Pith. sign in

REVIEW 4 major objections 5 minor 30 references

Splitting the input space into subdomains makes neural-network surrogates more accurate and faster to train when interface continuity is enforced with augmented Lagrange multipliers.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 17:16 UTC pith:TDIH7SEZ

load-bearing objection A useful small-scale demonstration of Lagrange-multiplier domain decomposition for data-driven NN surrogates, with the scalability claim out ahead of the evidence. the 4 major comments →

arxiv 2603.26396 v1 pith:TDIH7SEZ submitted 2026-03-27 cs.CE

Domain decomposition of large neural network surrogate models

Timm G\"odde , Eisso H. Atzema , Bojana Rosi\'c This is my paper
classification cs.CE MSC 65N5568T07
keywords domain decompositionneural network surrogateaugmented Lagrange multipliersinterface continuityL-BFGSfinite elementuncertainty propagationsurrogate modeling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to show that a single large neural network is not the best way to build surrogates for fields with local nonlinearities: splitting the input space into non-overlapping subdomains, training a small network on each, and forcing the networks to agree on interfaces recovers or beats global-network accuracy while reducing training time and parameters. Two enforcement schemes are compared: classic Lagrange multipliers with iteratively linearized constraints (LMA) and augmented Lagrange multipliers with a quadratic penalty (ALMA). On 2D and 3D elastic compression problems the interface constraints cut discontinuities dramatically, DDM approximations in nonlinear regions are more accurate than a global NN, and ALMA converges faster than LMA at a small accuracy cost. The paper concludes that ALMA is the better choice for large-scale engineering surrogates, with three vertically stacked subdomains in the 2D example giving the best accuracy-to-time balance, and a two-subdomain split reproducing 3D displacement statistics within a few percent.

Core claim

On its own terms, the paper establishes that non-overlapping domain decomposition is a viable and advantageous training strategy for data-driven neural network surrogates. By assigning each subdomain its own small network and an interface network, and by enforcing C0 and C1 continuity through Lagrange multiplier terms in the local loss, both proposed algorithms (LMA and ALMA) produce approximations with substantially reduced interface discontinuities compared to training without constraints. In the 2D compression benchmark the best DDM results halve the maximum relative error of a single-domain network while also reducing errors near the boundary where the global network struggles. ALMA is t

What carries the argument

The central object is a substructured (non-overlapping) domain decomposition of the input feature space: the domain is split into disjoint subdomains, each handled by a small feedforward network, with an extra network on every interface used to fix the shared values and normal derivatives. The C0/C1 interface conditions enter as constraints in each local loss. LMA enforces them with Lagrange multipliers after linearizing the network predictions around the current weights, requiring feasibility checks; ALMA replaces this with an augmented Lagrange term (λᵀQ + ρ‖Q‖²) that needs no linearization and uses the penalty ρ as multiplier step size. Local networks are trained with a quasi-Newton optim

Load-bearing premise

The claim that the augmented Lagrange method is the preferred choice for large problems rests on wall-clock measurements taken on a 32-core cluster with at most 10 subdomains in 2D and 2 subdomains in 3D, with the slowest LMA runs and least accurate ALMA runs left out of the headline comparisons.

What would settle it

Run ALMA on a large problem—hundreds of subdomains or high-dimensional stochastic inputs—and measure wall-clock time and interface error: if per-subdomain cost grows faster than linearly or interface errors exceed the tolerances used here, the scalability claim fails. Alternatively, find a problem where LMA with a feasible linearization converges in fewer outer iterations than ALMA with a tuned penalty, contradicting the claim that ALMA always converges faster.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, surrogate models for design optimization can be built from ensembles of small networks rather than one large one, with training parallelized over subdomains.
  • Interface continuity enforcement is what makes the decomposition usable: unconstrained local networks leave visible seams, and both Lagrange formulations remove them in the tested cases.
  • The augmented Lagrange variant offers a practical default: it needs no constraint linearization and its wall-clock time grows roughly linearly with subdomain count for vertical splits, while the classic Lagrange variant does not.
  • For stochastic problems, the method splits only the spatial domain while keeping the random parameter space global, which the 3D cylinder test shows is enough to reproduce mean and variance of displacement fields.
  • A practical recipe emerges: three vertically stacked subdomains with ALMA gave about 4% maximum relative error in about 10 hours on the 2D benchmark, compared to 3.2 minutes for a single network but with roughly double the error.

