Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Kernel manifolds: nonlinear-augmentation dimensionality reduction using reproducing kernel Hilbert spaces

T0 review · 2 major / 5 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 The paper claims that replacing polynomial correction terms with a regularized kernel-interpolated correction in a POD decoder yields better small-latent-dimension accuracy and lower training cost, with quadratic manifolds and GP/RBF closur

desk verdict A clean RKHS unification of QMs and GP/RBF closures whose empirical claims need a validation-protocol fix before they fully land. read the letter →

arxiv 2509.00224 v1 pith:UL5TR2TO submitted 2025-08-29 cs.CE cs.NAmath.NA

Alejandro N. Diaz , Jacob T. Needels , Irina K. Tezaur , Patrick J. Blonigan This is my paper · ORCID
classification cs.CEcs.NAmath.NA
keywords kernelmanifoldsquadraticdimensionalityreductionreproducingHilbertspaceproperorthogonaldecompositionnonlinear-augmentationradialbasisfunctionsreduced-ordermodeling
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

Classical POD approximates high-dimensional simulation snapshots with an affine subspace, but that subspace can become very wide for problems with sharp fronts or slowly decaying Kolmogorov n-width. This paper tries to fix that by keeping the POD encoder and adding a learned nonlinear correction to the decoder: the reconstructed state is qbar + V bq + Vbar n(bq), where n maps low-order POD coefficients to the missing high-order coefficients. The central claim is that choosing n as a regularized interpolant in a user-defined reproducing kernel Hilbert space makes this correction more expressive and cheaper to fit than the polynomial corrections used by quadratic manifolds, while containing quadratic manifolds and Gaussian-process/RBF closures as special cases. If the claim holds, nonlinear-augmentation manifolds become a kernel-selection problem rather than a polynomial-degree problem, and better reconstructions at small latent dimension should come without heavy neural-network training.

What carries the argument

The load-bearing object is the RKHS correction map n: R^r -> R^m, which the decoder uses to reconstruct the high-order POD coefficients that the linear part discards. Regularized kernel interpolation plus the Representer Theorem turn the infinite-dimensional optimization for n into the finite linear system (K(bQ,bQ) + lambda I)Omega = bP^T, so the learned decoder is fully determined by the chosen kernel, the regularization parameter, and the training snapshots.

What would settle it

Hold out a parameter value between two trained values, e.g. α = 3×10^-4 in the boundary-layer problem, tune the kernel and regularization using only training data, and compare kernel-manifold projection error with POD for r ≤ 12; if the kernel manifold is not visibly below POD at small r, or if an independent held-out value reverses the reported ordering of methods, the central transfer claim fails.

Watch

Extended reading notes

Core claim

This paper introduces kernel manifold dimensionality reduction: a decoder of the form g(bq) = qbar + V bq + Vbar n(bq), where the nonlinear correction n is not a polynomial or a neural network but the minimizer of a regularized least-squares fit in an RKHS. By the Representer Theorem, the optimal n has the closed form n(bq) = Omega^T K(bQ, bq), with coefficients solving (K(bQ,bQ) + lambda I)Omega = bP^T. This reduces the design of the nonlinear manifold to the choice of a kernel: polynomial kernels reproduce quadratic and feature-map manifolds, RBF kernels reproduce the GP/RBF closure approach, and the training system scales with the number of snapshots rather than with an explicit feature d

Load-bearing premise

The correction map fitted to training snapshots is assumed to keep working at parameters the training set never saw, and the paper supports that by selecting kernels and regularizers using the test errors rather than by an error bound.

Editorial extensions

If this is right

  • The nonlinear correction can be given arbitrary structure simply by choosing a kernel, so Gaussian, Matérn, inverse-quadratic, polynomial, or feature-map kernels can be tested without changing the training routine.
  • Training cost scales with the number of snapshots M rather than the feature dimension, making high-order polynomial-like corrections feasible without solving large coefficient systems.
  • Quadratic manifolds and Gaussian-process/RBF closure methods become special cases, so existing nonlinear-augmentation implementations can be unified under the kernel formulation.
  • On problems with slowly decaying Kolmogorov n-width, RBF-kernel manifolds can reach target projection accuracy at smaller latent dimension than POD or quadratic manifolds, which would make downstream reduced-order models cheaper to evaluate.
  • For very hard problems, such as the double Mach reflection test, nonlinear augmentation yields only slight improvement over POD because these manifolds still inherit the limits of their underlying linear basis.

