numbers from 50 to 5000' is internally inconsistent with Fig. 15, which shows ~7% velocity error at Re=50; that is a correctness/consistency concern, not a circularity, and does not affect this verdict.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a small number of hand-picked parameters (R*, label region, loss weights) and on the unverified assumptions that the 2D steady symmetric model is representative of the target flows and that the coarse-mesh labels and fine-mesh validation are reliable. No new physical entities are introduced.

free parameters (3)
  • R* (velocity profile tuning parameter) = 0.0875 m
    Appears in Eq. 22 for the approximate velocity profile; the approximate-label experiment (Section 4.6) depends on this value, which is inherited from prior work (Ref. [29]) and not derived in this paper.
  • Targeted label region (r bounds) = 0.04 m to 0.07 m
    In Section 4.5, labels are placed in the annular region just after the impeller tips based on prior experience about where PINN errors are largest; the six-datapoint result depends on this hand-picked region.
  • Data loss weight alpha_data = 25 (targeted) / 9 (approximate)
    Loss scaling factors in Tables 6 and 7 are chosen by hand or grid search, not derived; the balance between physics and data terms affects the reported error levels.
assumptions (5)
  • domain assumption Steady incompressible Navier-Stokes equations (Eq. 11) adequately describe the 2D stirred-tank flow.
    Used as the physics loss in all PINN models; if the flow is not steady or has unsteady features, the residual is incorrect (Section 3.2).
  • domain assumption The solution is symmetric about the axes, allowing the domain to be reduced to Omega_sym with the continuity BC of Eq. 16.
    Sections 4.5 and 4.6 rely on this symmetry; Figure 1b shows the reduced domain and Eq. 24-25 define the transformations.
  • domain assumption The high-fidelity FEM solution on the fine mesh is an accurate ground truth for validation.
    The fine mesh is stated to be at the Kolmogorov scale for Re=5000 but under-resolved for Re=6000 (Section 3.3.1), so validation at the highest Re is approximate.
  • domain assumption Labels from the coarse mesh are accurate enough to train a model that is then validated against the fine mesh.
    Training labels come from a 73k-element mesh, which is much coarser than the 4.9M-element test mesh; the paper does not quantify the coarse-mesh error.
  • ad hoc to paper The polynomial approximation in Eq. 22 with R*=0.0875 is a valid surrogate for the true velocity profile in the label region.
    The form and parameter are cited to prior work [29] and Figure 16 shows it diverges near the wall; the approximate-label result depends on this approximation being accurate in the impeller-tip region.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantifying data needs in surrogate modeling for flow fields in two-dimensional stirred tanks with physics-informed neural networks." pith.science (2026). https://pith.science/paper/XZ4XKKLT

@misc{pith2026250711640,
  author       = {Pith},
  title        = {Pith review of: Quantifying data needs in surrogate modeling for flow fields in two-dimensional stirred tanks with physics-informed neural networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XZ4XKKLT}},
  note         = {Machine review of arXiv:2507.11640}
}
abstract

Stirred tanks are vital in chemical and biotechnological processes, particularly as bioreactors. Although computational fluid dynamics (CFD) is widely used to model the flow in stirred tanks, its high computational cost$-$especially in multi-query scenarios for process design and optimization$-$drives the need for efficient data-driven surrogate models. However, acquiring sufficiently large datasets can be costly. Physics-informed neural networks (PINNs) offer a promising solution to reduce data requirements while maintaining accuracy by embedding underlying physics into neural network (NN) training. This study quantifies the data requirements of vanilla PINNs for developing surrogate models of a flow field in a 2D stirred tank. We compare these requirements with classical supervised neural networks and boundary-informed neural networks (BINNs). Our findings demonstrate that surrogate models can achieve prediction errors around 3% across Reynolds numbers from 50 to 5000 using as few as six datapoints. Moreover, employing an approximation of the velocity profile in place of real data labels leads to prediction errors of around 2.5%. These results indicate that even with limited or approximate datasets, PINNs can be effectively trained to deliver high accuracy comparable to high-fidelity data.

Figures

Figures reproduced from arXiv: 2507.11640 by the authors.

