Pith. sign in

REVIEW 3 major objections 7 minor 57 references

A Learning-based Domain Decomposition Method

T0 review · 3 major / 7 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 single neural operator trained on simple squares acts as a universal local solver in an additive Schwarz iteration, solving elliptic PDEs with discontinuous microstructures on complex domains at the level of its standalone error.

desk verdict A genuinely new combination—one pretrained neural operator as a shared additive-Schwarz subdomain solver—with a theory gap between the alternating Schwarz result and the implemented method, and an empirical section that needs baselines and error bars; still worth serious peer review. read the letter →

arxiv 2507.17328 v2 pith:AFBM3OU3 submitted 2025-07-23 cs.LG math-phmath.MP

Rui Wu , Nikola Kovachki , Burigede Liu This is my paper
classification cs.LGmath-phmath.MP MSC 65N5535J2568T07
keywords learning-baseddomaindecompositionneuraloperatorphysics-pretrainedadditiveSchwarzellipticPDEinterpolatedconvolutionmicrostructuregeneralizationresolutioninvariance
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 learning-based domain decomposition method: train one neural operator once on a simple square domain, then use it as the local solver in every subdomain of an overlapping domain decomposition while an additive Schwarz iteration reconciles the pieces. Its claim is that this single, never-retrained surrogate solves elliptic equations with discontinuous microstructures on large, non-convex domains — a square fifty times the training size, L-shaped and I-shaped geometries, grain counts from 10 to 1000 — with relative $L^2$ error of 0.039% to 0.645% for the solution and 3.89% to 8.10% for its gradient, matching the level of the operator's own standalone error. A theorem shows neural-operator approximations of the Schwarz local map can be made arbitrarily accurate over compact sets in spaces of continuously differentiable functions, while the implemented method is the additive (parallel) variant. If the claim holds, a large class of heterogeneous-material problems becomes solvable with no per-domain training and no per-resolution retraining.

What carries the argument

Three pieces carry the argument. First, the PPNO architecture: a U-Net-shaped neural operator whose interpolated convolution treats kernel values as nodal parameters of a linear spline, so no weight is tied to a specific discretization, with group normalization and projection layers keeping the computation on a fixed grid. Second, the kernel-based extension operator $V$, which maps boundary data $g$ on $\partial\Omega$ to a smooth interior field by inverse-square-distance weighting and feeds it as an input channel, so boundary conditions enter the operator as functions. Third, the additive Schwarz base operator $G(a,g)=\sum_{j=1}^{M}\hat{G}_{D,j}(a,g)\varphi_j$, which applies the single trained operator $\hat{G}_D$ to each congruent overlapping subdomain and assembles the results with a partition of unity $\{\varphi_j\}$. Theorem 1, the paper's main theoretical result, states that if the abstract problem satisfies the alternating Schwarz method on a compact set and the local solution operators are continuous and locally Lipschitz, an alternating Schwarz neural operator approximates the true solution arbitrarily well in $C^s$; the proof rests on a composition lemma showing a neural-operator approximation stays accurate after many self-compositions, a result the paper notes applies to autoregressive function-sequence models generally.

What would settle it

Run the same pretrained PPNO-driven additive Schwarz iteration on a non-convex domain holding a microstructure far outside its training distribution — for example a two-phase medium with contrast ratio $1{:}10^4$ — and compare the converged iterate with a fine FEM reference solution. If the relative $L^2$ error of $u$ does not stay at the PPNO's standalone level, or if the iteration reaches its tolerance at a fixed point visibly different from the FEM solution, the surrogate-convergence premise fails.

Watch

Extended reading notes

Core claim

The central discovery is that a physics-pretrained neural operator (PPNO), trained only on random 50-grain Voronoi microstructures in the unit square, serves as a universal local solver inside an additive Schwarz iteration for $- abla\cdot(a\nabla u)=0$ with random Dirichlet boundary data. Large composite domains are partitioned into congruent overlapping squares, the same trained operator is applied to each, and a partition of unity glues the local outputs; the iteration terminates when the successive $L^2$ difference falls below tolerance. The paper reports that the converged solution's accuracy is set by the local operator, not by the decomposition: relative $L^2$ errors of 0.645%, 0.388%, and 0.039% for $u$ (5.99%, 8.10%, 3.89% for $\nabla u$) on the $10\times10$ square, L-shaped, and I-shaped tests sit at the same order of magnitude as the standalone PPNO error. The construction is resolution-invariant because interpolated convolutions define kernels on function space rather than on a fixed grid, and the operator generalizes to microstructures unseen in training, including grain counts up to 1000, graded Voronoi, hexagonal, and fibre-composite patterns.

