acceptodds
Under review as a conference paper at ICLR 2027

PoreML: A Data Driven Framework for Learning Multiphase Flow in Porous Media

13 views

Abstract

Multiphase flow in porous microstructures is central to CO storage, fuel-cell operation, and flip-chip packaging. Predicting these flows remains challenging because wettability and complex pore geometry govern the nonlinear evolution of fluid interfaces. Machine learning holds substantial promise for advancing the field, but progress is hindered by scarce time-resolved 3D datasets and a lack of a unified workflow for training and evaluating models. To fill this gap, we introduce **PoreML**, an open-source framework unifying data generation, model training, and evaluation grounded in pore-scale physics. The framework comprises three core components. **(a)** A validated GPU-native [lattice Boltzmann solver](https://github.com/PoreML/bob) for multiphase flow in complex geometries supports reproducible data generation and future dataset expansion. **(b)** A [**3.3 TB dataset**](https://huggingface.co/datasets/PoreML/PoreML_data) contains **560 simulation runs** and **158,546 stored time steps** across four application-driven scenarios. These trajectories span synthetic structures and geometries derived from micro-CT scans of real materials, covering diverse wetting conditions and viscosity ratios. **(c)** A unified [learning framework](https://github.com/PoreML/poreml) evaluates one-step prediction and autoregressive rollouts. Its domain-specific evaluation protocols assess predictive accuracy and physical consistency. We evaluate five models of diverse architecture under these protocols. Two targeted challenges assess models’ suitability for practically relevant settings by testing transfer to larger domains and from synthetic to micro-CT-derived structures. PoreML provides a shared foundation for machine-learning research on multiphase flow in porous media, with the aim of empowering the community to develop reliable predictive models and advance the field.

Primary area:
datasets and benchmarks
Keywords:
AI for Science, Scientific Machine Learning, Porous Media, Multiphase Flow, Benchmark Dataset

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KEmoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a benchmark for flow in porous media, combining synthetic porous geometries and CT-reconstructed structures. It also presents poreML, a framework for evaluating pore-scale applications (e.g., rock, sandstone, and other porous media). The authors assess generalization across resolutions and geometries.

Critical strengths

Pore-scale flow prediction is timely and relevant across energy, materials, and environmental applications. The effort to curate paired synthetic and scanned geometries and to package evaluation pipelines is valuable, and a well-documented benchmark could benefit the community if released with clear protocols and baselines.

Critical weaknesses

I am not fully confident about the acceptance bar for benchmark-only submissions in ICLR, and from a technical standpoint the paper reads more as an integration of existing tools than as a standalone methodological advance. That may be acceptable for a benchmark track or dataset paper, but the manuscript should then emphasize reproducibility, baseline coverage, and metric design. I remain neutral and would weigh the other reviewers’ opinions heavily.

Questions for authors
  1. I would like a dedicated analysis of the physical and statistical plausibility of the synthetic geometries. Why were generative models, procedural modeling, or physics-based simulators not used (or not sufficient) for dataset construction? How do you verify physical reasonableness (e.g., connectivity/permeability ranges, mesh quality, and numerical stability)?
  2. What is novel (implementation, boundary conditions, porous coupling, multi-GPU) in the JAX-based GPU LBM solver? Please report accuracy and performance vs. established engines (e.g., Palabos) on identical geometries and refinement levels.
  3. Please clarify the specialized design choices in your geometry-aware metric and the concrete gains over existing methods and standard field-based metrics. In particular, explain what failure modes or geometric biases the proposed metric captures that prior geometry-aware or physics-based scores do not.
4 comments
AuthorsOpenReview ↗

Response to Reviewer KEmo (1/3): where the paper stands; reproducibility

We sincerely thank Reviewer KEmo for the thoughtful review and for recognizing that pore-scale flow prediction is "timely and relevant across energy, materials, and environmental applications". We agree that a benchmark should be judged on reproducibility, baseline coverage, and metric design. Your questions were an opportunity to strengthen each of them, above all the solver validation.

All changes in the revised manuscript are highlighted in blue. Below, W refers to the weaknesses and Q to the questions, numbered in the order they appear in the review.

Where the paper stands

We appreciate your openness to the other reviews, and first summarize the general feedback and what the new experiments show.

  • Summary from other reviewers. Reviewer kAB9 describes PoreML as "a missing and valuable data resource" with "physically meaningful evaluation"; Reviewer QU9t recognizes "strong technical contributions in terms of a solver and the dataset". The shared requests (a clearer reading of the results, quantitative evidence on the solver, a clearer case for the benchmark) are each answered by new experiments.
  • Insight from the experiments. We have now include a copy-last-frame baseline (Persistence) and a significance test. All five architectures learn bulk displacement: at step 64 they are significantly better than Persistence on phase volume in 43 of the 50 model–block pairs. None yet preserves interface curvature or topology: significantly better in 0 of 50 on either (new Section 4.4). Transfer to scanned geometries is largely absorbed; transfer to larger domains remains open (Section 4.3).
  • What this shows. To our knowledge, PoreML is the first open framework for these engineering problems, and the framework itself is what reveals where the technical challenge lies.

Reproducibility

You asked that a benchmark paper emphasize reproducibility, and we took this as a guide for the revision. Everything behind the paper is public and can be checked directly:

WhatWhere
ML framework demo notebook: the full workflow on a small subsetporeml/demo/demo.ipynb
Solver validation scripts: reproduce every validation case in the paperbob/demo
Framework: training, evaluation, and the Minkowski-functional metricsPoreML/poreml
DatasetPoreML/PoreML_data
Scripts that generated every simulation in the datasetbob/case

The trained checkpoints behind every reported number are released as well (Appendix K). The other reviews note the same effort: Reviewer QU9t points to the open-source release, and Reviewer kAB9 to the "released solver, geometries, splits, and checkpoints".

Baseline coverage: five published architectures from four families, each a verified port (Appendix H), with Persistence in every block, a significance test on every comparison (Table 2), and the inference cost of every model (Table 3). Metric design: see Q3.

AuthorsOpenReview ↗

Response to Reviewer KEmo (2/3): Q1

Q1. Plausibility of the generated geometries

Thank you for this question. The plausibility of the generated media decides how far the benchmark's conclusions carry, so we added a dedicated analysis (Appendix G.1, Table 16, Figure 18).