Reading between the lines

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

  • The same kernel correction map could be embedded in time-stepping reduced-order models to close the modal equations during online simulation; the paper lists this as future work and does not demonstrate it.
  • A parameterized or multi-task kernel that shares information across parameter values might improve transfer to unseen parameters; the paper trains one kernel interpolant on pooled snapshots, which is a cruder form of sharing.
  • Because the exact low-to-high mode map is not well-defined, the correction's value is inherently about average projection quality rather than recovering discarded modes individually; future theory could target expected-error bounds instead of pointwise map estimates.
  • A practical test of the method's robustness would be to fix kernel and regularization choices on a training-only validation split, then check whether the reported accuracy ordering across latent dimensions persists on a new parameter value.
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

2 major / 5 minor

Summary. The paper introduces 'kernel manifold' (KM) dimensionality reduction, a nonlinear-augmentation method that augments a POD decoder with a correction term n(bq) learned by regularized kernel interpolation in a reproducing kernel Hilbert space. The decoder has the form g(bq) = qbar + V bq + V n(bq), where n solves (K(bQ,bQ)+lambda I)Omega = bP^T. The authors show that the feature-map/quadratic-manifold formulation and the GP/RBF closure of [10] are special cases, and they report numerical comparisons against POD, Alternating QM, and Greedy QM on four test problems: advection-diffusion-reaction, an analytic surface-heating model, a 3D hyperelastic bracket, and double Mach reflection. The central mathematical derivation in Sections 3.1-3.3 is sound: the reduction of (3.5) to (3.7) is a valid orthogonality argument, and the Representer Theorem is applied correctly. The main weakness is empirical: the hyperparameters used in the reported error curves are selected using the same test data on which the methods are then evaluated, so the headline accuracy comparisons are optimistically biased.

Significance. If the empirical claims survive a clean validation protocol, this is a useful and reasonably general contribution. The unification of feature-map manifolds and RBF/GP closures within one RKHS framework is natural and clearly explained, and the method has a genuine practical advantage when high-order polynomial feature maps would make the FM normal equations large: the KM linear system (3.9) scales with the number of snapshots rather than the feature dimension. The derivation is machine-checkable in spirit and the special cases are shown honestly rather than assumed. The paper also reports a fair range of negative results (e.g., double Mach reflection, where nonlinear augmentation gives only marginal gains), which strengthens credibility. The main obstacle is that the numerical support for the central accuracy claim is contaminated by test-set hyperparameter tuning in Appendix A, and the abstract's monotonicity claim is unproved and imprecise. With a corrected validation protocol and a clarified monotonicity statement, the paper would be a solid contribution to nonlinear model reduction.

major comments (2)
  1. [Appendix A and Section 4] Hyperparameters are selected on the same test data later used for the headline error curves. In A.1 the RBF kernel and shape are chosen as the 'lowest error' on the alpha=1e-3 test set (Fig. 24); A.2 chooses Gaussian/epsilon=1e-1 from Fig. 26; A.3 chooses linear Matern from Fig. 28; A.4 chooses inverse quadratic/epsilon=1e-3 from Fig. 30. Regularization values for all methods, including baselines, are fixed from test-error sweeps in Figs. 25, 27, 29, and 31 and reported in Tables 2-5. The comparisons in Figs. 4, 9, 15, and 22 therefore report test-informed tuning performance, not unbiased out-of-sample generalization, and the KM gains over QM/POD are likely optimistic. Please re-run the comparison with a validation split (or nested CV) for hyperparameter selection and reserve the test set for final evaluation; if that is infeasible, state the protocol explicitly and label the reported er
  2. [Abstract and Section 3.2] The claim that KM error decreases monotonically as the latent dimension increases is stated without proof or formalization. In the decoder (3.10), when r increases by one and m is fixed, V and V both change: the first column of the old V becomes a linear mode and the last augmenting mode is dropped from the correction. The feasible sets in (3.5) are therefore not nested, so monotonicity of the projection error in r is not automatic. If the intended statement is monotonicity in m, it should be stated as a property of the training objective (larger codomain with fixed lambda) and distinguished from test projection error. Please either prove the intended statement for the appropriate variable or remove/qualify the claim.
minor comments (5)
  1. [Section 4.3] The text says 'Table 3 lists the regularization values' but the table is numbered Table 4.
  2. [Section 4.1] 'scipy.interpolate.solve ivp()' should be 'scipy.integrate.solve_ivp()'.
  3. [Section 3.3, Eq. (3.16)] The grouping Omega^T phi(bQ)^T G is dimensionally sensible but would benefit from an explicit dimension statement; the notation is easy to misread as Omega^T (phi(bQ)^T G phi(bq)) rather than (Omega^T phi(bQ)^T G) phi(bq).
  4. [Section 4.3 / Figure 14] The statement 'Kernel RBF has monotonically decreasing error as m increases' is made for this example; given the abstract's stronger global claim, please clarify here that this is an empirical observation rather than a consequence of the method.
  5. [Section 5] Typo: 'autencoders' should be 'autoencoders'.