Where Pith is reading between the lines

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

  • A fairer head-to-head would compare a global network with the same total parameter budget as all local and interface networks combined; the paper's global baseline is a single 80×2 network, so part of the accuracy gain may come from parameter distribution rather than decomposition per se.
  • For genuinely large runs, the number of Lagrange multipliers scales with interface points times stochastic samples (the paper notes this cost); the method's practical ceiling likely depends on compressing interface representations, e.g., by modeling interface gaps as random fields—an extension that would also replace the stochastic, run-dependent gap solutions with a single well-defined approximat
  • The robustness to missing data near interfaces suggests the same ALMA machinery could be applied to experimental data, such as digital image correlation fields with occluded zones, where the interface network bridges unsensed regions; the paper notes the application area but does not test real measurements.
  • The scalability conclusion would be strengthened by a test on more than 10 subdomains or a higher-dimensional stochastic input; as reported, the fastest ALMA configuration used only 2–3 subdomains, so 'large-scale' preference is extrapolated.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes two domain-decomposition strategies for training data-driven neural-network surrogate models in solid mechanics. The spatial domain is partitioned into non-overlapping subdomains, each served by a local feedforward NN, and interface continuity (C0 and C1) is enforced through either a Lagrange-multiplier formulation with iteratively linearized constraints (LMA) or an augmented-Lagrange formulation with a penalty term (ALMA). Both methods are applied to synthetic FEM data from a 2D linear-elastic compression problem and a 3D cylinder with stochastic material parameters. The reported results indicate that the constrained local models reduce interface discontinuities and improve accuracy in high-gradient regions compared with an unconstrained global NN, and that ALMA converges faster than LMA in the tested configurations. The abstract and conclusion further claim that ALMA's scalability makes it the preferred choice for large-scale problems.

Significance. If the central claims are sustained, the paper would be a useful contribution to the literature on domain decomposition for neural-network surrogate models, extending techniques previously developed for PINNs to purely data-driven regression. The paper contains several strengths: clear algorithmic descriptions (Algs. 1-4), a systematic sweep over the number of subdomains, a sensitivity study of ALMA to interface gaps and training-point density, and a 3D uncertainty-propagation demonstration. The convergence curves in Figs. 2 and 11 and the error maps in Figs. 7, 12 and 13 provide qualitative evidence that the interface constraints behave as intended. However, the evidence base is narrower than the claims: the 2D accuracy numbers are computed on training data, the headline comparisons exclude several poor-performing runs, a central optimizer statement is contradictory, and the scalability claim is explicitly deferred to future work in the Conclusion. These gaps concern load-bearing parts of the paper's contribution, so the manuscript needs substantial revision before the advertised claims are supported.