  • Procedural modeling is what we use, and it is common practice. Overlapping spheres$^{[1]},level−cutGaussianrandomfields, level-cut Gaussian random fields^{[2,3]},andcylindermatsforfibrousGDLs, and cylinder mats for fibrous GDLs^{[4]}$ are established stochastic models of porous media. They are available in the open-source PoreSpy toolkit$^{[5]},whichweuse,andflowsurrogatesarecommonlytrainedonsuchmedia, which we use, and flow surrogates are commonly trained on such media^{[6,7]}$.
  • Why not learned or process-based generators: GAN-based$^{[8,9]}$ and process-based$^{[10]}$ reconstructions are fitted to one material. They would place the test materials in the training distribution and remove the Geo-Shift challenge. The μ\muCT scans themselves serve as the realistic data.
  • Resolution and connectivity: generated and scanned samples pass the same checks. Throats under three voxels in radius are removed, the pore space must connect inlet to outlet, and for rock the critical throat radius, obtained by morphological drainage$^{[11]}$, must exceed five voxels. A resolution study confirms the lattice (see Q2): at 1283128^3 both our solver and LBPM reproduce their own 2563256^3 solution (IoU 0.98); at 64364^3 neither does (0.86, 0.85).
  • Physical range: the critical throat radius, which sets the capillary entry pressure, spans 5.1–12.4 voxels for generated rock and 5.1–13.9 for sandstone; 82–88% of generated pore and throat radii lie within the central 90% range of the scans. We report throat sizes and entry pressures, which control displacement at Ca=10−5\mathrm{Ca}=10^{-5}, and have not computed absolute permeability.
  • By design, not replicas: generated rock has pores 1.4 times larger at the median and fewer throats per pore. This controlled gap is what Geo-Shift measures.

References

[1] Torquato, S. (2002). Random heterogeneous materials: Microstructure and macroscopic properties. Springer. https://doi.org/10.1007/978-1-4757-6355-3

[2] Roberts, A. P., & Teubner, M. (1995). Transport properties of heterogeneous materials derived from Gaussian random fields: Bounds and simulation. Physical Review E, 51(5), 4141–4154. https://doi.org/10.1103/PhysRevE.51.4141

[3] Hyman, J. D., & Winter, C. L. (2014). Stochastic generation of explicit pore structures by thresholding Gaussian random fields. Journal of Computational Physics, 277, 16–31. https://doi.org/10.1016/j.jcp.2014.07.046

[4] Schulz, V. P., Becker, J., Wiegmann, A., Mukherjee, P. P., & Wang, C.-Y. (2007). Modeling of two-phase behavior in the gas diffusion medium of PEFCs via full morphology approach. Journal of The Electrochemical Society, 154(4), B419. https://doi.org/10.1149/1.2472547

[5] Gostick, J. T., Khan, Z. A., Tranter, T. G., Kok, M. D. R., Agnaou, M., Sadeghi, M., & Jervis, R. (2019). PoreSpy: A Python toolkit for quantitative analysis of porous media images. Journal of Open Source Software, 4(37), 1296. https://doi.org/10.21105/joss.01296

[6] Santos, J. E., Xu, D., Jo, H., Landry, C. J., Prodanović, M., & Pyrcz, M. J. (2020). PoreFlow-Net: A 3D convolutional neural network to predict fluid flow through porous media. Advances in Water Resources, 138, 103539. https://doi.org/10.1016/j.advwatres.2020.103539

[7] Wang, Y. D., Chung, T., Armstrong, R. T., & Mostaghimi, P. (2021). ML-LBM: Predicting and accelerating steady state flow simulation in porous media with convolutional neural networks. Transport in Porous Media, 138(1), 49–75. https://doi.org/10.1007/s11242-021-01590-6

[8] Mosser, L., Dubrule, O., & Blunt, M. J. (2017). Reconstruction of three-dimensional porous media using generative adversarial neural networks. Physical Review E, 96(4), 043309. https://doi.org/10.1103/PhysRevE.96.043309

[9] Mosser, L., Dubrule, O., & Blunt, M. J. (2018). Stochastic reconstruction of an oolitic limestone by generative adversarial networks. Transport in Porous Media, 125(1), 81–103. https://doi.org/10.1007/s11242-018-1039-9

[10] Øren, P.-E., & Bakke, S. (2002). Process based reconstruction of sandstones and prediction of transport properties. Transport in Porous Media, 46(2–3), 311–343. https://doi.org/10.1023/A:1015031122338

[11] Armstrong, R. T., Tavakkoli, O., Wang, Y. D., Li, Z., Mostaghimi, P., Berg, S., Ramstad, T., Rücker, M., & McClure, J. (2026). Lattice-Boltzmann for porous media: 100M+ GPU hours. Scientific Data, 13, 697. https://doi.org/10.1038/s41597-026-06823-1

AuthorsOpenReview ↗

Response to Reviewer KEmo (3/3): Q2; Q3

Q2. Solver novelty; accuracy and performance against established engines

Thank you for this request; it led to a new comparison in the revision. We claim no innovation in numerical methods; our strength lies in unifying best practices that few open solvers offer together: accurate wetting treatment, improved color-gradient estimation, and boundary conditions stable at low capillary number. Packaged as modern, easily installed software, the solver is an asset for the community, especially ML researchers: an open, GPU-native, validated implementation that lets anyone regenerate or extend the dataset.. As you suggested, and as Reviewer kAB9 also asked, we placed it against an established engine: the same cases run in LBPM$^{[1]}$ on identical geometries and at three refinement levels (new text in Appendix E.5, E.6, E.8).

All changes in the revised manuscript are highlighted in blue.

TestOursLBPM
Washburn imbibition, rate error against the analytic law, mean / worst3.0% / 6.2%5.2% / 11.5%
Porous drainage, 1283128^3 against own 2563256^3 solution, IoU0.982±0.0100.982\pm0.0100.985±0.0040.985\pm0.004
Porous drainage, 64364^3 against own 2563256^3 solution, IoU0.864±0.0200.864\pm0.0200.847±0.0270.847\pm0.027
Spurious current near the interface at 1283128^3, lattice units0.82×10−30.82\times10^{-3}2.6×10−32.6\times10^{-3}
Throughput at 1283128^3 / 2563256^3, million lattice updates per second675 / 652632 / 603

At 2563256^3 the two solvers overlap with an IoU of 0.985±0.0090.985\pm0.009, as closely as each overlaps its own 1283128^3 solution. Entry pressures in tube drainage differ by 3% on average. Throughput is within 8%.