Circularity Check

0 steps flagged · score 2.0 of 10

Core RKHS derivation is self-contained; no circular derivation steps. Empirical comparisons are weakened by test-set hyperparameter selection, but that is a validation flaw rather than derivation circularity.

full rationale

The derivation in Sections 3.1–3.3 is self-contained: the KM decoder (3.10) solves the regularized least-squares problem (3.5)/(3.7) in an RKHS, and the Representer Theorem (Theorem 3.1) directly yields the finite-dimensional solution n(bq) = Omega^T K(bQ,bq) with (K(bQ,bQ)+lambda I)Omega = bP^T. No quantity in this derivation is defined in terms of the reported test errors, and the equivalence of feature-map kernels to FM/QM decoders (3.15)–(3.17) and of RBF kernels to the GP/RBF closure (Remark 3.2) is shown by explicit construction, not assumed. The self-citations [10] and [12] are contextual and not load-bearing: [10] is used only in Remark 3.2 to state an equivalence that follows from the same Representer Theorem, and [12] is mentioned as future work. The main caveat is empirical: Appendix A selects the RBF kernel, shape parameter, and regularization by minimizing testing error (e.g., A.1: 'Among the combinations tested, the quadratic Matérn kernel with shape ϵ = 10−2 achieved the lowest error of E = 5.07 × 10−4 at regularization λ = 10−10. Therefore, we use the quadratic Matérn kernel...'), and Section 4 reuses those settings to report test errors. This is test-set leakage that makes the headline accuracy gains optimistic and can bias method comparisons, but it does not make the mathematical derivation circular. Remark 2.3 also honestly notes the exact map n* is not well-defined and no generalization bounds are supplied, so the empirical demonstration is the main support for the central accuracy claim and that support is contaminated by the Appendix A protocol. These are correctness/validation concerns, not definitional circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The kernel manifold method rests on standard RKHS theory (Representer Theorem, Moore-Aronszajn) and on the domain assumption that low-order POD coefficients determine high-order ones well enough to fit a kernel interpolant that generalizes. The free parameters are the kernel choice, shape parameter, regularization, and number of augmenting modes, all tuned on test error in Appendix A. No new physical entities are introduced.

free parameters (5)
  • RBF shape parameter epsilon = 1e-2 (advection-diffusion-reaction), 1e-1 (surface heating), 1e-1 (bracket), 1e-3 (Euler)
    Selected in Appendix A by sweeping the test error for each problem; the reported accuracy depends on these choices.
  • Regularization parameter lambda = Varies per method and problem, e.g., Kernel RBF lambda=1e-11, 1e-9, 1e-7, 1e-3 in the four examples
    Chosen via test-error sweeps (Figures 24-31), not via a principled a priori rule.
  • Kernel choice = Quadratic Matérn, Gaussian, linear Matérn, inverse quadratic across examples
    Picked based on which RBF gave lowest test error in appendix sweeps.
  • Number of augmenting modes m = Set to 5r, 10r, 4r depending on example
    Determined by examining error plateaus as m varies on test data (Figures 3, 8, 14, 21).
  • Input normalization setting = Applied in some examples, not in others
    Chosen per example, changes conditioning of the kernel matrix and therefore the solution.
assumptions (5)
  • standard math Representer Theorem (Theorem 3.1): minimizer of (3.2) has form s(x)=Omega^T K(X,x) and solves (K+lambda I)Omega = Y^T
    Used without proof (referenced [40, Theorem 9.3] for the scalar case); the paper extends to vector-valued RKHS by product space.
  • standard math Moore-Aronszajn theorem: each positive-definite kernel defines a unique RKHS
    Invoked in Section 3.3 to justify that the kernel choice determines the space of functions n can live in.
  • standard math Orthonormality and complementarity of POD mode matrices V and Vbar (V^T V = I_r, Vbar^T Vbar = I_m, V^T Vbar = 0)
    Used to reduce (3.5) to (3.7) by projecting the error onto the two orthonormal mode sets.
  • domain assumption The map from low-order modal coefficients bq to high-order modal coefficients bp is learnable as an RKHS function that generalizes out-of-sample
    Section 2.3 Remark 2.3 admits the exact map n* is not well-defined; the method assumes an approximate RKHS map is nevertheless useful for unseen data.
  • domain assumption Training data is representative of the test distribution in the sense that kernel interpolation of bq to bp transfers
    No error bounds are given; all numerical evidence is empirical projection error on test snapshots.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Kernel manifolds: nonlinear-augmentation dimensionality reduction using reproducing kernel Hilbert spaces." pith.science (2026). https://pith.science/paper/UL5TR2TO

