Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Equivariant U-Shaped Neural Operators for the Cahn-Hilliard Phase-Field Model

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read An equivariant U-shaped neural operator (E-UNO) learns Cahn-Hilliard phase-field evolution from short past histories and beats FNO and UNO baselines, cutting early-stage max relative L2 error by 34.32% while preserving thermodynamic trends.

desk verdict A plausible symmetry-aware neural operator for Cahn-Hilliard with consistent empirical gains, but the under-specified stochastic data-generation process currently muddies what is actually being learned. read the letter →

arxiv 2509.01293 v3 pith:V6HBRFQJ submitted 2025-09-01 cs.LG physics.comp-phphysics.flu-dyn

Xiao Xue , Marco F.P. ten Eikelder , Tianyue Yang , Yiqing Li , Kan He , Shuo Wang , Peter V. Coveney This is my paper
classification cs.LGphysics.comp-phphysics.flu-dyn
keywords neuraloperatorCahn-Hilliardequationphase-fieldmodelingD4equivarianceU-shapedsuper-resolutionsurrogatemodelmicrostructureevolution
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a neural operator that respects the symmetries of the Cahn-Hilliard equation—rotations and reflections of the square domain, the dihedral group D4—can learn the time evolution of phase separation from as few as five past snapshots and forecast several future frames accurately. The proposed Equivariant U-shaped Neural Operator (E-UNO) wraps a multi-resolution U-shaped Fourier neural operator in pre- and post-processing group transformations and an equivariance loss, so predictions commute with domain symmetries. Compared with the standard U-shaped Neural Operator and the Fourier Neural Operator, E-UNO reports systematically lower errors, with a 34.32% reduction in early-stage maximum relative L2 error, tighter error distributions, and free-energy trajectories that closely track thermodynamic dissipation. The model also transfers to a 200x200 grid after training on 100x100 without retraining, and runs at a nominal speedup of about 8000x over the reference numerical solver, with the hardware caveat noted by the authors. If these results hold, E-UNO is a fast, symmetry-aware surrogate for microstructure evolution that could accelerate phase-field simulations in materials and soft-matter applications.

What carries the argument

The central object is E-UNO itself: a U-shaped, multi-resolution encoder-decoder whose layers use global spectral (Fourier) convolution, wrapped by external D4 group transformations. The key identity is the equivariance loss Leq = Σ_{g∈D4} Σ_{x∈Ωh} ||G(φt)(x) − [g^{-1}·φ_{t+1}^{g}](x)||², which forces the predicted evolution to commute with rotations and reflections of the square domain. This pre/post-processing trick enforces physical symmetry without changing the internal architecture or inference cost.

What would settle it

Inspect the simulation setup and trajectory data: if all 300 runs start from exactly Φ=0 with no stochastic forcing, the unique-trajectory claim is false and the learned operator has no well-defined target; if stochastic forcing exists, Eq. (1) must be augmented. Separately, take any test field, rotate it by 90 degrees, predict, and compare against the 90-degree rotation of the prediction made on the original field—a mismatch beyond numerical tolerance would falsify the D4-equivariance claim.

Watch

Extended reading notes

Core claim

The paper claims that enforcing D4 equivariance—the eight rotations and reflections of the square domain—around a U-shaped Fourier neural operator yields a systematically better surrogate for the Cahn-Hilliard solution operator. E-UNO keeps the UNO backbone and imposes symmetry externally: each input is transformed by every group element, passed through the network, and inversely transformed before loss evaluation, so the learned map commutes with the group. Reported gains: early-stage max relative L2 error down 34.32% versus UNO, up to 11% median improvement, free-energy deviation down about 11%, and direct transfer from 100x100 training to 200x200 prediction without retraining. The same eq

Load-bearing premise

The load-bearing premise is that the 300 training trajectories are meaningful samples of a well-defined Cahn-Hilliard process, but Section 4.1 says each starts from identical uniform Φ=0 and varies only through stochastic terms in the dynamics that never appear in Eq. (1) and are never defined.

