Pith. sign in

REVIEW 4 major objections 5 minor 43 references

A Neural Operator based on Dynamic Mode Decomposition

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A neural operator that couples dynamic mode decomposition with a branch-trunk network reproduces heat, Laplace, and Burgers solutions, with error bounded by (L_G + 2L_H)ε.

desk verdict A modest hybrid idea undercut by an invalid error bound and missing empirical comparisons. read the letter →

arxiv 2507.01117 v1 pith:DXMT35VX submitted 2025-07-01 cs.LG

Nikita Sakovich , Dmitry Aksenov , Ekaterina Pleshakova , Sergey Gataullin This is my paper
classification cs.LG
keywords neuraloperatorsdynamicmodedecompositionpartialdifferentialequationsscientificmachinelearningdeepbranch-trunkarchitectureoperatorsurrogatemodeling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes a neural operator that fuses dynamic mode decomposition (DMD) with a branch-trunk deep network: DMD reduces training snapshots of a PDE solution to dominant spatial modes and their temporal or parametric dynamics, and these become the inputs of dedicated network branches, while a trunk network encodes the evaluation coordinate. The central claim is that this hybrid — one small architecture, no equation-specific tuning — approximates the solution operators of the heat, Laplace, and Burgers equations with reconstruction accuracy comparable to DeepONet and the Fourier neural operator, while adding physical interpretability and keeping prediction fast for new boundary and initial conditions. The quantitative anchor is Theorem 1: if the true solution map and the network are Lipschitz with constants $L_G$ and $L_H$, the operator error is at most $(L_G + 2L_H)\epsilon$, where $\epsilon$ is the error of the rank-$r$ truncation of the snapshot data. A sympathetic reader would take away that classical modal analysis can do the heavy lifting a purely data-driven operator would have to learn, and that the DMD truncation error is the quantity that governs accuracy.

What carries the argument

The central mechanism is the DMD-in-the-loop architecture of Eqs. (11)–(12). Dynamic mode decomposition — the working assumption that snapshots evolve under a linear operator $x_{k+1} = A x_k$, recovered by the singular value decomposition of the data matrix, truncation to rank $r$, and diagonalization of the projected operator $\tilde{A} = U_r^* X' V_r \Sigma_r^{-1}$ — supplies spatial modes $\Phi$ and scalar dynamics that feed dedicated 'modes' and 'dynamics' branch networks; a trunk network encodes the prediction coordinate, and the branches are combined through Hadamard products and summation in the style of a DeepONet. The second load-bearing piece is Theorem 1, whose bound $(L_G + 2L_H)\epsilon$ turns the rank-truncation error $\epsilon$ into a certificate on operator error: the proof uses the Eckart–Young theorem to replace the DMD modes of the data by those of its optimal rank-$r$ approximation, and Lipschitz continuity of both the target map $G$ and the network $H$ to control the propagation of $\epsilon$. The architecture is what connects the two: because the DMD modes and dynamics are explicit inputs rather than learned features, the truncation error of DMD is the only approximation error the bound must track.

What would settle it

Generate a test family of initial or boundary conditions whose true solutions are dominated by components orthogonal to the first ten DMD modes of the training data on the same 10×10 grid (for example, high-frequency boundary values), train with DMD rank fixed at 10, and measure test error: if the error jumps far above training-era levels, the DMD information bottleneck is refuted as the assumption that carries the method. A second, cheaper check is quantitative: measure the Lipschitz constants $L_G$ and $L_H$ and the Frobenius truncation error $\epsilon$ on the actual test snapshots, and test whether observed reconstruction errors respect the claimed $(L_G + 2L_H)\epsilon$ bound.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a new class of DMD-enhanced neural operator $G_\theta^{\mathrm{DMD}}: \mathcal{A} \to \mathcal{U}$ described by Eqs. (11)–(12), in which the prediction at a point $x$ is a sum over trunk basis functions of products of branch outputs: branches for DMD spatial modes $\varphi_k$, branches for DMD dynamics $\psi_k$, branches for the differential-operator values, and a trunk for the grid coordinates, combined by Hadamard products. The DMD decomposition is computed once from the training function values (Algorithm 2, step 6) with rank 10, so the network only learns how to combine physically meaningful modes and dynamics with the geometry rather than discovering latent structure from scratch. The theoretical guarantee is Theorem 1: if the solution map $G$ is $L_G$-Lipschitz and the network $H$ is $L_H$-Lipschitz, the error of the DMD operator against $G$ is at most $(L_G + 2L_H)\epsilon$, where $\epsilon$ is the Frobenius-norm error of the optimal rank-$r$ SVD approximation, and the proof invokes the Eckart–Young theorem to show that the DMD modes of the truncated data coincide with those of the full data. Experiments on the Laplace, heat, and Burgers equations — 10×10 grids, 1000 samples, DMD rank 10 — report training and test losses that fall by one to two orders of magnitude over 90 epochs, with the largest residual errors near boundaries and corners, and the abstract claims high reconstruction accuracy in comparative analysis with DeepONet and FNO.

