Pith. sign in

REVIEW 3 major objections 7 minor 48 references

A manifold-aware Neural ODE surrogate model for stochastic induction heating with anisotropic electrical conductivity

T0 review · 3 major / 7 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that SPD-aware CMNN-NODE surrogates, trained on a subset of Monte Carlo FEM solutions, reproduce full temperature fields of stochastic induction welding on unseen cases and long-horizon rollouts with mean L2 errors around 0

desk verdict The surrogate accuracy is real but narrower than the abstract's 'full-field' framing: it holds on a single-integration-point subsample, and the paper's own stability analysis contradicts its 'temporally stable' wording. read the letter →

arxiv 2608.01945 v1 pith:YNLMATOA submitted 2026-08-03 cs.CE

Wouter J. Schuttert , Mohammed Iqbal Abdul Rasheed , Bojana Rosić This is my paper
classification cs.CE
keywords inductionweldingstochasticconductivitySPDmanifoldsurrogatemodellingNeuralODEConstitutiveNetworkuncertaintyquantificationcarbon-fibrethermoplastics
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

This paper argues that a surrogate model can replace expensive paired electromagnetic–thermal finite-element simulations for predicting full temperature fields in induction welding of carbon-fibre thermoplastics, including variability introduced by manufacturing scatter in electrical conductivity. The conductivity tensor is treated as a random symmetric positive-definite (SPD) quantity, with magnitude and orientation uncertainty modelled independently, and the surrogate embeds this geometry via a Constitutive Manifold Neural Network. Temporal evolution is learned either as a residual Euler step or as a Neural ODE integrated with explicit RK4 or implicit Adams–Moulton schemes. On 5,000 training realisations, all formulations reproduce unseen Monte Carlo trajectories with mean errors around 0.3–0.4 °C, while the Neural ODE variants show several-fold lower run-to-run variability. The wider point is that geometry-aware, time-continuous surrogates can carry uncertainty quantification for multiphysics processes at a fraction of the cost.

What carries the argument

The StrAng layer decomposes the SPD conductivity tensor into eigenvalues (strength) and a planar rotation (orientation), applies logarithmic maps, and flattens them into a Euclidean vector, hard-constraining physical admissibility. The evolution map is the residual update u_{k+1} = u_k + F_θ(u_k, σ). For Neural ODEs, F_θ is replaced by an integrated vector field f_ξ; Euler, RK4, and Adams–Moulton define different discretisations of the same learned field, with the implicit scheme requiring fixed-point iteration.

What would settle it

Run the same training pipeline on data where the FEM solver stores 2×2 integration-point temperatures, or evaluate the trained surrogates on conductivity realisations sampled with larger angular spread or eigenvalue variance than the training range. If mean L2 errors grow by more than the 0.3–0.4 °C baseline, or if the measured per-step amplification leads to visible divergence by 20 s, the central reproducibility claim would fail; comparing KL divergence at t = 10 s on out-of-distribution inputs would settle generalisation.

Watch

Extended reading notes

Core claim

Using the random SPD decomposition σ = exp(W) V exp(Y) Vᵀ exp(Wᵀ), the paper generates 20,000 Monte Carlo solutions of the coupled eddy-current and heat equations, then trains CMNN surrogates to map (current temperature field, manifold-embedded conductivity) to the next temperature increment. On held-out realisations and 10-second rollouts, all three integrator choices reproduce the reference temperature statistics with mean L2 errors of 0.33–0.44 °C. The strongest observed difference is not mean accuracy but consistency: repeated training runs of the Euler network scatter roughly four times more in test MSE than the Neural ODE models, so the continuous-time formulation is the more reproduci

Load-bearing premise

The surrogate is trained on temperature snapshots at one integration point per element while the FEM reference uses 2×2 integration points, and the electromagnetic heat source is never an explicit input; if either the reduced spatial information or the implicit source encoding is insufficient for states outside the training distribution, the reported long-horizon accuracy will not transfer.