Why LBPM. It is a popular choice in the pore-scale community and underlies numerous studies$^{[2,3,4]}$. It also implements the same model class as our solver, so a difference reflects the implementation and not the model.

Q3. Design of the geometry-aware metrics

Thank you for asking what the metrics add; the question helped us state their value more sharply.

Design. The descriptors are the four Minkowski functionals of the fluid body, established state variables of two-fluid flow in porous media$^{[4]}$. Our contribution is to turn them into rollout metrics (Appendix B): relative errors for M0M_0–M2M_2; for M3M_3, the miscount of the Euler characteristic χ\chi per 10510^5 voxels of injected phase, because χ\chi passes through zero; and M2M_2 integrated over the fluid–fluid interface only, where curvature sets the capillary pressure.

What they capture that other scores do not. Number of the 50 model–block pairs in which a model is significantly better than Persistence at step 64 (paired test over the evaluation runs, 95%):

ϕ\phi MAEIoUSaturationM0M_0M1M_1M2M_2Curvature distributionM3M_3
1230434330000

Bulk scores say the models are far better than a frozen frame; on curvature and topology no model is, in any block. Neither the field error nor the bulk scores can make this distinction.

Failure modes captured (Appendix I). A speck, a spurious island or bubble, adds 1 to χ\chi. A tunnel, a loop around an engulfed fiber or through a voxel-thin bridge, subtracts 1. Each moves only a few voxels, yet changes whether a fluid body is connected, and so whether it can flow or is trapped.

Thank you again. Your questions made the solver validation and the geometry analysis considerably stronger, and we are glad to clarify any remaining point.

References

[1] McClure, J. E., Li, Z., Berrill, M., & Ramstad, T. (2021). The LBPM software package for simulating multiphase flow on digital images of porous rocks. Computational Geosciences, 25, 871–895. https://doi.org/10.1007/s10596-020-10028-9

[2] Armstrong, R. T., Tavakkoli, O., Wang, Y. D., Li, Z., Mostaghimi, P., Berg, S., Ramstad, T., Rücker, M., & McClure, J. (2026). Lattice-Boltzmann for porous media: 100M+ GPU hours. Scientific Data, 13, 697. https://doi.org/10.1038/s41597-026-06823-1

[3] Armstrong, R. T., McClure, J. E., Berrill, M. A., Rücker, M., Schlüter, S., & Berg, S. (2016). Beyond Darcy's law: The role of phase topology and ganglion dynamics for two-fluid flow. Physical Review E, 94(4), 043113. https://doi.org/10.1103/PhysRevE.94.043113

[4] Armstrong, R. T., McClure, J. E., Robins, V., Liu, Z., Arns, C. H., Schlüter, S., & Berg, S. (2019). Porous media characterization using Minkowski functionals: Theories, applications and future directions. Transport in Porous Media, 130, 305–335. https://doi.org/10.1007/s11242-018-1201-4

AuthorsOpenReview ↗

A kind follow-up on our response, with thanks

Dear reviewer KEmo,

Thank you again for your thoughtful review. Your questions prompted us to compare our solver directly with LBPM and clarify what the geometry-aware metrics reveal. If you have a chance to revisit our response, we’d appreciate your considering an update of your score if you feel the new results address your concerns.

Best regards, The authors

Reviewer_kAB9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PoreML, a sim-to-learn framework for 3D two-phase flow in porous media. It contributes (i) an open-source JAX solver with geometric wetting, validated on eight analytical/experimental cases (ii) a 3.3 TB dataset of 560 time-resolved 3D simulations across four application scenarios (primary drainage, capillary trapping, GDL water intrusion, flip-chip underfill), pairing generated and µCT-derived geometries over a grid of contact angles and viscosity ratios (iii) a benchmark evaluating five architectures (UNet3D, FNO3D, P3D, Transolver++, AB-UPT) on one-step and 64-step autoregressive prediction, using field errors and Minkowski functionals, plus two generalization challenges (Geo-Shift: generated→µCT; Scale-Up: 128³→256³). The main finding is that low field error does not imply correct phase topology; an appendix attributes Euler-characteristic errors to spurious “specks” and “tunnels”.

Critical strengths

A missing and valuable data resource. To my knowledge, no public resource provides time-resolved, full-field 3D multiphase configurations inside porous solids at this scale, with paired synthetic/µCT geometries and released solver, geometries, splits, and checkpoints. This is the paper’s main contribution and is consequential for engineering applications (CCS, PEMFC, packaging). Physically meaningful evaluation. Using Minkowski functionals (in particular the Euler characteristic) exposes failure modes invisible to MAE, and the speck/tunnel analysis (App. I) is a concrete, actionable insight for future model design.