Editorial extensions

If this is right

  • Cahn-Hilliard microstructure evolution can be forecast from five past frames with mean relative L2 errors below 0.02 across configurations, and longer input histories further reduce error.
  • Enforcing D4 equivariance is a lightweight add-on: it cuts early-stage maximum relative L2 error by 34.32% and median error by up to 11% versus UNO at unchanged inference time.
  • E-UNO preserves thermodynamic consistency, tracking the normalized free-energy dissipation curve with about an 11% smaller deviation than UNO.
  • A model trained on a 100x100 grid predicts directly on a 200x200 grid without retraining, with localized errors mostly below 10%, demonstrating mesh-independent super-resolution.
  • The nominal acceleration over the reference solver is about 8000x (0.03 s inference versus 240 s simulation), though the authors note the comparison is not hardware-neutral.

Reading between the lines

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

  • Beyond the paper: the same external D4 pre/post-processing could be attached to any operator architecture and should extend to cubic symmetry in 3D or to translation-equivariant losses on periodic domains wherever the PDE's invariance group is known.
  • Beyond the paper: the equivariance loss likely acts as a strong physics-informed data augmenter; a training-curve comparison against pure data augmentation with the same D4 transforms would separate 'symmetry as prior' from 'symmetry as extra samples'.
  • Beyond the paper: because each trajectory reportedly starts from uniform Φ=0 with variability from undefined stochastic terms, the actual target is a stochastic map, not the deterministic Cahn-Hilliard solution operator; defining those terms and the corresponding probability measure would make the learning problem and generalization claims well-posed.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. E-UNO is a U-shaped Fourier neural operator augmented with a D4-equivariance loss and an optional H1 gradient loss, proposed as a surrogate for the Cahn-Hilliard equation. Given five past snapshots of the phase field, it predicts three future snapshots; the authors compare it against UNO and FNO on 300 COMSOL simulations, report lower relative L2 errors especially at early stages, reproduce the free-energy decay, and demonstrate super-resolution from 100x100 to 200x200. The paper also studies input/output sequence lengths and the effect of the H1 loss.

Significance. If the empirical claims are fully supported, E-UNO would be a practically useful surrogate: fast inference (~0.03 s per case), improved accuracy over standard operators at equal inference cost, physically consistent free-energy behavior, and resolution transfer. The manuscript has several strengths: it compares against two standard baselines, includes an E-FNO ablation in Appendix A.2, reports hyperparameters and timing, and explicitly assesses the H1-loss variant. The main weaknesses are that the data-generation process is not specified consistently with Eq. (1), the equivariance loss in Eq. (22) is misstated, and the super-resolution comparison appears to mix different stochastic realizations. These issues must be resolved before the central claims can be evaluated.

major comments (4)
  1. [Sec. 4.1; Eq. (1)] Section 4.1 states that every simulation starts from uniform Phi=0 and variability is introduced by 'a distinct random seed controlling the stochastic terms in the dynamics,' but Eq. (1) is deterministic and Phi=0 is a stationary state under homogeneous Neumann conditions. No stochastic term is defined in Section 2 or 4.1, and Section 4.6 refers instead to 'randomness in the initialization process.' The target operator in Eq. (13) is therefore unspecified: it cannot be the deterministic CH solution operator, and if it is a stochastic map the D4-equivariance loss is unjustified unless the noise is included in the model input and transformed consistently. Please define the data-generation model precisely (including covariance and symmetry properties of any noise), or replace it with random initial conditions.
  2. [Sec. 3.3, Eq. (22)] Equation (22) does not express the equivariance constraint described in Sec. 3.3 and Figure 5. As written, L_eq compares the prediction G(phi_t)(x) with a transformed ground-truth field [g^{-1}·phi_{t+1} g](x), which would force the output to match a rotated/reflected target rather than make the operator commute with D4. The notation for the group action is also undefined. The correct constraint should relate G(g·phi_t) and g·G(phi_t), e.g. sum_g ||G(g·phi_t) - g·G(phi_t)||^2 with the action defined as (g·phi)(x)=phi(g^{-1}x). This must be fixed because the equivariance loss is the paper's main contribution.
  3. [Sec. 4.6] Section 4.6 reports super-resolution by training on 100x100 and testing on 200x200, but states that 'due to the randomness in the initialization process, the initial conditions at low resolution cannot be exactly reproduced at higher resolutions.' If the fine-grid reference is a different stochastic realization, the pointwise error maps in Fig. 11 measure initial-condition mismatch as well as discretization error, so they do not support the claim of mesh-independent generalization. Clarify whether the same physical realization is used (e.g., by downsampling the fine initial condition) or replace the pointwise comparison with ensemble-level statistical metrics.
  4. [Sec. 4.4; Appendix A.2] The improvements attributed to the D4-equivariance loss in Sec. 4.4 and Appendix A.2 are not compared with the standard data-augmentation baseline: training UNO or FNO on D4-transformed input-output pairs without a separate equivariance term. Because the equivariance loss is evaluated by transforming inputs and inverse-transforming outputs, it uses the same data as augmentation, so the observed gains could be due to increased effective training diversity rather than to the symmetry regularizer. Adding this baseline would make the claim that 'symmetry enforcement' is responsible for the improvement precise.
minor comments (6)
  1. [Sec. 4.1] 'interface thickness lambda=0.01' mislabels the parameter; in Eq. (5), lambda is the surface energy density and epsilon is the interface thickness.
  2. [Fig. 7] The text refers to Fig. 7a for input-length effects and Fig. 7b for output-length effects, but the caption describes panel (a) as t_out and panel (b) as t_in. The references should be corrected.
  3. [Eq. (22)] Define the group action on fields explicitly, e.g. (g·phi)(x)=phi(g^{-1}x); the current notation [g^{-1}·phi_{t+1} g](x) is not meaningful.
  4. [Sec. 4.4] 'Averaged over all time windows, the max relative L2 error D(t) is reduced by 34.32%' is ambiguous because D(t) in Eq. (26) is a time-dependent error. Clarify whether this is the maximum over time, averaged over windows, and report standard deviations or confidence intervals.
  5. [Sec. 4.4] 'Based on previous study, we used 5 input timeframes' lacks a reference.
  6. [Fig. 9] The caption contains 'Both E-UNO and UNO unrecognizable error,' which appears to be a typo (probably 'negligible error').