Editorial extensions

If this is right

  • Scaling uncertainty in conductivity dominates orientation uncertainty in driving temperature spread; orientation adds a small, spatially structured correction.
  • Temperature uncertainty grows over the 10 s heating interval because conductivity randomness accumulates through the transient heat equation.
  • A surrogate trained on 5,000 Monte Carlo realisations reproduces the statistics of the 20,000-sample reference, with inference around 10⁻³ s per sample for Euler and 10⁻² s for Neural ODEs, enabling tractable full-field uncertainty quantification.
  • Neural ODE integrators improve run-to-run reproducibility across training restarts, while Euler gives marginally lower mean error; no single scheme wins on all metrics.
  • Long-horizon rollout errors grow over time, especially during the initial 0.5 s gradient-formation stage, and learned updates are locally expansive, so stability must be assessed empirically.

Reading between the lines

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

  • Because the surrogate learns an implicit driving source with no explicit electromagnetic solve, the same architecture could transfer to other weakly coupled multiphysics problems where the hidden source is encoded in state trajectories.
  • If the single-integration-point limitation is the main accuracy bottleneck, retraining on 2×2-integration-point outputs might close the gap to full-order fields without changing the architecture.
  • The observed local amplification suggests a spectral regulariser that penalises Jacobian spectral radii above unity could make even Euler rollouts as reproducible as Neural ODEs, a testable modification.
  • The homogeneous random conductivity likely overstates spatial variability; extending to random tensor fields should reduce predicted uncertainty and is the natural next step.
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

3 major / 7 minor

Summary. The manuscript proposes a surrogate modeling pipeline for stochastic induction heating of carbon-fibre-reinforced thermoplastic laminates. It introduces a spectral decomposition of the electrical conductivity tensor into lognormal eigenvalues and von Mises orientation uncertainty, generates 20,000 Monte Carlo realisations of a coupled electromagnetic-thermal FEM model, and trains CMNN-based Neural ODE surrogates (Euler, RK4, Adams-Moulton) to predict temperature-field increments. Reported results show test-set MSE on the order of 1e-3 and mean L2 errors around 0.3-0.4 degree C, with the Neural ODE variants reducing run-to-run variance relative to the Euler baseline. A stability analysis reveals local expansion of the learned flow map. The paper concludes that the framework is 'temporally stable' and suitable as a cheap replacement for full Monte Carlo FEM.

Significance. If the accuracy and reproducibility findings survive scrutiny, the surrogate could substantially reduce the cost of uncertainty quantification for induction welding. The paper deserves credit for using 20,000 Monte Carlo realisations, held-out test data, five repeated training runs, and explicit stability probes. The stochastic conductivity model is informed by prior literature rather than fitted to the target temperature field, which reduces circularity. However, the central 'full-field replacement' claim is weakened by training on subsampled quadrature points, and the 'temporally stable' characterization is contradicted by the paper's own Figure 11. The manifold-awareness contribution is not isolated by an ablation. These issues prevent acceptance in the current form.

