pith. sign in

arxiv: 2606.11963 · v1 · pith:KRPVJLATnew · submitted 2026-06-10 · 💻 cs.LG · physics.comp-ph

HAMNO: A Hierarchical Adaptive Multi-scale Neural Operator with Physics-Informed Learning for Dynamical Systems

Mostafa Bamdad , Mohammad Sadegh Eshaghi , Timon Rabczuk This is my paper

Pith reviewed 2026-06-27 10:09 UTC · model grok-4.3

classification 💻 cs.LG physics.comp-ph
keywords neural operatorsphysics-informed learningmulti-scale modelingdynamical systemsAllen-Cahn equationCahn-Hilliard equationSwift-Hohenberg equationhierarchical architecture
0
0 comments X

The pith

HAMNO uses data-dependent gating to adaptively balance local and global features in neural operators for multi-scale dynamical systems.

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

The paper presents HAMNO as a neural operator that processes functions through hierarchical encoder-decoder layers while using a gating mechanism to choose between convolutional local details and spectral global dependencies at every point in space. This design targets the challenge of learning time-evolving partial differential equations that mix fine-scale structures with long-range effects. A physics-informed version adds penalties on both the strong-form PDE residual and a weak-form integral residual to the training objective. Tests on three non-periodic phase-field equations show gains in rollout accuracy, performance with scarce data, and robustness to new initial conditions compared with baseline operators. If the improvements hold, the approach offers a route to more stable learned simulators for nonlinear physical evolution.

Core claim

HAMNO achieves improved predictive accuracy over standard neural-operator baselines across long-horizon rollout, data-limited training, out-of-distribution initial-condition shifts, and random-seed variations on the Allen-Cahn, Cahn-Hilliard, and Swift-Hohenberg equations; the PI-HAMNO extension, which augments the loss with domain-integrated strong-form residuals and weak-form integrals obtained via centroid quadrature, further increases stability, physical consistency, and data efficiency.

What carries the argument

The data-dependent gating mechanism that adaptively balances local convolutional representations and global spectral operators at each spatial location inside a hierarchical encoder-decoder structure.

If this is right

  • Long-horizon predictions of the tested phase-field models become more accurate than with prior neural operators.
  • Training succeeds with smaller datasets while retaining physical fidelity.
  • Models remain accurate when initial conditions lie outside the training distribution.
  • Incorporating strong- and weak-form residuals yields more stable and consistent long-time behavior.

Where Pith is reading between the lines

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

  • The gating idea may transfer to other multi-scale PDE families such as fluid or wave equations without major redesign.
  • Spatially varying resolution learned by the gates could be inspected to reveal which regions require fine versus coarse treatment.
  • Hybrid use with traditional solvers might allow selective replacement of expensive subdomains by the learned operator.

Load-bearing premise

The data-dependent gating mechanism successfully adapts to balance local convolutional and global spectral information at each spatial location to resolve fine-scale features while preserving long-range dependencies.

What would settle it

If HAMNO fails to produce lower rollout error than standard neural-operator baselines on the Allen-Cahn equation under data-limited training and long time horizons, the accuracy claim would not hold.

Figures

Figures reproduced from arXiv: 2606.11963 by Mohammad Sadegh Eshaghi, Mostafa Bamdad, Timon Rabczuk.

Figure 1
Figure 1. Figure 1: Schematic overview of HAMNO and its physics-informed extension PI-HAMNO, [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Rollout relative L 2 errors for the full neural-operator comparison on the three￾dimensional phase-field benchmarks: (a) AC, (b) CH, and (c) SH. The vertically stacked layout compares long-horizon predictive behavior across the governing equations while preserving readability. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Rollout relative L 2 errors of the U-shaped baselines for the three-dimensional phase￾field benchmarks: (a) AC, (b) CH, and (c) SH. The comparison highlights the effect of hier￾archical encoder–decoder structure without adaptive local–global fusion. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Training histories of all compared models, showing (a) total loss and (b) test relative [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Predicted phase-field solutions at t = 100 for the AC, CH, and SH benchmarks, showing reference fields, model predictions, and the HAMNO absolute error. A qualitative comparison is shown in [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Free-energy decay and mass conservation for the CH benchmark under different [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
read the original abstract