Load-bearing premise

The load-bearing premise is that the rank-truncated DMD modes and dynamics computed once from the training snapshots retain all the solution structure the network needs, so anything the rank-10 truncation discards is also information the surrogate cannot recover for new boundary or initial conditions.

Editorial extensions

If this is right

  • Reconstruction accuracy becomes predictable before training: under the theorem, test error is controlled by the DMD rank-truncation error and the two Lipschitz constants, so a user who measures these on a small holdout set can certify accuracy in advance.
  • The same wiring ports across equations: because the mode and dynamics branches are equation-agnostic inputs, one architecture should handle any problem family whose snapshots admit a low-rank DMD representation, which is what the three experiments demonstrate.
  • Data needs shrink relative to purely data-driven operators: the DMD pass fixes the dominant degrees of freedom once, so the network only learns a combination rule rather than discovering latent structure from scratch.
  • Small networks are not only cheaper but also accuracy-optimal under the bound, since an over-parameterized network with a large Lipschitz constant increases the (L_G + 2L_H) prefactor and worsens the worst-case error.
  • Inference stays fast: DMD modes are computed once at train time and a new solution is a single forward pass, matching the speed advantage claimed for neural operators over repeated classical solving.

Reading between the lines

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

  • A direct test the paper leaves implicit: hold the network architecture fixed and vary only the DMD rank; if test error tracks the truncation error epsilon as the theorem predicts, the bound is doing real work, and if not, one of the proof's assumptions is the weak point.
  • The largest reported errors sit near boundaries and corners in all three experiments, which points to a testable variant the paper does not explore: keep the DMD operator in the interior and let a classical boundary-layer solver own the edges, then compare total error.
  • An adaptive-rank extension is the natural next step: replace the fixed rank-10 DMD by a singular-value energy criterion so branch input width varies with the data, which the modular architecture can absorb by construction.
  • The method is most likely to pay off on parametric families where linear dynamics already describe the physics decently — thermal transients, mild-convection flows — and least on shock-dominated problems, where rank-10 linear reconstruction cannot represent the sharp moving fronts the experiments already show as boundary errors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a hybrid neural operator that combines Dynamic Mode Decomposition (DMD) with a branch-trunk deep network for approximating PDE solution operators. The model extracts DMD modes and dynamics from training snapshots and feeds them to specialized branches, with a trunk net over spatial coordinates. The main formal claim is Theorem 1, an error bound of (L_G+2L_H)epsilon for the DMD-neural operator. Empirical sections report training and test losses for the Laplace, heat, and Burgers equations, and the abstract asserts high reconstruction accuracy relative to DeepONet and FNO. The paper also includes data-generation algorithms, architecture diagrams, a code repository link, and a software patent reference.

Significance. If the theorem and comparisons were valid, the paper would make a modest contribution: a lightweight hybrid operator with interpretable DMD modes and a rigorous error bound. The architectural idea is reasonable, and the availability of source code and data-generation algorithms is a positive feature. However, the theoretical result as stated does not follow from the assumptions, and the empirical comparison promised in the abstract is not reported. As it stands, the paper does not establish either of its two central claims, so its current significance is limited.

