pith. sign in

arxiv: 2605.20250 · v1 · pith:BR2E46SGnew · submitted 2026-05-18 · 💻 cs.LG · physics.comp-ph· physics.flu-dyn

Physics-informed convolutional neural networks for fluid flow through porous media

Rafa{\l} Topolnicki , Pawe{\l} D{\l}otko , Maciej Matyka This is my paper

Pith reviewed 2026-05-21 08:16 UTC · model grok-4.3

classification 💻 cs.LG physics.comp-phphysics.flu-dyn
keywords physics-informed neural networksporous media flowconvolutional neural networksfluid dynamicsLattice-Boltzmann methodvelocity field predictiontortuosity
0
0 comments X

The pith

A convolutional neural network predicts pore-scale fluid velocity fields directly from porous sample geometry while enforcing physical laws via a custom loss function.

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

The authors build a convolutional encoder-decoder network that maps the geometry of a porous medium to the fluid velocity field inside its pores. Training combines direct velocity matching with penalties that enforce incompressibility, zero velocity inside solids, periodic boundaries, and consistency with the sample's tortuosity. This setup is tested on unseen geometries, different porosities, and realistic structures. The resulting fields then serve as starting guesses for Lattice-Boltzmann simulations, cutting the number of iterations needed for convergence in most cases. The work addresses the repeated, expensive solves that arise in design or optimization tasks involving flow through complex pores.

Core claim

We present a neural-network-based framework for predicting pore-scale velocity fields directly from sample geometry. The method uses a convolutional encoder-decoder architecture with skip connections. Physical consistency is encouraged through a custom loss function combining velocity reconstruction with incompressibility, no-flow conditions inside solids, periodicity constraints, and agreement with the global tortuosity index. We demonstrate that the predicted velocity fields can be used as initial conditions for Lattice-Boltzmann simulations, accelerating solver convergence in over 90 percent of tested cases.

What carries the argument

Convolutional encoder-decoder network with skip connections driven by a multi-term loss that adds physics-based penalties for divergence-free flow, solid boundaries, periodicity, and tortuosity matching to the velocity reconstruction error.

If this is right

  • The network produces usable velocity fields on samples whose obstacle shapes, boundary conditions, and porosities differ from those seen in training.
  • Feeding the network output as an initial guess reduces Lattice-Boltzmann iteration count in more than 90 percent of the tested porous samples.
  • Different CNN backbones yield varying robustness, and the contribution of each loss term to final accuracy can be measured.
  • The framework supports both direct prediction and warm-start use for traditional solvers.

Where Pith is reading between the lines

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

  • The same warm-start approach could shorten convergence for other grid-based flow solvers that accept an initial velocity guess.
  • Extending the input to include three-dimensional geometries would allow the method to address more realistic engineering problems.
  • If the loss weights prove stable across domains, the trained model might serve as a fast surrogate for repeated flow queries in optimization loops.

Load-bearing premise

A fixed weighted sum of velocity reconstruction, incompressibility, no-flow, periodicity, and tortuosity terms produces fields whose accuracy and generalization do not depend critically on the exact choice of weights or on the distribution of training geometries and porosities.

What would settle it

Train the network on one family of obstacle shapes and porosities, then evaluate velocity prediction error against full Navier-Stokes solutions on a large collection of samples whose obstacle geometry and porosity lie well outside the training range.

Figures

Figures reproduced from arXiv: 2605.20250 by Maciej Matyka, Pawe{\l} D{\l}otko, Rafa{\l} Topolnicki.

Figure 1
Figure 1. Figure 1: Example porous structures and corresponding LBM solutions - color [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Statistical summaries of the LBM solutions for the dataset used to train the model. (a) Distribution of the trivial porosity (defined as the ratio of void area [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: a) Example porous structure (top row) and corresponding LBM so [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Explanation of the loss term Lperio defined in equation (14). a) The structure S and translation vector T. b) Velocity field predicted for structure S – this velocity field is denoted as CNN(S ). c) Velocity field after imposing translation T. d) Structure S after translation by vector T: i.e. T[S ]. e) Veloc￾ity field CNN(T[S ]) predicted for translated structure T[S ]. Finally Lperio is computed as integ… view at source ↗
Figure 5
Figure 5. Figure 5: (left) Reference velocity field obtained from the LBM simulation. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: (left) Reference velocity field obtained from LBM simulation. (mid [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: a) Tortuosity predicted by the best-performing model as the function [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: LBM computed and neural-network predicted velocity fields for ex [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: LBM computed and neural-network predicted velocities fields for ex [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The same as Figure 9 but with linear obstacle placed in the center of [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: LBM computed and neural-network predicted velocities fields for [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The graph presents the number of steps of the LBM when starting [PITH_FULL_IMAGE:figures/full_fig_p012_13.png] view at source ↗
read the original abstract