Load-bearing premise

The load-bearing premise is that the additive iteration using the neural-network surrogate — instead of the exact local solver — converges to a fixed point close to the true solution on subdomains unlike anything the operator was trained on; the paper's convergence theorem does not cover this case (it treats the alternating method for continuously differentiable functions), so today the premise rests on the reported experiments alone.

Editorial extensions

If this is right

  • Converged accuracy inherits from the local operator: relative $L^2$ errors on the large-domain tests (0.039%–0.645% for $u$, 3.89%–8.10% for $\nabla u$) sit at the same order as the standalone PPNO, so the decomposition adds little or no extra error.
  • One offline pretraining suffices for many online problems: the same square-trained operator was reused for all test geometries, boundary conditions, and decompositions without retraining.
  • The method is resolution-invariant: the same trained weights evaluate at any discretization by design, and test error stays low across resolutions.
  • Out-of-distribution microstructures stay within bounded error: 10–1000 grains, graded Voronoi, hexagonal, and fibre-composite patterns were solved accurately by an operator trained only on 50-grain random Voronoi cells.
  • The composition lemma is not specific to Schwarz methods: any autoregressive operator model on compact sets in $C^s$ keeps its accuracy after repeated self-composition, which the paper points to for applications such as learned weather forecasting.

Reading between the lines

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

  • My inference: the paper's theory covers the alternating Schwarz method in spaces of continuously differentiable functions, while the numerics use the additive variant; if surrogate error propagates like the local operator's Lipschitz constant raised to the iteration count, the observed rapid convergence is expected, and failure would show up first on extreme-contrast or highly non-smooth microstru
  • My inference: the same pretrained local operator could serve as a reusable cell-problem solver in multiscale homogenization, replacing per-cell numerical solves — a link the paper mentions as future work but does not develop.
  • My inference: because the theory allows general Lipschitz subdomains, a direct test of the method on rotated, sheared, or otherwise mapped subdomain tiles would show whether the elliptic equation's length-scale invariance generalizes to genuine geometry invariance, which would unlock unstructured decompositions.
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 paper proposes a learning-based domain decomposition method (L-DDM) in which a single pretrained neural operator (PPNO), trained on a unit square with random Voronoi microstructures and random Dirichlet boundary conditions, is reused as a surrogate local solver inside an overlapping additive Schwarz iteration. The authors prove an existence theorem for neural operator approximation of the alternating Schwarz method in spaces of continuously differentiable functions, introduce a resolution-invariant convolutional architecture (PPNO), benchmark it on a square domain, and demonstrate the full L-DDM on a large square, an L-shaped domain, and an I-shaped domain with graded microstructure, reporting relative L2 errors for the solution and its gradient.

Significance. If the central claim holds, the paper would make a useful contribution to operator-learning-based domain decomposition: reusing one pretrained operator across many congruent subdomains is an attractive way to handle large, geometrically complex domains without retraining per subdomain, and the resolution-invariant interpolated convolution is a principled architectural idea. The theoretical results on composition of neural operator approximations (Lemma 2, Theorem 3, and Theorem 1) are nontrivial and of independent interest for auto-regressive operator models. The empirical generalization study across grain counts and microstructure classes is broad and visually convincing. However, the advertised claims go beyond what is actually established: the implemented additive surrogate method is not covered by the theory, the elliptic H^1 setting is explicitly outside the theoretical framework, and the numerical evidence for the full DDM consists of single runs without baselines or error bars.