Critical weaknesses
  1. Limited ML insight and incomplete baseline design.
  • The benchmark mainly reports that existing architectures, trained once with a single seed, produce topological errors. (a) There is no trivial persistence (copy-last-frame) baseline. At Ca = 10⁻⁵ consecutive frames may differ little, so the one-step numbers (φ MAE 0.002–0.012), and the Scale-Up ratios below 1 (e.g., 0.72×), are hard to interpret without it. (b) Probabilistic/generative surrogates are not evaluated (including P3D’s flow-matching variant, which the authors explicitly drop), although abrupt, sensitive events such as Haines jumps and snap-off arguably call for them; neither are pretrained PDE foundation models or topology-aware objectives. (c) No seed variance is reported, so many “best-in-column” differences in Table 2 may not be significant. As a result, the benchmark offers limited guidance on what works.
  1. Narrow physical coverage and unclear practical value as a surrogate benchmark.
  • All porous-media runs share one capillary number (10⁻⁵) and contact angles 120–150° only, so viscous/stable displacement regimes and near-neutral/mixed wettability, central in applications, are absent; underfill viscosity ratios are orders of magnitude below real values. Rollouts cover only 64 of ~100–700 frames per run. Surrogate inference cost vs. the LBM solver is not reported, even though cost is the stated motivation. Solver validation against 3D experiments is qualitative (filling order, regime morphology) rather than quantitative.
Questions for authors
  1. (Clarification, likely to affect my assessment) How do all models compare with a copy-last-frame baseline for one-step and 64-step metrics, in-distribution and under Scale-Up? Does it explain the <1 Scale-Up ratios?
  2. (Clarification) With u_in = 1.25×10⁻⁵ l.u., how large are spurious currents near interfaces relative to the imposed flow? Are the stored velocity/pressure fields physically meaningful learning targets at this Ca, or dominated by numerical artifacts?
  3. (Clarification) Is the 5,000-step frame interval fine enough to resolve Haines jumps and snap-off, or are these events aliased between frames? Please report the fraction of frame transitions containing such events.
  4. (Clarification) What are the wall-clock inference times of the surrogates vs. the solver at 128³ and 256³?
  5. (Clarification) Could you report mean ± std over ≥3 seeds for at least the main table?
  6. (Related work) Please position against Wang et al. (2026, arXiv:2603.12516), which learns 3D pore-scale interface evolution from 4D velocimetry; this contradicts the statement that evolving rock-fluid surrogates remain 2D.
  7. (Future work, not affecting score) Plans to extend Ca and wettability ranges, and to provide a probabilistic evaluation protocol?
7 comments
AuthorsOpenReview ↗

Response to Reviewer kAB9 (1/5): overview; W1(a), Q1

We sincerely thank Reviewer kAB9 for the detailed and technically rigorous review, and for recognizing PoreML as "a missing and valuable data resource". That is exactly our aim: data and a framework that help the community develop this field further. Your questions shaped much of this revision; the first, in particular, made the paper much clearer about where current models stand.

What we did in response

  • Ran a Persistence baseline through the full protocol (Tables 2 and 4, Figure 5, new Section 4.4).
  • Added the run spread and a paired significance test to Table 2; Sections 4.2–4.4 state only significant comparisons (Appendix A.1).
  • Retrained all five models from scratch with three seeds on drainage (Tables R1 and R2 in this response), and timed the solver and all models (Table 3, Appendix A.2).
  • Compared the solver with LBPM on identical geometries (Appendix E.5, E.6, E.8).
  • Measured spurious currents, the Young–Laplace consistency of the stored pressure, and the frames spanned by Haines jumps and snap-off (Appendix F.8, F.9).
  • Revised the related work; stated coverage, plans and next steps (Section 2, Conclusion, Appendix J).

All changes in the revised manuscript are highlighted in blue. Below, W refers to the weaknesses and Q to the questions, numbered in the order they appear in the review.

W1(a), Q1. Persistence baseline and the Scale-Up ratios

Thank you for suggesting this baseline. It strengthens our main finding and makes it easier to interpret: learned models are not yet good at predicting interface curvature and topology (new Section 4.4).

Number of the 50 model–block pairs (5 models ×\times 10 blocks) in which a model is significantly better than Persistence (paired test over the evaluation runs, 95%; see W1(c)):

M0M_0 volumeM1M_1 areaM2M_2 curvatureM3M_3 topologyppϕ\phi
After one step4039002313
After 64 steps433000512
  • Bulk displacement is learned. In distribution, all 20 model–campaign pairs keep a significant lead on phase volume to step 64: the best model misses it by 2–10%, Persistence by 21–58%.
  • Curvature and topology are not, from the first step. No model is significantly better than Persistence in any block. One step from a true frame, the models are significantly worse in 37 (M2M_2) and 29 (M3M_3) of the 50 pairs; after 64 steps, in 45 and 48.
  • The fields are learned in part. The pressure lead holds in 23 pairs after one step and in 5 after 64.

Why the Scale-Up ratios fell below 1. With 8×8\times the voxels, the same 64 frames change a smaller share of the pore space, so even a frozen frame scores better: Persistence's own ϕ\phi error falls to 0.48–0.62 of its small-domain value. Dividing a model's error by its own small-domain error hid this.

How to read Figure 5. Each bar is a model's step-64 ϕ\phi MAE divided by Persistence's on the same runs (hatched: in distribution; solid: after the shift); beyond the dashed line at 1, a model is worse than predicting no motion. Under Scale-Up every solid bar lies at 1.2–2.3, significantly worse than Persistence in all 15 pairs. Under Geo-Shift a bar moves by a factor of 0.83–1.15, indistinguishable from no change in 12 of the 15 pairs.

AuthorsOpenReview ↗

Response to Reviewer kAB9 (2/5): W1(b); W1(c), Q5

W1(b). Probabilistic, pretrained, and topology-aware models

We appreciate these suggestions and share all three directions; the Conclusion names them as next steps.

  • P3D's generative variant. We would like to clarify that we did not set aside a dynamics model. In the original work$^{[1]}$ the flow-matching variant is a sampler: it draws equilibrium states of turbulent channel flow from noise, with no input state, and is assessed on flow statistics. Time stepping uses the deterministic model, which we benchmark.
  • Generative surrogates for dynamics are a promising route for abrupt events, at an inference slow-down roughly proportional to the number of diffusion steps$^{[2]}$. Appendix J sets out the probabilistic protocol we plan: prediction ensembles scored against ensembles of perturbed reference runs.
  • Pretrained models are a good direction. Many are currently built for 2D fields, for example Poseidon$^{[3]}$ (formulated with d=2d=2), and cannot take 3D volumes directly. Reviewer QU9t asked about the same family, and our answer to both reviews is the same.
  • Topology-aware objectives[4]^{[4]}. The new Section 4.4 provides the evidence for them: all five models share a voxel-wise regression loss, and one step from a true frame they are already significantly worse than Persistence on the Euler characteristic in 29 of the 50 pairs, and better in none.

W1(c), Q5. Seeds and the significance of the best-in-column entries

Thank you for this point; it led us to test every comparison in Table 2 and to repeat training with further seeds.