major comments (4)
  1. [Section 2.4.2, Eq. (14)] The triangle inequality in Eq. (14) is not valid as written. It bounds ||G(u) - G_DMD(u)|| by ||G(u) - G(u_r)|| + ||G_DMD(u_r) - G_DMD(u)||, but inserting u_r requires also the middle term ||G(u_r) - G_DMD(u_r)||, which equals ||G(u_r) - H(u_r, Phi_r(u_r), Phi_r(u_r)^T u_r)||. This is the neural-operator approximation error at the rank-r projection. The Lipschitz assumptions on G and H and the Eckart-Young theorem do not control it. Taking u = u_r (so epsilon = 0) and H identically zero, which satisfies the Lipschitz assumption on H, makes the theorem assert ||G(u)|| <= 0 for an arbitrary nonzero G, so the stated (L_G + 2 L_H) epsilon bound is false as stated.
  2. [Sections 2.4.1-2.4.2 and Algorithm 2] The object analyzed in Theorem 1 is not the object implemented. The theorem defines G_DMD(u) = H(u, Phi_r(u), Phi_r(u)^T u) using the rank-r SVD of the input u, whereas Algorithm 2 computes DMD modes and dynamics once from the training snapshots (step 6) and then uses those fixed features m and d for all training and test predictions (steps 14 and 23). The DMD modes from Algorithm 1 are Phi = X' V_r Sigma_r^{-1} W, which are not the left singular vectors Phi_r(u) and need not be orthogonal, and the dynamics are eigenvalues rather than Phi_r(u)^T u. Consequently, even a corrected version of the theorem would not apply to the reported experiments, and the test losses do not measure generalization to genuinely new inputs.
  3. [Section 3.4 and Abstract] The abstract's claim of high reconstruction accuracy relative to DeepONet and FNO is not supported by the reported experiments. Section 3.4 states that plots of train and test loss for each model are provided, but no numerical results for DeepONet or FNO are given, and the content of Figures 9-10 is not discussed quantitatively. Tables 2, 4, and 6 contain only the proposed model's own losses. Without comparative MSE or relative L2 errors on a common test set, the central empirical claim cannot be evaluated.
  4. [Section 5] The conclusion contains a truncated sentence, 'achieving 0.8-2.1', with no units, no definition of the metric, and no comparison baseline. This is the only quantitative accuracy statement in the conclusion and it is incomplete, so it provides no support for the paper's effectiveness claim.
minor comments (5)
  1. [References] References [4] and [22] are the same work and should not be duplicated.
  2. [Algorithm 2] The symbol N_train is used both for the number of training epochs and for the size of the training split; these should be given distinct names.
  3. [Equation (11)] The inputs m_0,...,m_n, d_0,...,d_k, u_1,...,u_m in Eq. (11) are never defined, and the transition to the compact form in Eq. (12) is not explained.
  4. [Section 3 and Tables 2, 4, 6] The text says metrics are MSE and maximum absolute error, but the tables report only a scalar 'Loss'; the precise loss definition, normalization, and test-set aggregation should be stated.
  5. [Section 2.4.3] The layer counts, widths, activation functions, and DMD rank selection used in the experiments are not specified, which is insufficient for reproducibility despite the provided repository link.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular reduction; Theorem 1 has an invalid proof, not a circular one, and the only self-citation (patent [43]) is not load-bearing.

full rationale

Walking the claimed derivation chain, the DMD features in Algorithm 2 step 6 are computed from the training snapshots and then fed into the branch/trunk network, which is a standard two-stage operator-learning pipeline rather than a fitted parameter renamed as a prediction: the reported test losses are on held-out grid points and the DMD representation is fixed before training. The only first-principles claim, Theorem 1 in Section 2.4.2, is not circular: G_DMD is defined as H(u, Φ_r(u), Φ_r(u)^T u) and the proof attempts to bound it from Lipschitz assumptions plus Eckart-Young, so the conclusion is not identical to an input by construction. However, the proof has a genuine gap: the triangle inequality in Eq. (14) omits the middle term ||G(u_r) - G_DMD(u_r)||, and the stated bound is false for H ≡ 0 with u = u_r; that is a correctness error, not a circularity. The abstract and Section 3.4 promise a comparative analysis against DeepONet and FNO, but this version reports no baseline tables, which is missing external evidence rather than circular reasoning. The only self-citation is the software registration patent [43], mentioned in the introduction and not used to justify any premise; accordingly the score is 2 at most.