major comments (3)
  1. [§2.5, Eqs. (16)–(18), Algorithm 1] The implemented method is the additive Schwarz iteration with approximate neural local solvers, but no convergence or error-propagation analysis is provided for this iteration. Theorem 1 in §2.3 (Eq. (12)) covers the alternating Schwarz operator with exact local solution operators acting on C^s(Ω), and §2.5 explicitly states that analogous results for the additive method are not pursued. The central claim in §4.2 that the final accuracy is "primarily limited by the underlying neural operator" therefore lacks support: there is no contraction estimate for the exact additive operator (which depends on overlap, number of subdomains, and coefficient contrast) and no perturbation bound relating the local surrogate error to the global fixed-point error. This is the load-bearing gap between the theory and the numerical method. Please add either a perturbation analysis for the additive surrogate iteration or a careful empirical study of error propagation over many trials with per-iteration true errors; absent that, the stated interpretation of the reported accuracies is not justified.
  2. [§2.4, Eq. (13)] The theoretical framework does not apply to the elliptic PDE setting that is actually solved numerically. Equation (13) is posed with a ∈ L^∞(Ω; R^{d×d}_{sym,≻0}) and u ∈ H^1(Ω), while Theorem 1 requires U = C^s(Ω̄) and the alternating Schwarz property in that space. The manuscript concedes in §2.4 that for d > 1 the Sobolev embedding H^1 ⊂ C(Ω̄) fails, that Hölder continuity of a would be needed for continuity of u, and that establishing the alternating Schwarz property on C(Ω̄) is "outside the scope of the current work." Since the numerical examples are two-dimensional with discontinuous piecewise-constant coefficients, the solutions are not in C(Ω̄) in general. Consequently, Theorem 1 does not provide a rigorous foundation for the reported experiments. This is a correctness-risk concern rather than a circularity concern; I recommend either extending the theory to H^1 (or a suitable regularity class) or explicitly presenting the DDM results as empirical heuristics that are not covered by the theorem.
  3. [§4.2, Figures 8, 10, 12] The full-domain experiments lack baselines, repeated trials, and error bars. Each of the three main test cases (square, L-shaped, I-shaped) is demonstrated with a single boundary-condition realization and a single microstructure draw, and the text reports only one final relative L2 error per field. Table 1 compares PPNO against FNO, ViT, CNN, and FCNN on the standalone square benchmark, but no comparison is made for the full L-DDM against classical additive Schwarz with exact local solvers, against other ML-enhanced DDM methods, or against the PPNO applied directly on the whole domain. The abstract's claim that the approach "outperforms current state-of-the-art methods on these challenging problems" is therefore not supported by the DDM experiments. Please report means and standard deviations over at least several boundary conditions and microstructure realizations per geometry, and include at least one classical DDM baseline (e.g., overlapping additive Schwarz with exact local FEM solves) reporting both accuracy and wall-clock time.
minor comments (7)
  1. [§2.2, Definition 2 and Lemma 1] The text switches between "alternating Schwarz property" and "uniform additive Schwarz property"; Definition 2 defines the uniform alternating Schwarz property, and Lemma 1 concludes the same, so the occurrences of "additive" are typos that should be corrected.
  2. [§4.2, Figure captions] Several figure captions reference panels that do not exist in the figure: Figure 8 lists "(g)" while the figure shows (a)–(f), and Figures 10 and 12 similarly list "(g)" for the gradient error plot. Please renumber the panels consistently.
  3. [Appendix C] Two different figures are both numbered "Figure C.13"; they should be renumbered sequentially.
  4. [§4.1, Eq. (45)] Equation (45) defines "Relative L2 Error" using squared L2 norms in the numerator and denominator; the standard relative L2 error uses the unsquared norms. If the squared definition is intended, please call it a relative squared-L2 error or clarify the convention.
  5. [§3.5, Eq. (39)] The boundary-kernel exponent -2 in Eq. (39) is presented as a fixed choice, but no sensitivity study is given for this exponent; a brief note on why inverse-square weighting was chosen would help.
  6. [Throughout] There are several typos and minor language issues, including "remainder fo this paper" in Section 1, "Lipchitz" in Appendix A, "nueral" in Section 4, and "the the" in Section 4.1. A careful proofread is needed.
  7. [Table 1] The text states that the benchmark table includes inference time, but Table 1 has no inference-time column; either add the column or remove the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical transfer argument and the PPNO-in-Schwarz results are validated against independent FEM ground truth, and the cited universal approximation theorem is external and parameter-free.