  • Run spread and a paired test. Table 2 reports each mean with its standard deviation over the evaluation runs. All models are scored on the same runs, so we compare them run by run (paired bootstrap, 95%) and underline every value that cannot be told apart from the best in its column (Section 4.2, Appendix A.1).
  • Of the 120 best-in-column entries, 65 stand apart and 55 share the lead. Sections 4.2–4.4 no longer rank individual models; every comparison they state is significant, and the main findings do not rest on a ranking.
  • Training seeds. We retrained all five models from scratch with three seeds on drainage, the campaign the rebuttal period allowed us to complete. Each seed repeats the full pipeline (frame-to-frame training, then push-forward fine-tuning) and is evaluated in distribution, under Geo-Shift, and under Scale-Up. Table R1 summarizes the study; Table R2 (part 5/5) gives all values.

Table R1. Drainage, three seeds, each trained from scratch. Seed spread: standard deviation over seeds as a share of the mean, over all models and metrics (median / max).

Best learned model as in Table 2Seed spread, one stepSeed spread, step 64
In-distribution12 of 12 columns2.4% / 3.8%5.8% / 15.3%
Geo-Shift12 of 12 columns2.5% / 3.9%5.6% / 11.9%
Scale-Up12 of 12 columns2.3% / 3.7%5.1% / 9.8%
  • The seed spread is small: about 2.4% of the mean after one step and 5–6% at step 64.
  • The leaders are stable: the three-seed mean selects the same best learned model as Table 2 in all 36 columns. In 9 columns another model lies within one seed standard deviation of the best, and the paired test marks most of these pairs (7 of 11) as tied as well.
  • The main finding is far above seed noise: against the Persistence values of Table 2, every model in every split is worse on M2M_2 and M3M_3 after one step by more than 6 seed standard deviations, and on M3M_3 at step 64 by more than 8.

References

[1] Holzschuh, B., Kohl, G., Redinger, F., & Thuerey, N. (2026). P3D: Highly scalable 3D neural surrogates for physics simulations with global context. International Conference on Learning Representations. https://openreview.net/forum?id=8UdCE5nhFl

[2] Kohl, G., Chen, L.-W., & Thuerey, N. (2026). Benchmarking autoregressive conditional diffusion models for turbulent flow simulation. Neural Networks, 199, 108641. https://doi.org/10.1016/j.neunet.2026.108641

[3] Herde, M., Raonić, B., Rohner, T., Käppeli, R., Molinaro, R., de Bézenac, E., & Mishra, S. (2024). Poseidon: Efficient foundation models for PDEs. Advances in Neural Information Processing Systems, 37, 72525–72624. https://doi.org/10.52202/079017-2311

[4] Hu, X., Li, F., Samaras, D., & Chen, C. (2019). Topology-preserving deep image segmentation. Advances in Neural Information Processing Systems, 32. https://papers.nips.cc/paper/8803-topology-preserving-deep-image-segmentation

AuthorsOpenReview ↗

Response to Reviewer kAB9 (3/5): W2; Q3; Q4; Q7

W2. Physical coverage

We appreciate these observations and respond to each.

  • Capillary number. We chose Ca=10−5\mathrm{Ca}=10^{-5} because flow in this range is the most challenging to simulate, and it is where established pore-scale studies work: capillary forces dominate, Haines jumps set in between Ca=3.1×10−4\mathrm{Ca}=3.1\times10^{-4} and 3.8×10−53.8\times10^{-5} [1]^{[1]}, and lowering Ca\mathrm{Ca} can increase the number of solver time steps tenfold$^{[2]}.Zacharoudiouetal.. Zacharoudiou et al.^{[1]}$ simulate 8.5×10−6≤Ca≤10−28.5\times10^{-6}\le \mathrm{Ca}\le10^{-2} and "give particular attention to the low Ca flow regime, characterised by Haines jumps". Akai et al.[3]^{[3]} simulate oil recovery at Ca\mathrm{Ca} of order 10−510^{-5} and match the measured remaining oil saturation and water effective permeability. A benchmark is most useful where the problem is hard; this is also the technical challenge Reviewer QU9t asked us to make explicit.
  • Contact angle. We acknowledge that 120∘120^\circ–150∘150^\circ is a limited range and will extend it; this is reflected in the Conclusion and Appendix J.
  • Underfill viscosity ratio. We acknowledge this: Appendix F.6 presents M=5M=5–3030 as a qualitative regime and not a material match.
  • 64 of 100–700 frames. We would like to clarify that the rollouts cover the whole run: the 64-frame windows start at frame 0 and at every 64th frame after it, so every stage of the displacement is scored.

Q3. Does the 5,000-step interval resolve Haines jumps and snap-off?

Thank you for this question; it led us to measure how many stored frames each event spans and to narrow our claim (new Appendix F.9, Table 15). Of the Haines jumps and snap-off events, 65–81% span five or more frame transitions and only 1.1–2.2% fall within one.

CampaignEventTransitions containing an eventEvents spanning ≥5\ge5 transitionsEvents in one transition
DrainageHaines jump38%81%2.2%
GDLHaines jump51%77%1.1%
TrappingSnap-off11%65%1.7%

Haines jumps overlap 38–51% of transitions because capillary-dominated drainage is a near-continuous sequence of pore-filling events.

Narrowed claim. The dataset records the change in phase configuration and connectivity across most events over several stored states. It does not resolve dynamics faster than the frame interval, and events shorter than one interval may be under-counted; Appendix F.9 now states this.

Q4. Inference time against the solver

Following your suggestion, we have added this comparison (Table 3 in Section 4.2; Appendix A.2, Table 5). Per stored frame on NVIDIA H200, the models are 89–2,741$\times$ faster than the solver at 1283128^3 and 131–4,338$\times$ at 2563256^3 in GPU time. At 2563256^3 the solver used four GPUs, so the wall-clock speed-up is 33–1,085$\times$. A 64-step rollout at 2563256^3 takes 6 s to 3.2 min instead of 7.0 GPU-hours.

Q7. Future plans

We appreciate your proposed suggestions and add those in the laset section, in the Conclusion and Appendix J: releases that span capillary- to viscous-dominated displacement and intermediate- and mixed-wet conditions, and a probabilistic evaluation protocol.