Figure 1
Figure 1. Depiction of a simplified three-dimensional stirred tank geometry (A) and the simplified [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Geometrical dimensions of the 2D problem domain. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. High-fidelity solution computed on a finite element mesh with 4,914,370 elements for velocity [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Effect of the number of labeled points per Re [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Normalized velocity magnitude error for predictions by a classical NN lacking physics in [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Comparison of mean normalized errors in velocity and pressure for physics-based vanilla [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Comparison of PINN model accuracy with labels for both velocity and pressure versus [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Normalized velocity magnitude error at Re [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Comparison of mean normalized errors in velocity and pressure for a labeled data-enhanced [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Prediction accuracy of surrogate models with varying Re [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Velocity magnitude field (a) and normalized velocity magnitude error (b) at Re [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Robustness of PINN models with respect to random label placement for two different [PITH_FULL_IMAGE:figures/full_fig_p014_12.png]
Figure 13
Figure 13. Figure 13: Mean velocity and pressure errors as a function of the number of labeled points per Re [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 14
Figure 14. Figure 14: Comparison of label allocation strategies in training datasets. Locations of strategically [PITH_FULL_IMAGE:figures/full_fig_p015_14.png]
Figure 15
Figure 15. Figure 15: Comparison of mean normalized errors for velocity and pressure between models with [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]
Figure 16
Figure 16. Figure 16: Velocity magnitude profile along the reactor radius at [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]
Figure 17
Figure 17. Figure 17: (a) Violin plot showing velocity magnitude prediction error distributions across all Re [PITH_FULL_IMAGE:figures/full_fig_p017_17.png]
Figure 18
Figure 18. Figure 18: (a) Violin plot showing pressure prediction error distributions across all Re [PITH_FULL_IMAGE:figures/full_fig_p018_18.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 25 canonical work pages

  1. [1]

    Physics Informed Neural Net- works for Modeling of 3D Flow-Thermal Problems with Sparse Domain Data

    Saakaar Bhatnagar, Andrew Comerford, and Araz Banaeizadeh. Physics Informed Neural Net- works for Modeling of 3D Flow-Thermal Problems with Sparse Domain Data. Journal of Machine Learning for Modeling and Computing , 2024. ISSN 2689-3967. doi: 10.1615/JMachLearnModel- Comput.2024051540

  2. [2]

    Hemo- dynamics modeling with physics-informed neural networks: A progressive boundary complexity approach

    Xi Chen, Jianchuan Yang, Xu Liu, Yong He, Qiang Luo, Mao Chen, and Wenqi Hu. Hemo- dynamics modeling with physics-informed neural networks: A progressive boundary complexity approach. Computer Methods in Applied Mechanics and Engineering , 438:117851, April 2025. ISSN 0045-7825. doi: 10.1016/j.cma.2025.117851

  3. [3]

    Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s Next

    Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s Next. Journal of Scientific Computing , 92(3):88, July 2022. ISSN 1573-

  4. [4]

    Eulerian formulation of the tensor-based morphology equations for strain-based blood damage modeling

    Nico Dirkes, Fabian Key, and Marek Behr. Eulerian formulation of the tensor-based morphology equations for strain-based blood damage modeling. Computer Methods in Applied Mechanics and Engineering, 426:116979, June 2024. ISSN 0045-7825. doi: 10.1016/j.cma.2024.116979. 21

  5. [5]

    Application of CFD to Analyze the Hydrodynamic Behaviour of a Biore- actor with a Double Impeller

    Mohammadreza Ebrahimi, Melih Tamer, Ricardo Martinez Villegas, Andrew Chiappetta, and Farhad Ein-Mozaffari. Application of CFD to Analyze the Hydrodynamic Behaviour of a Biore- actor with a Double Impeller. Processes, 7(10):694, October 2019. ISSN 2227-9717. doi: 10.3390/pr7100694

  6. [6]

    Physics-informed neural networks for solving Reynolds-averaged Navier-Stokes equations

    Hamidreza Eivazi, Mojtaba Tahani, Philipp Schlatter, and Ricardo Vinuesa. Physics-informed neural networks for solving Reynolds-averaged Navier-Stokes equations. Physics of Fluids , 34(7): 075117, July 2022. ISSN 1070-6631, 1089-7666. doi: 10.1063/5.0095270

  7. [7]

    Faroughi, Nikhil M

    Salah A. Faroughi, Nikhil M. Pawar, C´ elio Fernandes, Maziar Raissi, Subasish Das, Nima K. Kalantari, and Seyed Kourosh Mahjour. Physics-Guided, Physics-Informed, and Physics-Encoded Neural Networks and Operators in Scientific Computing: Fluid and Solid Mechanics. Journal of Computing and Information Science in Engineering , pages 1–45, January 2024. ISS...