Accurate simulation of fluid flow in porous media is challenging due to complex pore-space geometries and the computational cost of solving the Navier-Stokes equations. This difficulty is particularly important when repeated simulations are required, as standard numerical solvers may converge slowly in intricate porous domains. We present a neural-network-based framework for predicting pore-scale velocity fields directly from sample geometry. The method uses a convolutional encoder-decoder architecture with skip connections to preserve spatial detail while extracting multi-scale features. Physical consistency is encouraged through a custom loss function combining velocity reconstruction with incompressibility, no-flow conditions inside solids, periodicity constraints, and agreement with the global tortuosity index. We analyze the influence of the corresponding loss weights and quantify the contribution of individual loss components to prediction accuracy. Several CNN backbones are evaluated to identify architectures providing accurate and robust predictions. The generalization ability of the trained model is tested on samples outside the training distribution, including changes in obstacle geometry, boundary conditions, porosity, and realistic porous structures. Finally, we demonstrate a practical use of the predicted velocity fields as initial conditions for Lattice-Boltzmann simulations. This warm-start strategy accelerates solver convergence, reducing the number of iterations in over 90% of tested cases.

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

2 major / 2 minor

Summary. The manuscript presents a physics-informed convolutional neural network framework for predicting pore-scale velocity fields in porous media directly from geometry. It employs a convolutional encoder-decoder architecture with skip connections and a custom loss function that combines velocity reconstruction, incompressibility (divergence-free), no-flow conditions inside solids, periodicity constraints, and agreement with a global tortuosity index. The authors analyze loss-weight influences, compare multiple CNN backbones, test generalization on out-of-distribution samples (different obstacle geometries, porosities, boundary conditions, and realistic structures), and demonstrate that the predicted fields serve as effective initial conditions for Lattice-Boltzmann simulations, reducing iterations in over 90% of tested cases.

Significance. If the quantitative results and generalization claims hold, the work offers a practical contribution to accelerating repeated fluid-flow simulations in complex porous domains by supplying fast, physically consistent warm starts for traditional solvers. The multi-term physics-informed loss and the empirical LBM acceleration demonstration are clear strengths; the loss-component analysis and out-of-distribution tests further support the central claim when properly documented.

major comments (2)
  1. [paragraph describing the custom loss function] The tortuosity-matching term in the loss is defined from the predicted velocity field itself. If tortuosity is computed directly from the same velocity that the network is trained to match, the term risks becoming a self-referential constraint rather than an independent physical anchor. Please clarify the exact computation of tortuosity (including any auxiliary fields or averaging) and demonstrate independence, for example by reporting accuracy with and without this term.
  2. [section on loss weight analysis and generalization tests] The central claim that the weighted combination of velocity reconstruction, divergence-free, no-flow, periodicity, and tortuosity terms yields accurate and generalizable velocity fields rests on the assumption that performance is not dominated by the particular choice of loss weights or the training distribution of geometries and porosities. The manuscript analyzes weight influence but does not show that the chosen weights remain effective across substantially different test distributions; this is load-bearing for the generalization results.
minor comments (2)
  1. [loss function description] Explicit equations for each loss term (with weighting coefficients) would improve clarity and reproducibility.
  2. [results figures] Figure captions should state the exact error metric (e.g., relative L2 velocity error) and the number of test samples used for each reported statistic.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will revise the manuscript to incorporate clarifications and additional analyses where appropriate.

read point-by-point responses

  1. Referee: The tortuosity-matching term in the loss is defined from the predicted velocity field itself. If tortuosity is computed directly from the same velocity that the network is trained to match, the term risks becoming a self-referential constraint rather than an independent physical anchor. Please clarify the exact computation of tortuosity (including any auxiliary fields or averaging) and demonstrate independence, for example by reporting accuracy with and without this term.

