Pith. sign in

REVIEW 5 major objections 7 minor 20 references

Physics-Guided Generative AI for Property-Targeted 3D Porous Media Design

T0 review · 5 major / 7 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read A physics-guided latent diffusion model designs 3D porous media to match target porosity and directional permeability.

desk verdict Solid engineering methods paper: joint denoiser–decoder refinement is a real, usable idea, but Palabos gains on real data are thinner than the abstract implies and the surrogate also picks what gets verified. read the letter →

arxiv 2607.24274 v1 pith:RL55BWLW submitted 2026-07-27 cs.LG cs.AI

Peng Wang This is my paper · ORCID
classification cs.LGcs.AI
keywords physics-guidedgenerativeAIporousmediainversedesignlatentdiffusionmodelspermeabilitycontrolstructure-propertymodellingvariationalautoencodersmicro-CT
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

Designing porous materials for filtration, batteries, fuel cells, or scaffolds is hard because many different pore geometries can share the same porosity or permeability, while tiny geometric changes can wreck transport. This paper presents a generative pipeline that first compresses binary 3D pore volumes into a compact latent space that already knows about porosity and directional permeability, then uses a conditional diffusion model to sample latents for a requested property vector, and finally decodes them into voxel structures. An independently trained neural surrogate predicts the same properties from the voxels and supplies gradients that jointly refine the diffusion denoiser and the decoder, so generated samples better match the targets without running a slow fluid simulator inside the loop. On both synthetic and real micro-CT 100³ volumes the method improves lattice-Boltzmann-verified multi-property scores and permeability correlations relative to property-aware VAE optimisation and frozen-decoder latent diffusion. The practical claim is that closed-loop property feedback inside latent diffusion gives a scalable route to inverse design of complex porous geometries.

What carries the argument

Surrogate-guided denoiser–decoder joint refinement: property mismatch from a frozen voxel-to-property network is back-propagated through the decoded volume into both the diffusion denoiser and the VAE decoder, aligning generated latents with physically meaningful geometries.

What would settle it

Generate a held-out set of target porosity and directional-permeability vectors, run the refined model and both baselines, measure lattice-Boltzmann errors, and check whether the proposed method still shows lower multi-property scores and higher permeability correlations; if the surrogate gains reverse or vanish under simulation, the central claim fails.

Watch

Extended reading notes

Core claim

Jointly refining a conditional latent diffusion denoiser and a property-aware VAE decoder under frozen surrogate property loss produces 3D porous structures whose porosity and directional permeabilities match prescribed targets more closely, under lattice-Boltzmann verification, than either property-aware VAE latent optimisation or latent diffusion with a frozen decoder.

Load-bearing premise

Gradients from the frozen neural property predictor stay faithful enough that gains on the surrogate still appear when the same structures are checked with a full lattice-Boltzmann fluid simulator.

Editorial extensions

If this is right

  • Inverse design of porous electrodes, filters, and scaffolds can be driven by target porosity and directional permeability without online fluid simulation at every step.
  • Conditional latent diffusion plus a differentiable property surrogate becomes a practical template for other geometry-to-transport inverse problems.
  • Directional permeability can be controlled independently of bulk porosity by closed-loop refinement rather than post-hoc latent search alone.
  • Real micro-CT trained models can propose candidate microstructures that already sit closer to simulator-verified targets than pure reconstruction or frozen-decoder baselines.
  • Extending the same feedback loop to tortuosity, reactive surface area, or multiphase transport is a direct next step the paper itself flags.

Reading between the lines

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

  • The same joint-refinement pattern could be tried on other ill-posed microstructure problems where a cheap surrogate exists but the high-fidelity solver is non-differentiable.
  • Because the surrogate is trained independently and frozen, any systematic bias it shares with the generator will not be corrected by Palabos post-checks; multi-fidelity or uncertainty-aware surrogates would be a natural stress test.
  • If decoder mismatch is the dominant error source, similar gains might appear in other latent-diffusion material generators that currently freeze the autoencoder after pre-training.
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