  8. [8]

    Garcia-Ochoa, V

    F. Garcia-Ochoa, V. E. Santos, and E. Gomez. 2.15 - Stirred Tank Bioreactors. In Murray Moo- Young, editor, Comprehensive Biotechnology (Second Edition) , pages 179–198. Academic Press, Burlington, January 2011. ISBN 978-0-08-088504-9. doi: 10.1016/B978-0-08-088504-9.00108-2

Show all 44 references
  1. [9]

    Geometry-aware PINNs for Turbulent Flow Prediction, December 2024

    Shinjan Ghosh, Julian Busch, Georgia Olympia Brikis, and Biswadip Dey. Geometry-aware PINNs for Turbulent Flow Prediction, December 2024

  2. [10]

    A physics- informed deep learning framework for inversion and surrogate modeling in solid mechanics

    Ehsan Haghighat, Maziar Raissi, Adrian Moure, Hector Gomez, and Ruben Juanes. A physics- informed deep learning framework for inversion and surrogate modeling in solid mechanics. Com- puter Methods in Applied Mechanics and Engineering , 379:113741, June 2021. ISSN 0045-7825. d...

  3. [11]

    C. Haringa. Through the Organism’s eyes: The interaction between hydrodynamics and metabolic dynamics in industrial-scale fermentation processes. 2017. doi: 10.4233/uuid:441ec955-cd8d-4ae0- b2f0-98fbf91a570a

  4. [12]

    Flow field reconstruction from sparse sensor mea- surements with physics-informed neural networks

    Mohammad Yasin Hosseini and Yousef Shiri. Flow field reconstruction from sparse sensor mea- surements with physics-informed neural networks. Physics of Fluids , 36(7):073606, July 2024. ISSN 1070-6631. doi: 10.1063/5.0211680

  5. [13]

    A General Review of the Current Development of Mechanically Agitated Vessels

    Marek Jaszczur and Anna M lynarczykowska. A General Review of the Current Development of Mechanically Agitated Vessels. Processes, 8(8):982, August 2020. ISSN 2227-9717. doi: 10.3390/pr8080982

  6. [14]

    NSFnets (Navier-Stokes Flow nets): Physics-informed neural networks for the incompressible Navier-Stokes equa- tions

    Xiaowei Jin, Shengze Cai, Hui Li, and George Em Karniadakis. NSFnets (Navier-Stokes Flow nets): Physics-informed neural networks for the incompressible Navier-Stokes equa- tions. Journal of Computational Physics , 426:109951, February 2021. ISSN 00219991. doi: 10.1016/j.jcp.20...

  7. [15]

    Joshi, Nandkishor K

    Jyeshtharaj B. Joshi, Nandkishor K. Nere, Chinmay V. Rane, B. N. Murthy, Channamallikarjun S. Mathpati, Ashwin W. Patwardhan, and Vivek V. Ranade. CFD simulation of stirred tanks: Com- parison of turbulence models. Part I: Radial flow impellers. The Canadian Journal of Chemica...

  8. [16]

    Fotiadis

    Isaac Elias Lagaris, Aristidis Likas, and Dimitrios I. Fotiadis. Artificial neural networks for solving ordinary and partial differential equations.IEEE Transactions on Neural Networks, 9(5):987–1000,

  9. [17]

    Lorenzen, A

    F. Lorenzen, A. Zargaran, and U. Janoske. Potential of physics-informed neural networks for solving fluid flow problems with parametric boundary conditions. Physics of Fluids , 36(3):037143, March 2024. ISSN 1070-6631. doi: 10.1063/5.0193952

  10. [18]

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence , 3(3):218–229, March 2021. ISSN 2522-5839. doi: 10.1038/s42256- 021-...