major comments (3)
  1. [Section 5.3 (also Section 5.2)] The surrogate is trained and evaluated on temperature values at a single integration point per element (dimension 2688), while the FEM solver enforces accuracy at a 2x2 integration-point configuration. The paper explicitly acknowledges in Section 5.3 that this removes fine-scale information and may obscure local temperature effects. No comparison is provided between the subsampled representation and the full 2x2 FEM field. Consequently, the reported MSE/L2/KL values are relative to a reduced reference, not to the full-order solution, and the abstract/conclusion claim of replacing full-field Monte Carlo FEM is not supported. A quantitative comparison of the reduced representation against the full 2x2 field, or retraining on full quadrature data, is needed.
  2. [Section 6 vs Section 5.2/Figure 11] The conclusion calls the framework 'temporally stable', but the paper's own stability analysis shows the learned flow map is locally expansive for all integrators: the per-step amplification rate r peaks around 0.12 s^-1 with a tail to 0.8 s^-1, the one-step amplification factor exceeds unity, and the initial-condition deviation grows monotonically over the rollout. This directly contradicts the 'temporally stable' characterization. Either remove the claim or add a spectral/stability constraint to the training objective and re-evaluate.
  3. [Section 4, Tables 3-4] All surrogate variants use the StrAng manifold layer, so the paper cannot substantiate the claim that manifold-awareness improves learning or generalisation. A Euclidean baseline (e.g., vectorised sigma input) under otherwise identical architecture and training budget is required to isolate the effect of the CMNN/StrAng component. Without this ablation, the central methodological novelty is not demonstrated.
minor comments (7)
  1. [Section 2, first sentence] Typo: 'eletrically' should be 'electrically'.
  2. [Figure 5(a)] The colorbar contains the placeholder 'dunno'; replace with a proper unit label.
  3. [Table 3] The labels 't[5-15]' and 't[0-20]' are unexplained; Section 5.2 states validation horizon t in [2.5,7.5] s and test horizon t in [0,10] s. Use consistent notation for step indices versus seconds.
  4. [Figure 12] Panel (i) is captioned 'Kullbach leibner Divergence' while panels (g)-(h) are described as errors in Std; the caption appears mismatched. Also 'Kullbach leibner' should be 'Kullback-Leibler'.
  5. [Eq. (21)] There is an extra closing parenthesis in 'u(x, t, omega))'.
  6. [References [30,48]] Both contain 'Accesed' instead of 'Accessed'.
  7. [Section 6] 'in principal' should be 'in principle'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: surrogate accuracy is tested on held-out FEM realizations, and self-citations are used as stated building blocks rather than as assumed conclusions.

full rationale

The derivation chain is not circular. The stochastic conductivity model (Eqs. 22-26) is defined independently of the temperature output, with its distribution parameters taken from external literature [27, 42]. The surrogate is trained to approximate the FEM one-step increment (Eqs. 37-48) on a subset of 5000 of 20,000 Monte Carlo realizations, and is then evaluated on held-out test realizations and long-horizon rollouts; the reported MSE/L2/KL metrics in Tables 3-4 compare against FEM reference solutions on these held-out cases, so the empirical claim is a genuine generalization test rather than a fit-to-prediction. The CMNN StrAng layer and NODE integrators are cited from prior work, including the authors' own [10, 19], but the paper re-states the active equations (75-78, 55-58) and does not invoke those citations to prohibit alternatives or to define the target quantity in terms of the prediction. The Section 5.3 admission that training uses a single integration point per element while the FEM solver uses a 2x2 configuration is an honest limitation and a correctness risk, not a circular step. The self-citations present are not load-bearing: the cited stochastic tensor model and CMNN architecture are components whose equations are reproduced, and the paper's central accuracy claim rests on held-out evaluation, not on the authority of the self-citations.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The central claim rests on the stochastic material model, the one-way coupled FEM model, and the learned one-step map. Surrogate accuracy is measured against the same FEM model, so the physical realism of the uncertainty magnitudes is not independently verified. No new physical entities are introduced; the StrAng layer is a mathematical embedding.