full rationale

The derivation chain is not circular. The central theoretical result (Theorem 1) is an approximation/transfer theorem: it assumes the exact alternating Schwarz operator G converges uniformly on a compact set K (Definition 2, Eq. (6)) and that the exact local solution operators are Lipschitz (Assumption 1), then invokes the universal approximation theorem for neural operators [12, 43] to obtain a surrogate operator Ghat whose n-fold composition is epsilon-close to u (Eq. (12)). Convergence of the exact Schwarz method is an input assumption, not a consequence of the neural network, so the conclusion does not reduce to the assumption; the universal approximation result in [43] (Lanthaler–Li–Stuart) is an external, parameter-free theorem, and although [12] includes two of the present authors, the existence result is not being smuggled in as an unverified premise. The numerical claims are validated against independently computed finite-element ground truth (Matlab PDE toolbox), not against the operator's own training data or fitted values. The PPNO is trained on square 50-grain domains with random periodic Dirichlet data (Eq. (41)) and then tested on square, L-shaped, and I-shaped domains with in- and out-of-distribution microstructures; the reported global relative L2 errors (0.039%–0.645% for u, 3.89%–8.10% for the gradient) are computed against the FEM truth, not against quantities used to fit the model. The paper does concede that the implemented additive method lacks the theory of Theorem 1 (§2.5: 'Similar results to our own can also be derived for the additive method. We do not pursue this in the current work'), but that is a gap in support, not circularity. No fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported from prior self-citations in a load-bearing way.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests primarily on the empirical generalization of the trained operator and on an existence theorem that is proven only for an idealized alternating Schwarz setting on smooth function spaces. The free parameters are mainly experimental design choices; none are fitted to force the final DDM errors.

free parameters (4)
  • boundary kernel exponent = 2 (inverse-square distance)
    Hand-chosen in Eq. (39); controls how boundary values are diffused into the domain and is not learned or optimized.
  • boundary condition generation parameters = N=15, k=2.5, and sampling ranges in Eq. (41)
    Defines the distribution of Dirichlet data used in pretraining and testing; the claimed generalization to random BCs is relative to this specific family.
  • overlap ratio = 31.25% (10x10 square experiment)
    Hand-chosen overlap of subdomains, affects convergence and error of the additive Schwarz iteration.
  • stopping tolerance = not specified
    Algorithm 1 stops when the successive error S < epsilon_tol, but the value of epsilon_tol is never stated, making the reported iteration counts and final errors hard to interpret.
assumptions (6)
  • domain assumption The abstract Dirichlet problem (1) has a unique solution and the solution map (a,g) -> u is continuous.
    Assumed in Section 2.1 for Theorem 1; standard for the elliptic case but not for general P.
  • domain assumption The local solution operators Gj are continuous and Gj(a, .) are locally Lipschitz (Assumption 1).
    Used in Theorem 1 and its proof in Appendix A.
  • domain assumption Compact sets K admit the uniform alternating Schwarz property (Definition 2).
    Necessary for Theorem 1; the paper does not prove this for the elliptic problem in C(Omega).
  • standard math Universal approximation theorems for neural operators between spaces of continuous functions hold (as in [12,43]).
    Basis of the approximation steps in Theorem 1 and Appendix B.
  • standard math The mollification operator M_delta is uniformly bounded between C^s spaces (from [43, Lemma A.2]).
    Used in Theorem 3 to gain compactness; the authors note this does not extend to Sobolev spaces.
  • domain assumption Matlab finite element solver provides accurate reference solutions.
    The ground truth for both training and testing is generated by this solver; no convergence study with respect to FEM mesh is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Learning-based Domain Decomposition Method." pith.science (2026). https://pith.science/paper/AFBM3OU3

@misc{pith2026250717328,
  author       = {Pith},
  title        = {Pith review of: A Learning-based Domain Decomposition Method},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AFBM3OU3}},
  note         = {Machine review of arXiv:2507.17328}
}
read the original abstract

