pith. sign in

arxiv: 2503.04901 · v1 · submitted 2025-03-06 · ⚛️ physics.comp-ph · cs.NE

Multiscale Analysis of Woven Composites Using Hierarchical Physically Recurrent Neural Networks

Ehsan Ghane , Marina A. Maia , Iuri B.C.M. Rocha , Martin Fagerstr\"om , Mohsen Mirakhalaf This is my paper

Pith reviewed 2026-05-23 00:54 UTC · model grok-4.3

classification ⚛️ physics.comp-ph cs.NE
keywords multiscale homogenizationwoven compositesphysically recurrent neural networkssurrogate modelingelasto-plastic behaviorcyclic loadingphysics-informed neural networks
0
0 comments X

The pith

Hierarchical physically recurrent neural networks avoid nonphysical predictions in multiscale woven composite modeling by embedding physics at both yarn and meso-to-macro scales.

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

The paper proposes a two-level surrogate approach for homogenizing woven composites. At the finer scale, physically recurrent neural networks are trained on micromechanical data to represent the elasto-plastic response of individual yarns. At the coarser scale, a physics-encoded model combines the trained yarn surrogates with the matrix constitutive law by placing physical quantities directly into the latent representation. This structure is intended to eliminate the nonphysical outputs and weak extrapolation that appear in purely data-driven recurrent or transformer models, particularly when the material experiences repeated cyclic loading. The result is presented as a faster, more interpretable replacement for conventional multiscale homogenization.

Core claim

Adopting HPRNNs for both scale transitions can avoid nonphysical behavior often observed in predictions from pure data-driven recurrent neural networks and transformer networks. This results in better generalization under complex cyclic loading conditions.

What carries the argument

The Hierarchical Physically Recurrent Neural Network (HPRNN), in which PRNN surrogates capture yarn nonlinearity at the microscale and a second physics-encoded network integrates those surrogates with the matrix model at the meso-to-macro scale by embedding physical properties into the latent space.

If this is right

  • Multiscale simulations of woven composites become computationally cheaper while retaining physical consistency at both scales.
  • Predictions remain stable and physically admissible under repeated cyclic loading paths that expose weaknesses in purely data-driven surrogates.
  • The framework supplies an explainable surrogate that can be inspected at the level of embedded physical quantities rather than opaque weights.
  • The same hierarchical structure can be reused for different woven architectures once the yarn-level PRNNs are retrained.

Where Pith is reading between the lines

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

  • The same two-level construction might be tested on other textile or braided composites whose yarns exhibit similar elasto-plastic nonlinearity.
  • If the latent-space embedding proves robust, the method could reduce the volume of micromechanical data needed for new matrix-yarn combinations.
  • Engineering workflows that require many load cycles, such as fatigue assessment, would become feasible at the structural scale without full-field micromechanics at every step.

Load-bearing premise

Embedding physical properties directly into the latent space of the meso-to-macro model will correctly integrate the trained yarn surrogates with the matrix constitutive model without introducing new inconsistencies or requiring extensive additional calibration.

What would settle it

A direct comparison showing that the HPRNN still produces nonphysical stress or strain values or fails to generalize under the same complex cyclic loading paths where pure data-driven networks already fail would falsify the central claim.

Figures

Figures reproduced from arXiv: 2503.04901 by Ehsan Ghane, Iuri B.C.M. Rocha, Marina A. Maia, Martin Fagerstr\"om, Mohsen Mirakhalaf.

Figure 1
Figure 1. Figure 1: Hierarchical structure of woven composites and the transition across two scales with PRNN and the proposed [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of woven composite structures and associated simulations. Microscale images captured by E. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Samples of scaled (normalized between 1 and -1) shear strain components [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Details of the computational model at micro- and mesoscale: (a) Top view of the mesoscale model, (b) X-Z [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Architecture of PRNN used as the transition between micro- to mesoscales. It consists of a feed-forward [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Multi-scale structure of a woven composite. (a) Photograph of a carbon fiber woven composite taken by E. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: HPRNN architecture, featuring a standard feed-forward encoder for strain decomposition, a material layer [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prediction of stress components by the weft-PRNN with five bulk points for three randomly selected samples [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Comparison of micro-to-mesoscale transition for weft yarns under two standard loading scenarios not included [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of von Mises stress mean absolute errors for the mesoscale network (HPRNN) configured with [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of validation mean absolute errors for the mesoscale network configured with different numbers [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Prediction of stress components by the mesoscale HPRNN, GRU, and Transformer models on three randomly [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: (a) GRU and (b) transformer training curves over 200 and 437 random loading simulations, respectively. [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Mean error and error standard deviation comparison for 20 random loading samples from the test set, [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: The relative error distribution of GRU (black) and HPRNN (red) models is normalized with respect to the [PITH_FULL_IMAGE:figures/full_fig_p021_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: In-plain stress vs. strain behavior in testing against (a) unseen random loading and extrapolation over unseen [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Mean error and error standard deviation comparison for five in-plane shear cyclic loading samples from [PITH_FULL_IMAGE:figures/full_fig_p022_17.png] view at source ↗
read the original abstract