major comments (4)
  1. [Abstract; §4.1, Figs. 3 and 6; §5] The claim that ALMA's scalability makes it 'the preferred choice for large-scale problems' is not supported by the experiments. All runs are on a single 32-core node with at most M=10 subdomains in 2D and M=2 in 3D. Fig. 3b shows total wall-clock time growing roughly linearly with M, not sublinearly; no strong- or weak-scaling study is provided. Multi-directional LMA runs did not finish within 168 h and are omitted from Figs. 3a/6, while multi-directional ALMA runs are omitted from Fig. 6 because their errors exceeded 30%. The Conclusion itself states that larger problems are needed 'to further assess its scalability' and that 'larger problems would highlight also the method's computational advantages.' Because the scalability preference is a key differentiator of ALMA, this extrapolation is load-bearing. Please either add scaling experiments (e.g., weak scaling with fixed work per subdo
  2. [§4.1, Eq. (45), Figs. 6-7] The accuracy assessment uses the modified relative error e_rel = |u_hat - u|/(|u|+1). Since the outputs are normalized to [-1,1], the denominator ranges only from 1 to 2, so the reported 'maximum relative error' is not a standard relative error and can differ from the conventional definition by up to a factor of 2, especially where u is small. More importantly, the 2D errors are computed from the same FEM dataset used for training; no held-out test split is described for the 2D problem. The 3D section does mention test samples, but the 2D accuracy claims in Figs. 6 and 7 are therefore vulnerable to overfitting. Please recompute the 2D accuracy on an independent test set, report the standard relative error with a small floor in the denominator, and include the excluded multi-directional ALMA configurations (or state explicitly that the headline comparison covers only converged, low-error
  3. [§4.1, Fig. 2; §3.1; Table 2] There is a direct contradiction about the optimizer used for LMA. The text in §4.1 states that 'LMA utilizes the Nadam optimizer' [30], while §3.1, Algorithms 1 and 3, and Table 2 specify L-BFGS for the primal updates. This is not a cosmetic inconsistency: the convergence behavior in Fig. 2, the interpretation of 'LBFGS restarts' in §4.2, and the reproducibility of the method all depend on which optimizer is actually used. Please correct the text/algorithms so that the stated optimizer matches the implementation, and report the hyperparameters actually used for LMA.
  4. [Algorithms 1-4; §4.1] The algorithmic description is not sufficiently specified to reproduce the reported timings and convergence. The main loop uses tolerances epsilon_pr, epsilon_lambda, epsilon_ij, epsilon_l, and the inner loops use D_m and P_m, but no numerical values are given for any of these quantities. Fig. 3 is described as 'total computational time over 100 training iterations,' yet Algorithms 1 and 2 terminate by convergence checks, not by a fixed 100-iteration count. It is unclear how the 100 iterations relate to the outer loop k, the dual loop d, and the primal loop p. In addition, Algorithm 2 line 7 sets D_m = 0 inside the loop when lambda = 0, which appears to disable the dual loop under that condition and needs clarification; Algorithm 3 lists rho as a parameter but never uses it, while Algorithm 4 lists epsilon_l but never uses it. Please provide the hyperparameter values and correct the pseu
minor comments (5)
  1. [Throughout] There are numerous typographical and formatting issues, e.g., 'G¨ odde', 'T able 1', 'B. Rosi´ c', 'domain decomposition methods (DDM)' used inconsistently, and 'the third update' versus 'the fourth dual update' in §4.2. A careful proofread is needed.
  2. [§2, Eqs. (10)-(11)] The notation for the symmetric gradient is inconsistent: Eq. (10) writes ∇s u(ω) while Eq. (11) defines a mapping ∇s : u1(x)u2(ω) -> (∇s u1(x))u2(ω). Please define the symbol once and use it consistently.
  3. [§3, Eq. (21)] The index in the local objective is written 'i,...,M'; it should be 'i=1,...,M'.
  4. [§4.1, Fig. 2 caption] The caption says 'iteration of the loop outside the given loops,' which is confusing. Please specify which outer iteration is shown and why that iteration was chosen.
  5. [§4.2, Table 4 and Fig. 11] The text states that 'the fourth dual update is used to show θ_i convergence,' while Fig. 11b caption says 'dual update ALMA iteration 2.' Please align the text and captions.

Circularity Check

0 steps flagged

No circular derivation: central claims are empirical measurements on FEM-generated data, not fitted inputs relabeled as predictions.