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

The central claim rests on the unproved sufficiency of DMD features, Lipschitz assumptions in a flawed theorem, and finite difference data whose accuracy is not quantified. The method introduces no new physical entities, but several hyperparameters and modeling assumptions are left unspecified, reducing the independence of the evidence.

free parameters (5)
  • DMD rank r = 10
    Chosen by hand for all three experiments (stated as rank=10 in Algorithms 3-5). The paper admits rank selection requires preliminary investigation.
  • Number of DMD modes s = not stated
    The architecture depends on s, but no value or sensitivity analysis is given.
  • Learning rate eta = not stated
    Hyperparameter in Algorithm 2, but no value or schedule is reported.
  • Regularization coefficient lambda = not stated
    Used in the loss function, but no value is given.
  • Network widths and depth = not stated
    The architecture diagram is conceptual; implementation details are absent.
assumptions (5)
  • standard math Existence of optimal rank-r SVD approximation u_r with error epsilon
    Used as Assumption 1 in Theorem 1, relying on the Eckart-Young theorem. This is a standard mathematical fact.
  • domain assumption Solution operator G is L_G-Lipschitz
    Assumption 3 of Theorem 1. No justification is given for the specific PDEs considered.
  • ad hoc to paper Neural network H is L_H-Lipschitz in all inputs
    Assumption 4 of Theorem 1. This is a typical assumption for bounded-parameter networks, but the constant L_H is unknown and not estimated.
  • domain assumption DMD modes and dynamics are sufficient to determine the solution operator
    Core premise of the method, used in Algorithm 2. It is unproved and not ablated.
  • domain assumption Finite difference generators produce accurate ground truth
    The data generation relies on explicit finite difference schemes (Algorithms 3-5) with stated stability conditions, but no convergence checks or grid refinement studies are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Neural Operator based on Dynamic Mode Decomposition." pith.science (2026). https://pith.science/paper/DXMT35VX

@misc{pith2026250701117,
  author       = {Pith},
  title        = {Pith review of: A Neural Operator based on Dynamic Mode Decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DXMT35VX}},
  note         = {Machine review of arXiv:2507.01117}
}
read the original abstract

The scientific computation methods development in conjunction with artificial intelligence technologies remains a hot research topic. Finding a balance between lightweight and accurate computations is a solid foundation for this direction. The study presents a neural operator based on the dynamic mode decomposition algorithm (DMD), mapping functional spaces, which combines DMD and deep learning (DL) for spatiotemporal processes efficient modeling. Solving PDEs for various initial and boundary conditions requires significant computational resources. The method suggested automatically extracts key modes and system dynamics using them to construct predictions, reducing computational costs compared to traditional numerical methods. The approach has demonstrated its efficiency through comparative analysis of performance with closest analogues DeepONet and FNO in the heat equation, Laplaces equation, and Burgers equation solutions approximation, where it achieves high reconstruction accuracy.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [1]

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

    work page 2023

  2. [2]

    12556–12569, 2023

    Hao, Z., Wang, Z., Su, H., Ying, C., Dong, Y., Liu, S., Cheng, Z., Song, J., Zhu, J., ”GNOT: A general neural operator transformer for operator learning,” In: International Conference on Machine Learning, pp. 12556–12569, 2023

    work page 2023

  3. [3]

    Azizzadenesheli, K., Kovachki, N., Li, Z., Liu-Schiaffini, M., Kossaifi, J., Anand- kumar, A., ”Neural operators for accelerating scientific simulations and design,” Nature Reviews Physics, vol. 6, pp. 320–328, 2024

    work page 2024

  4. [5]

    S., ”SSNO: Spatio-Spectral Neural Operator for Functional Space Learning of Partial Differential Equations,”IEEE Access, vol

    Rafiq, M., Rafiq, G., Jung, H.-Y., Choi, G. S., ”SSNO: Spatio-Spectral Neural Operator for Functional Space Learning of Partial Differential Equations,”IEEE Access, vol. 10, pp. 15084–15095, 2022. doi: 10.1109/ACCESS.2022.3148401

    work page arXiv 2022

  5. [6]

    Raonic, B., Molinaro, R., Rohner, T., Mishra, S., Bezenac, E., ”Convolutional neural operators,” In:ICLR 2023 Workshop on Physics for Machine Learning, 2023

    work page 2023

  6. [7]

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A., ”Neural operator: Graph kernel network for partial differential equations,”arXiv preprint arXiv:2003.03485, 2020

    arXiv 2003

  7. [8]

    26 18247–18256, 2023

    Wei, M., Zhang, X., ”Super-resolution neural operator,” In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 26 18247–18256, 2023

    work page 2023

  8. [9]

    Rahman, M., Ross, Z., Azizzadenesheli, K., ”U-NO: U-shaped neural operators,” arXiv preprint arXiv:2204.11127, 2022

    arXiv 2022

Show all 43 references
  1. [10]