5 major / 7 minor

Summary. The manuscript proposes a pipeline for inverse design of 3D porous media: a property-aware VAE (pVAE) learns a latent space over 100³ binary pore/solid volumes; a conditional latent diffusion model samples latents conditioned on target porosity and directional permeability; an independently trained, frozen voxel-space surrogate supplies differentiable property feedback used to jointly refine the diffusion denoiser and the pVAE decoder (Eqs. 17–20, Algorithm 1). Generated candidates are binarised and verified with Palabos LBM simulations. On a 17k-sample synthetic dataset and a 5.5k-sample real micro-CT dataset, the method reports better Palabos-verified multi-property scores, lower mean permeability error, and markedly higher target–verified Pearson correlations than pVAE latent-optimisation and frozen-decoder latent-diffusion baselines (Tables I, III), with ablations on latent-optimisation steps (Tables V–VI) and latent resolution (Fig. 8).

Significance. If the results hold, the contribution is a practically useful one: a closed-loop, differentiable refinement mechanism for simulator-consistent inverse design of porous media, evaluated end-to-end with an actual physics solver rather than only a learned proxy. Specific strengths: (i) final metrics are Palabos-verified in raw physical units on two datasets, not surrogate-reported; (ii) the controllability gains are large and consistent — e.g., Table III real-data permeability correlations of 0.758/0.727/0.740 versus 0.048/−0.017/0.171 for the pVAE baseline; (iii) ablations disentangle latent optimisation from the proposed joint refinement (Tables V–VI) and justify the 5³ latent resolution (Fig. 8); (iv) the comparison protocol (same datasets, targets, surrogate, and verification pipeline across methods) is methodologically fair in design. The main risks to significance are the undisclosed candidate-selection protocol in Algorithm 1 step 19, the unreported accuracy of the surrogate that drives refinement, and the absence of any statistical-variability reporting.

major comments (5)
  1. [Algorithm 1, step 19; §III-E] The candidate-selection protocol is load-bearing but undisclosed. Step 19 states that sampled candidates are 'evaluated with the surrogate' and 'selected geometries' are exported for Palabos verification. The number of candidates sampled per target, the selection criterion (e.g., top-k by Eq. 23 score), and whether the identical selection rule was applied to all three methods are not stated anywhere. This matters because the same frozen surrogate sψ supplies the refinement gradients (Eq. 18) and screens candidates: joint refinement moves the decoder toward regions of voxel space where sψ scores well, and surrogate-based top-k selection then preferentially keeps candidates where sψ is most optimistic (optimizer's curse). Palabos verification removes surrogate error from the reported numbers themselves, but not from *which* candidates get measured. I note that the strongest version of this
  2. [§III-C (surrogate), Eqs. 15–16; §IV-C-a] The accuracy of sψ on held-out data is never reported. Every refinement gradient in Eq. (8)/(18) flows through this network, yet no test-set MAE/R² per property (nF, log Kx, log Ky, log Kz) is given for either dataset. Without this, the reader cannot judge whether the refinement is exploiting genuine structure–property signal or surrogate blind spots. Please report held-out surrogate errors, and ideally show that Lprop improvements correlate with Palabos improvements on a per-sample basis (a scatter of surrogate error vs. LBM error for refined outputs would directly address the gradient-fidelity question).
  3. [Table III; §IV-E-2; Abstract] The headline real-data result is much weaker than the abstract suggests. Under the no-latent-optimisation setting, the proposed method's mean score (0.306) is statistically indistinguishable from the pVAE baseline (0.309), and its porosity error (0.106) is 4–5× worse than either baseline (0.019, 0.024). The clear wins are in median score and in correlations. The abstract's blanket claim of 'improved target-property matching' on real micro-CT therefore needs qualification, and the post-hoc explanation of the porosity/permeability trade-off (§IV-E-2, §IV-F-1) should be supported quantitatively — e.g., by reporting the score with permeability-only terms, or a per-property Pareto analysis. Relatedly, the pVAE baseline with a single latent-optimisation step is close to a strawman for the no-optimisation comparison; the fair head-to-head is Table VI, which should be foregrounded rather than re
  4. [§IV-E; Tables I, III, V–VI] No variability information is given anywhere: the number of targets evaluated per table, the number of generation seeds, and run-to-run variance are all absent. This is load-bearing for interpretation — e.g., in Table V the proposed method's mean score is non-monotonic in optimisation steps (1.046 at 10 steps, 1.821 at 30, 0.980 at 50), which is either a meaningful instability of the joint refinement (contradicting the 'more stable optimisation space' claim in §IV-F-2) or noise. Report target counts, seed counts, and confidence intervals or per-target distributions for the main tables, and explain the Table V 30-step result.
  5. [§IV-A, §IV-B, §IV-D] Reproducibility details needed to assess the verification pipeline are missing: the source of the micro-CT scans (material, instrument, voxel resolution); Palabos simulation settings (boundary conditions, convergence criteria, lattice resolution relative to the 100³ input); the binarisation threshold applied to the decoder's soft occupancy output before export (§III-E); the value of δ in Eqs. (21)–(23) and (29)–(32); and the surrogate architecture/training hyperparameters. Since the entire claim rests on Palabos-verified numbers, the verification configuration must be specified precisely.
minor comments (7)
  1. [§IV-B] 'a latent representation of size 32×5×5×5, with 32 the batch size' — 32 is presumably the channel dimension, not the batch size; please clarify.
  2. [Tables V–VI] Row labels are inconsistent across step counts: 'Latent diffusion' vs. 'Diffusion seed' and 'Ours' vs. 'Joint decoder' in Table VI. Use one naming convention throughout.
  3. [Fig. 8] Axis labels, units, and the metric being plotted are not described in the caption; 'best overall performance' at 5³ cannot be verified from the figure as presented.
  4. [§III-E, Eqs. 21–23] The surrogate-side score r (Eq. 23) is defined but never reported in any table; either report it alongside Palabos metrics (useful for diagnosing surrogate–simulator gaps) or state explicitly that it is used only for candidate screening.
  5. [Throughout] Typos and spacing: 'indepent' (§II-B), 'optimsation' (§III-E), 'follows aforward designparadigm' (§II-A), inconsistent 'pV AE' spacing, 'Naiffet al.' and similar missing spaces before citations.
  6. [References] Refs. [11]–[13] (camouflage generation, antenna-array synthesis, time-series VAEs) are only loosely related to the points they support; consider replacing with more directly relevant porous-media or inverse-design literature.
  7. [Tables II and IV] State how the eight showcased targets were chosen ('randomly selected' per the figure captions — say so in the table captions too), to pre-empt cherry-picking concerns.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Palabos-verified gains are not true by construction; surrogate feedback is a standard differentiable proxy with independent LBM reporting.