@misc{pith2026250900224,
  author       = {Pith},
  title        = {Pith review of: Kernel manifolds: nonlinear-augmentation dimensionality reduction using reproducing kernel Hilbert spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UL5TR2TO}},
  note         = {Machine review of arXiv:2509.00224}
}
read the original abstract

This paper generalizes recent advances on quadratic manifold (QM) dimensionality reduction by developing kernel methods-based nonlinear-augmentation dimensionality reduction. QMs, and more generally feature map-based nonlinear corrections, augment linear dimensionality reduction with a nonlinear correction term in the reconstruction map to overcome approximation accuracy limitations of purely linear approaches. While feature map-based approaches typically learn a least-squares optimal polynomial correction term, we generalize this approach by learning an optimal nonlinear correction from a user-defined reproducing kernel Hilbert space. Our approach allows one to impose arbitrary nonlinear structure on the correction term, including polynomial structure, and includes feature map and radial basis function-based corrections as special cases. Furthermore, our method has relatively low training cost and has monotonically decreasing error as the latent space dimension increases. We compare our approach to proper orthogonal decomposition and several recent QM approaches on data from several example problems.

Figures

Figures reproduced from arXiv: 2509.00224 by the authors.

Figure 1
Figure 1. Solution profile along x = 0.75 for different diffusion values at different timesteps. To collect training data Q, we simulate (4.1) for ℓ = 10 logarithmically spaced values of α ∈ [10−6 , 10−1 ]. This results in a training dataset Q = h q (1) 0 . . . q (1) Nt . . . q (ℓ) 0 . . . q (ℓ) Nt i ∈ R N×M, (4.3) where N = 65, 536 and M = ℓ(Nt + 1) = 5, 010 and where q (k) j corresponds to the state of (4.2) at time step j … view at source ↗
Figure 2
Figure 2. First 500 normalized singular values for the 2D advection-diffusion-reaction example. For this example, the Kernel RBF approach uses the quadratic Mat´ern kernel with shape parameter ϵ = 10−2 . We do not use the input normalization discussed in Remark 3.3 for any of the KMs computed. The Kernel QM, Kernel RBF, Alternating QM, and Greedy QM approaches use the regularization values in [PITH_FULL_IMAGE:figures/full_fi… view at source ↗
Figure 3
Figure 3. Relative projection error as a function of number of augmenting modes m for several different latent dimension sizes r for the 2D advection-diffusion-reaction example. Next, we compare the different approaches by varying the latent dimension size r. As indicated above, we take the number of augmenting modes to be m = 5r for the Kernel QM, Kernel RBF, and Alternating QM approaches [PITH_FULL_IMAGE:figures/full_fig_p… view at source ↗
Figures from the paper (28 more)
Figure 4
Figure 4. Figure 4: Relative projection error and train time as a function of the latent dimension size r for different dimensionality reduction approaches. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Plot of state and relative projection error at different time steps for r = 6 and m = 30. Top row: full state; second row: POD projection error; third row: Alternating QM projection error; bottom row: Kernel RBF projection error. 4.2 High-speed aerodynamics – analytic …
Figure 6
Figure 6. Figure 6: Field quantity s for several different parameters µ on flared-cone geometry. For a given parameter µ, we compute s on a uniform grid for 100 equispaced z values in [0, 2] and 100 equispaced θ values in [0, 2π], resulting in snapshots q(µ) ∈ R N , N = 10, 000. To collec…
Figure 7
Figure 7. Figure 7: Normalized singular values for the surface heating example. For this example, the Kernel RBF approach uses the Gaussian kernel with shape parameter ϵ = 10−1 . The Kernel QM and Kernel RBF approaches both use the input normalization discussed in Remark 3.3, and the Kern…
Figure 8
Figure 8. Figure 8: Relative projection error as a function of number of augmenting modes m for several different latent dimension sizes r for the surface heating example. We now compare the different approaches by computing the error as a function of the latent space dimension r. From […
Figure 9
Figure 9. Figure 9 [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Plot of full state and relative projection errors for r = 10, m = 100 for the surface heating example. 4.3 Solid mechanics – 3D flexible bracket The third example is a moderately sized 3D hyper-elastic solid dynamics problem, referred to herein as the flexible bracket…
Figure 11
Figure 11. Figure 11: Geometry and mesh for the 3D flexible bracket example, along with a one-dimensional (1D) cartoon illustrating the problem setup. moving throughout the bracket, leading to truly nonlinear 3D dynamics. A linear version of this problem was considered in [18]. To generate…
Figure 12
Figure 12. Figure 12: Displacement magnitude solutions to the flexible bracket problem, scaled by a factor of five for visualization purposes, at the final time t = 20.0 × 10−3 seconds for different initial z–velocities µ. 0 500 1000 1500 Index k 10 15 10 12 10 9 10 6 10 3 10 0 k / 1 Displ…
Figure 13