References

[1] Zacharoudiou, I., Boek, E. S., & Crawshaw, J. (2018). The impact of drainage displacement patterns and Haines jumps on CO2\mathrm{CO_2} storage efficiency. Scientific Reports, 8, 15561. https://doi.org/10.1038/s41598-018-33502-y

[2] Armstrong, R. T., Tavakkoli, O., Wang, Y. D., Li, Z., Mostaghimi, P., Berg, S., Ramstad, T., Rücker, M., & McClure, J. (2026). Lattice-Boltzmann for porous media: 100M+ GPU hours. Scientific Data, 13, 697. https://doi.org/10.1038/s41597-026-06823-1

[3] Akai, T., Alhammadi, A. M., Blunt, M. J., & Bijeljic, B. (2019). Modeling oil recovery in mixed-wet rocks: Pore-scale comparison between experiment and simulation. Transport in Porous Media, 127(2), 393–414. https://doi.org/10.1007/s11242-018-1198-8

Reviewer_QU9tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors present a framework, PoreML, for modeling the physics of porous microstructures which are ubiquitous in multiple application scenario from carbon sequestration to chip-flip applicatons. The authors present (a) A GPU accelerated Lattice-Boltzmann solver for modeling the physics (b) 560 simulation runs using the solver mentioned before and thus collected a total of 158K snapshots of data by running them using 3000 GPU-hourse (c) comparison of a few architectures for evaluation of the models on the data.

Critical strengths
  1. Strong technical contributions in terms of a solver and the dataset using 560 simulations in 3D.
  2. The code is open-sourced that way helping several researchers in the future.
  3. the validation results of multiple models show good performance on the dataset.
Critical weaknesses
  1. Limited dataset of just 560 simulations (although cover overall physics trends) can easily be skipped by the models that are being trained.
  2. The four scenarios are well addressed by the existing models to a great extent.... which raises the question as to why do we need a benchmark dataset for this scenarios? Perhaps does that mean that this is just a simple application of AI to engineering paper? This needs to be addressed.
  3. The key attributes that make this problem technically challenging are not discussed. It seems the focus by the authors of the paper is on benchmarking for a domain than on the challenges and reasons of why we need this kind of dataset. Today, datasets are created rather very easily.
Questions for authors
  1. Why should one benchmark on your dataset? applications are good, but they might be very small dataset.
  2. Is the goal of the authors to have others benchmark on your datasets or to publish training dataset useful for training?
  3. What is the rationale to only choose these chosen architecture? did you consider something like CNO? or levelset based arhitectures? or foundation models like Poseiden?
6 comments
AuthorsOpenReview ↗

Response to Reviewer QU9t (1/3): W2

We sincerely thank Reviewer QU9t for the review and for recognizing the "strong technical contributions in terms of a solver and the dataset". Your comments on motivation led us to state more clearly what the benchmark shows and why it is needed.

All changes in the revised manuscript are highlighted in blue. Below, W refers to the weaknesses and Q to the questions, numbered in the order they appear in the review.

W2. Why a benchmark, if existing models already do well?

Thank you for raising this concern. The first version did not state our position clearly, and Reviewer kAB9 raised the related point of limited ML insight. In response to both reviews, the revision adds one consistent baseline to every block, Persistence (copy the last frame), tests every comparison for significance, and reflect all results in the baseline in a new Section 4.4.

Persistence baseline takes the last observed frame and uses it as the prediction for every future step. For a 64-step rollout, this means holding the initial frame fixed for all 64 steps. This simple rule is a strong baseline in our capillary-dominated simulations. A meniscus can stay pinned at a pore throat for several frames before suddenly advancing, while much of the surrounding fluid remains nearly stationary. Copying the frame preserves the shape and connectivity of those interfaces exactly as observed. It misses subsequent displacement, but it never introduces a spurious droplet or bridge. A learned model therefore has to predict enough of the actual change to offset any errors it introduces into regions that barely move.

Models agree on the bulk quantities. Existing models predict phase volume (M0M_0) and interfacial area (M1M_1) well. After 64 steps they are significantly better than Persistence in 43 and 30 of the 50 model–block pairs (5 models ×\times 10 blocks). In distribution the best model misses the phase volume by 2–10%, Persistenc emodel by 21–58%.

The picture is not coherent across the quantities that define the flow. Number of the 50 pairs in which a model is significantly better than Persistence:

M0M_0 volumeM1M_1 areaM2M_2 curvatureM3M_3 topologyppϕ\phi
After one step4039002313
After 64 steps433000512

On interface curvature (M2M_2) and topology (M3M_3) no model is significantly better than a frozen frame in any block, and after 64 steps the models are significantly worse in 45 and 48 of the 50 pairs. On the larger domain every model's ϕ\phi error is significantly worse than Persistence's, by a factor of 1.2–2.3.

Curvature and topology are not secondary. In porous media, saturation alone does not fix the state of the flow: fluid topology changes at constant saturation$^{[1]},andcapillarypressureiswelldescribedbyageometricstatefunctionofsaturation,interfacialarea,andtheEulercharacteristic, and capillary pressure is well described by a geometric state function of saturation, interfacial area, and the Euler characteristic^{[2]}.ThefourMinkowskifunctionalsaretheestablisheddescriptorsofthisstate. The four Minkowski functionals are the established descriptors of this state^{[3]}$. Topology decides whether a fluid is connected and can flow, or is trapped: how much CO2\mathrm{CO_2} stays in the rock, whether product water blocks a fuel-cell layer. A surrogate with the right volume and the wrong connectivity answers neither.

This is why the benchmark is needed. Field error and bulk scores alone would have reported success; PoreML shows which part of the problem is solved and makes the open part measurable. We have reflected the above insight in a extra section 4.4 What Do Models Learn Beyound a Persistence Baseline?

References

[1] Armstrong, R. T., McClure, J. E., Berrill, M. A., Rücker, M., Schlüter, S., & Berg, S. (2016). Beyond Darcy's law: The role of phase topology and ganglion dynamics for two-fluid flow. Physical Review E, 94(4), 043113. https://doi.org/10.1103/PhysRevE.94.043113

[2] McClure, J. E., Armstrong, R. T., Berrill, M. A., Schlüter, S., Berg, S., Gray, W. G., & Miller, C. T. (2018). Geometric state function for two-fluid flow in porous media. Physical Review Fluids, 3(8), 084306. https://doi.org/10.1103/PhysRevFluids.3.084306

[3] Armstrong, R. T., McClure, J. E., Robins, V., Liu, Z., Arns, C. H., Schlüter, S., & Berg, S. (2019). Porous media characterization using Minkowski functionals: Theories, applications and future directions. Transport in Porous Media, 130, 305–335. https://doi.org/10.1007/s11242-018-1201-4

AuthorsOpenReview ↗

Response to Reviewer QU9t (2/3): W3

W3. What makes the problem technically challenging?