  11. [19]

    McClenny and Ulisses M

    Levi D. McClenny and Ulisses M. Braga-Neto. Self-adaptive physics-informed neural net- works. Journal of Computational Physics , 474:111722, February 2023. ISSN 0021-9991. doi: 10.1016/j.jcp.2022.111722

  12. [20]

    Finite basis physics-informed neural networks (FBPINNs): A scalable domain decomposition approach for solving differential equa- tions

    Ben Moseley, Andrew Markham, and Tarje Nissen-Meyer. Finite basis physics-informed neural networks (FBPINNs): A scalable domain decomposition approach for solving differential equa- tions. Advances in Computational Mathematics , 49(4):62, August 2023. ISSN 1019-7168, 1572-

  13. [21]

    Raissi, P

    M. Raissi, P. Perdikaris, and G.E. Karniadakis. Physics-informed neural networks: A deep learn- ing framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics , 378:686–707, February 2019. ISSN 002199...

  14. [22]

    Hidden fluid me- chanics: Learning velocity and pressure fields from flow visualizations

    Maziar Raissi, Alireza Yazdani, and George Em Karniadakis. Hidden fluid me- chanics: Learning velocity and pressure fields from flow visualizations. Sci- ence, 367(6481):1026–1030, 2020. doi: 10.1126/science.aaw4741. URL https://www.science.org/doi/abs/10.1126/science.aaw4741

  15. [23]

    CFD simulation of a Rushton turbine stirred-tank using open-source software with critical evaluation of MRF-based rotation modeling

    Alfred Reid, Riccardo Rossi, Ciro Cottini, and Andrea Benassi. CFD simulation of a Rushton turbine stirred-tank using open-source software with critical evaluation of MRF-based rotation modeling. Meccanica, July 2024. ISSN 1572-9648. doi: 10.1007/s11012-024-01824-z

  16. [24]

    Rosseburg, J

    A. Rosseburg, J. Fitschen, J. Wutz, T. Wucherpfennig, and M. Schl¨ uter. Hydrodynamic inhomo- geneities in large scale stirred tanks – Influence on mixing time. Chemical Engineering Science , 188:208–220, October 2018. ISSN 0009-2509. doi: 10.1016/j.ces.2018.05.008

  17. [25]

    PFNN: A penalty-free neural network method for solving a class of second-order boundary-value problems on complex geometries

    Hailong Sheng and Chao Yang. PFNN: A penalty-free neural network method for solving a class of second-order boundary-value problems on complex geometries. Journal of Computational Physics , 428:110085, 2021. ISSN 00219991. doi: 10.1016/j.jcp.2020.110085

  18. [26]

    Jagtap, and George Em Karniadakis

    Khemraj Shukla, Ameya D. Jagtap, and George Em Karniadakis. Parallel physics-informed neural networks via domain decomposition. Journal of Computational Physics , 447:110683, December

  19. [27]

    Steinfurth, A

    B. Steinfurth, A. Hassanein, N. A. K. Doan, and F. Scarano. Physics-informed neural networks for dense reconstruction of vortex rings from particle tracking velocimetry. Physics of Fluids , 36 (9):095110, September 2024. ISSN 1070-6631. doi: 10.1063/5.0212585

  20. [28]

    Shape-optimization of extrusion-dies via parameterized physics-informed neural networks

    Steffen Tillmann, Daniel Hilger, Norbert Hosters, and Stefanie Elgeti. Shape-optimization of extrusion-dies via parameterized physics-informed neural networks. PAMM, 23(4):e202300203,

  21. [29]

    A model hierarchy for predicting the flow in stirred tanks with physics-informed neu- ral networks

    Veronika Tr´ avn ´ ıkov´ a, Daniel Wolff, Nico Dirkes, Stefanie Elgeti, Eric von Lieres, and Marek Behr. A model hierarchy for predicting the flow in stirred tanks with physics-informed neu- ral networks. Advances in Computational Science and Engineering , 2(2):91–129, 2024. d...