    Authors: We appreciate this observation on the tortuosity term. Tortuosity is computed as a global scalar τ = L_path / L_straight, where L_path is obtained by integrating the magnitude of the velocity field along streamlines (or equivalently via a volume-averaged path length derived from the velocity components) and L_straight is the domain length in the flow direction; no auxiliary fields beyond the velocity are used, and averaging is performed over the fluid domain. While derived from the predicted velocity, the term acts as an independent macroscopic constraint because it enforces agreement with a known physical property of porous media that is not directly optimized by the local velocity reconstruction loss. To demonstrate its contribution, we will add an ablation study in the revised manuscript comparing velocity prediction accuracy (MSE, divergence error, and boundary condition satisfaction) with and without the tortuosity term. revision: yes

  2. Referee: The central claim that the weighted combination of velocity reconstruction, divergence-free, no-flow, periodicity, and tortuosity terms yields accurate and generalizable velocity fields rests on the assumption that performance is not dominated by the particular choice of loss weights or the training distribution of geometries and porosities. The manuscript analyzes weight influence but does not show that the chosen weights remain effective across substantially different test distributions; this is load-bearing for the generalization results.

    Authors: We thank the referee for this point on robustness. The manuscript reports loss-weight sensitivity on the training distribution and applies the selected weights to out-of-distribution tests covering different obstacle geometries, porosities, boundary conditions, and realistic structures, where the model maintains accuracy. We agree, however, that an explicit check of whether these fixed weights remain near-optimal on the new distributions would strengthen the generalization claims. In the revision we will add results evaluating the chosen weights on the OOD test sets and include a short sensitivity table for a subset of the OOD cases. revision: yes

Circularity Check

0 steps flagged

No circularity: composite loss uses independent physical constraints

full rationale

The paper presents a supervised CNN for velocity prediction from geometry, augmented by a composite loss containing velocity reconstruction (presumably against ground-truth fields), divergence-free enforcement, solid no-flow, periodicity, and tortuosity matching. The tortuosity term matches a scalar property of the predicted field to an external global index (precomputed from data or geometry), which does not reduce the output to a function of itself by construction. No equation or training step is shown to be tautological; the authors explicitly analyze loss-weight sensitivity and test generalization on out-of-distribution samples, indicating the method retains independent content. The warm-start use for Lattice-Boltzmann is a downstream application, not a self-referential derivation. No self-citations or imported uniqueness theorems appear load-bearing in the provided description.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim depends on several tunable loss weights whose values are not stated, on the assumption that the chosen physics penalties are sufficient proxies for the full Navier-Stokes solution, and on the training data distribution being representative enough for the reported generalization.

free parameters (1)
  • loss weights for velocity reconstruction, incompressibility, no-flow, periodicity, and tortuosity terms
    Abstract states that the influence of corresponding loss weights is analyzed, implying they are chosen hyperparameters that affect prediction accuracy.
axioms (2)
  • domain assumption Fluid is incompressible and flow satisfies no-slip or no-flow inside solid obstacles
    Invoked in the custom loss function description.
  • domain assumption Periodic boundary conditions apply to the domain
    Included as a constraint in the loss.

pith-pipeline@v0.9.0 · 5760 in / 1677 out tokens · 36044 ms · 2026-05-21T08:16:55.358784+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 7 internal anchors

  1. [1]

    , author Pinto, W.G

    author Alguacil, A. , author Pinto, W.G. , author Bauerheim, M. , author Jacob, M.C. , author Moreau, S. , year 2021 . title Effects of boundary conditions in fully convolutional networks for learning spatio-temporal dynamics . https://arxiv.org/abs/2106.11160, http://arxiv.org/abs/2106.11160 arXiv:2106.11160

    work page arXiv 2021

  2. [2]

    , year 2002

    author Anderson, J.D. , year 2002 . title Computational fluid dynamics: the basics with applications . publisher McGraw-Hill New York

    work page 2002

  3. [3]