Figure 13. Figure 13: Singular values for bracket displacement, velocity, and acceleration fields. For this example, we compute separate nonlinear-augmentation manifolds for the displacement, velocity, and acceleration fields. For each field, the Kernel RBF approach uses the linear Mat´ern…
Figure 14
Figure 14. Figure 14: Relative projection error as a function of number of augmenting modes m for several different latent dimension sizes r for the 3D bracket example. Next, in [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]
Figure 15
Figure 15. Figure 15: Relative projection error and train time as a function of the latent dimension size r for different dimensionality reduction approaches for the 3D bracket example. Finally, in [PITH_FULL_IMAGE:figures/full_fig_p021_15.png]
Figure 16
Figure 16. Figure 16: Plots of full states and relative projection errors for the displacement, velocity and acceleration fields at the final time t = 20.0 × 10−3 seconds. Top row: full state; second row: POD projection error; third row: Alternating QM projection error; bottom row: Kernel …
Figure 17
Figure 17. Figure 17: Cartoon of double Mach reflection computational domain. When the shock interacts with the wedge, a complicated shock interaction occurs, as seen in [PITH_FULL_IMAGE:figures/full_fig_p023_17.png]
Figure 18
Figure 18. Figure 18: Pressure at t = 0.25 for wedge half-angle α = 30◦. To parametrize the problem, we consider the pressure fields p(x, y) computed for different wedge half￾angles α. This results in a range of different shock structures that are difficult to capture accurately with a low…
Figure 19
Figure 19. Figure 19: plots the initial pressure for varying wedge angles. Notice that each initial pressure contains a sharp front. The training set is comprised of α = 5◦ , 15◦ , 25◦ , 35◦ , 45◦ , and all testing is done with α = 30◦ . (a) Training, α = 5◦ (b) Training, α = 45◦ (c) Testi…
Figure 20
Figure 20. Figure 20: Normalized singular values for double Mach reflection example. For this example, the Kernel RBF approach uses the inverse quadratic RBF kernel with shape param￾eter ϵ = 10−3 . The Kernel QM and Kernel RBF approaches do use the input normalization discussed in Remark 3…
Figure 21
Figure 21. Figure 21: plots the relative projection error as a function of the number of augmenting modes for r = 10, 20, 30 for Kernel QM, Kernel RBF, and Alternating QM. For r = 10, Alternating QM outperforms both Kernel QM and Kernel RBF, yielding smaller relative errors and plateaus in…
Figure 22
Figure 22. Figure 22: Relative projection error and train time as a function of the latent dimension size r for different dimensionality reduction approaches for the 2D Euler double Mach reflection example. Lastly, [PITH_FULL_IMAGE:figures/full_fig_p026_22.png]
Figure 23
Figure 23. Figure 23: Plot of state and relative projection error at different time steps for r = 10, m = 40 for the 2D Euler double Mach reflection example. Top row: full state; second row: POD projection error; third row: Alternating QM projection error; bottom row: Kernel RBF projection…
Figure 24
Figure 24. Figure 24: Projection error as a function of regularization λ for different RBFs and shape parameters for the boundary layer example, r = 10, m = 20. Next we examine the effect of the regularization parameter λ on the different KM and QM approaches for different latent space dim…
Figure 25
Figure 25. Figure 25: Projection error as a function of regularization λ for different latent dimension sizes for the boundary layer example, m = 2r. A.2 High-speed aerodynamics – analytic model of surface heating As in the previous numerical example, we first compare KMs with different RB…
Figure 26
Figure 26. Figure 26: shows that over the regularization range tested each KM attains a similar minimum error for each RBF and shape parameter combination. For each shape value, as the regularization increases, the error also increases from a minimum value to the POD projection error. The …
Figure 27
Figure 27. Figure 27: Mean projection error as a function of regularization λ for different latent dimension sizes for the surface heating example, m = 10r. A.3 Solid mechanics - 3D flexible bracket As in the previous examples, we first examine the effect of different RBFs and shape parame…
Figure 28
Figure 28. Figure 28: Displacement (r = 10), velocity (r = 30), and acceleration (r = 30) projection errors as a function of regularization λ for different RBFs and shape parameters for the flexible bracket example, m = 5r. Next we examine the effect of the regularization value λ on the Ke…
Figure 29
Figure 29. Figure 29: Displacement, velocity, and acceleration projection errors as a function of regularization λ for different latent dimension sizes for the flexible bracket example, m = 5r. A.4 2D Euler – double Mach reflection [PITH_FULL_IMAGE:figures/full_fig_p032_29.png]
Figure 30
Figure 30. Figure 30: Projection error as a function of regularization λ for different RBFs and shape parameters for the double Mach reflection example, r = 10, m = 20 [PITH_FULL_IMAGE:figures/full_fig_p033_30.png]
Figure 31
Figure 31. Figure 31: plots relative projection error as a function of regularization λ for the Kernel QM, Kernel RBF, Alternating QM, and Greedy QM approaches for several different values of r. For this comparison, we take m = 2r for the Kernel QM, Kernel RBF, and Alternating QM approache…