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: E-UNO is an empirically trained surrogate evaluated on held-out data; equivariance is a soft regularizer, not a derived prediction.

full rationale

The paper contains no derivation chain that reduces its predictions to its inputs. The target operator is learned from COMSOL-generated trajectories (Section 4.1) and evaluated on a held-out 10% test split (Section 4.2); accuracy claims in Sections 4.4–4.6 are test-set comparisons against FNO/UNO baselines, not fitted parameters renamed as predictions. The D4-equivariance term (Eq. 22) is a training regularizer; it is not used to synthesize test outputs. Self-citations (e.g., [8,10,12,15,28,32,50]) are background/future-work references and do not carry the E-UNO argument. Two manuscript issues affect correctness but not circularity: Section 4.1 says variability comes from 'stochastic terms in the dynamics' while Section 4.6 attributes differences to 'randomness in the initialization process'; and Eq. 22 as typeset appears to compare G(phi_t) to transformed targets rather than expressing the commutation G(g phi_t)=g G(phi_t) described in the text. These are inconsistencies in problem definition and notation, not reductions of the reported predictions to the training data. Therefore no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the dataset faithfully representing deterministic Cahn-Hilliard dynamics, on the D4 equivariance of the solution operator, and on several hand-chosen hyperparameters (loss weighting, mode counts, temporal fragment selection) that are not systematically varied.

free parameters (3)
  • equivariance loss weight = not reported (implicitly 1 in Eq. 24)
    The total loss is L_data + L_eq without a weighting coefficient; this hand-set choice affects the balance between accuracy and symmetry and is not ablated.
  • Fourier mode counts per layer = [[32,32],[16,16],[8,8],[4,4],[8,8],[16,16],[32,32]]
    Mode truncation k_max per layer is a hand-chosen architectural hyperparameter that controls spectral resolution; no sensitivity study is given.
  • temporal fragment selection = 30 fragments per simulation, focused on rapid interfacial evolution
    The extraction of 30 temporal windows per simulation is a manual choice that may oversample the dynamic regime where equivariance benefits are largest.
assumptions (4)
  • domain assumption The training data are generated by the deterministic Cahn-Hilliard equation (1)
    The paper's theory uses deterministic CH, but Section 4.1 mentions stochastic terms in the dynamics, creating ambiguity about what operator is actually learned.
  • domain assumption The solution operator of the Cahn-Hilliard equation on a square domain commutes with the dihedral group D4
    True for homogeneous Neumann conditions on a square, but not stated or validated explicitly; the equivariance loss relies on this.
  • domain assumption The U-shaped neural operator (UNO) backbone is a valid and well-specified model
    UNO is used as a baseline and backbone but no primary citation or formal architecture definition is given in the paper.
  • ad hoc to paper Adding the equivariance loss and H1 gradient loss does not change the target functional mapping
    The paper introduces these modifications heuristically; no proof or analysis shows they preserve the learned operator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Equivariant U-Shaped Neural Operators for the Cahn-Hilliard Phase-Field Model." pith.science (2026). https://pith.science/paper/V6HBRFQJ