    , author Mishra, A

    author Cossins, T. , author Mishra, A. , author Haese, R.R. , year 2023 . title The feasibility of enhanced pore space utilization in co2 storage reservoirs using an artificially emplaced si-gel flow barrier . journal Scientific Reports volume 13 , pages 9334

    work page 2023

  4. [4]

    title DRP-1129v2 dataset

    author Digital Porous Media Portal , year 2024 . title DRP-1129v2 dataset . https://doi.org/10.17612/dpm/DRP-1129v2, doi:10.17612/dpm/DRP-1129v2

    work page doi:10.17612/dpm/drp-1129v2 2024

  5. [5]

    , author Koza, Z

    author Duda, A. , author Koza, Z. , author Matyka, M. , year 2011 . title Hydraulic tortuosity in arbitrary porous media flow . journal Physical Review E—Statistical, Nonlinear, and Soft Matter Physics volume 84 , pages 036319

    work page 2011

  6. [6]

    , year 2015

    author Fong, C.W. , year 2015 . title Permeability of the blood--brain barrier: molecular mechanism of transport of drugs and physiologically important compounds . journal The Journal of membrane biology volume 248 , pages 651--669

    work page 2015

  7. [7]

    , author Sun, L

    author Gao, H. , author Sun, L. , author Wang, J.X. , year 2021 . title Phygeonet: Physics-informed geometry-adaptive convolutional neural networks for solving parameterized steady-state pdes on irregular domain . journal Journal of Computational Physics volume 428 , pages 110079 . :https://doi.org/10.1016/j.jcp.2020.110079

    work page doi:10.1016/j.jcp.2020.110079 2021

  8. [8]

    , author Matyka, M

    author Graczyk, K.M. , author Matyka, M. , year 2020 . title Predicting porosity, permeability, and tortuosity of porous media from images by deep learning . journal Scientific reports volume 10 , pages 21488

    work page 2020

  9. [9]

    , author Strzelczyk, D

    author Graczyk, K.M. , author Strzelczyk, D. , author Matyka, M. , year 2023 . title Deep learning for diffusion in porous media . journal Scientific Reports volume 13 , pages 9769

    work page 2023

  10. [10]

    , author Wang, Y

    author Guo, C. , author Wang, Y. , author Han, Y. , author Ji, M. , author Wu, Y. , year 2024 . title Unsteady flow-field forecasting leveraging a hybrid deep-learning architecture . journal Physics of Fluids volume 36

    work page 2024

  11. [11]

    , author Li, W

    author Guo, X. , author Li, W. , author Iorio, F. , year 2016 . title Convolutional neural networks for steady flow approximation , in: booktitle Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining , pp. pages 481--490

    work page 2016

  12. [12]

    , author Shu, C

    author Guo, Z. , author Shu, C. , year 2013 . title Lattice Boltzmann method and its application in engineering . volume volume 3 . publisher World Scientific

    work page 2013

  13. [13]

    Deep Residual Learning for Image Recognition

    author He, K. , author Zhang, X. , author Ren, S. , author Sun, J. , year 2015 . title Deep residual learning for image recognition . https://arxiv.org/abs/1512.03385, http://arxiv.org/abs/1512.03385 arXiv:1512.03385

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  14. [14]

    , author Sandler, M

    author Howard, A. , author Sandler, M. , author Chu, G. , author Chen, L.-C. , author Chen, B. , author Tan, M. , author Wang, W. , author Zhu, Y. , author Pang, R. , author Vasudevan, V. , author Le, Q.V. , author Adam, H. , year 2019 . title Searching for MobileNetV3 . https://arxiv.org/abs/1905.02244, http://arxiv.org/abs/1905.02244 arXiv:1905.02244

    work page arXiv 2019

  15. [15]

    Densely Connected Convolutional Networks

    author Huang, G. , author Liu, Z. , author van der Maaten, L. , author Weinberger, K.Q. , year 2018 . title Densely connected convolutional networks . https://arxiv.org/abs/1608.06993, http://arxiv.org/abs/1608.06993 arXiv:1608.06993

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  16. [16]