Thank you for this question; it led us to state the challenge explicitly in the paper (Section 4.4, Conclusion). Three properties make the problem demanding for every architecture.

The physical state lives in the interface. Curvature sets the capillary pressure, and topology decides whether a fluid is connected or trapped. Current architectures already capture the bulk displacement; the interface is the natural next step, because a voxel-wise loss gives it little weight. To make this actionable, we traced where the topology error comes from, and we invite you to read this short analysis (Appendix I). Its key points, for the GDL campaign:

  • Specks are small islands or bubbles; each adds 1 to the Euler characteristic χ\chi. They account for one third to two thirds of the Euler discrepancy.
  • Tunnels are loops closed around a fiber or through a thin bridge; each subtracts 1. Predictions hold more tunnels than the reference in 82–98% of windows.
  • Both involve only a few voxels, so they change the field error very little.

The Conclusion names the directions we see from here: topology-aware objectives$^{[1]},generativesurrogatesforabruptevents, generative surrogates for abrupt events^{[2]},andpretrainedmodelsextendedfrom2Dto3D, and pretrained models extended from 2D to 3D^{[3]}$.

The dynamics are intermittent, three-dimensional, and large. Each frame holds 1.3–18.9 million voxels. Most of the fluid body moves quasi-statically, while displacement events such as Haines jumps advance the interface orders of magnitude faster. A model must capture both regimes, hold them through 64 autoregressive steps, and transfer to domains with eight times the voxels.

The data are not easy to generate. The simulations need a purpose-built, validated solver and domain knowledge: the interface must be resolved in the narrowest throats$^{[4]}$; a standard wetting condition can realize 40.7∘40.7^\circ for a prescribed 60∘60^\circ [5]^{[5]}; pseudopotential models give diffuse interfaces under viscosity contrast$^{[6]}$; and the largest public campaign we know of, with more than 10810^8 GPU-hours, releases time series and not 3D fluid configurations$^{[7]}.InthisrevisionwealsocomparedoursolverwiththeestablishedLBPMpackage. In this revision we also compared our solver with the established LBPM package^{[8]}$ on identical geometries at three resolutions. The solutions agree (IoU 0.985 at 2563256^3) and our spurious currents are about three times lower (Appendix E.5, E.6, E.8).

References

[1] Hu, X., Li, F., Samaras, D., & Chen, C. (2019). Topology-preserving deep image segmentation. Advances in Neural Information Processing Systems, 32. https://papers.nips.cc/paper/8803-topology-preserving-deep-image-segmentation

[2] Kohl, G., Chen, L.-W., & Thuerey, N. (2026). Benchmarking autoregressive conditional diffusion models for turbulent flow simulation. Neural Networks, 199, 108641. https://doi.org/10.1016/j.neunet.2026.108641

[3] Herde, M., Raonić, B., Rohner, T., Käppeli, R., Molinaro, R., de Bézenac, E., & Mishra, S. (2024). Poseidon: Efficient foundation models for PDEs. Advances in Neural Information Processing Systems, 37, 72525–72624. https://doi.org/10.52202/079017-2311

[4] Li, Z., McClure, J. E., Middleton, J., Varslot, T., & Sheppard, A. P. (2020). Discretization limits of lattice-Boltzmann methods for studying immiscible two-phase flow in porous media. International Journal for Numerical Methods in Fluids, 92, 1162–1197. https://doi.org/10.1002/fld.4822

[5] Akai, T., Bijeljic, B., & Blunt, M. J. (2018). Wetting boundary condition for the color-gradient lattice Boltzmann method: Validation with analytical and experimental data. Advances in Water Resources, 116, 56–66. https://doi.org/10.1016/j.advwatres.2018.03.014

[6] Yang, J., & Boek, E. S. (2013). A comparison study of multi-component lattice Boltzmann models for flow in porous media applications. Computers & Mathematics with Applications, 65, 882–890. https://doi.org/10.1016/j.camwa.2012.11.022

[7] Armstrong, R. T., Tavakkoli, O., Wang, Y. D., Li, Z., Mostaghimi, P., Berg, S., Ramstad, T., Rücker, M., & McClure, J. (2026). Lattice-Boltzmann for porous media: 100M+ GPU hours. Scientific Data, 13, 697. https://doi.org/10.1038/s41597-026-06823-1

[8] McClure, J. E., Li, Z., Berrill, M., & Ramstad, T. (2021). The LBPM software package for simulating multiphase flow on digital images of porous rocks. Computational Geosciences, 25, 871–895. https://doi.org/10.1007/s10596-020-10028-9

AuthorsOpenReview ↗

Response to Reviewer QU9t (3/3): Q1, Q2; W1; Q3

Q1, Q2. Why benchmark on PoreML; benchmark or training set?

We are glad to clarify: PoreML serves both purposes. It is designed first as a benchmark: frozen splits, held-out geometries, a fixed one-step and 64-step protocol, and two transfer challenges (to μ\muCT scans and to larger domains). The same trajectories and the open solver serve training. Reviewer kAB9 notes that, to their knowledge, no public resource of this kind existed.

W1. Size of the dataset

Thank you for raising this concern. Size matters for a benchmark, so we looked at it carefully. Here we would respectfully beg to differ, and offer the numbers:

ResourceTotal sizeSnapshots
The Well$^{[1]}$15 TB over 16 datasets (6.9 GB–5.1 TB each)3D dataset with the most snapshots: 2,700×50=135,0002{,}700\times50=135{,}000 at 64364^3
BubbleML$^{[2]}$0.78 TB, 79 simulations69,500 in 2D and 141 in 3D
PoreML3.3 TB, 560 simulations158,546, all 3D, 1.3–18.9 million voxels each

PoreML alone is 3.5 times the average dataset of The Well, and it holds more 3D snapshots than that collection's largest 3D dataset by count, on grids 5 to 72 times larger. Models are always tested on geometries they have not seen. The dataset can also grow: the solver and the scripts that generated every run are public.