Neural operators provide a powerful framework for learning solution mappings of partial differential equations directly in function space. However, many existing architectures still struggle to represent nonlinear time-dependent systems that involve multi-scale structures, long-range interactions, and stable long-time evolution. In this work, we introduce the Hierarchical Adaptive Multi-scale Neural Operator (HAMNO), a neural-operator architecture that combines local convolutional representations, global spectral operators, and hierarchical encoder-decoder processing. The central component of HAMNO is a data-dependent gating mechanism that adaptively balances local and global information at each spatial location, allowing the model to resolve fine-scale features while preserving long-range dependencies. We further develop a physics-informed extension, PI-HAMNO, based on a multi-objective loss strategy that combines data fitting with strong- and weak-form physics constraints. The strong-form term penalizes the domain-integrated squared PDE residual in physical coordinates, while the weak-form term is constructed by multiplying the governing residual by finite-element test functions and evaluating the resulting element integrals using centroid-based tetrahedral quadrature. The framework is evaluated on non-periodic Allen-Cahn (AC), Cahn-Hilliard (CH), and Swift-Hohenberg (SH) equations defined on cubic domains. Across long-horizon rollout, data-limited training, out-of-distribution initial-condition shifts, and random-seed variations, HAMNO improves predictive accuracy over standard neural-operator baselines, while PI-HAMNO further enhances stability, physical consistency, and data efficiency. The implementation is publicly available at https://github.com/MBamdad/HAMNO .

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The manuscript introduces HAMNO, a hierarchical neural operator architecture that integrates local convolutional representations, global spectral operators, and a data-dependent gating mechanism within an encoder-decoder structure to learn solution operators for nonlinear time-dependent PDEs. It further proposes PI-HAMNO, which augments training with a multi-objective loss combining data fidelity, strong-form PDE residuals integrated over the domain, and weak-form residuals discretized via finite-element test functions and centroid-based tetrahedral quadrature. Evaluations on non-periodic Allen-Cahn, Cahn-Hilliard, and Swift-Hohenberg equations on cubic domains report improved predictive accuracy over standard neural-operator baselines in long-horizon rollouts, data-limited regimes, out-of-distribution initial conditions, and across random seeds; PI-HAMNO additionally improves stability and physical consistency.

Significance. If the reported gains in accuracy, stability, and data efficiency hold under the described conditions, the adaptive gating and combined strong/weak-form physics constraints would represent a practical advance for neural operators applied to multi-scale dynamical systems. The public code release strengthens reproducibility.

minor comments (3)
  1. Abstract: the claim of improvement 'across long-horizon rollout, data-limited training, out-of-distribution initial-condition shifts, and random-seed variations' would be strengthened by explicit quantitative metrics (e.g., relative L2 errors or rollout horizons) rather than qualitative statements; these appear in later sections but should be previewed with numbers.
  2. The description of the weak-form term (centroid-based tetrahedral quadrature on cubic domains) is clear in principle but would benefit from a brief statement on how the test functions are chosen and whether the quadrature order is fixed or adaptive.
  3. Ensure that all baseline neural-operator architectures (FNO, DeepONet, etc.) are cited with exact references and that the same training protocol (optimizer, epochs, data splits) is used for fair comparison.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive summary of our work and the recommendation of minor revision. No specific major comments were provided in the report, so we have no points requiring rebuttal or clarification at this stage. We will address any minor suggestions in the revised manuscript.

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces HAMNO as a hierarchical neural operator architecture combining convolutional, spectral, and encoder-decoder components with a data-dependent gating mechanism, plus PI-HAMNO using a multi-objective loss with strong- and weak-form residuals. All central claims concern empirical improvements in rollout accuracy, data efficiency, and stability on standard phase-field PDEs (Allen-Cahn, Cahn-Hilliard, Swift-Hohenberg). No load-bearing mathematical derivation exists that reduces by construction to fitted parameters, self-definitions, or self-citation chains. The architecture is specified independently of the target metrics, and validation relies on external benchmarks and public code rather than internal tautologies.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review limits identification of free parameters; the main assumptions are in the architecture design and loss construction.

axioms (1)
  • domain assumption The centroid-based tetrahedral quadrature accurately evaluates the weak-form integrals for the physics constraints.
    Mentioned in the abstract as the method for weak-form term.

pith-pipeline@v0.9.1-grok · 5825 in / 1146 out tokens · 29413 ms · 2026-06-27T10:09:40.841489+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

33 extracted references · 21 canonical work pages

  1. [1]