    , author Ritschel, T

    author Innamorati, C. , author Ritschel, T. , author Weyrich, T. , author Mitra, N.J. , year 2019 . title Learning on the edge: Investigating boundary filters in cnns . journal International Journal of Computer Vision volume 128 , pages 773--782 . :10.1007/s11263-019-01223-y

    work page doi:10.1007/s11263-019-01223-y 2019

  17. [17]

    Adam: A Method for Stochastic Optimization

    author Kingma, D.P. , author Ba, J. , year 2017 . title Adam: A method for stochastic optimization . https://arxiv.org/abs/1412.6980, http://arxiv.org/abs/1412.6980 arXiv:1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  18. [18]

    , author Kataja, M

    author Koponen, A. , author Kataja, M. , author Timonen, J. , year 1996 . title Tortuous flow in porous media . journal Physical Review E volume 54 , pages 406

    work page 1996

  19. [19]

    , author Kataja, M

    author Koponen, A. , author Kataja, M. , author Timonen, J. , year 1997 . title Permeability and effective porosity of porous media . journal Physical Review E volume 56 , pages 3319

    work page 1997

  20. [20]

    , author Li, Z

    author Kovachki, N. , author Li, Z. , author Azizzadenesheli, K. , author Liu, B. , author Bhattacharya, K. , author Stuart, A. , author Anandkumar, A. , year 2023 . title Neural Operator: Learning Maps Between Function Spaces . journal Journal of Machine Learning Research volume 24 , pages 1--97

    work page 2023

  21. [21]

    , author Matyka, M

    author Koza, Z. , author Matyka, M. , author Khalili, A. , year 2009 . title Finite-size anisotropy in statistically uniform porous media . journal Physical Review E—Statistical, Nonlinear, and Soft Matter Physics volume 79 , pages 066306

    work page 2009

  22. [22]

    , author Krause, M.J

    author Lehmann, M. , author Krause, M.J. , author Amati, G. , author Sega, M. , author Harting, J. , author Gekle, S. , year 2022 . title Accuracy and performance of the lattice boltzmann method with 64-bit, 32-bit, and customized 16-bit number formats . journal Physical Review E volume 106 , pages 015308

    work page 2022

  23. [23]

    Fourier Neural Operator for Parametric Partial Differential Equations

    author Li, Z. , author Kovachki, N. , author Azizzadenesheli, K. , author Liu, B. , author Bhattacharya, K. , author Stuart, A. , author Anandkumar, A. , year 2021 . title Fourier Neural Operator for Parametric Partial Differential Equations . https://arxiv.org/abs/2010.08895, http://arxiv.org/abs/2010.08895 arXiv:2010.08895

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  24. [24]

    , author Wu, Z

    author Lin, Y. , author Wu, Z. , author You, S. , author Yang, C. , author Wang, Q. , author Yin, W. , author Qiu, T. , year 2025 . title Deep learning-based prediction of velocity and temperature distributions in metal foam with hierarchical pore structure . journal Green Chemical Engineering volume 6 , pages 209--222

    work page 2025

  25. [25]

    , author Mao, H

    author Liu, Z. , author Mao, H. , author Wu, C.-Y. , author Feichtenhofer, C. , author Darrell, T. , author Xie, S. , year 2022 . title A ConvNet for the 2020s . https://arxiv.org/abs/2201.03545, http://arxiv.org/abs/2201.03545 arXiv:2201.03545

    work page arXiv 2022

  26. [26]

    , author Khalili, A

    author Matyka, M. , author Khalili, A. , author Koza, Z. , year 2008 . title Tortuosity-porosity relation in porous media flow . journal Physical Review E—Statistical, Nonlinear, and Soft Matter Physics volume 78 , pages 026306

    work page 2008

  27. [27]

    , author Koza, Z

    author Matyka, M. , author Koza, Z. , year 2012 . title How to calculate tortuosity easily? , in: booktitle AIP Conference Proceedings 4 , organization American Institute of Physics . pp. pages 17--22

    work page 2012

  28. [28]

    On the difficulty of training Recurrent Neural Networks

    author Pascanu, R. , author Mikolov, T. , author Bengio, Y. , year 2013 . title On the difficulty of training recurrent neural networks . https://arxiv.org/abs/1211.5063, http://arxiv.org/abs/1211.5063 arXiv:1211.5063

    work page internal anchor Pith review Pith/arXiv arXiv 2013

  29. [29]