Recent developments in mechanical, aerospace, and structural engineering have driven a growing need for efficient ways to model and analyse structures at much larger and more complex scales than before. While established numerical methods like the Finite Element Method remain reliable, they often struggle with computational cost and scalability when dealing with large and geometrically intricate problems. In recent years, neural network-based methods have shown promise because of their ability to efficiently approximate nonlinear mappings. However, most existing neural approaches are still largely limited to simple domains, which makes it difficult to apply to real-world PDEs involving complex geometries. In this paper, we propose a learning-based domain decomposition method (L-DDM) that addresses this gap. Our approach uses a single, pre-trained neural operator-originally trained on simple domains-as a surrogate model within a domain decomposition scheme, allowing us to tackle large and complicated domains efficiently. We provide a general theoretical result on the existence of neural operator approximations in the context of domain decomposition solution of abstract PDEs. We then demonstrate our method by accurately approximating solutions to elliptic PDEs with discontinuous microstructures in complex geometries, using a physics-pretrained neural operator (PPNO). Our results show that this approach not only outperforms current state-of-the-art methods on these challenging problems, but also offers resolution-invariance and strong generalization to microstructural patterns unseen during training.

Figures

Figures reproduced from arXiv: 2507.17328 by the authors.

Figure 1
Figure 1. Schematic illustration of the overlapping domain decomposition. The global [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. A schematic comparison of Schwarz methods for iteration [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Schematic of the PPNO architecture. The workflow proceeds in three main [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: (a) An example of the random Voronoi microstructure. (b) An example of the [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: (a) Relative L2 test error of pre-trained PPNO model on 4 different resolutions. [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 6
Figure 6. Figure 6: (a-d) Voronoi crystal microstructure of 10 random grains, 100 grains, hexagonal [PITH_FULL_IMAGE:figures/full_fig_p025_6.png]
Figure 7
Figure 7. Figure 7: (a) The random Dirichlet boundary conditions applied to the 4 edges of the [PITH_FULL_IMAGE:figures/full_fig_p027_7.png]
Figure 8
Figure 8. Figure 8: (a-b) The truth solution u and learning-based solution ˆu field for the 10 × 10 rectangular domain, with boundary conditions prescribed in Fig.7c applied. (c) The iterative (black line) and successive (red line) error versus the number of iterations of the uˆ field. (d…
Figure 9
Figure 9. Figure 9: (a) The random Dirichlet boundary conditions applied to the 6 edges of the [PITH_FULL_IMAGE:figures/full_fig_p028_9.png]
Figure 10
Figure 10. Figure 10: (a-b) The truth solution u and learning-based solution ˆu field for the L-shaped domain, with boundary conditions prescribed in Fig.9c applied. (c) The iterative (black line) and successive (red line) error versus the number of iterations of the ˆu field. (e-f) The tr…
Figure 11
Figure 11. Figure 11: (a) The Dirichlet boundary conditions applied to the 12 edges of the I-shaped [PITH_FULL_IMAGE:figures/full_fig_p030_11.png]
Figure 12
Figure 12. Figure 12: (a-b) The truth solution u and learning-based solution ˆu field for the I-shaped domain, with boundary conditions prescribed in Fig.11c applied. (c) The iterative (black line) and successive (red line) error versus the number of iterations of the ˆu field. (d-e) The t…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 36 canonical work pages

  1. [1]

    Bullo, A

    F. Bullo, A. D. Lewis, Geometric control of mechanical systems: model- ing, analysis, and design for simple mechanical control systems, Vol. 49, Springer, 2019

    work page 2019

  2. [2]

    Leng, Z.-g

    J.-x. Leng, Z.-g. Wang, W. Huang, Y. Shen, K. An, Multidisci- plinary design optimization processes for efficiency improvement of air- craft: State-of-the-art review, International Journal of Aeronautical and Space Sciences (2024) 1–23

    work page 2024

  3. [3]

    O. A. Bauchau, J. I. Craig, Structural analysis: with applications to aerospace structures, Vol. 163, Springer Science & Business Media, 2009

    work page 2009

  4. [4]