full rationale

The paper's derivation chain (Sections 2–3) develops a stochastic weak form and then proposes two algorithms, LMA and ALMA, with interface constraints. The only self-citation is [21] (Rosić's PhD thesis), used for the standard stochastic weak-form/tensor-product setup in Eqs. (7)–(12); this is background material and is not load-bearing for the paper's DDM contribution. The headline claims—improved interface continuity, better accuracy in nonlinear regions, and faster ALMA convergence—are supported by measured experiments in Section 4 (Figs. 3, 6, 7) and are not derived from the authors' own equations or from a fitted parameter renamed as a prediction. The paper explicitly excludes some runs (LMA multi-directional splits exceeding 168 h; ALMA multi-directional splits with >30% error), and the Conclusion states: 'In the future, ALMA will be applied to larger-scale problems to further assess its scalability... larger problems would highlight also the method's computational advantages.' This undercuts the abstract's scalability claim, but that is an evidence/overclaim concern, not circularity: no quantity is defined in terms of the conclusion it is used to support. Under the required standard of exhibiting a specific reduction Eq. X = Eq. Y by construction, no circular step can be identified.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The method is mostly standard optimization machinery (Lagrangian, augmented Lagrangian, L-BFGS). The novel part—interface neural networks and the alternating training schedule—introduces several hand-chosen hyperparameters and an auxiliary network with no independent validation. The comparison also depends entirely on synthetic FEM data generated under linear-elastic assumptions.

free parameters (5)
  • penalty parameter rho (ALMA) = 1e-3 for 2D ALMA; not reported for 3D
    Controls the augmented-Lagrange constraint enforcement in Eq. (40). The text calls it tunable, but no tuning or ablation study is provided, and 3D does not report a value.
  • dual update learning rate alpha = not reported
    Used in Eq. (34) and Algorithm 2 to update Lagrange multipliers. Its value affects convergence but is never specified; Algorithm 2 line 12 uses alpha while ALMA text and Eq. (42) use rho.
  • convergence tolerances eps_pr, eps_lambda, eps_ij, eps_l = not reported
    Terminate primal, dual, interface, and linearization loops in Algorithms 1-4. No values or units are given, making the stopping criteria irreproducible.
  • local NN architecture and activation = 40x2 swish (2D); 80x4 swish (3D)
    Chosen by hand in Tables 2 and 4; no architecture sensitivity study is reported, though accuracy and runtime likely depend strongly on these choices.
  • interface collocation points = 10 per interface (2D); 73 spatial points x 1000 samples (3D)
    Interface accuracy and training cost depend on this discretization, but no convergence study over the number of interface points is provided.
axioms (4)
  • standard math Lax-Milgram gives a unique weak solution of the linear-elastic boundary value problem.
    Invoked in Section 2 to justify well-posedness of the FEM data used as the surrogate target.
  • domain assumption Strong duality / Slater condition holds for the linearized Lagrangian in LMA.
    Section 3.1 states that dual ascent requires Slater's condition, but no proof is given that the strongly non-convex, linearized NN constraint problem satisfies it.
  • ad hoc to paper The Taylor linearization of NN interface constraints is accurate over each L-BFGS step.
    Equations (35)-(36) linearize the constraints around the current parameters; convergence is asserted to follow if updates are small, but this is not guaranteed and no safeguard proof is supplied.
  • domain assumption Non-overlapping partition with C0/C1 continuity on interfaces is sufficient for surrogate accuracy.
    Section 3 assumes interface continuity (Eqs. 22-23) is the right condition to recover global accuracy; no theoretical bound connecting interface residual to global approximation error is given.
invented entities (1)
  • Interface neural networks u_hat_{theta_ij} no independent evidence
    purpose: Represent the trace of the solution on each interface and act as reference predictions for both adjacent local networks during constraint enforcement.
    Algorithm 1 introduces separate interface NNs trained by minimizing Eq. (28). Their accuracy is checked only through final error maps, not against an independent benchmark, so the entity has no falsifiable handle outside the paper's own pipeline.

pith-pipeline@v1.3.0-alltime-deepseek · 19377 in / 15264 out tokens · 151260 ms · 2026-08-02T17:16:02.361494+00:00 · methodology