full rationale

The paper’s load-bearing claim is empirical: joint denoiser–decoder refinement guided by a frozen structure-to-property surrogate improves Palabos LBM-verified multi-property matching and directional-permeability correlations relative to pVAE latent optimisation and frozen-decoder latent diffusion (Tables I, III, V–VI). Walking the chain: (i) property labels y come from Palabos (Eqs. 1, 4); (ii) models train in normalised/log space c (Eq. 2) with losses on reconstruction, KL, diffusion noise prediction, and surrogate MSE (Eqs. 11, 13, 16–20); (iii) final reported errors invert normalisation and use Palabos-measured nF and Ki (Eqs. 28–34), not surrogate outputs. The surrogate sψ is trained independently and frozen during refinement (III-C/D, Alg. 1); Lprop backpropagates through sψ only as a differentiable stand-in for expensive non-differentiable LBM. That is ordinary supervised inverse design with a proxy, not a definitional loop: nothing forces Palabos errors to equal the refinement objective, and the paper’s own tables show systematic surrogate–Palabos gaps (e.g. Table IV porosity). Candidate screening by the surrogate (Alg. 1 step 19) is shared with baselines under the same evaluator protocol (IV-C) and is an evaluation-bias concern, not circularity by construction. No self-citation uniqueness theorem, no fitted constant renamed as a first-principles prediction, and no renaming of a known identity. Score 0.