free parameters (3)
  • Eigenvalue lognormal means and standard deviations (lambda1=38300±1460, lambda2=6.04±3.27, lambda3=0.507±0.285 S/m) = From [27]
    Adopted from the PhD thesis [27] to set stochastic conductivity magnitude; all quantitative uncertainty results depend directly on these values.
  • von Mises concentration parameter kappa = 2000
    Chosen in Section 5.1 so the angular spread (standard deviation about 1.3 degrees) matches the global absolute fibre misalignment reported in [42].
  • Network hyperparameters (4 hidden layers, width 12, SiLU, batch size 128, learning rate 7.5e-5) = As listed in Section 5.2
    Selected by Bayesian hyperparameter search on validation loss; surrogate accuracy depends on them, though they are not physical parameters.
assumptions (7)
  • domain assumption Magneto-quasi-static approximation: displacement currents are negligible and E = -j 2 pi f A in the conducting domain.
    Invoked in Section 2, Eqs. (1)-(10), to derive the reduced electromagnetic equation; justified by wavelength much larger than the domain, but still an approximation.
  • domain assumption One-way coupling: the electromagnetic heat source Q depends on sigma and A but not on temperature, and thermal material properties are temperature-independent.
    Used throughout, e.g., Eqs. (30)-(33) where M and K are time-independent and P(sigma) is the only stochastic load. Stated in Section 2.
  • domain assumption Spatial homogeneity of the random conductivity: sigma(omega) is a single SPD tensor per ply, not a spatial random field.
    Assumed in Section 2 and conceded in Section 5.3 to likely overestimate effective variability relative to a spatial random field.
  • domain assumption Gaussian log-eigenvalues and von Mises orientation distributions fully describe manufacturing variability.
    Introduced in Eqs. (23)-(27); parameters are taken from literature, and no experiment in this paper validates these distributional forms.
  • domain assumption The neural one-step map u_{k+1} = u_k + F_theta(u_k, sigma) with Q implicit is a valid discrete evolution operator.
    Core modeling choice in Eqs. (38)-(41); the paper notes this can fail for stiff or unstable learned dynamics and relies on empirical validation.
  • domain assumption Eigendecomposition uniqueness is fixed by the alignment rule in Appendix A, valid for rotations up to about pi/4.
    Required for the StrAng layer to be well-defined; the adopted orientation spread is much smaller, so the rule is adequate for this study.
  • standard math The COMSOL FEM discretization and BDF2 time integration produce ground-truth solutions.
    Section 2.1 relies on COMSOL with a GMRES/PARDISO solver and BDF2 integration; no manufactured-solution verification is shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A manifold-aware Neural ODE surrogate model for stochastic induction heating with anisotropic electrical conductivity." pith.science (2026). https://pith.science/paper/YNLMATOA

@misc{pith2026260801945,
  author       = {Pith},
  title        = {Pith review of: A manifold-aware Neural ODE surrogate model for stochastic induction heating with anisotropic electrical conductivity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YNLMATOA}},
  note         = {Machine review of arXiv:2608.01945}
}
read the original abstract

Induction welding plays a central role in enabling lightweight, integrated structures made from fibre-reinforced thermoplastic composites. From a modelling perspective, the induction welding process can be approximated by one-way coupled electromagnetic and heat-transfer equations. In practice, material parameters such as electrical conductivity vary significantly resulting from the deviations in the placement of fibres and hence the fibre-fibre contacts in the mesostructure are governed by consolidation quality of the material. Explicit representation of this variability on the macroscopic scale is essential to capture the closed current loops required for the induction heating process. To address this, a stochastic material model is introduced that respects the symmetric positive-definite (SPD) nature of the conductivity tensor and separates scaling and orientation uncertainties, forming the basis of a surrogate framework. The resulting stochastic conductivity model is first used to quantify the uncertainty in the induction heating process through extensive Monte Carlo simulations, providing detailed insight into the induced currents and the resulting temperature field. Subsequently, to enable efficient uncertainty propagation, SPD-aware surrogate models are trained with a subset of the simulation data, consisting of labelled material states and temperature fields. The surrogates are formulated as Constitutive Manifold Neural Networks (CMNNs) that explicitly respect the underlying SPD manifold structure and are integrated with a Neural Ordinary Differential Equation (NODE) framework to capture temporal dynamics. Several NODE integration schemes are evaluated and compared.

Figures

Figures reproduced from arXiv: 2608.01945 by the authors.