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

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karni- adakis. Learning nonlinear operators via deeponet based on the universal ap- proximation theorem of operators.Nature Machine Intelligence, 3:218–229, 2021. https://doi.org/10.1038/s42256-021-00302-5

    work page doi:10.1038/s42256-021-00302-5 2021

  2. [2]

    Lu Lu, Xuhui Meng, Shengze Cai, Zhiping Mao, Somdatta Goswami, Zhongqiang Zhang, and George Em Karniadakis. A comprehensive and fair comparison of two neural oper- ators (with practical extensions) based on fair data.Computer Methods in Applied Me- chanics and Engineering, 393:114778, 2022. https://doi.org/10.1016/j.cma.2022.114778

    work page doi:10.1016/j.cma.2022.114778 2022

  3. [3]

    The parametric complexity of operator learning.SIAM Journal on Mathematics of Data Science, 5(4):1072–1101,

    Nikola Kovachki, Samuel Lanthaler, and Siddhartha Mishra. The parametric complexity of operator learning.SIAM Journal on Mathematics of Data Science, 5(4):1072–1101,

  4. [4]

    https://doi.org/10.1137/23M156382X

    work page doi:10.1137/23m156382x

  5. [5]

    Operator approximation of the wave equation based on deep learning of green’s function.arXiv preprint arXiv:2307.13902, 2023

    Ziad Aldirany, R´ emi Cottereau, Mathieu Laforest, and Serge Prudhomme. Operator approximation of the wave equation based on deep learning of green’s function.arXiv preprint arXiv:2307.13902, 2023. https://arxiv.org/abs/2307.13902

    arXiv 2023

  6. [6]

    Abdelmalik, Thomas J

    Dhairya Patel, Deep Ray, Mohamed R. Abdelmalik, Thomas J. R. Hughes, and Assad A. Oberai. Variationally mimetic operator networks.Computer Methods in Applied Mechan- ics and Engineering, 419:116536, 2024. https://doi.org/10.1016/j.cma.2023.116536

    work page doi:10.1016/j.cma.2023.116536 2024

  7. [7]

    Factorized fourier neural operators.International Conference on Learning Representations, 2023

    Alasdair Tran, Alexander Mathews, Lexing Xie, and Cheng Soon Ong. Factorized fourier neural operators.International Conference on Learning Representations, 2023. https://arxiv.org/abs/2111.13802

    arXiv 2023

  8. [8]

    Multiwavelet-based operator learning for differential equations.Advances in Neural Information Processing Systems, 34:24048– 24062, 2021

    Gaurav Gupta, Xiongye Xiao, and Paul Bogdan. Multiwavelet-based operator learning for differential equations.Advances in Neural Information Processing Systems, 34:24048– 24062, 2021. https://arxiv.org/abs/2109.13459

    arXiv 2021

  9. [9]

    Tapas Tripura and Souvik Chakraborty. Wavelet neural operator for solv- ing parametric partial differential equations in computational mechanics prob- lems.Computer Methods in Applied Mechanics and Engineering, 404:115783, 2023. https://doi.org/10.1016/j.cma.2022.115783

    work page doi:10.1016/j.cma.2022.115783 2023

  10. [10]

    Laplace neural op- erator for solving differential equations.Nature Machine Intelligence, 6:631–640, 2024

    Qianying Cao, Somdatta Goswami, and George Em Karniadakis. Laplace neural op- erator for solving differential equations.Nature Machine Intelligence, 6:631–640, 2024. https://doi.org/10.1038/s42256-024-00844-4

    work page doi:10.1038/s42256-024-00844-4 2024

  11. [11]

    Geometry-aware fourier neural operator for partial differential equations on complex geometries.arXiv preprint arXiv:2207.05209, 2022

    Zongyi Li, Weiyang Huang, Burigede Liu, and Anima Anandkumar. Geometry-aware fourier neural operator for partial differential equations on complex geometries.arXiv preprint arXiv:2207.05209, 2022. https://arxiv.org/abs/2207.05209

    arXiv 2022

  12. [12]

    Choose a transformer: Fourier or galerkin.Advances in Neural Information Processing Systems, 34:24924–24940, 2021

    Shuhao Cao. Choose a transformer: Fourier or galerkin.Advances in Neural Information Processing Systems, 34:24924–24940, 2021. https://arxiv.org/abs/2105.14995

    arXiv 2021

  13. [13]

    Transformer for partial differen- tial equations’ operator learning.Transactions on Machine Learning Research, 2023