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Stuart, A., Bhattacharya, K., Anandkumar, A., ”Multipole graph neural operator for parametric partial differential equations,” In:Advances in Neural Information Processing Systems, vol. 33, pp. 6755–6766, 2020

    2020

  2. [11]

    Raonic, B., Molinaro, R., De Ryck, T., Rohner, T., Bartolucci, F., Alaifari, R., Mishra, S., B´ ezenac, E., ”Convolutional neural operators for robust and accurate learning of PDEs,”Advances in Neural Information Processing Systems, vol. 36, pp. 77187–77200, 2023

    2023

  3. [12]

    1–8, 2024

    Michalowska, K., Goswami, S., Karniadakis, G., Riemer-Sørensen, S., ”Neural operator learning for long-time integration in dynamical systems with recurrent neural networks,” In:2024 International Joint Conference on Neural Networks (IJCNN), pp. 1–8, 2024

    2024

  4. [13]

    Yang, Y., Gao, A., Castellanos, J., Ross, Z., Azizzadenesheli, K., Clayton, R., ”Seismic wave propagation and inversion with neural operators,”The Seismic Record, vol. 1, pp. 126–134, 2021

    2021

  5. [14]

    Zhang, Z., ”ModNO: Multi-operator learning with distributed neural operators,” Computer Methods in Applied Mechanics and Engineering, vol. 431, p. 117229, 2024

    2024

  6. [15]

    Tran, A., Mathews, A., Xie, L., Ong, C., ”Factorized Fourier neural operators,” arXiv preprint arXiv:2111.13802, 2021

    2021 arXiv

  7. [16]

    Kovachki, N., Lanthaler, S., Mishra, S., ”On universal approximation and error bounds for Fourier neural operators,”Journal of Machine Learning Research, vol. 22, pp. 1–76, 2021

    2021

  8. [17]

    Hao, W., Liu, X., Yang, Y., ”Newton informed neural operator for solving nonlinear partial differential equations,” In:Advances in Neural Information Processing Systems, vol. 37, pp. 120832–120860, 2024

    2024

  9. [18]

    Lu, L., Jin, P., Karniadakis, G., ”Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators,”arXiv preprint arXiv:1910.03193, 2019

    1910 arXiv

  10. [19]

    Lu, L., Jin, P., Pang, G., Zhang, Z., Karniadakis, G., ”Learning nonlinear opera- tors via DeepONet based on the universal approximation theorem of operators,” Nature Machine Intelligence, vol. 3, pp. 218–229, 2021. 27

    2021

  11. [20]

    Jin, P., Meng, S., Lu, L., ”MIONet: Learning multiple-input operators via tensor product,”SIAM Journal on Scientific Computing, vol. 44, pp. A3490–A3514, 2022

    2022

  12. [21]

    Yin, M., Charon, N., Brody, R., Lu, L., Trayanova, N., Maggioni, M., ”Dimon: Learning solution operators of partial differential equations on a diffeomorphic family of domains,”arXiv preprint arXiv:2402.07250, 2024

    2024 arXiv

  13. [22]

    Sirota, D.D., Gushchin, K.A., Khan, S.A., Kostikov, S.L., Butov, K.A., ”Neu- ral operators for hydrodynamic modeling of underground gas storage facili- ties,”Russian Technological Journal, vol. 12, no. 6, pp. 102–112, 2024. doi: 10.32362/2500-316X-2024-12-6-102-112. EDN: YBOYBL

    2024 doi

  14. [23]

    Li, Z., Huang, D., Liu, B., Anandkumar, A., ”Fourier neural operator with learned deformations for PDEs on general geometries,”Journal of Machine Learning Research, vol. 24, pp. 1–26, 2023

    2023

  15. [24]

    Tripura, T., Chakraborty, S., ”Wavelet neural operator: a neural operator for parametric partial differential equations,”arXiv preprint arXiv:2205.02191, 2022

    2022 arXiv

  16. [25]

    Tripura, T., Chakraborty, S., ”Wavelet neural operator for solving parametric partial differential equations in computational mechanics problems,”Computer Methods in Applied Mechanics and Engineering, vol. 404, p. 115783, 2023

    2023

  17. [26]

    Cai, S., Mao, Z., Wang, Z., Yin, M., Karniadakis, G., ”Physics-informed neural networks (PINNs) for fluid mechanics: A review,”Acta Mechanica Sinica, vol. 37, pp. 1727–1738, 2021

    2021

  18. [27]