0 comments
read the original abstract

Neural networks (NNs) have gained significant attention across various engineering disciplines, particularly in design optimization, where they are used to build surrogate models for high-dimensional regression problems. Despite their power as global approximators, NNs often fail to accurately capture local nonlinearities without relying on a large number of training parameters. To address these limitations, in this paper we propose domain decomposition methods (DDM), which divide the input feature space into multiple local subdomains, each modeled by a simpler NN, trained in parallel. To recover the accuracy of a global approximation, interface constraints are introduced in the local loss functions to enforce continuity between subdomains. The interface constraints are enforced with two different approaches, by utilizing Lagrange multiplier or augmented Lagrange multiplier methods. Both approaches are validated using synthetic data from 2D and 3D linear compression problems, numerically solved using the finite element method. The study investigates computational time and accuracy across varying numbers of subdomains to identify optimal partitioning strategies. Compared to unconstrained approximations, both methods significantly improve continuity across subdomain interfaces. Also, the use of DDMs improves approximation accuracy in nonlinear regions when compared to standard global NN training. The augmented Lagrange method outperforms the standard Lagrange formulation by converging faster due to lower convergence requirements, albeit with a slightly lower accuracy. Its scalability makes it the preferred choice for large-scale problems, as the faster convergence outweighs the minor loss in accuracy. Overall, these results highlight the augmented Lagrange method as a promising DDM approach for training efficient and scalable NN surrogate models.

Figures

Figures reproduced from arXiv: 2603.26396 by Bojana Rosi\'c, Eisso H. Atzema, Timm G\"odde.

Figure 1
Figure 1. Figure 1: a) Schematic of the 2D specimen clamped at the bottom and compressed by a predefined displacement at the top. b) Interpolation plot of the normalized horizontal displacement field of the specimen. c) Schematic view of possible split sampling domains (3) from top to bottom, in blue, red and yellow, with in green and black possible interface points. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Shows the convergence of the θi and dual updates of the ALMA and LMA methods. The subfigure captions indicate the iteration of the loop outside the given loops. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Full convergence times for different numbers of domains for the LMA and ALMA. The circles indicate domains split only horizontally, while the triangles are denoting runs for domain splits in both directions. For example for four domains, the circle is a run split into four by one domains and the triangle symbol run is two by two domains. The triangle symbol runs are left out in a) since they did not finish… view at source ↗
Figure 4
Figure 4. Figure 4: Schematic of a (3,3) decomposition of a rectangular domain. the interface happens to align with the transition between linear regime and boundaries. Hence, the lowest NN only has to approximate the boundary part, while the others approximate the linear parts. Meaning that it is beneficial to train the boundary separately from the rest of the domain. When comparing the total simulation times of the performe… view at source ↗
Figure 5
Figure 5. Figure 5: Convergence times of a single iteration for different numbers of domains for the LMA and ALMA. The circles indicate domains split only horizontally, while the triangles are denoting runs for domain splits in both directions. For example for four domains, the circle is a run split into four by one domains and the triangle symbol run is two by two domains. splits, this growth cannot be seen in the LMA with c… view at source ↗
Figure 6
Figure 6. Figure 6: Final relative errors for different numbers of domains for the LMA and ALMA. The circles indicate one directional splits, while the triangles denote multi-directional splits. Multi-directional splits of the ALMA are left out since they have higher relative errors. The black x denotes a reference solution for a single NN trained on the whole domain. 15 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FEM solution and approximations of the solution split into three subdomains. One can view the FEM solution in a), the ALMA and LMA approximations in b) and c), the relative error of each of these methods in e) and f), and the relative error without continuity constraints enforced in d). 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Relative errors for ALMA approximations with different gap sizes between three vertical domains in which the interface constraints are enforced halfway across the gaps. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Relative errors for three vertically stacked domains for ALMA. The number of points per local domain decrease from a) to d) with the exact number of points per dom as denoted above. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: a) Schematic of the 3D Cylinder clamped at the bottom and compressed at the top. b) Resultant displacement field of the FEM simulation for one random set of material parameters. c) Data points on the cylinder split into two domains with the upper domain in blue, the lower domain in yellow and the interface in black. to a larger relative error in a larger area near the boundary when compared to using a lar… view at source ↗
Figure 11
Figure 11. Figure 11: Shows the convergence of the θi and dual updates of ALMA for the three-dimensional problem. The subfigure captions indicate the iteration of the loop outside the given loops. The accuracy of the approximation is assessed by comparing the model and data statistics by computing the mean and standard deviation across the entire domain. The mean and standard deviations are: E(u(x)) ≈ 1 Nu X Nu i=1 ui(x), and … view at source ↗
Figure 12
Figure 12. Figure 12: Interpolation plots of the relative errors in mean and standard deviation of the displacement components u approximated by the NNs compared to data on a section through the cylinder. Values are normalized to allow optical comparison to the two-dimensional problem. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Interpolation plots of the relative errors in mean and standard deviation of the displacement components u approximated by the NNs compared to data on the interface. Values are normalized to allow optical comparison to the two-dimensional problem. the interface at 0, which is shown in more detail in [PITH_FULL_IMAGE:figures/full_fig_p022_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Kernel density estimator distributions of an interface point of the three-dimensional problem for each displacement component. 5 Conclusion Two algorithms are proposed to decompose global NN approximations into simpler, local models. Each local NN is independently optimized using a local loss function and the L-BFGS optimization algorithm. Interface continuity between local NN domains is enforced through … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 5 canonical work pages