Figure 1
Figure 1. Schematic of the induction heating process. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of an SPD matrix as an ellipse (left) and the application of a planar rotation [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Neural network framework for the induction heating process with arbitrarily oriented coil in teal on top of a [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Simulations domains effectively providing a local linearisation of the underlying manifold structure. Finally, ϱ flat combines these tangent-space representations into a single Euclidean vector, making the result compatible with standard neural network layers. Since th…
Figure 5
Figure 5. Figure 5: Deterministic simulation results. composite commonly used in aerospace applications [8], with a symmetric layup [45◦ , −45◦ ]s. Each ply has a thickness of 0.14 mm. For simplicity heat capacity cp, density ϱ and thermal conductivities κ are assumed to be constant with …
Figure 6
Figure 6. Figure 6: Independent distributions of the second eigenvalue of electrical conductivity for all four plies with [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: Independent distributions of the rotation angle in the XY plane for all four plies with scaling-orientation [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Statistics for different types of uncertainties in the electrical conductivity during induction heating. [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: Probability density functions obtained by kernel density estimates at 5s and 10s for points marked in Fig.(5b). [PITH_FULL_IMAGE:figures/full_fig_p025_9.png]
Figure 10
Figure 10. Figure 10: Loss history for scaling and orientation uncertainty. [PITH_FULL_IMAGE:figures/full_fig_p026_10.png]
Figure 11
Figure 11. Figure 11: Stability analysis of the learned surrogates. [PITH_FULL_IMAGE:figures/full_fig_p029_11.png]
Figure 12
Figure 12. Figure 12: Errors in estimation of mean and std. for the different networks on rollouts from 0s to 10s. [PITH_FULL_IMAGE:figures/full_fig_p030_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 1 canonical work pages

  1. [1]

    Assessment of Thermoplastic Composite Joining by Resis- tance, Induction, and Ultrasonic Welding - NASA Technical Reports Server (NTRS),

    S. G. Miller, J. J. Pinakidis, P. J. Heimann, A. C. Bergan, P. H. Johnston, J. Su, F. Leone, K. N. Se- gal, W. Mulhearn, J. Chiu, W. Seneviratne, M. Walthers, U. Palliyaguru, S. Slater, M. Wadsworth, J. Hertel, C. Skocik, and J. Stratton, “Assessment of Thermoplastic Composite Joining by Resis- tance, Induction, and Ultrasonic Welding - NASA Technical Rep...

    2025

  2. [2]

    Current status of carbon fibre and carbon fibre composites recycling,

    J. Zhang, V. S. Chevali, H. Wang, and C.-H. Wang, “Current status of carbon fibre and carbon fibre composites recycling,”Composites Part B: Engineering, vol. 193, p. 108053, July 2020

    2020

  3. [3]

    A Study On development of Induction Welding of Thermoplastic Composites,

    P. Velmurugan, J. Manohar, C. R. Kannan, S. Manivannan, J. Vairamuthu, and B. Stalin, “A Study On development of Induction Welding of Thermoplastic Composites,”IOP Conference Series: Materials Science and Engineering, vol. 988, p. 012109, Dec. 2020

    2020

  4. [4]

    Assembling of Carbon Fi- bre/PEEK Composites: Comparison of Ultrasonic, Induction, and Transmission Laser Welding,

    A. Korycki, C. Garnier, M. Bonmatin, E. Laurent, and F. Chabert, “Assembling of Carbon Fi- bre/PEEK Composites: Comparison of Ultrasonic, Induction, and Transmission Laser Welding,” Materials 2022: Fusion Bonding/Welding of Polymer Composites, vol. 15, p. 6365, Sept. 2022

    2022

  5. [5]

    Fusion Bonding/Welding of Thermoplastic Com- posites,

    A. Yousefpour, M. Hojjati, and J.-P. Immarigeon, “Fusion Bonding/Welding of Thermoplastic Com- posites,”Journal of Thermoplastic Composite Materials, vol. 17, pp. 303–341, July 2004

    2004

  6. [6]

    Worral, E

    C. Worral, E. Kellar, and C. Vacogne,Joining of fibre-reinforced polymer composites. Northbridge Road, Berkhamsted, Hertfordshire: Composites UK Audley House, Jan. 2020

    2020

  7. [7]

    Simulating the in- duction heating of cross-ply C/PEKK laminates – sensitivity and effect of material vari- ability,

    W. J. B. Grouve, F. Sacchetti, E. J. Vruggink, and R. Akkerman, “Simulating the in- duction heating of cross-ply C/PEKK laminates – sensitivity and effect of material vari- ability,”Advanced Composite Materials, vol. 30, pp. 409–430, Sept. 2021. _eprint: https://doi.org/10.1080/09243046.2020.1783078. 32

    arXiv 2021

  8. [8]