Q3. Choice of architectures; CNO, Poseidon, level-set models

Thank you for these suggestions. We chose five published models that run natively on 3D volumes of 2 to 17 million voxels and span the main families: convolutional (UNet3D), spectral operator (FNO3D), hybrid convolution–transformer (P3D), and point-based transformers (Transolver++, AB-UPT). Each is a verified port of the authors' code (Appendix H). The models you suggest are valuable, and we looked at each:

  • CNO[3]^{[3]} is formulated and benchmarked in 2D ("we will focus here on the two-dimensional case"); its authors describe the 3D extension as "conceptually straightforward but computationally demanding".
  • Poseidon[4]^{[4]} is pretrained on 2D fluid dynamics (the paper sets d=2d=2).
  • Level-set models. Learned surrogates built on a level-set interface that we are aware of, such as Bubbleformer$^{[5]}$, are so far demonstrated on 2D data. Our models already predict an implicit interface: the phase field ϕ\phi, whose zero level set is the fluid–fluid interface.

We see all three as promising directions. Their current 2D scope also illustrates the challenge that PoreML targets: 3D dynamics in real application geometries. We hope PoreML helps extend them to 3D; a new model plugs into the same splits, protocol, and metrics.

Thank you again for helping us state the motivation more clearly. We are glad to clarify any remaining point.

References

[1] Ohana, R., McCabe, M., Meyer, L., Morel, R., Agocs, F. J., Beneitez, M., Berger, M., Burkhart, B., Dalziel, S. B., Fielding, D. B., Fortunato, D., Goldberg, J. A., Hirashima, K., Jiang, Y.-F., Kerswell, R. R., Maddu, S., Miller, J., Mukhopadhyay, P., Nixon, S. S., … Ho, S. (2024). The Well: A large-scale collection of diverse physics simulations for machine learning. Advances in Neural Information Processing Systems, 37, 44989–45037. https://doi.org/10.52202/079017-1430

[2] Hassan, S. M. S., Feeney, A., Dhruv, A., Kim, J., Suh, Y., Ryu, J., Won, Y., & Chandramowlishwaran, A. (2023). BubbleML: A multiphase multiphysics dataset and benchmarks for machine learning. Advances in Neural Information Processing Systems, 36, 418–449. https://doi.org/10.52202/075280-0021

[3] Raonić, B., Molinaro, R., De Ryck, T., Rohner, T., Bartolucci, F., Alaifari, R., Mishra, S., & de Bézenac, E. (2023). Convolutional neural operators for robust and accurate learning of PDEs. arXiv. https://doi.org/10.48550/arXiv.2302.01178

[4] Herde, M., Raonić, B., Rohner, T., Käppeli, R., Molinaro, R., de Bézenac, E., & Mishra, S. (2024). Poseidon: Efficient foundation models for PDEs. Advances in Neural Information Processing Systems, 37, 72525–72624. https://doi.org/10.52202/079017-2311

[5] Hassan, S. M. S., Zou, X., Dhruv, A., Ganesan, V., & Chandramowlishwaran, A. (2025). Bubbleformer: Forecasting boiling with transformers. arXiv. https://doi.org/10.48550/arXiv.2507.21244

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response and rebuttal summary

To all reviewers and the Area Chair

We sincerely thank the three reviewers and the Area Chair for the time and care they gave to our submission. Your comments shaped this revision, and the paper is clearer and better supported because of them.

Paper summary. PoreML is the first end-to-end framework enabling machine to combine a validated solver for transient 3D two-phase flow in porous media, a public dataset, and a common ML training and evaluation protocol. Porous-media researchers can contribute new geometries and flow conditions, and ML researchers can test new models on them, so both communities can build on each other's results. Everything is public, and we warmly invite you to try it:

ResourceContent
Demo notebookA worked example of training and evaluating a model on a small subset of the data.
PoreML dataset560 simulations comprising 158,546 three-dimensional snapshots.
ML frameworkCode for training models and comparing their predictions on fixed splits using geometry-aware metrics.
GPU solverThe open-source lattice-Boltzmann solver used to generate the dataset.
Solver validation scriptsRunnable examples that reproduce the validation cases reported in the paper.
Simulation case scriptsScripts to regenerate the published trajectories or adapt a case to a new geometry or flow condition.

Where reviewers agree. We are grateful that all three reviewers see PoreML as a valuable contribution: "a missing and valuable data resource" with "physically meaningful evaluation" (kAB9), "strong technical contributions in terms of a solver and the dataset" (QU9t), on a "timely and relevant" problem (KEmo).

Shared concerns. The reviews also raise two main concerns. Both were fair, and we addressed each with new experiments:

ConcernWhat we addedWhere
What do current models actually learn, and why is a benchmark needed? (kAB9, QU9t, KEmo)Persistence baseline, significance test on every comparison, three-seed retrainingTables 2, 4, R1, R2; Section 4.4; Appendix I
Are the solver and the data reliable? (kAB9, KEmo)Comparison with the established LBPM solver, spurious-current and pressure checks, geometry fidelityAppendix E, F, G

Main insight from the rebuttal Following Reviewer kAB9's suggestion, we added Persistence, the simplest possible predictor: it assumes nothing changes and copies the last known frame forward. A learned model is only useful on a quantity if it beats this. The table counts, out of 50 model–block pairs, how often a learned model does so significantly:

VolumeAreaCurvatureTopologyppϕ\phi
After one step4039002313
After 64 steps433000512
  • Current strength: models predict bulk displacement. They beat Persistence on fluid volume and interfacial area in most cases.
  • Current limitation: no model beats Persistence on interface curvature or topology, in any case. These two quantities set the capillary pressure and decide whether a fluid is connected or trapped.
  • Research need: training objectives that preserve interface geometry and topology, which a voxel-wise loss does not reward.

This baseline made our main message much clearer, and we thank the reviewers for prompting it. Section 4.4 gives the full analysis, and Appendix I traces the topology errors to their source.

The revision leaves PoreML stronger on every point the reviews raised. And the benchmark delivers a clear scientific finding: current models run up to three orders of magnitude faster than the solver and capture bulk displacement, yet none preserves the interface geometry and connectivity that govern the physics, poinintg out some of the central challenge for future research on this topic.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.