  1. [1]

    Robust optimization – A comprehensive survey

    H.-G. Beyer and B. Sendhoff. “Robust optimization – A comprehensive survey”. In:Computer Methods in Applied Mechanics and Engineering196.33–34 (July 2007), pp. 3190–3218.doi:10. 1016/j.cma.2007.03.003

    2007

  2. [2]

    Chan.Theory and Applications of Monte Carlo Simulations

    V. Chan.Theory and Applications of Monte Carlo Simulations. London: IntechOpen, Mar. 2013. doi:10.5772/45892

    doi:10.5772/45892 2013

  3. [3]

    Forrester, A

    A. Forrester, A. Sobester, and A. Keane.Engineering Design via Surrogate Modelling. John Wiley and Sons, Ltd, 2008. Chap. 4, pp. 109–116.isbn: 9780470770801

    2008

  4. [4]

    The Homogeneous Chaos

    N. Wiener. “The Homogeneous Chaos”. In:American Journal of Mathematics60.4 (Oct. 1938), pp. 897–936.doi:10.2307/2371268

    doi:10.2307/2371268 1938

  5. [5]

    Weighted discrete least-squares polynomial approximation using randomized quadratures

    T. Zhou, A. Narayan, and D. Xiu. “Weighted discrete least-squares polynomial approximation using randomized quadratures”. In:Journal of Computational Physics298 (Oct. 2015), pp. 787–800.doi: 10.1016/j.jcp.2015.06.042

    work page doi:10.1016/j.jcp.2015.06.042 2015

  6. [6]

    Universal approximation bounds for superpositions of a sigmoidal function

    A. R. Barron. “Universal approximation bounds for superpositions of a sigmoidal function”. In: IEEE Transactions on Information Theory39.3 (May 1993), pp. 930–945.doi:10 . 1109 / 18 . 256500

    1993

  7. [7]

    Toselli and O

    A. Toselli and O. Widlund.Domain Decomposition Methods - Algorithms and Theory. Vol. 34. Springer Series in Computational Mathematics. Berlin, Heidelberg: Springer, 2005.isbn: 978-3- 540-20696-5.doi:10.1007/b137868

    doi:10.1007/b137868 2005

  8. [8]

    ¨Uber einige Abbildungsaufgaben

    H.A. Schwarz. “ ¨Uber einige Abbildungsaufgaben”. In:Journal f¨ ur die reine und angewandte Math- ematik1869.70 (1869), pp. 105–120.doi:10.1515/crll.1869.70.105

    work page doi:10.1515/crll.1869.70.105

  9. [9]

    On the Schwarz Alternating Method I

    P. L. Lions. “On the Schwarz Alternating Method I”. In:First Proceedings of Domain Decomposition Methods for Partial Differential Equations. Ed. by R. Glowinski, G.H. Golub, G.A. Meurant, and J. Periaux. Philadelphia, USA: SIAM, 1988. Chap. 1, pp. 1–42

    1988

  10. [10]