    Characterisation of Or- thotropic Electrical Conductivity of Unidirectional C/PAEK Thermoplastic Composites,

    Y. Buser, G. Bieleman, W. Grouve, S. Wijskamp, and R. Akkerman, “Characterisation of Or- thotropic Electrical Conductivity of Unidirectional C/PAEK Thermoplastic Composites,” inPro- ceedings of the 20th European Conference on Composite Materials - Composites Meet Sustainability, pp. 83–91, Ecole Polytechnique Fédérale de Lausanne, 2022

    2022

Show all 48 references
  1. [9]

    Stochastic Model and Generator for Random Fields with Symmetry Properties: Application to the Mesoscopic Modeling of Elastic Random Media,

    J. Guilleminot and C. Soize, “Stochastic Model and Generator for Random Fields with Symmetry Properties: Application to the Mesoscopic Modeling of Elastic Random Media,”Multiscale Modeling and Simulation, vol. 11, pp. 840–870, Aug. 2013

    2013

  2. [10]

    Stochastic modelling of symmetric positive definite material tensors,

    S. K. Shivanand, B. Rosić, and H. G. Matthies, “Stochastic modelling of symmetric positive definite material tensors,”Journal of Computational Physics, vol. 505, p. 112883, May 2024

    2024

  3. [11]

    Modeling variability of the electrical conduc- tivity tensor for the induction welding of composites,

    W. J. Schuttert, M. I. Abdul Rasheed, and B. Rosić, “Modeling variability of the electrical conduc- tivity tensor for the induction welding of composites,”PAMM, vol. 24, no. 3, p. e202400058, 2024. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/pamm.202400058

    2024 doi

  4. [12]

    Cheng, N

    Z. Cheng, N. Takahashi, and B. Forghani,Modeling and application of electromagnetic and thermal field in electrical engineering. Modeling and Application of Electromagnetic and Thermal Field in Electrical Engineering, Science Press, Beijing and Springer Nature Singapore Pte, 2019

    2019

  5. [13]

    Surrogate models for uncertainty quantification: An overview,

    B. Sudret, S. Marelli, and J. Wiart, “Surrogate models for uncertainty quantification: An overview,” in2017 11th European Conference on Antennas and Propagation (EUCAP),pp.793–797, Mar.2017

    2017

  6. [14]

    A review of uncertainty quantifi- cation in deep learning: Techniques, applications and challenges,

    M.Abdar, F.Pourpanah, S.Hussain, D.Rezazadegan, L.Liu, M.Ghavamzadeh, P.Fieguth, X.Cao, A. Khosravi, U. R. Acharya, V. Makarenkov, and S. Nahavandi, “A review of uncertainty quantifi- cation in deep learning: Techniques, applications and challenges,”Information Fusion, vol. 76...

    2021

  7. [15]

    Deep UQ: Learning deep neural network surrogate models for high dimensional uncertainty quantification,

    R. K. Tripathy and I. Bilionis, “Deep UQ: Learning deep neural network surrogate models for high dimensional uncertainty quantification,”Journal of Computational Physics, vol. 375, pp. 565–588, Dec. 2018

    2018

  8. [16]

    A Riemannian Framework for Tensor Computing,

    X. Pennec, P. Fillard, and N. Ayache, “A Riemannian Framework for Tensor Computing,”Interna- tional Journal of Computer Vision, vol. 66, pp. 41–66, Jan. 2006

    2006

  9. [17]

    A Riemannian Network for SPD Matrix Learning,

    Z. Huang and L. Van Gool, “A Riemannian Network for SPD Matrix Learning,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 31, Feb. 2017

    2017

  10. [18]

    Riemannian batch normal- ization for SPD neural networks,

    D. Brooks, O. Schwander, F. Barbaresco, J.-Y. Schneider, and M. Cord, “Riemannian batch normal- ization for SPD neural networks,” inAdvances in Neural Information Processing Systems(H. Wal- lach, H. Larochelle, A. Beygelzimer, F. d. Alché-Buc, E. Fox, and R. Garnett, eds.), vo...