    M. Umer, P. Olejnik, Approximate analytical approaches to nonlinear differential equations: A review of perturbation, decomposition and coefficient methods in engineering, Archives of Computational Methods in Engineering (2025) 1–32

    work page 2025

  5. [5]

    Stepinski, T

    T. Stepinski, T. Uhl, W. Staszewski, Advanced structural damage de- tection: from theory to engineering applications (2013). 32

    work page 2013

  6. [6]

    Gilbarg, N

    D. Gilbarg, N. S. Trudinger, D. Gilbarg, N. Trudinger, Elliptic partial differential equations of second order, Vol. 224, Springer, 1977

    work page 1977

  7. [7]

    O. C. Zienkiewicz, R. L. Taylor, The finite element method for solid and structural mechanics, Elsevier, 2005

    work page 2005

  8. [8]

    K. H. Huebner, D. L. Dewhirst, D. E. Smith, T. G. Byrom, The finite element method for engineers, John Wiley & Sons, 2001

    work page 2001

Show all 57 references
  1. [9]

    Raissi, P

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational physics 378 (2019) 686–707

    2019

  2. [10]

    Yu, et al., The deep ritz method: a deep learning-based numerical algorithm for solving variational problems, Communications in Mathe- matics and Statistics 6 (1) (2018) 1–12

    B. Yu, et al., The deep ritz method: a deep learning-based numerical algorithm for solving variational problems, Communications in Mathe- matics and Statistics 6 (1) (2018) 1–12

    2018

  3. [11]

    Kharazmi, Z

    E. Kharazmi, Z. Zhang, G. E. Karniadakis, Variational physics- informed neural networks for solving partial differential equations, arXiv preprint arXiv:1912.00873 (2019)

    2019 arXiv

  4. [12]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, A. Anandkumar, Neural operator: Learning maps between function spaces with applications to pdes, Journal of Machine Learning Research 24 (89) (2023) 1–97

    2023

  5. [13]

    L. Lu, P. Jin, G. E. Karniadakis, Deeponet: Learning nonlinear opera- tors for identifying differential equations based on the universal approx- imation theorem of operators, arXiv preprint arXiv:1910.03193 (2019)

    2019 arXiv

  6. [14]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Fourier neural operator for parametric par- tial differential equations, arXiv preprint arXiv:2010.08895 (2020)

    2020 arXiv

  7. [15]

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

    2020 arXiv

  8. [16]

    Z. Li, H. Zheng, N. Kovachki, D. Jin, H. Chen, B. Liu, K. Azizzade- nesheli, A. Anandkumar, Physics-informed neural operator for learning partial differential equations, ACM/IMS Journal of Data Science 1 (3) (2024) 1–27. 33

    2024

  9. [17]

    Z. Li, N. Kovachki, C. Choy, B. Li, J. Kossaifi, S. Otta, M. A. Nabian, M. Stadler, C. Hundt, K. Azizzadenesheli, et al., Geometry-informed neural operator for large-scale 3d pdes, Advances in Neural Information Processing Systems 36 (2023) 35836–35854

    2023

  10. [18]

    Calvello, N

    E. Calvello, N. B. Kovachki, M. E. Levine, A. M. Stuart, Continuum attention for neural operators, arXiv preprint arXiv:2406.06486 (2024)

    2024

  11. [19]

    N. Liu, Y. Yu, Neural interpretable pdes: Harmonizing fourier insights with attention for scalable and interpretable physics discovery, arXiv preprint arXiv:2505.23106 (2025)

    2025 arXiv

  12. [20]

    Goswami, A

    S. Goswami, A. Bora, Y. Yu, G. E. Karniadakis, Physics-informed deep neural operator networks, in: Machine learning in modeling and simu- lation: methods and applications, Springer, 2023, pp. 219–254

    2023

  13. [21]

    Z. Li, D. Z. Huang, B. Liu, A. Anandkumar, Fourier neural operator with learned deformations for pdes on general geometries, Journal of Machine Learning Research 24 (388) (2023) 1–26

    2023

  14. [22]

    H. A. Schwarz, Ueber einen Grenz¨ ubergang durch alternirendes Ver- fahren, Z¨ urcher u. Furrer Berlin, 1870

  15. [23]