    On the Schwarz Alternating Method III: A Variant for Nonoverlapping Subdomains

    P. L. Lions. “On the Schwarz Alternating Method III: A Variant for Nonoverlapping Subdomains”. In:Third International Symposium on Domain Decomposition Methods for Partial Differential Equations. Ed. by T. F. Chan, R. Glowinski, J. P´ eriaux, and O. B. Widlund. Philadelphia, USA: SIAM, 1990. Chap. 11, pp. 202–223

    1990

  11. [11]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equa- tions

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equa- tions”. In:Journal of Computational Physics378 (Feb. 2019), pp. 686–707.doi:10.1016/j.jcp. 2018.10.045. 24

    doi:10.1016/j.jcp 2019

  12. [12]

    Domain Decomposition Algorithms for Neural Network Approximation of Partial Differential Equations

    H. H. Kim and H. J. Yang. “Domain Decomposition Algorithms for Neural Network Approximation of Partial Differential Equations”. In:Domain Decomposition Methods in Science and Engineering XXVII. Ed. by Z. Dost´ al, T. Kozubek, A. Klawonn, U. Langer, L. F. Pavarino, J.ˇS ´ ıstek, and O. B. Widlund. Cham: Springer Nature Switzerland, 2024, pp. 27–37.doi:10.1...

    work page doi:10.1007/978-3-031-50769- 2024

  13. [13]

    D3M: A Deep Domain Decomposition Method for Partial Differential Equations

    K. Li, K. Tang, T. Wu, and Q. Liao. “D3M: A Deep Domain Decomposition Method for Partial Differential Equations”. In:IEEE Access8 (Dec. 2019), pp. 5283–5294.doi:10.1109/ACCESS. 2019.2957200

    arXiv 2019

  14. [14]

    Deep Domain Decomposition Method: Elliptic Problems

    W. Li, X. Xiang, and Y. Xu. “Deep Domain Decomposition Method: Elliptic Problems”. In:Pro- ceedings of The First Mathematical and Scientific Machine Learning Conference. Ed. by L. Jianfeng and W. Rachel. Vol. 107. PMLR, July 2020, pp. 269–286.url:https://proceedings.mlr.press/ v107/li20a.html

    2020

  15. [15]

    Deep Ritz method with adaptive quadrature for linear elasticity

    M. Liu, Z. Cai, and K. Ramani. “Deep Ritz method with adaptive quadrature for linear elasticity”. In:Computer Methods in Applied Mechanics and Engineering415 (2023), p. 116229.issn: 0045- 7825.doi:10.1016/j.cma.2023.116229

    arXiv 2023

  16. [16]

    Finite basis physics-informed neural networks (FBPINNs): a scalable domain decomposition approach for solving differential equations

    B. Moseley, A. Markham, and T. Nissen-Meyer. “Finite basis physics-informed neural networks (FBPINNs): a scalable domain decomposition approach for solving differential equations”. In:Adv Comput Math49 (July 2023), pp. 1–39.doi:10.1007/s10444-023-10065-9

    doi:10.1007/s10444-023-10065-9 2023

  17. [17]

    Extended Physics-Informed Neural Networks (XPINNs): A Generalized Space-Time Domain Decomposition Based Deep Learning Framework for Nonlinear Partial Differential Equations

    A. D. Jagtap and G. E. Karniadakis. “Extended Physics-Informed Neural Networks (XPINNs): A Generalized Space-Time Domain Decomposition Based Deep Learning Framework for Nonlinear Partial Differential Equations”. In:CiCP28.5 (Nov. 2020), pp. 2002–2041.doi:10.4208/cicp.OA- 2020-0164

    doi:10.4208/cicp.oa- 2020

  18. [18]

    Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems

    A. D. Jagtap, E. Kharazmi, and G. E. Karniadakis. “Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems”. In: Computer Methods in Applied Mechanics and Engineering365 (June 2020), p. 113028.doi:10. 1016/j.cma.2020.113028

    arXiv 2020

  19. [19]

    Partitioned neural network approximation for partial differential equations enhanced with Lagrange multipliers and localized loss functions