    2019

  11. [19]

    Constitutive manifold neural networks,

    W. J. Schuttert, M. I. Abdul Rasheed, and B. Rosić, “Constitutive manifold neural networks,” Journal of Computational Physics, vol. 559, p. 114912, Aug. 2026

    2026

  12. [20]

    Neural Ordinary Differential Equations,

    T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neural Ordinary Differential Equations,”Advances in Neural Information Processing Systems, no. 31, 2018

    2018

  13. [21]

    DeepResidualLearningforImageRecognition,

    K.He, X.Zhang, S.Ren, andJ.Sun, “DeepResidualLearningforImageRecognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778, 2016. 33

    2016

  14. [22]

    A review of joining techniques for thermoplastic composite materials,

    A. Siddique, Z. Iqbal, Y. Nawab, and K. Shaker, “A review of joining techniques for thermoplastic composite materials,”Journal of Thermoplastic Composite Materials, vol. 36, pp. 3417–3454, June 2022

    2022

  15. [23]

    In-Situ Monitoring and Control of Induction Welding in Thermoplastic Composites Using High Definition Fiber Optic Sensors,

    N. Rahim, J. Pandher, N. Coppola, V. Penumetsa, and M. Tooren, “In-Situ Monitoring and Control of Induction Welding in Thermoplastic Composites Using High Definition Fiber Optic Sensors,” in CAMX 2019, NA SAMPE, 2019

    2019

  16. [24]

    Heating mechanisms in induction welding of thermoplastic composites,

    D. Barazanchy and M. van Tooren, “Heating mechanisms in induction welding of thermoplastic composites,”Journal of Thermoplastic Composite Materials, vol. 36, pp. 473–492, Feb. 2023

    2023

  17. [25]

    D. J. Griffiths,Introduction to electrodynamics. Upper Saddle River, NJ: Prentice Hall, 3. ed., reprinted with corrections ed., 1999

    1999

  18. [26]

    Zangwill,Modern Electrodynamics

    A. Zangwill,Modern Electrodynamics. Cambridge: Cambridge University Press, Dec. 2012. ISBN: 9781139034777

    2012

  19. [27]

    Y. M. Buser,Predicting induction heating of tape-based thermoplastic composites: The role of or- thotropic electrical conductivity. PhD thesis, University of Twente, Enschede, Feb. 2025

    2025

  20. [28]

    K. V. Mardia and P. E. Jupp,Directional Statistics. Chichester: John Wiley & Sons, Ltd, 2000

    2000

  21. [29]

    Bathe,Finite Element Procedures

    K.-J. Bathe,Finite Element Procedures. Prentice Hall, Pearson Education, Inc, 1996

    1996

  22. [30]

    COMSOL Multiphsyics®v. 6.2

    “COMSOL Multiphsyics®v. 6.2..” https://www.comsol.com/. [Accesed 6/8/2025]

    2025

  23. [31]

    On the use of the magnetic vector potential in the finite-element analysis of three-dimensional eddy currents,

    O. Biro and K. Preis, “On the use of the magnetic vector potential in the finite-element analysis of three-dimensional eddy currents,”IEEE Transactions on Magnetics, vol. 25, pp. 3145–3159, July 1989

    1989

  24. [32]

    GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems,

    Y. Saad and M. H. Schultz, “GMRES: A Generalized Minimal Residual Algorithm for Solving Nonsymmetric Linear Systems,”SIAM Journal on scientfic and statistical computing, vol. 7, no. 3, pp. 856–869, 1986

    1986

  25. [33]

    Block-implicit multigrid solution of Navier-Stokes equations in primitive variables,

    S. P. Vanka, “Block-implicit multigrid solution of Navier-Stokes equations in primitive variables,” Journal of Computational Physics, vol. 65, pp. 138–158, July 1986

    1986

  26. [34]

    Hairer and G

    E. Hairer and G. Wanner,Solving Ordinary Differential Equations II, vol. 14 ofSpringer Series in Computational Mathematics. Berlin, Heidelberg: Springer, 1996