@misc{pith2026250901293,
  author       = {Pith},
  title        = {Pith review of: Equivariant U-Shaped Neural Operators for the Cahn-Hilliard Phase-Field Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V6HBRFQJ}},
  note         = {Machine review of arXiv:2509.01293}
}
read the original abstract

Phase separation in binary mixtures, governed by the Cahn-Hilliard equation, plays a central role in interfacial dynamics across materials science and soft matter. While numerical solvers are accurate, they are often computationally expensive and lack flexibility across varying initial conditions and geometries. Neural operators provide a data-driven alternative by learning solution operators between function spaces, but current architectures often fail to capture multiscale behavior and neglect underlying physical symmetries. Here we show that an equivariant U-shaped neural operator (E-UNO) can learn the evolution of the phase-field variable from short histories of past dynamics, achieving accurate predictions across space and time. The model combines global spectral convolution with a multi-resolution U-shaped architecture and regulates translation equivariance to align with the underlying physics. E-UNO outperforms standard Fourier neural operator and U-shaped neural operator baselines, particularly on fine-scale and high-frequency structures. By encoding symmetry and scale hierarchy, the model generalizes better, requires less training data, and yields physically consistent dynamics. This establishes E-UNO as an efficient surrogate for complex phase-field systems.

Figures

Figures reproduced from arXiv: 2509.01293 by the authors.

Figure 1
Figure 1. Ginzburg-Landau free energy potential W = W(Φ). The free energy potential has a double well structure with minima at the pure phases represented by Φ = ±1. The corresponding chemical potential takes the form: µ = λ  Φ(Φ2 − 1) ϵ − ϵ∆2Φ  . (6) The equilibrium solution of the Cahn-Hilliard system (1) is governed by a constant chemical potential µ = const, where const does not depend on (x, t) ∈ Ω×T . Inserting the ch… view at source ↗
Figure 2
Figure 2. In one spatial dimension (coordinate s) the Cahn-Hilliard equation has an equilbrium solution with an analytical expression. Panel (a) shows the equilibrium profile Φeq = Φeq(s), and panel (b) shows the corresponding free energy density. In both panels the interface width parameter is varied. implies that pure phases have no mobility and the right-hand side of (1) vanishes there. In some scenarios, thermodynamic con… view at source ↗
Figure 3
Figure 3. Schematic illustration of the neural operator architecture. The mapping from input functions [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Schematic of the Fourier Neural Operator (FNO) architecture. The mapping from input functions [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Architectures of the U-shaped Neural Operator and its Equivariant Extension (E-UNO). Panel [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Spatiotemporal evolution of the predicted field compared to the ground truth at three dimensionless time instances [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Panel a Box plots of the error D for training and test sets at three time steps (tout = 1∆t, 2∆t, and 3∆t). The boxes represent the interquartile range, the horizontal line inside each box shows the median, and whiskers indicate the range of the data excluding outliers…
Figure 8
Figure 8. Figure 8: Temporal evolution of the spatially integrated relative [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Comparison of predicted phase-field evolution errors for E-UNO, UNO, and FNO from [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Temporal evolution of the normalized free energy, [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Super-resolution prediction of phase-field evolution using E-UNO, compared to the ground truth reference, for [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MENO: MeanFlow-Enhanced Neural Operators for Dynamical Systems

    cs.LG 2026-04 unverdicted novelty 6.0 of 10

    MENO restores multi-scale structure in neural-operator PDE surrogates via one-step improved MeanFlow, claiming up to 2× better power-spectrum accuracy and up to 14× faster inference than DDIM enhancement.