Multiscale homogenization of woven composites requires detailed micromechanical evaluations, leading to high computational costs. Data-driven surrogate models based on neural networks address this challenge but often suffer from big data requirements, limited interpretability, and poor extrapolation capabilities. This study introduces a Hierarchical Physically Recurrent Neural Network (HPRNN) employing two levels of surrogate modeling. First, Physically Recurrent Neural Networks (PRNNs) are trained to capture the nonlinear elasto-plastic behavior of warp and weft yarns using micromechanical data. In a second scale transition, a physics-encoded meso-to-macroscale model integrates these yarn surrogates with the matrix constitutive model, embedding physical properties directly into the latent space. Adopting HPRNNs for both scale transitions can avoid nonphysical behavior often observed in predictions from pure data-driven recurrent neural networks and transformer networks. This results in better generalization under complex cyclic loading conditions. The framework offers a computationally efficient and explainable solution for multiscale modeling of woven composites.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The manuscript introduces a Hierarchical Physically Recurrent Neural Network (HPRNN) framework for multiscale homogenization of woven composites. It employs Physically Recurrent Neural Networks (PRNNs) trained on micromechanical data to surrogate the nonlinear elasto-plastic response of warp and weft yarns, followed by a second-scale physics-encoded meso-to-macro model that integrates the yarn surrogates with the matrix constitutive law by embedding physical properties directly into the latent space. The central claim is that this hierarchical physics-informed approach avoids nonphysical behavior typical of pure data-driven RNNs and transformers, yielding improved generalization under complex cyclic loading while remaining computationally efficient and explainable.

Significance. If the quantitative validation and consistency proofs hold, the work would supply a scalable surrogate that reduces the cost of detailed micromechanical evaluations in composite analysis while preserving physical consistency across scales. The explicit embedding of physical properties into the latent space, if shown to enforce thermodynamic admissibility or interface equilibrium without extra calibration, would be a concrete strength over purely data-driven alternatives.

major comments (2)
  1. [Abstract] Abstract (paragraph on second scale transition): the claim that embedding physical properties directly into the latent space 'correctly integrates the trained yarn surrogates with the matrix constitutive model without introducing new inconsistencies' is load-bearing for the central claim, yet the description supplies neither the explicit loss terms, invariance constraints, nor thermodynamic admissibility conditions used in the embedding; without these it is impossible to determine whether the coupling reduces to quantities already fitted in the two-stage training.
  2. [Abstract] Abstract: the assertion of 'better generalization under complex cyclic loading conditions' and avoidance of nonphysical behavior is presented without any reported error metrics, error bars, baseline comparisons (e.g., against standard RNNs or FE^{2}), or specific validation cases on cyclic histories, which prevents assessment of whether the data support the central claim.
minor comments (1)
  1. The abstract refers to an 'explainable solution' but does not indicate which architectural features (e.g., latent-space interpretability or physics constraints) are intended to deliver this property.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We address each point below and will revise the manuscript accordingly to improve clarity and support for the central claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract (paragraph on second scale transition): the claim that embedding physical properties directly into the latent space 'correctly integrates the trained yarn surrogates with the matrix constitutive model without introducing new inconsistencies' is load-bearing for the central claim, yet the description supplies neither the explicit loss terms, invariance constraints, nor thermodynamic admissibility conditions used in the embedding; without these it is impossible to determine whether the coupling reduces to quantities already fitted in the two-stage training.

    Authors: We agree that the abstract is too concise on this point. The explicit loss terms, invariance constraints, and thermodynamic admissibility conditions are detailed in Section 3.2 of the manuscript (physics-encoded meso-to-macro model), where physical properties are embedded via additional penalty terms in the training objective that enforce interface equilibrium and constitutive consistency beyond the two-stage fitting. To address the concern, we will revise the abstract to briefly reference these elements and the relevant section. revision: yes

  2. Referee: [Abstract] Abstract: the assertion of 'better generalization under complex cyclic loading conditions' and avoidance of nonphysical behavior is presented without any reported error metrics, error bars, baseline comparisons (e.g., against standard RNNs or FE^{2}), or specific validation cases on cyclic histories, which prevents assessment of whether the data support the central claim.

    Authors: The quantitative results supporting these claims, including error metrics with error bars, baseline comparisons against standard RNNs and FE^{2}, and specific cyclic loading validation cases, are reported in Sections 5 and 6 with accompanying figures. The abstract provides a high-level summary of these findings. We will revise the abstract to include key quantitative indicators (e.g., relative errors under cyclic histories) to better substantiate the claim. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The abstract describes a two-stage process: PRNNs are trained directly on micromechanical data to surrogate yarn behavior, followed by a separate physics-encoded meso-to-macro model that integrates the surrogates with the matrix law via latent-space embedding. No equations, loss terms, or self-citations are supplied that would reduce the claimed generalization or avoidance of nonphysical behavior to a fitted quantity by construction. The physics embedding is presented as an independent modeling choice rather than a tautology or renamed input. This leaves the central claim self-contained against external data and benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on the availability of micromechanical training data for yarns and on the premise that physics can be directly embedded into the latent space without further derivation; no explicit free parameters or invented entities are named in the abstract.