  22. [30]

    PACMANN: Point Adaptive Collocation Method for Artificial Neural Networks, November 2024

    Coen Visser, Alexander Heinlein, and Bianca Giovanardi. PACMANN: Point Adaptive Collocation Method for Artificial Neural Networks, November 2024

  23. [31]

    Understanding and Mitigating Gradient Flow Pathologies in Physics-Informed Neural Networks

    Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and Mitigating Gradient Flow Pathologies in Physics-Informed Neural Networks. SIAM Journal on Scientific Computing , 43(5): A3055–A3081, January 2021. ISSN 1064-8275. doi: 10.1137/20M1318043

  24. [32]

    An Expert’s Guide to Training Physics-informed Neural Networks, August 2023

    Sifan Wang, Shyam Sankaran, Hanwen Wang, and Paris Perdikaris. An Expert’s Guide to Training Physics-informed Neural Networks, August 2023. arXiv:2308.08468

  25. [33]

    Gradient Alignment in Physics-informed Neural Networks: A Second-Order Optimization Perspective, February 2025

    Sifan Wang, Ananyae Kumar Bhartari, Bowen Li, and Paris Perdikaris. Gradient Alignment in Physics-informed Neural Networks: A Second-Order Optimization Perspective, February 2025

  26. [34]

    A comprehensive study of non- adaptive and residual-based adaptive sampling for physics-informed neural networks

    Chenxi Wu, Min Zhu, Qinyang Tan, Yadhu Kartha, and Lu Lu. A comprehensive study of non- adaptive and residual-based adaptive sampling for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering , 403:115671, January 2023. ISSN 0045-7825. doi: 10...

  27. [35]

    Hy- brid Modeling of Fed-Batch Cell Culture Using Physics-Informed Neural Network

    Shu Yang, William Fahey, Brendha Truccollo, Jill Browning, Reza Kamyar, and Huiyi Cao. Hy- brid Modeling of Fed-Batch Cell Culture Using Physics-Informed Neural Network. Industrial & Engineering Chemistry Research, September 2024. ISSN 0888-5885. doi: 10.1021/acs.iecr.4c01459

  28. [36]

    MultiAdam: Parameter-wise scale-invariant optimizer for multiscale training of physics-informed neural net- works

    Jiachen Yao, Chang Su, Zhongkai Hao, Songming Liu, Hang Su, and Jun Zhu. MultiAdam: Parameter-wise scale-invariant optimizer for multiscale training of physics-informed neural net- works. In Proceedings of the 40th International Conference on Machine Learning , volume 202 of I...

  29. [37]

    Non-intrusive reduced-order modeling for fluid prob- lems: A brief review

    Jian Yu, Chao Yan, and Mengwu Guo. Non-intrusive reduced-order modeling for fluid prob- lems: A brief review. Proceedings of the Institution of Mechanical Engineers, Part G: Jour- nal of Aerospace Engineering , 233(16):5896–5912, December 2019. ISSN 0954-4100. doi: 10.1177/095...

  30. [38]

    Gambaruto

    Chengxi Zeng, Tilo Burghardt, and Alberto M. Gambaruto. Training dynamics in Physics- Informed Neural Networks with feature mapping, February 2024

  31. [39]

    Physics-Informed Neural Networks with Complementary Soft and Hard Constraints for Solving Complex Boundary Navier-Stokes Equations, November 2024

    Chuyu Zhou, Tianyu Li, Chenxi Lan, Rongyu Du, Guoguo Xin, Pengyu Nan, Hangzhou Yang, Guoqing Wang, Xun Liu, and Wei Li. Physics-Informed Neural Networks with Complementary Soft and Hard Constraints for Solving Complex Boundary Navier-Stokes Equations, November 2024. 24

  32. [1998]

    doi: 10.1109/72.712178

    ISSN 10459227. doi: 10.1109/72.712178

  33. [2021]

    doi: 10.1016/j.jcp.2021.110683

    ISSN 00219991. doi: 10.1016/j.jcp.2021.110683

  34. [2023]

    doi: 10.1002/pamm.202300203

    ISSN 1617-7061. doi: 10.1002/pamm.202300203

  35. [7691]

    doi: 10.1007/s10915-022-01939-z

  36. [9044]

    doi: 10.1007/s10444-023-10065-9

Pith tools

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