Reference graph

Works this paper leans on

53 extracted references · 48 canonical work pages · cited by 1 Pith paper

  1. [1]

    Anderson, G.B

    D.M. Anderson, G.B. McFadden, and A.A. Wheeler. Diffuse-interface methods in fluid mechanics. Annual review of fluid mechanics, 30:139–165, 1998

    work page 1998

  2. [2]

    J. Kim. Phase-field models for multi-component fluid flows.Communications in Computational Physics, 12:613–661, 2012

    work page 2012

  3. [3]

    Wight and J

    C.L. Wight and J. Zhao. Solving Allen-Cahn and Cahn-Hilliard equations using the adaptive physics informed neural networks.arXiv preprint arXiv:2007.04542, 2020

    arXiv 2007

  4. [4]

    J. D. van der Waals. The thermodynamic theory of capillarity under the hypothesis of a continuous variation of density.Journal of Statistical Physics, 20:200–244, 1979. 17

    work page 1979

  5. [5]

    Cahn and J.E

    J.W. Cahn and J.E. Hilliard. Free energy of a nonuniform system. I. Interfacial free energy.The Journal of chemical physics, 28:258–267, 1958

    work page 1958

  6. [6]

    Gurtin, D

    M.E. Gurtin, D. Polignone, and J. Vinals. Two-phase binary fluids and immiscible fluids described by an order parameter.Math. Models Methods Appl. Sci., 6:815–831, 1996

    work page 1996

  7. [7]

    Lowengrub and L

    J. Lowengrub and L. Truskinovsky. Quasi–incompressible cahn–hilliard fluids and topological transi- tions.Proceedings of the Royal Society of London. Series A: Mathematical, Physical and Engineering Sciences, 454(1978):2617–2654, 1998

    work page 1978

  8. [8]

    ten Eikelder, K.G

    M.F.P. ten Eikelder, K.G. Van Der Zee, I. Akkerman, and D. Schillinger. A unified framework for Navier–Stokes Cahn–Hilliard models with non-matching densities.Mathematical Models and Methods in Applied Sciences, 33:175–221, 2023

    work page 2023

Show all 53 references
  1. [9]

    Abels, H

    H. Abels, H. Garcke, and A. Poiatti. Mathematical analysis of a diffuse interface model for multi-phase flows of incompressible viscous fluids with different densities.Journal of Mathematical Fluid Mechanics, 26:29, 2024

    2024

  2. [10]

    AunifiedframeworkforN-phaseNavier–StokesCahn–HilliardAllen–Cahnmixture models with non-matching densities.Journal of Fluid Mechanics, 1013:A26, 2025

    M.F.P.tenEikelder. AunifiedframeworkforN-phaseNavier–StokesCahn–HilliardAllen–Cahnmixture models with non-matching densities.Journal of Fluid Mechanics, 1013:A26, 2025

    2025

  3. [11]

    Knopf and Y

    P. Knopf and Y. Liu. A thermodynamically consistent free boundary model for two-phase flows in an evolving domain with moving contact lines, variable contact angles and bulk-surface interaction.arXiv preprint arXiv:2507.01618, 2025

    2025 arXiv

  4. [12]

    ten Eikelder, E.H

    M.F.P. ten Eikelder, E.H. van Brummelen, and D. Schillinger. Compressible n-phase fluid mixture models.arXiv preprint arXiv:2503.24225, 2025

    2025 arXiv

  5. [13]

    Kay and R

    D. Kay and R. Welford. Efficient numerical solution of cahn–hilliard–navier–stokes fluids in 2d.SIAM Journal on Scientific Computing, 29:2241–2257, 2007

    2007

  6. [14]

    Minjeaud

    S. Minjeaud. An unconditionally stable uncoupled scheme for a triphasic Cahn–Hilliard/Navier–Stokes model.Numerical Methods for Partial Differential Equations, 29:584–618, 2013

    2013

  7. [15]

    ten Eikelder and D

    M.F.P. ten Eikelder and D. Schillinger. The divergence-free velocity formulation of the consistent Navier-Stokes Cahn-Hilliard model with non-matching densities, divergence-conforming discretization, and benchmarks.J. Comput. Phys., 513:113148, 2024

    2024

  8. [16]

    Garcke and S

    H. Garcke and S. Yayla. Long-time dynamics for a cahn–hilliard tumor growth model with chemotaxis. Zeitschrift für angewandte Mathematik und Physik, 71:123, 2020

    2020

  9. [17]