Discussion (0). Continue with ORCID 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. Hybrid coupling with operator inference and the overlapping Schwarz alternating method

    math.NA 2025-11 conditional novelty 5.0 of 10

    Overlapping Schwarz iteration can glue operator-inference ROMs to each other and to full-order finite-element models for 3D nonlinear solid dynamics, with online speedups up to 106x.

Reference graph

Works this paper leans on

48 extracted references · 42 canonical work pages · cited by 1 Pith paper

  1. [10]

    S. A. de Parga, R. Tezaur, C. G. Hern´ andez, and C. Farhat. Nonlinear projection-based model order reduction with machine learning regression for closure error modeling in the latent space. arXiv preprint arXiv:2507.00634, 2025

    work page arXiv 2025

  2. [1]

    Norma.jl

    Alejandro Mota. Norma.jl

    work page

  3. [2]

    P. J. Baddoo, B. Herrmann, B. J. McKeon, and S. L. Brunton. Kernel learning for robust dynamic mode decomposition: linear and nonlinear disambiguation optimization. Proceedings of the Royal Society A , 478(2260):20210830, 2022

    work page 2022

  4. [3]

    Barnett and C

    J. Barnett and C. Farhat. Quadratic approximation manifold for mitigating the kolmogorov barrier in nonlinear projection-based model order reduction. Journal of Computational Physics , 464:111348, 2022

    work page 2022

  5. [4]

    Barnett, C

    J. Barnett, C. Farhat, and Y. Maday. Neural-network-augmented projection-based model order reduc- tion for mitigating the Kolmogorov barrier to reducibility. J. Comput. Phys. , 492:Paper No. 112420, 20, 2023

    work page 2023

  6. [5]

    Batlle, M

    P. Batlle, M. Darcy, B. Hosseini, and H. Owhadi. Kernel methods are competitive for operator learning. Journal of Computational Physics , 496:112549, 2024. 33 Kernel Manifolds Diaz, Needels, Tezaur and Blonigan

    work page 2024

  7. [6]

    Benner, P

    P. Benner, P. Goyal, J. Heiland, and I. Pontes Duff. A quadratic decoder approach to nonintrusive reduced-order modeling of nonlinear dynamical systems. PAMM, 23(1):e202200049, 2023

    work page 2023

  8. [7]

    Bonneville, Y

    C. Bonneville, Y. Choi, D. Ghosh, and J. L. Belof. Gplasdi: Gaussian process-based interpretable latent space dynamics identification through deep autoencoder. Computer Methods in Applied Mechanics and Engineering, 418:116535, 2024

    work page 2024

Show all 48 references
  1. [8]

    Bonneville, X

    C. Bonneville, X. He, A. Tran, J. S. Park, W. Fries, D. A. Messenger, S. W. Cheung, Y. Shin, D. M. Bortz, D. Ghosh, J.-S. Chen, J. Belof, and Y. Choi. A comprehensive review of latent space dynam- ics identification algorithms for intrusive and non-intrusive reduced-order-mode...