    K. Jang, K. Kim, and H. H. Kim. “Partitioned neural network approximation for partial differential equations enhanced with Lagrange multipliers and localized loss functions”. In:CoRR(Dec. 2023). doi:10.48550/ARXIV.2312.14370

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.14370 2023

  20. [20]

    Compact Operators. Spectral Decomposition of Self-Adjoint Compact Operators

    H. Brezis. “Compact Operators. Spectral Decomposition of Self-Adjoint Compact Operators”. In: Functional Analysis, Sobolev Spaces and Partial Differential Equations. New York, NY: Springer New York, 2011, pp. 157–179.isbn: 978-0-387-70914-7.doi:10.1007/978-0-387-70914-7_6

    work page doi:10.1007/978-0-387-70914-7_6 2011

  21. [21]

    Variational formulations and functional approximation algorithms in stochastic plas- ticity of materials

    B. V. Rosic. “Variational formulations and functional approximation algorithms in stochastic plas- ticity of materials”. PhD thesis. Faculty of Engineering, Kragujevac, 2012.doi:10 . 2298 / KG 20121116ROSIC

    2012

  22. [22]

    Bathe.Finite Element Procedures in Engineering Analysis

    K. Bathe.Finite Element Procedures in Engineering Analysis. Prentice-Hall, Inc., Englewood Cliffs, New Jersey 07632, USA, 1982.doi:10.1002/nag.1610070412

    work page doi:10.1002/nag.1610070412 1982

  23. [23]

    Nocedal and S

    J. Nocedal and S. J. Wright.Numerical Optimization. 2e. New York, NY, USA: Springer, 2006

    2006

  24. [24]

    PyTorch: An Imperative Style, High-Performance Deep Learn- ing Library

    A. Paszke, S. Gross, and F. Massa. “PyTorch: An Imperative Style, High-Performance Deep Learn- ing Library”. In:Advances in Neural Information Processing Systems 32 (NeurIPS 2019). 2019, pp. 8024–8035.url:http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style- high-performance-deep-learning-library.pdf

    2019

  25. [25]

    Abadi, A

    M. Abadi, A. Agarwal, and P. and Barham.TensorFlow: Large-Scale Machine Learning on Hetero- geneous Systems. Software available from tensorflow.org. 2015.url:https://www.tensorflow. org/

    2015

  26. [26]

    Distributed optimization and statisti- cal learning via the alternating direction method of multipliers

    S. Boyd, N. Parikh, E. Chu, B. Peleato, and J. Eckstein. “Distributed optimization and statisti- cal learning via the alternating direction method of multipliers”. In:Foundations and Trends in Machine Learning3 (Jan. 2011), pp. 1–122.doi:10.1561/2200000016

    doi:10.1561/2200000016 2011

  27. [27]

    Lagrange Multipliers Revisited

    M. Slater. “Lagrange Multipliers Revisited”. In:Cowles Foundation Discussion Papers. Jan. 1959, p. 304.url:https://elischolar.library.yale.edu/cowles-discussion-paper-series/304

    1959

  28. [28]

    On the Problem of Local Minima in Backpropagation

    M. Gori and A. Tesi. “On the Problem of Local Minima in Backpropagation”. In:IEEE Transactions on Pattern Analysis and Machine Intelligence14.1 (Jan. 1992), pp. 76–86.doi:10 . 1109 / 34 . 107014. 25

    1992

  29. [29]

    A Limited Memory Algorithm for Bound Constrained Optimization

    R. H. Byrd, P. Lu, J. Nocedal, and C. Zhu. “A Limited Memory Algorithm for Bound Constrained Optimization”. In:SIAM Journal on Scientific Computing16.5 (Feb. 1995), pp. 1190–1208.doi: 10.1137/0916069

    doi:10.1137/0916069 1995

  30. [30]

    An overview of gradient descent optimization algorithms

    S. Ruder. “An overview of gradient descent optimization algorithms”. In:CoRRabs/1609.04747 (Sept. 2016).url:http://arxiv.org/abs/1609.04747. 26

    Pith/arXiv arXiv 2016