    1996

  27. [35]

    Scientific Computation, Dordrecht: Springer Netherlands, 2010

    O.P.LeMaîtreandO.M.Knio,Spectral Methods for Uncertainty Quantification: With Applications to Computational Fluid Dynamics. Scientific Computation, Dordrecht: Springer Netherlands, 2010

    2010

  28. [36]

    Solving unsymmetric sparse systems of linear equations with PAR- DISO,

    O. Schenk and K. Gärtner, “Solving unsymmetric sparse systems of linear equations with PAR- DISO,”Future Generation Computer Systems, vol. 20, pp. 475–487, Apr. 2004

    2004

  29. [37]

    Continuous-in-Depth Neural Networks,

    A. F. Queiruga, N. B. Erichson, D. Taylor, and M. W. Mahoney, “Continuous-in-Depth Neural Networks,” Aug. 2020. arXiv:2008.02389 [cs]

    2020 arXiv

  30. [38]

    Hairer, G

    E. Hairer, G. Wanner, and S. P. Norsett,Solving Ordinary Differential Equations I, vol. 8 ofSpringer Series in Computational Mathematics. Berlin, Heidelberg: Springer, 1993. 34

    1993

  31. [39]

    C. T. Kelley,Iterative methods for linear and nonlinear equations.Society for Industrial and Applied Mathematics, 1995

    1995

  32. [40]

    Geometries and Interpolations for Symmetric Positive Definite Matrices,

    A. Feragen and A. Fuster, “Geometries and Interpolations for Symmetric Positive Definite Matrices,” inModeling, Analysis, and Visualization of Anisotropy(T. Schultz, E. Özarslan, and I. Hotz, eds.), (Cham), pp. 85–113, Springer International Publishing, 2017

    2017

  33. [41]

    Multiphysics simulations: Challenges and opportunities,

    D. E. Keyes, L. C. McInnes, C. Woodward, W. Gropp, E. Myra, M. Pernice, J. Bell, J. Brown, A. Clo, J. Connors, E. Constantinescu, D. Estep, K. Evans, C. Farhat, A. Hakim, G. Hammond, G. Hansen, J. Hill, T. Isaac, X. Jiao, K. Jordan, D. Kaushik, E. Kaxiras, A. Koniges, K. Lee, ...

    2013

  34. [42]

    Characterising microstructural organisation in unidirectional composites,

    S. Gomarasca, D. M. J. Peeters, B. Atli-Veltin, and C. Dransfeld, “Characterising microstructural organisation in unidirectional composites,”Composites Science and Technology, vol. 215, p. 109030, Oct. 2021

    2021

  35. [43]

    Optuna: A Next-generation Hyperparam- eter Optimization Framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A Next-generation Hyperparam- eter Optimization Framework,” inProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’19, (New York, NY, USA), pp. 2623–2631, Associ- a...

    2019

  36. [44]

    Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,

    S. Elfwing, E. Uchibe, and K. Doya, “Sigmoid-weighted linear units for neural network function approximation in reinforcement learning,”Neural Networks, vol. 107, pp. 3–11, Nov. 2018

    2018

  37. [45]

    Auto-Encoding Variational Bayes,

    D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,”Proceedings of the International Conference on Learning Representations, 2014

    2014

  38. [46]

    Stiff Neural Ordinary Differential Equa- tions,

    S. Kim, W. Ji, S. Deng, Y. Ma, and C. Rackauckas, “Stiff Neural Ordinary Differential Equa- tions,”Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 31, p. 093122, Sept. 2021. arXiv:2103.15341 [math]

    2021 arXiv

  39. [47]

    Learning Stable Deep Dynamics Models,

    G. Manek and J. Z. Kolter, “Learning Stable Deep Dynamics Models,”Advances in neural infor- mation processing systems, vol. 32, 2019

    2019

  40. [48]

    Enlighten Project

    “Enlighten Project.” https://www.tpc-enlighten.nl. [Accesed 6/8/2025]. 35

    2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.