    Ebenbeck and H

    M. Ebenbeck and H. Garcke. Analysis of a Cahn–Hilliard–Brinkman model for tumour growth with chemotaxis.Journal of Differential Equations, 266:5998–6036, 2019

    2019

  10. [18]

    Bertozzi, S

    A. Bertozzi, S. Esedoglu, and A. Gillette. Analysis of a two-scale Cahn–Hilliard model for binary image inpainting.Multiscale Modeling & Simulation, 6:913–936, 2007

    2007

  11. [19]

    Khain and L.M

    E. Khain and L.M. Sander. Generalized Cahn-Hilliard equation for biological applications.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 77:051129, 2008

    2008

  12. [20]

    Elson, E

    E.L. Elson, E. Fried, J.E. Dolbow, and G.M. Genin. Phase separation in biological membranes: inte- gration of theory and experiment.Annual review of biophysics, 39:207–226, 2010

    2010

  13. [21]

    Berry, C.P

    J. Berry, C.P. Brangwynne, and M. Haataja. Physical principles of intracellular organization via active and passive phase transitions.Reports on Progress in Physics, 81:046601, 2018. 18

    2018

  14. [22]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

    1901

  15. [23]

    Voulodimos, N

    A. Voulodimos, N. Doulamis, A. Doulamis, and E. Protopapadakis. Deep learning for computer vision: A brief review.Computational intelligence and neuroscience, 2018:7068349, 2018

    2018

  16. [24]

    Ding.Data-driven design of fault diagnosis and fault-tolerant control systems

    S.X. Ding.Data-driven design of fault diagnosis and fault-tolerant control systems. Springer, 2014

    2014

  17. [25]

    Brunton, B.R

    S.L. Brunton, B.R. Noack, and P. Koumoutsakos. Machine learning for fluid mechanics.Annual Review of Fluid Mechanics, 52:477–508, 2020

    2020

  18. [26]

    Karniadakis, I.G

    G.E. Karniadakis, I.G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang. Physics-informed machine learning.Nature Reviews Physics, 3:422–440, 2021

    2021

  19. [27]

    Cheng, M

    S. Cheng, M. Bocquet, W. Ding, T.S. Finn, R. Fu, J. Fu, Y. Guo, E. Johnson, S. Li, C. Liu, et al. Machine learning for modelling unstructured grid data in computational physics: a review.Information Fusion, page 103255, 2025

    2025

  20. [28]

    M. Wang, X. Xue, M. Gao, and P.V. Coveney. Quantum-informed machine learning for the prediction of chaotic dynamical systems.arXiv preprint arXiv:2507.19861, 2025

    2025 arXiv

  21. [29]

    Fukami, Y

    K. Fukami, Y. Nabae, K. Kawai, and K. Fukagata. Synthetic turbulent inflow generator using machine learning.Physical Review Fluids, 4:064603, 2019

    2019

  22. [30]

    Xue, H.-D

    X. Xue, H.-D. Yao, and L. Davidson. Synthetic turbulence generator for lattice boltzmann method at the interface between rans and les.Physics of Fluids, 34:055118, 2022

    2022

  23. [31]

    X. Yang, S. Zafar, J.-X. Wang, and H. Xiao. Predictive large-eddy-simulation wall modeling via physics- informed neural networks.Physical Review Fluids, 4:034602, 2019

    2019

  24. [32]

    X. Xue, S. Wang, H.-D. Yao, L. Davidson, and P.V. Coveney. Physics informed data-driven near-wall modelling for lattice boltzmann simulation of high reynolds number turbulent flows.Communications Physics, 7:338, 2024

    2024

  25. [33]

    Cheng, F

    M. Cheng, F. Fang, C. C. Pain, and I.M. Navon. Data-driven modelling of nonlinear spatio-temporal fluid flows using a deep convolutional generative adversarial network.Computer Methods in Applied Mechanics and Engineering, 365:113000, 2020

    2020

  26. [35]

    L. Lu, P. Jin, G. Pang, Z. Zhang, and G.E. Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3:218–229, 2021

    2021

  27. [36]

    Q. Cao, S. Goswami, and G.E. Karniadakis. Laplace neural operator for solving differential equations. Nature Machine Intelligence, 6:631–640, 2024

    2024

  28. [37]