    2024 arXiv

  2. [9]

    Cocola, J

    J. Cocola, J. Tencer, F. Rizzi, E. Parish, and P. Blonigan. Hyper-reduced autoencoders for efficient and accurate nonlinear model reductions, 2023

    2023

  3. [11]

    A. N. Diaz, Y. Choi, and M. Heinkenschloss. A fast and accurate domain-decomposition nonlinear manifold reduced order model. Comput. Methods Appl. Mech. Engrg. , 425:116943, 2024

    2024

  4. [12]

    A. N. Diaz, S. A. McQuarrie, J. T. Tencer, and P. J. Blonigan. Interpretable and flexible non-intrusive reduced-order models using reproducing kernel hilbert spaces. arXiv preprint arXiv:2506.10224 , 2025

    2025

  5. [13]

    W. D. Fries, X. He, and Y. Choi. Lasdi: Parametric latent space dynamics identification. Computer Methods in Applied Mechanics and Engineering , 399:115436, 2022

    2022

  6. [14]

    Geelen, L

    R. Geelen, L. Balzano, S. Wright, and K. Willcox. Learning physics-based reduced-order models from data using nonlinear manifolds. Chaos: An Interdisciplinary Journal of Nonlinear Science, 34(3):033122, 03 2024

    2024

  7. [15]

    Geelen, S

    R. Geelen, S. Wright, and K. Willcox. Operator inference for non-intrusive model reduction with quadratic manifolds. Computer Methods in Applied Mechanics and Engineering , 403:115717, 2023

    2023

  8. [16]

    Ghattas and K

    O. Ghattas and K. Willcox. Learning physics-based models from data: Perspectives from inverse problems and model reduction. Acta Numerica, 30:445–554, 2021

    2021

  9. [17]

    Goyal and P

    P. Goyal and P. Benner. Generalized quadratic embeddings for nonlinear dynamics using deep learning. Physica D: Nonlinear Phenomena , 463:134158, 2024

    2024

  10. [18]

    Gruber and I

    A. Gruber and I. Tezaur. Variationally consistent hamiltonian model reduction. SIAM Journal on Applied Dynamical Systems , 24(1):376–414, 2025

    2025

  11. [19]

    Gubisch and S

    M. Gubisch and S. Volkwein. Chapter 1: Proper Orthogonal Decomposition for linear-quadratic optimal control. In P. Benner, A. Cohen, M. Ohlberger, and K. Willcox, editors, Model Reduction and Approx- imation: Theory and Algorithms , Computational Science and Engineering, pages...

  12. [20]

    M. D. Gunzburger, J. S. Peterson, and J. N. Shadid. Reduced-order modeling of time-dependent pdes with multiple parameters in the boundary data. Computer Methods in Applied Mechanics and Engineering, 196(4):1030–1047, 2007

    2007

  13. [21]

    Hinze and S

    M. Hinze and S. Volkwein. Proper orthogonal decomposition surrogate models for nonlinear dynamical systems: Error estimates and suboptimal control. In P. Benner, V. Mehrmann, and D. C. Sorensen, editors, Dimension Reduction of Large-Scale Systems , Lecture Notes in Computation...

    2005

  14. [22]

    S. Jain, P. Tiso, J. B. Rutzmoser, and D. J. Rixen. A quadratic manifold for model order reduction of nonlinear structural dynamics. Computers & Structures , 188:80–94, 2017

    2017

  15. [23]

    Jalalian, J

    Y. Jalalian, J. F. O. Ramirez, A. Hsu, B. Hosseini, and H. Owhadi. Data-efficient kernel methods for learning differential equations and their solution operators: Algorithms and error analysis. arXiv preprint arXiv:2503.01036, 2025

    2025 arXiv

  16. [24]

    Y. Kim, Y. Choi, D. Widemann, and T. Zohdi. A fast and accurate physics-informed neural network reduced order model with shallow masked autoencoder. Journal of Computational Physics , 451:Paper No. 110841, 29, 2022

    2022

  17. [25]

    Kramer, B

    B. Kramer, B. Peherstorfer, and K. Willcox. Learning nonlinear reduced models from data with operator inference. Annual Review of Fluid Mechanics , 56:521–548, 2024

    2024

  18. [26]

    Lance, A

    B. Lance, A. M. Krueger, B. A. Freno, and R. M. Wagnild. Validation Study of the Multi-Fidelity Toolkit. 2022

    2022

  19. [27]

    B. W. Larsen, T. G. Kolda, A. R. Zhang, and A. H. Williams. Tensor decomposition meets rkhs: Efficient algorithms for smooth and misaligned data. arXiv preprint arXiv:2408.05677 , 2024

    2024 arXiv

  20. [28]

    Lee and K

    K. Lee and K. T. Carlberg. Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders. Journal of Computational Physics , 404:108973, 32, 2020

    2020

  21. [29]

    Lowery, J

    M. Lowery, J. Turnage, Z. Morrow, J. D. Jakeman, A. Narayan, S. Zhe, and V. Shankar. Kernel neural operators (knos) for scalable, memory-efficient, geometrically-flexible operator learning. arXiv preprint arXiv:2407.00809, 2024