    Lions, et al., On the schwarz alternating method

    P.-L. Lions, et al., On the schwarz alternating method. i, in: First international symposium on domain decomposition methods for partial differential equations, Vol. 1, Paris, France, 1988, p. 42

    1988

  16. [24]

    B. F. Smith, Domain decomposition methods for partial differential equations, in: Parallel Numerical Algorithms, Springer, 1997, pp. 225– 243

    1997

  17. [25]

    Toselli, O

    A. Toselli, O. Widlund, Domain decomposition methods-algorithms and theory, Vol. 34, Springer Science & Business Media, 2004

    2004

  18. [26]

    C. I. of Mathematical Sciences. Ultracomputer Research Laboratory, M. Dryja, O. Widlund, An Additive Variant of the Schwarz Alternating Method for the Case of Many Subregions, Computer science technical report, Ultracomputer Research Laboratory, Univ., Courant Inst. of Mathema...

    1987

  19. [27]

    Klawonn, M

    A. Klawonn, M. Lanser, J. Weber, Machine learning and domain de- composition methods-a survey, Computational Science and Engineering 1 (1) (2024) 2. 34

    2024

  20. [28]

    Heinlein, A

    A. Heinlein, A. Klawonn, M. Lanser, J. Weber, Combining machine learning and domain decomposition methods for the solution of par- tial differential equations—a review, GAMM-Mitteilungen 44 (1) (2021) e202100001

    2021

  21. [29]

    A. D. Jagtap, E. Kharazmi, G. E. Karniadakis, Conservative physics- informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems, Computer Methods in Applied Mechanics and Engineering 365 (2020) 113028

    2020

  22. [30]

    A. D. Jagtap, G. E. Karniadakis, Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equa- tions, Communications in Computational Physics 28 (5) (2020)

    2020

  23. [31]

    Stiller, F

    P. Stiller, F. Bethke, M. B¨ ohme, R. Pausch, S. Torge, A. Debus, J. Vorberger, M. Bussmann, N. Hoffmann, Large-scale neural solvers for partial differential equations, in: Driving Scientific and Engineer- ing Discoveries Through the Convergence of HPC, Big Data and AI: 17th S...

    2020

  24. [32]

    Bischof, M

    R. Bischof, M. A. Kraus, Mixture-of-experts-ensemble meta-learning for physics-informed neural networks, in: Proceedings of 33. Forum Bauinformatik, 2022

    2022

  25. [33]

    Z. Hu, A. D. Jagtap, G. E. Karniadakis, K. Kawaguchi, Augmented physics-informed neural networks (apinns): A gating network-based soft domain decomposition methodology, Engineering Applications of Artificial Intelligence 126 (2023) 107183

    2023

  26. [34]

    Dwivedi, N

    V. Dwivedi, N. Parashar, B. Srinivasan, Distributed physics informed neural network for data-efficient solution to partial differential equa- tions, arXiv preprint arXiv:1907.08967 (2019)

    2019 arXiv

  27. [35]

    Basir, I

    S. Basir, I. Senocak, A generalized schwarz-type non-overlapping do- main decomposition method using physics-constrained neural networks, arXiv preprint arXiv:2307.12435 (2023)

    2023 arXiv

  28. [36]

    Moseley, A

    B. Moseley, A. Markham, T. Nissen-Meyer, Finite basis physics- informed neural networks (fbpinns): a scalable domain decomposition 35 approach for solving differential equations, Advances in Computational Mathematics 49 (4) (2023) 62

    2023

  29. [37]

    Dolean, A

    V. Dolean, A. Heinlein, S. Mishra, B. Moseley, Multilevel domain decomposition-based architectures for physics-informed neural net- works, Computer Methods in Applied Mechanics and Engineering 429 (2024) 117116

    2024

  30. [38]

    K. Li, K. Tang, T. Wu, Q. Liao, D3m: A deep domain decomposition method for partial differential equations, Ieee Access 8 (2019) 5283– 5294

    2019

  31. [39]