    Shen, J.-J

    Z.-H. Shen, J.-J. Wang, J.-Y. Jiang, S.X. Huang, Y.-H. Lin, C.-W. Nan, L.-Q. Chen, and Y. Shen. Phase-field modeling and machine learning of electric-thermal-mechanical breakdown of polymer-based dielectrics.Nature Communications, 10:1843, 2019

    2019

  29. [38]

    Y. Feng, Q. Wang, D. Wu, Z. Luo, X. Chen, T. Zhang, and W. Gao. Machine learning aided phase field method for fracture mechanics.International Journal of Engineering Science, 169:103587, 2021. 19

    2021

  30. [39]

    Teichert, A.R

    G.H. Teichert, A.R. Natarajan, A. Van der Ven, and K. Garikipati. Machine learning materials physics: Integrable deep neural networks enable scale bridging by learning free energy functions.Computer Methods in Applied Mechanics and Engineering, 353:201–216, 2019

    2019

  31. [40]

    Oommen, K

    V. Oommen, K. Shukla, S. Desai, R. Dingreville, and G.E. Karniadakis. Rethinking materials sim- ulations: Blending direct numerical simulations with neural operators.npj Computational Materials, 10:145, 2024

    2024

  32. [41]

    Cohen and M

    T. Cohen and M. Welling. Group equivariant convolutional networks. InInternational conference on machine learning, pages 2990–2999. PMLR, 2016

    2016

  33. [42]

    Thomas, T

    N. Thomas, T. Smidt, S. Kearnes, L. Yang, L. Li, K. Kohlhoff, and P. Riley. Tensor field net- works: Rotation-and translation-equivariant neural networks for 3d point clouds.arXiv preprint arXiv:1802.08219, 2018

    2018 arXiv

  34. [43]

    Schütt, O

    K. Schütt, O. Unke, and M. Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. InInternational conference on machine learning, pages 9377–9388. PMLR, 2021

    2021

  35. [44]

    Novick-Cohen

    A. Novick-Cohen. The Cahn–Hilliard equation.Handbook of differential equations: evolutionary equa- tions, 4:201–228, 2008

    2008

  36. [45]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, and A. Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24:1–97, 2023

    2023

  37. [46]

    Evans.Partial differential equations, volume 19

    L.C. Evans.Partial differential equations, volume 19. American mathematical society, 2022

    2022

  38. [47]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

    2010 arXiv

  39. [48]

    D.J. Eyre. Systems of Cahn–Hilliard equations.SIAM Journal on Applied Mathematics, 53:1686–1712, 1993

    1993

  40. [49]

    Multi-componentCahn–Hilliardsystemwithdifferentboundaryconditions in complex domains.Journal of Computational Physics, 323:1–16, 2016

    Y.Li, J.-I.Choi, andJ.Kim. Multi-componentCahn–Hilliardsystemwithdifferentboundaryconditions in complex domains.Journal of Computational Physics, 323:1–16, 2016

    2016

  41. [50]

    ten Eikelder, K.G

    M.F.P. ten Eikelder, K.G. van der Zee, and D. Schillinger. Thermodynamically consistent diffuse- interface mixture models of incompressible multicomponent fluids.Journal of Fluid Mechanics, 990:A8, 2024

    2024

  42. [51]

    H. Garcke. On Cahn—Hilliard systems with elasticity.Proceedings of the Royal Society of Edinburgh Section A: Mathematics, 133:307–331, 2003

    2003

  43. [52]

    Oudich, P

    H. Oudich, P. Carrara, and L. De Lorenzis. Phase-field modeling of elastic microphase separation. arXiv preprint arXiv:2505.01389, 2025

    2025 arXiv

  44. [53]

    Golovin, S.H

    A.A. Golovin, S.H. Davis, and A.A. Nepomnyashchy. A convective Cahn-Hilliard model for the forma- tion of facets and corners in crystal growth.Physica D: Nonlinear Phenomena, 122:202–230, 1998

    1998

  45. [54]

    Balakrishna and W.C

    A.R. Balakrishna and W.C. Carter. Combining phase-field crystal methods with a Cahn-Hilliard model for binary alloys.Physical Review E, 97:043304, 2018. 20 Appendix A. Supplementary Results Appendix A.1. Gradient-Enhanced Results By incorporating the spatial gradient term in E...

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.