    Rasht-Behesht, M., Huber, C., Shukla, K., Karniadakis, G., ”Physics-informed neural networks (PINNs) for wave propagation and full waveform inversions,” Journal of Geophysical Research: Solid Earth, vol. 127, p. e2021JB023120, 2022

    2022

  19. [28]

    Wang, S., Yu, X., Perdikaris, P., ”When and why PINNs fail to train: A neu- ral tangent kernel perspective,”Journal of Computational Physics, vol. 449, p. 110768, 2022

    2022

  20. [29]

    Yu, B., Others, ”The deep Ritz method: a deep learning-based numerical algo- rithm for solving variational problems,”Communications in Mathematics and Statistics, vol. 6, pp. 1–12, 2018

    2018

  21. [30]

    Sirignano, J., Spiliopoulos, K., ”DGM: A deep learning algorithm for solving partial differential equations,”Journal of Computational Physics, vol. 375, pp. 1339–1364, 2018

    2018

  22. [31]

    Cao, Q., Goswami, S., Karniadakis, G., ”Laplace neural operator for solving differential equations,”Nature Machine Intelligence, vol. 6, pp. 631–640, 2024. 28

    2024

  23. [32]

    Xiong, W., Huang, X., Zhang, Z., Deng, R., Sun, P., Tian, Y., ”Koopman neu- ral operator as a mesh-free solver of non-linear partial differential equations,” Journal of Computational Physics, vol. 513, p. 113194, 2024

    2024

  24. [33]

    Lu, J., Jiang, J., Bai, Y., ”Deep Embedding Koopman Neural Operator Based Nonlinear Flight Training Trajectory Prediction Approach,”Mathematics, vol. 12, p. 2162, 2024

    2024

  25. [34]

    Iwata, T., Kawahara, Y., ”Neural dynamic mode decomposition for end-to-end modeling of nonlinear dynamics,”Journal of Computational Dynamics, vol. 10, pp. 268–280, 2023

    2023

  26. [35]

    Tu, J., ”Dynamic mode decomposition: Theory and applications,”Princeton University, 2013

    2013

  27. [36]

    Proctor, J., Brunton, S., Kutz, J., ”Dynamic mode decomposition with control,” SIAM Journal on Applied Dynamical Systems, vol. 15, pp. 142–161, 2016

    2016

  28. [37]

    Schmid, P., ”Dynamic mode decomposition and its variants,”Annual Review of Fluid Mechanics, vol. 54, pp. 225–254, 2022

    2022

  29. [38]

    Chen, S., Liu, Z., Zhang, W., Yang, J., ”A hard-constraint wide-body physics- informed neural network model for solving multiple cases in forward problems for partial differential equations,”Applied Sciences, vol. 14, p. 189, 2023

    2023

  30. [39]

    Zhang, Z., Wang, Q., Zhang, Y. et al., ”Physics-informed neural networks with hybrid Kolmogorov-Arnold network and augmented Lagrangian function for solving partial differential equations,”Scientific Reports, vol. 15, p. 10523, 2025. doi:10.1038/s41598-025-92900-1

    2025 doi

  31. [40]

    Li, Y., Gao, W., Ying, S., ”RBF-Assisted Hybrid Neural Network for Solving Partial Differential Equations,”Mathematics, vol. 12, p. 1617, 2024

    2024

  32. [41]

    Wang, J., Li, Y., Wu, A., Chen, Z., Huang, J., Wang, Q., Liu, F., ”Multi-Step Physics-Informed Deep Operator Neural Network for Directly Solving Partial Differential Equations,”Applied Sciences, vol. 14, p. 5490, 2024

    2024

  33. [42]

    Sun, K., Feng, X., ”A second-order network structure based on gradient enhanced physics-informed neural networks for solving parabolic partial differ- ential equations,”Entropy, vol. 25, p. 674, 2023

    2023

  34. [43]

    2025663626; Rospatent Federal Service on Intellectual Property: Moscow, Russia, 2025

    Sakovich, N.D., Aksenov, D.A., Pleshakova, E.S., Gataullin, S.T., ”A neural operator based on dynamic mode decomposition analysis for approximation of partial differential equations,” Certificate of State Registration of Computer Program No. 2025663626; Rospatent Federal Servi...

    2025

  35. [44]

    Eckart, C., Young, G., ”The approximation of one matrix by another of lower rank,”Psychometrika, vol. 1, pp. 211–218, 1936. 30

    1936

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.