    Zijie Li, Kazem Meidani, and Amir Barati Farimani. Transformer for partial differen- tial equations’ operator learning.Transactions on Machine Learning Research, 2023. https://arxiv.org/abs/2205.13671

    arXiv 2023

  14. [14]

    Shields, and George Em Kar- niadakis

    Katiana Kontolati, Somdatta Goswami, Michael D. Shields, and George Em Kar- niadakis. Learning nonlinear operators in latent spaces for real-time predictions 32 of complex dynamics in physical systems.Nature Communications, 15:5101, 2024. https://doi.org/10.1038/s41467-024-49411-w

    work page doi:10.1038/s41467-024-49411-w 2024

  15. [15]

    Loglo-fno: Efficient learning of local and global features in fourier neural operators.Transactions on Machine Learning Research, 2025

    Marimuthu Kalimuthu, David Holzm¨ uller, and Mathias Niepert. Loglo-fno: Efficient learning of local and global features in fourier neural operators.Transactions on Machine Learning Research, 2025. https://doi.org/10.48550/arXiv.2504.04260

    work page doi:10.48550/arxiv.2504.04260 2025

  16. [16]

    Gege Wen, Zongyi Li, Kamyar Azizzadenesheli, Anima Anandkumar, and Sally M. Benson. U-fno: An enhanced fourier neural operator-based deep-learning model for multiphase flow.Advances in Water Resources, 163:104180, 2022. https://doi.org/10.1016/j.advwatres.2022.104180

    work page doi:10.1016/j.advwatres.2022.104180 2022

  17. [17]

    Md Ashiqur Rahman and Zachary E. Ross. U-no: U-shaped neural operators.Transac- tions on Machine Learning Research, 2022. https://arxiv.org/abs/2204.11127

    arXiv 2022

  18. [18]

    U-net: Convolutional networks for biomedical image segmentation.Medical Image Computing and Computer-Assisted Intervention, pages 234–241, 2015

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation.Medical Image Computing and Computer-Assisted Intervention, pages 234–241, 2015. https://arxiv.org/abs/1505.04597

    Pith/arXiv arXiv 2015

  19. [19]

    Multi-head neural operator for modelling in- terfacial dynamics.International Journal of Mechanical Sciences, 285:111363, 2026

    Mohammad Sadegh Eshaghi, Navid Valizadeh, Cosmin Anitescu, Yizheng Wang, Xi- aoying Zhuang, and Timon Rabczuk. Multi-head neural operator for modelling in- terfacial dynamics.International Journal of Mechanical Sciences, 285:111363, 2026. https://doi.org/10.1016/j.ijmecsci.2026.111363

    work page doi:10.1016/j.ijmecsci.2026.111363 2026

  20. [20]

    , author Katori, Y

    Katarzyna Micha lowska, Somdatta Goswami, George Em Karniadakis, and Signe Riemer- Sørensen. Neural operator learning for long-time integration in dynamical systems with recurrent neural networks.International Joint Conference on Neural Networks, pages 1–8, 2024. https://doi.org/10.1109/IJCNN60899.2024.10650331

    work page doi:10.1109/ijcnn60899.2024.10650331 2024

  21. [21]

    Nows: Neural operator warm starts for accelerating iter- ative solvers.arXiv preprint arXiv:2511.02481, 2025

    Mohammad Sadegh Eshaghi, Cosmin Anitescu, Navid Valizadeh, Yizheng Wang, Xiaoy- ing Zhuang, and Timon Rabczuk. Nows: Neural operator warm starts for accelerating iter- ative solvers.arXiv preprint arXiv:2511.02481, 2025. https://arxiv.org/abs/2511.02481

    Pith/arXiv arXiv 2025

  22. [22]

    Variational physics-informed neural operator (VINO) for solving partial differential equations.Computer Methods in Applied Mechanics and Engineering, 437:117785, 2025

    Mohammad Sadegh Eshaghi, Cosmin Anitescu, Manish Thombre, Yizheng Wang, Xiaoy- ing Zhuang, and Timon Rabczuk. Variational physics-informed neural operator (VINO) for solving partial differential equations.Computer Methods in Applied Mechanics and Engineering, 437:117785, 2025. https://doi.org/10.1016/j.cma.2025.117785

    work page doi:10.1016/j.cma.2025.117785 2025

  23. [23]