    , author Esteva, M

    author Prodanovic, M. , author Esteva, M. , author Ketcham, R. , author Chang, B. , author Turhan, C. , author Gentle, J. , author Khan, S. , author Belcher, V. , year 2025 . title Digital Porous Media Portal (DPMP) for publication, analysis, and simulation of porous media images . publisher Digital Porous Media Portal . :10.17612/FGMN-D889

    work page doi:10.17612/fgmn-d889 2025

  30. [30]

    U-Net: Convolutional Networks for Biomedical Image Segmentation

    author Ronneberger, O. , author Fischer, P. , author Brox, T. , year 2015 . title U-net: Convolutional networks for biomedical image segmentation . https://arxiv.org/abs/1505.04597, http://arxiv.org/abs/1505.04597 arXiv:1505.04597

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  31. [31]

    Very Deep Convolutional Networks for Large-Scale Image Recognition

    author Simonyan, K. , author Zisserman, A. , year 2015 . title Very deep convolutional networks for large-scale image recognition . https://arxiv.org/abs/1409.1556, http://arxiv.org/abs/1409.1556 arXiv:1409.1556

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  32. [32]

    , year 2001

    author Succi, S. , year 2001 . title The lattice Boltzmann equation: for fluid dynamics and beyond . publisher Oxford university press

    work page 2001

  33. [33]

    , author Le, Q.V

    author Tan, M. , author Le, Q.V. , year 2019 . title EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks , in: booktitle Proceedings of the 36th International Conference on Machine Learning . pages 6105--6114

    work page 2019

  34. [34]

    , author Szafran, R.G

    author Tomczak, T. , author Szafran, R.G. , year 2019 . title A new gpu implementation for lattice-boltzmann simulations on sparse geometries . journal Computer Physics Communications volume 235 , pages 258--278

    work page 2019

  35. [35]

    , author Chang, B

    author Turhan, C . , author Chang, B. , author Mohamed, A. , author Esteva, M. , author Ketcham, R. , author McClure, J. , author Prodanovic, M. , year 2024 . title Digital Porous Media Portal for image curation, characterization, visualization, and transport simulation in porous media , in: booktitle International Symposium of the Society of Core Analyst...

    work page 2024

  36. [36]

    , author Gong, Z

    author Zhao, X. , author Gong, Z. , author Zhang, Y. , author Yao, W. , author Chen, X. , year 2023 . title Physics-informed convolutional neural networks for temperature field prediction of heat source layout without labeled data . journal Engineering Applications of Artificial Intelligence volume 117 , pages 105516

    work page 2023

  37. [37]

    , author Han, X

    author Zhong, L. , author Han, X. , author Yuan, X. , author Liu, Y. , author Zou, J. , author Wang, Q. , year 2021 . title Permeability variation and its impact on oil recovery from unconsolidated sand heavy-oil reservoirs during steamflooding process . journal SPE Reservoir Evaluation & Engineering volume 24 , pages 159--173

    work page 2021

  38. [38]

    , author Carlson, R.E

    author Fritsch, F.N. , author Carlson, R.E. , year 1980 . title Monotone Piecewise Cubic Interpolation . journal SIAM Journal on Numerical Analysis volume 17 , pages 238--246 . doi:10.1137/0717021

    work page doi:10.1137/0717021 1980

  39. [39]

    , author Butland, J

    author Fritsch, F.N. , author Butland, J. , year 1984 . title A Method for Constructing Local Monotone Piecewise Cubic Interpolants . journal SIAM Journal on Scientific and Statistical Computing volume 5 , pages 300--304 . doi:10.1137/0905021

    work page doi:10.1137/0905021 1984

  40. [40]

    , author Kovachki, N

    author Li, Z. , author Kovachki, N. , author Azizzadenesheli, K. , author Liu, B. , author Bhattacharya, K. , author Stuart, A. , author Anandkumar, A. , year 2021 . title Fourier Neural Operator for Parametric Partial Differential Equations . journal International Conference on Learning Representations (ICLR)

    work page 2021