Assumptions & free parameters 6 free parameters · 6 assumptions · 1 invented entities

The central empirical claim rests on standard deep generative modelling assumptions, LBM-as-ground-truth permeability, and several hand-chosen training hyperparameters and modelling choices (log-permeability normalisation, frozen independent surrogate, joint refinement objective). No new physical entities are postulated; physics enters as simulation labels and a learned surrogate of those labels.

free parameters (6)
  • Refinement loss weights λ_diff, λ_prop, λ_rec, λ_anchor = 1.0, 0.1, 0.1, 0.001
    Set by authors for denoiser-decoder fine-tuning; directly control the property vs reconstruction vs anchor trade-off that drives the main claimed gains.
  • pVAE / diffusion latent spatial resolution = 5×5×5 (with 32 channels)
    Chosen via ablation on real data (3³–6³); 5³ adopted for all main results and thus conditions reported performance.
  • Denoiser and decoder refinement learning rates = 1e-5 (denoiser), 1e-6 (decoder)
    Hand-set asymmetric rates for joint updates; affect whether refinement improves or destabilizes decoder manifold.
  • Property normalisation statistics μ, σ on nF and log K_i = computed from training set (values not tabulated)
    Dataset-derived scales used in all losses and conditioning (Eq. 2); not physics constants but fit to the training corpus.
  • pVAE loss weights λ_rec, λ_KL, λ_lat = not numerically specified in text
    Balance reconstruction, prior, and latent property head; shape the ‘physically informative’ latent space the diffusion model consumes.
  • Latent optimisation step counts and λ_z (baselines/ablation) = 10, 30, 50 steps (λ_z unspecified numerically)
    Control strength of post-hoc latent search used in comparisons; tables show metrics depend strongly on 10/30/50 steps.
assumptions (6)
  • domain assumption Palabos LBM directional permeability on thresholded 100³ binary voxels is a sufficient ground-truth label for training and final verification of inverse-design quality.
    Invoked throughout §III-A/E and §IV-D; all ‘physics-verified’ claims reduce to agreement with this solver under the paper’s boundary/setup choices.
  • ad hoc to paper An independently trained residual CNN surrogate on normalised/log properties yields gradients useful for improving true simulator metrics when the surrogate is frozen.
    Core of §III-B and refinement objective (17)–(18); without this, closed-loop refinement has no justified link to Palabos errors.
  • domain assumption Log-transform plus standardisation of permeability (Eq. 2) is the correct condition space for stable multi-property matching across orders of magnitude.
    Stated in §III-A as necessary for training stability; all generative conditioning and Lprop live in this space.
  • domain assumption Standard VAE ELBO-style training plus a latent property head produces a latent manifold suitable for conditional diffusion and decoding of valid pore geometry.
    §III-C pVAE objective (11); shared encoder/decoder backbone for all methods.
  • standard math Gaussian latent diffusion with FiLM conditioning on (c, t) is an appropriate generative model for the pVAE posterior latents of porous media.
    §III-C equations (12)–(14), following latent diffusion practice [20].
  • domain assumption Synthetic procedural geometries plus cropped micro-CT subvolumes with stride-25 extraction are representative enough to support claims about real porous-media inverse design.
    §IV-A dataset construction; generalisation beyond these corpora is assumed when stating a ‘scalable route’ in the abstract/conclusion.
invented entities (1)
  • Denoiser–decoder joint refinement mechanism under frozen surrogate property feedback
    purpose: Reduce mismatch between diffusion-sampled latents and the pVAE decoder’s training distribution while steering porosity/permeability.
    Named as a main contribution and implemented via trainable subset in Algorithm 1 and loss (17); it is a procedural training stage, not a new physical object. independent_evidence is false beyond the paper’s own ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-Guided Generative AI for Property-Targeted 3D Porous Media Design." pith.science (2026). https://pith.science/paper/RL55BWLW

@misc{pith2026260724274,
  author       = {Pith},
  title        = {Pith review of: Physics-Guided Generative AI for Property-Targeted 3D Porous Media Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RL55BWLW}},
  note         = {Machine review of arXiv:2607.24274}
}
read the original abstract

Inverse design of three-dimensional porous media is central to applications in filtration, catalysis, energy storage, fuel cells, thermal management, and biomedical scaffolds, but remains challenging because many distinct pore geometries can share similar porosity or permeability while small structural changes can strongly affect transport behaviour. This paper proposes a physics-guided generative AI framework for property-targeted porous media design, combining a property-aware variational autoencoder, a conditional latent diffusion model, and an independently trained differentiable structure-to-property surrogate. The framework learns a compact, physically informative latent design space, generates porous structures conditioned on target porosity and directional permeability, and refines generated samples using property-level feedback during denoising and decoding. Experiments on procedurally generated structures and real micro-CT porous-media datasets show improved target-property matching, directional permeability control, and property correlation compared with representative property-aware variational-autoencoder and latent-diffusion baselines. The results demonstrate a scalable route towards controllable inverse design of complex porous geometries and establish a foundation for simulation-informed generative AI tools in engineering and advanced materials discovery.