axioms (2)
  • domain assumption Micromechanical simulations provide sufficient and representative data for training PRNNs on warp and weft yarn elasto-plastic behavior.
    Invoked in the description of the first scale transition.
  • ad hoc to paper Embedding physical properties into the latent space of the meso-to-macro model preserves physical consistency across scale transitions.
    Central premise of the second scale transition; not derived in the abstract.
invented entities (1)
  • Hierarchical Physically Recurrent Neural Network (HPRNN) no independent evidence
    purpose: Two-level surrogate that combines yarn PRNNs with physics-encoded meso-to-macro integration.
    New architectural construct introduced to address nonphysical predictions of standard RNNs.

pith-pipeline@v0.9.0 · 5722 in / 1501 out tokens · 36828 ms · 2026-05-23T00:54:07.266966+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

12 extracted references · 12 canonical work pages · 2 internal anchors

  1. [1]

    A review on data-driven constitutive laws for solids

    Jan Niklas Fuhg, Govinda Anantha Padmanabha, Nikolaos Bouklas, Bahador Bahmani, WaiChing Sun, Nikolaos N Vlassis, Moritz Flaschel, Pietro Carrara, and Laura De Lorenzis. A review on data-driven constitutive laws for solids. arXiv preprint arXiv:2405.03658,

    work page arXiv

  2. [2]

    Improving bayesian networks multifidelity surrogate construction with basis adaptation

    Xiaoshu Zeng, Gianluca Geraci, Alex Gorodetsky, John Jakeman, Michael S Eldred, and Roger G Ghanem. Improving bayesian networks multifidelity surrogate construction with basis adaptation. In AIAA Scitech 2023 F orum, page 0917,

    work page 2023

  3. [3]

    Physics-informed neural networks and extensions

    Maziar Raissi, Paris Perdikaris, Nazanin Ahmadi, and George Em Karniadakis. Physics-informed neural networks and extensions. arXiv preprint arXiv:2408.16806,

    work page arXiv

  4. [4]

    Learning solution of nonlinear constitutive material models using physics-informed neural networks: Comm-pinn

    Shahed Rezaei, Ahmad Moeineddin, and Ali Harandi. Learning solution of nonlinear constitutive material models using physics-informed neural networks: Comm-pinn. arXiv preprint arXiv:2304.06044,

    work page arXiv

  5. [5]

    Hard encoding of physics for learning spatiotemporal dynamics

    Chengping Rao, Hao Sun, and Yang Liu. Hard encoding of physics for learning spatiotemporal dynamics. arXiv preprint arXiv:2105.00557,

    work page arXiv

  6. [6]

    Fourier Neural Operator for Parametric Partial Differential Equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895,

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  7. [7]

    Epi-ckans: Elasto-plasticity informed kolmogorov-arnold networks using chebyshev polynomials

    Farinaz Mostajeran and Salah A Faroughi. Epi-ckans: Elasto-plasticity informed kolmogorov-arnold networks using chebyshev polynomials. arXiv preprint arXiv:2410.10897,

    work page arXiv

  8. [8]

    KAN: Kolmogorov-Arnold Networks

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljaˇci´c, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024b. Salah A Faroughi, Nikhil M Pawar, Celio Fernandes, Maziar Raissi, Subasish Das, Nima K Kalantari, and Seyed Kourosh Mahjour. Physics-guided, physics-informed, and ...

    work page internal anchor Pith review Pith/arXiv arXiv

  9. [9]

    doi:https://doi.org/10.1016/j.cma.2016.10.022

    ISSN 0045-7825. doi:https://doi.org/10.1016/j.cma.2016.10.022. MM Shokrieh, R Ghasemi, and R Mosalmani. A general micromechanical model to predict elastic and strength properties of balanced plain weave fabric composites. Journal of Composite Materials, 51(20):2863–2878,

    work page doi:10.1016/j.cma.2016.10.022 2016

  10. [10]

    Fine-tuning large language models for domain adaptation: Exploration of training strategies, scaling, model merging and synergistic capabilities

    Wei Lu, Rachel K Luu, and Markus J Buehler. Fine-tuning large language models for domain adaptation: Exploration of training strategies, scaling, model merging and synergistic capabilities. arXiv preprint arXiv:2409.03444,

    work page arXiv

  11. [11]

    https://www.e-xstream.com/products/digimat/tools?fe=1

    Digimat-FE. https://www.e-xstream.com/products/digimat/tools?fe=1. accessed: 24.10.2016. Jeff Bezanson, Alan Edelman, Stefan Karpinski, and Viral B Shah. Julia: A fresh approach to numerical computing. SIAM review, 59(1):65–98,

    work page 2016

  12. [12]

    Bezanson , author A

    URL https://doi.org/10.1137/141000671. Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch

    work page doi:10.1137/141000671