    W. Li, X. Xiang, Y. Xu, Deep domain decomposition method: Elliptic problems, in: Mathematical and Scientific Machine Learning, PMLR, 2020, pp. 269–286

    2020

  32. [40]

    Adams, J

    R. Adams, J. Fournier, Sobolev Spaces, Pure and Applied Mathematics, Academic Press, 2003

    2003

  33. [41]

    Pathak, S

    J. Pathak, S. Subramanian, P. Harrington, S. Raja, A. Chattopad- hyay, M. Mardani, T. Kurth, D. Hall, Z. Li, K. Azizzadenesheli, et al., Fourcastnet: A global data-driven high-resolution weather model us- ing adaptive fourier neural operators, arXiv preprint arXiv:2202.11214 (2022)

    2022 arXiv

  34. [42]

    R. Lam, A. Sanchez-Gonzalez, M. Willson, P. Wirnsberger, M. Fortu- nato, F. Alet, S. Ravuri, T. Ewalds, Z. Eaton-Rosen, W. Hu, et al., Learning skillful medium-range global weather forecasting, Science 382 (6677) (2023) 1416–1421

    2023

  35. [43]

    Lanthaler, Z

    S. Lanthaler, Z. Li, A. M. Stuart, Nonlocality and nonlinearity im- plies universality in operator learning, arXiv preprint arXiv:2304.13221 (2023)

    2023 arXiv

  36. [44]

    Gilbarg, N

    D. Gilbarg, N. Trudinger, Elliptic Partial Differential Equations of Sec- ond Order, Classics in Mathematics, Springer Berlin Heidelberg, 2001

    2001

  37. [45]

    P. L. Lions, On the schwarz alternating method. i, in: First Inter- national Symposium on Domain Decomposition Methods for Partial Differential Equations, 1988, pp. 1–42

    1988

  38. [46]

    P. L. Lions, On the Schwarz Alternating Method. II, Stochastic In- terpretation and Order Proprieties, T. F. Chan, et al., Eds., Domain Decomposition Methods, 1989, Ch. 4. 36

    1989

  39. [47]

    P. L. Lions, On the schwarz alternating method. iii, a variant for nonoverlapping subdomains, in: Third International Symposium on Do- main Decomposition Methods for Partial Differential Equations, 1980, pp. 202–223

    1980

  40. [48]

    Liu-Schiaffini, J

    M. Liu-Schiaffini, J. Berner, B. Bonev, T. Kurth, K. Azizzadenesheli, A. Anandkumar, Neural operators with localized integral and differen- tial kernels, arXiv preprint arXiv:2402.16845 (2024)

    2024 arXiv

  41. [49]

    Ronneberger, P

    O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, Spri...

    2015

  42. [50]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT Press, 2016, http://www.deeplearningbook.org

    2016

  43. [51]

    Paszke, Pytorch: An imperative style, high-performance deep learn- ing library, arXiv preprint arXiv:1912.01703 (2019)

    A. Paszke, Pytorch: An imperative style, high-performance deep learn- ing library, arXiv preprint arXiv:1912.01703 (2019)

    2019 arXiv

  44. [52]

    DeVore, G

    R. DeVore, G. Lorentz, Constructive Approximation, Grundlehren der mathematischen Wissenschaften, Springer-Verlag, 1993

    1993

  45. [53]

    T. M. Inc., Partial differential equation toolbox, version: 23.2 (r2023b) (2023). URL https://www.mathworks.com

    2023

  46. [54]

    D. E. Rumelhart, G. E. Hinton, R. J. Williams, Learning representa- tions by back-propagating errors, nature 323 (6088) (1986) 533–536

    1986

  47. [55]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, Advances in neural information processing systems 25 (2012)

    2012

  48. [56]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

    2020 arXiv

  49. [57]

    Bhattacharya, N

    K. Bhattacharya, N. B. Kovachki, A. Rajan, A. M. Stuart, M. Traut- ner, Learning homogenization for elliptic operators, SIAM Journal on Numerical Analysis 62 (4) (2024) 1844–1873. 37 Appendix A. Approximation of the Schwarz Method Proof of Lemma 1. Let ˜L > 0 denote the Lipsch...

    2024

Pith tools

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