Figures

Figures reproduced from arXiv: 2607.24274 by the authors.

Figure 1
Figure 1. Overview of the proposed physical-property-guided latent diffusion framework for porous media inverse design. (a) A [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Porous structures generated by the pVAE method for randomly selected synthetic targets. Each column corresponds to [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Porous structures generated by the latent diffusion method for randomly selected synthetic targets. Each column [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Porous structures generated by the proposed method for randomly selected synthetic targets. Each column corresponds [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Porous structures generated by the pVAE method for randomly selected real micro-CT targets. Each column corresponds [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Porous structures generated by the latent diffusion method for randomly selected real micro-CT targets. Each column [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Porous structures generated by the proposed method for randomly selected real micro-CT targets. Each column [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Comparison of the impact of latent space dimension [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 3 linked inside Pith

  1. [1]

    New 2d to 3d reconstruction of heterogeneous porous media via deep generative adversarial networks (gans),

    H. Amiri, H. V ogel, and O. Plümper, “New 2d to 3d reconstruction of heterogeneous porous media via deep generative adversarial networks (gans),”Journal of Geophysical Research: Machine Learning and Computation, vol. 1, no. 3, p. e2024JH000178, 2024

    2024

  2. [2]

    Diffusion model-based generation of three-dimensional multiphase pore-scale images,

    L. Zhu, B. Bijeljic, and M. J. Blunt, “Diffusion model-based generation of three-dimensional multiphase pore-scale images,”Transport in Porous Media, vol. 152, no. 3, p. 22, 2025

    2025

  3. [3]

    Controlled latent diffusion models for 3d porous media reconstruction,

    D. Naiff, B. P. Schaeffer, G. Pires, D. Stojkovic, T. Rapstine, and F. Ramos, “Controlled latent diffusion models for 3d porous media reconstruction,”arXiv preprint arXiv:2503.24083, 2025

    arXiv 2025

  4. [4]

    Microlib: A library of 3d microstructures generated from 2d micrographs using slicegan,

    S. Kench, I. Squires, A. Dahari, and S. J. Cooper, “Microlib: A library of 3d microstructures generated from 2d micrographs using slicegan,” Scientific Data, vol. 9, no. 1, p. 645, 2022

    2022

  5. [5]

    Reconstruction of 3d random media from 2d images: generative adversarial learning approach,

    E. Kononov, M. Tashkinov, and V . V . Silberschmidt, “Reconstruction of 3d random media from 2d images: generative adversarial learning approach,”Computer-Aided Design, vol. 158, p. 103498, 2023

    2023

  6. [6]

    Generative adversarial networks for reconstruction of three-dimensional porous media from two-dimensional slices,

    D. V olkhonskiy, E. Muravleva, O. Sudakov, D. Orlov, E. Burnaev, D. Koroteev, B. Belozerov, and V . Krutko, “Generative adversarial networks for reconstruction of three-dimensional porous media from two-dimensional slices,”Physical Review E, vol. 105, no. 2, p. 025304, 2022

    2022

  7. [7]

    Using physics informed generative adversarial networks to model 3d porous media,

    Z. Ren and S. Srinivasan, “Using physics informed generative adversarial networks to model 3d porous media,”arXiv preprint arXiv:2409.11541, 2024

    arXiv 2024

  8. [8]

    Synthesizing controlled microstructures of porous media using generative adversarial networks and reinforcement learning,

    P. C. Nguyen, N. N. Vlassis, B. Bahmani, W. Sun, H. Udaykumar, and S. S. Baek, “Synthesizing controlled microstructures of porous media using generative adversarial networks and reinforcement learning,” Scientific reports, vol. 12, no. 1, p. 9034, 2022