    Navaneeth and Souvik Chakraborty

    N. Navaneeth and Souvik Chakraborty. Physics-informed wavelet neural op- erator for learning solution operators of parametric partial differential equa- tions.Computer Methods in Applied Mechanics and Engineering, 418:116546, 2024. https://doi.org/10.1016/j.cma.2023.116546

    work page doi:10.1016/j.cma.2023.116546 2024

  24. [24]

    Physics-informed latent neural operator for real-time simulations of parametric pdes.arXiv preprint arXiv:2501.08428, 2025

    Sharmila Karumuri, Lori Graham-Brady, and Somdatta Goswami. Physics-informed latent neural operator for real-time simulations of parametric pdes.arXiv preprint arXiv:2501.08428, 2025. https://arxiv.org/abs/2501.08428

    arXiv 2025

  25. [25]

    Penco: A physics–energy–numerics–consistent operator for 3d phase field modeling.Computer Methods in Applied Mechanics and Engineering, 454:118862,

    Mostafa Bamdad, Mohammad Sadegh Eshaghi, Cosmin Anitescu, Navid Valizadeh, and Timon Rabczuk. Penco: A physics–energy–numerics–consistent operator for 3d phase field modeling.Computer Methods in Applied Mechanics and Engineering, 454:118862,

  26. [26]

    https://doi.org/10.1016/j.cma.2026.118862. 33

    work page doi:10.1016/j.cma.2026.118862 2026

  27. [27]

    The phase field method for geometric moving inter- faces and their numerical approximations

    Qiang Du and Xiaobing Feng. The phase field method for geometric moving inter- faces and their numerical approximations. InGeometric Partial Differential Equations – Part I, volume 21 ofHandbook of Numerical Analysis, pages 425–508. Elsevier, 2020. https://doi.org/10.1016/bs.hna.2019.05.001

    work page doi:10.1016/bs.hna.2019.05.001 2020

  28. [28]

    Phase-field modeling of microstructure evolution: Recent applications, perspectives and challenges.Progress in Materials Sci- ence, 123:100810, 2022

    Damien Tourret, Hong Liu, and Javier Llorca. Phase-field modeling of microstructure evolution: Recent applications, perspectives and challenges.Progress in Materials Sci- ence, 123:100810, 2022. https://doi.org/10.1016/j.pmatsci.2021.100810

    work page doi:10.1016/j.pmatsci.2021.100810 2022

  29. [29]

    Xiaoying Zhuang, Shuwei Zhou, G. D. Huynh, P. Areias, and Timon Rabczuk. Phase field modeling and computer implementation: A review.Engineering Fracture Mechanics, 262:108234, 2022. https://doi.org/10.1016/j.engfracmech.2022.108234

    work page doi:10.1016/j.engfracmech.2022.108234 2022

  30. [30]

    npj Comput

    David Montes de Oca Zapiain, James A. Stewart, and R´ emi Dingreville. Ac- celerating phase-field-based microstructure evolution predictions via surrogate mod- els trained by machine learning methods.npj Computational Materials, 7:3, 2021. https://doi.org/10.1038/s41524-020-00471-8

    work page doi:10.1038/s41524-020-00471-8 2021

  31. [31]

    Andreas Krischok, Basavesh Yaraguntappa, and Marc-Andr´ e Keip. Fast implicit up- date schemes for cahn–hilliard-type gradient flow in the context of fourier-spectral meth- ods.Computer Methods in Applied Mechanics and Engineering, 431:117220, 2024. https://doi.org/10.1016/j.cma.2024.117220

    work page doi:10.1016/j.cma.2024.117220 2024

  32. [32]

    Applications of scientific machine learning for the analysis of functionally graded porous beams.Neurocomputing, 619:129119, 2025

    Mohammad Sadegh Eshaghi, Mostafa Bamdad, Cosmin Anitescu, Yizheng Wang, Xi- aoying Zhuang, and Timon Rabczuk. Applications of scientific machine learning for the analysis of functionally graded porous beams.Neurocomputing, 619:129119, 2025. https://doi.org/10.1016/j.neucom.2024.129119

    work page doi:10.1016/j.neucom.2024.129119 2025

  33. [33]

    Fourier-spectral method for the phase-field equations.Mathemat- ics, 8(8):1385, 2020

    Sungha Yoon, Darae Jeong, Chaeyoung Lee, Hyundong Kim, Sangkwon Kim, Hyun Geun Lee, and Junseok Kim. Fourier-spectral method for the phase-field equations.Mathemat- ics, 8(8):1385, 2020. https://doi.org/10.3390/math8081385. 34

    work page doi:10.3390/math8081385 2020