    2024 arXiv

  22. [30]

    C. A. Micchelli and M. Pontil. On learning vector-valued functions. Neural computation, 17(1):177–204, 2005

    2005

  23. [31]

    Mota and J

    A. Mota and J. A. Zimmerman. A variational, finite-deformation constitutive model for piezoelectric materials. International Journal for Numerical Methods in Engineering , 85(6):752–767, 2011

    2011

  24. [32]

    Ohlberger and S

    M. Ohlberger and S. Rave. Reduced basis methods: Success, limitations and future challenges. Pro- ceedings of the Conference Algoritmy , pages 1–12, 2016

    2016

  25. [33]

    Parish, M

    E. Parish, M. Yano, I. Tezaur, and T. Iliescu. Residual-based stabilized reduced-order models of the transient convection–diffusion–reaction equation obtained through discrete and continuous projection. Archives of Computational Methods in Engineering , 32:1885–1929, 2025

    1929

  26. [34]

    Peherstorfer and K

    B. Peherstorfer and K. Willcox. Data-driven operator inference for nonintrusive projection-based model reduction. Comput. Methods Appl. Mech. Engrg. , 306:196–215, 2016

    2016

  27. [35]

    E. Qian, B. Kramer, B. Peherstorfer, and K. Willcox. Lift & learn: Physics-informed machine learning for large-scale nonlinear dynamical systems. Physica D: Nonlinear Phenomena , 406:132401, 2020

    2020

  28. [36]

    Romor, G

    F. Romor, G. Stabile, and G. Rozza. Non-linear manifold reduced-order models with convolutional autoencoders and reduced over-collocation method. Journal of Scientific Computing , 94(3):Paper No. 74, 2023

    2023

  29. [37]

    J. B. Rutzmoser, D. J. Rixen, P. Tiso, and S. Jain. Generalization of quadratic manifolds for reduced order modeling of nonlinear structural dynamics. Computers & Structures , 192:196–209, 2017

    2017

  30. [38]

    Salvador, L

    M. Salvador, L. Dede, and A. Manzoni. Non intrusive reduced order modeling of parametrized pdes by kernel pod and neural networks. Computers & Mathematics with Applications , 104:1–13, 2021. 35 Kernel Manifolds Diaz, Needels, Tezaur and Blonigan

    2021

  31. [39]

    G. Santin. Approximation with kernel methods, 2018. Lecture Notes WS 2017/18, Department of Mathematics, University Stuttgart, Germany

    2018

  32. [40]

    Santin and B

    G. Santin and B. Haasdonk. 9 kernel methods for surrogate modeling. In P. Benner, S. Grivet-Talocia, A. Quarteroni, G. Rozza, W. Schilders, and L. M. Silveira, editors, Model Order Reduction. Volume 1: System- and Data-Driven Methods and Algorithms , pages 311–354. Walter de G...

    2021

  33. [41]

    Sch¨ olkopf, A

    B. Sch¨ olkopf, A. Smola, and K.-R. M¨ uller. Kernel principal component analysis. In International conference on artificial neural networks , pages 583–588. Springer, 1997

    1997

  34. [42]

    Schwerdtner, P

    P. Schwerdtner, P. Mohan, A. Pachalieva, J. Bessac, D. O’Malley, and B. Peherstorfer. Online learning of quadratic manifolds from streaming data for nonlinear dimensionality reduction and nonlinear model reduction. Proceedings of the Royal Society A , 481(2314):20240670, 2025

    2025

  35. [43]

    Schwerdtner and B

    P. Schwerdtner and B. Peherstorfer. Greedy construction of quadratic manifolds for nonlinear dimen- sionality reduction and nonlinear model reduction. arXiv preprint arXiv:2403.06732 , 2024

    2024 arXiv

  36. [44]

    L. F. Shampine and M. W. Reichelt. The MATLAB ODE suite. SIAM Journal on Scientific Computing , 18(1):1–22, 1997

    1997

  37. [45]

    Virtanen, R

    P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wilson, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, R. Kern, E. Larson, C. J. Carey,˙I. Polat, Y. Feng,...

    2020

  38. [46]

    Wendland

    H. Wendland. Scattered data approximation, volume 17. Cambridge university press, 2004

    2004

  39. [47]

    Woodward and P

    P. Woodward and P. Colella. The numerical simulation of two-dimensional fluid flow with strong shocks. Journal of computational physics , 54(1):115–173, 1984

    1984

  40. [48]

    G. B. Wright. Radial basis function interpolation: numerical and analytical developments . University of Colorado at Boulder, 2003. 36

    2003

Pith tools

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