    2022

Show all 20 references
  1. [9]

    Conditional diffusion-based microstructure reconstruction,

    C. Düreth, P. Seibert, D. Rücker, S. Handford, M. Kästner, and M. Gude, “Conditional diffusion-based microstructure reconstruction,”Materials Today Communications, vol. 35, p. 105608, 2023

    2023

  2. [10]

    Da-vegan: Differentiably augmenting vae-gan for microstructure reconstruction from extremely small data sets,

    Y . Zhang, P. Seibert, A. Otto, A. Raßloff, M. Ambati, and M. Käst- ner, “Da-vegan: Differentiably augmenting vae-gan for microstructure reconstruction from extremely small data sets,”Computational Materials Science, vol. 232, p. 112661, 2024

    2024

  3. [11]

    Camox: A diffusion-based method with few-shot learning for environment-guided camouflage pattern generation,

    T. T. P. Nguyen, T. Gulrez, J. B. Culpepper, S. L. Phung, and H. T. Le, “Camox: A diffusion-based method with few-shot learning for environment-guided camouflage pattern generation,”IEEE Transactions on Artificial Intelligence, 2025

    2025

  4. [12]

    Two-order deep learning for generalized synthesis of radiation patterns for antenna arrays,

    Z. Zhou, Z. Wei, J. Ren, Y . Yin, G. F. Pedersen, and M. Shen, “Two-order deep learning for generalized synthesis of radiation patterns for antenna arrays,”IEEE Transactions on Artificial Intelligence, vol. 4, no. 5, pp. 1359–1368, 2022

    2022

  5. [13]

    Itf-vae: Variational auto-encoder using interpretable continuous time series features,

    H. Klopries and A. Schwung, “Itf-vae: Variational auto-encoder using interpretable continuous time series features,”IEEE Transactions on Artificial Intelligence, 2025

    2025

  6. [14]

    Deep generative modeling for mechanistic-based learning and design of metamaterial systems,

    L. Wang, Y .-C. Chan, F. Ahmed, Z. Liu, P. Zhu, and W. Chen, “Deep generative modeling for mechanistic-based learning and design of metamaterial systems,”Computer Methods in Applied Mechanics and Engineering, vol. 372, p. 113377, 2020. 15

    2020

  7. [15]

    Deep learning-aided inverse design of porous metamaterials,

    P. T. Nguyen, Y . Heider, D. M. Kochmann, and F. Aldakheel, “Deep learning-aided inverse design of porous metamaterials,”Computer Methods in Applied Mechanics and Engineering, vol. 449, no. 11849, p. 118499, 2026

    2026

  8. [16]

    Pore-gnn: A graph neural network-based framework for predicting flow properties of porous media from micro-ct images,

    M. K. Alzahrani, A. Shapoval, Z. Chen, and S. S. Rahman, “Pore-gnn: A graph neural network-based framework for predicting flow properties of porous media from micro-ct images,”Advances in Geo-Energy Research, vol. 10, no. 1, pp. 39–55, 2023

    2023

  9. [17]

    Multiscale simulation of spatially correlated microstructure via a latent space representation,

    R. E. Jones, C. M. Hamel, D. Bolintineanu, K. Johnson, R. B. de Macedo, J. Fuhg, N. Bouklas, and S. Kramer, “Multiscale simulation of spatially correlated microstructure via a latent space representation,”International Journal of Solids and Structures, vol. 301, p. 112966, 2024

    2024

  10. [18]

    Investigat- ing the impact of deformation on foam permeability through ct scans and the lattice-boltzmann method,

    N. T. Phu, U. Navrath, Y . Heider, J. Carmai, and B. Markert, “Investigat- ing the impact of deformation on foam permeability through ct scans and the lattice-boltzmann method,”PAMM, vol. 24, no. 1, p. e202300154, 2024

    2024

  11. [19]

    Synthetic porous microstructures: Automatic design, simulation, and permeability analysis,

    T. Lavigne, C. A. S. Afanador, A. Obeidat, and S. Urcun, “Synthetic porous microstructures: Automatic design, simulation, and permeability analysis,”arXiv preprint arXiv:2502.14518, 2025

    2025 arXiv

  12. [20]

    High- resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.