Pith. sign in

REVIEW 3 major objections 5 minor 30 references

Neural operator preconditioning from mixed dataset for the Helmholtz equations: Application to transcranial ultrasound

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A 64×64-trained U-Net preconditioner solves 512×512 skull-ultrasound Helmholtz problems in minutes, where plain GMRES takes hours or stalls.

desk verdict A genuine, honestly-reported extension of the author's own learned-preconditioning framework; the headline 64x64-to-512x512 head-CT result is worth taking seriously, but the grid-transfer premise rests on only four CT slices and deserves scrutiny before the strong generalization claim is accepted. read the letter →

arxiv 2607.16906 v1 pith:XHZW6YV7 submitted 2026-07-18 math.NA cs.NA

Yanfei Xiang This is my paper · ORCID
classification math.NAcs.NA MSC 65F1065N2268T07
keywords NeuraloperatorpreconditionerHelmholtzequationTranscranialultrasoundFGMRESU-NetMixedtrainingdatasetOut-of-distributiongeneralizationResidual-basedloss
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a U-Net trained on coarse 64×64 Helmholtz problems, with a physics-informed loss based on the relative residual, can act as a nonlinear preconditioner that makes FGMRES converge on realistic 512×512 transcranial ultrasound problems—grids with 64 times more unknowns than anything seen in training. The decisive design choice is dataset composition: mixing skull-like velocity fields with randomized velocities and source configurations (the D0 dataset) yields the fastest and most robust solves. If the claim holds, a single learned, matrix-free approximate inverse can replace slow or failing classical solvers for a sequence of patient-specific Helmholtz systems, reaching arbitrary accuracy. On four adult head-CT problems the paper reports about 12 minutes to reach a relative residual of 10^-3, versus 12.5 hours for GMRES (which solves only one of the four), and about 46 minutes to reach 10^-6, where the prior learned optimizer stagnates near 10^-4.

What carries the argument

The load-bearing mechanism is the U-Net as a learned approximate inverse of the discrete Helmholtz operator, trained with a residual (backward-error) loss so that A(c)Nθ(c,σ)≈I. Its 5-channel input encodes real and imaginary parts of the source, the velocity field, and the real and imaginary parts of the PML damping; the 2-channel output is the predicted complex wavefield. Inside FGMRES, the preconditioner application z_j = Nθ(v_j, c, σ) is computed in 32-bit precision while the rest of the iteration runs in 64-bit, making the preconditioner nonlinear yet still valid within the flexible Arnoldi relation. Because the network is convolutional and applied matrix-free, no Helmholtz matrix assemb

What would settle it

Apply the D0-trained U-Net as an FGMRES preconditioner to ten held-out adult head-CT slices at 512×512 with the same focused-arc source; if any case fails to reach η=10^-3 within 5120 iterations, or if the same anatomy discretized at 256, 512, and 1024 shows iteration counts growing steeply with resolution, the claimed coarse-to-fine transfer and strong out-of-distribution generalization would be contradicted.

Watch

Extended reading notes

Core claim

The central discovery is that a convolutional U-Net, trained only on 64×64 grids to minimize the normalized residual ||b − A(c)Nθ(b,c,σ)||/||b||, transfers to 512×512 operators with realistic skull velocity fields and still gives FGMRES a useful search direction. Used as a nonlinear right preconditioner in FGMRES—each Krylov vector is passed through the network in 32-bit arithmetic—the method solves all four tested head-CT transcranial ultrasound problems to 10^-3, and to 10^-6 when the network is trained on the mixed D0 dataset, while GMRES converges on only one problem within 5120 iterations and the learned optimizer stagnates. The paper attributes this to the mixed training datasets: idea

Load-bearing premise

A U-Net whose weights were fitted to approximate the inverse Helmholtz operator on 64×64 grids remains a useful preconditioner when the same fixed-receptive-field network is applied to 512×512 operators—64 times more unknowns—with the paper's evidence for this transfer resting on four head-CT slices.

Editorial extensions

If this is right

  • One coarse-trained network can accelerate a whole family of parametric Helmholtz solves across resolutions, removing the need to retrain per mesh.
  • Practical 2D transcranial ultrasound planning becomes computationally feasible on CPUs: four head-CT problems to 10^-3 in about 12 minutes and to 10^-6 in about 46 minutes.
  • The matrix-free character sidesteps the memory and setup costs of classical algebraic preconditioners (Jacobi, incomplete LU, algebraic multigrid) on fine grids.
  • Because FGMRES corrects the network's errors, the final solution accuracy is not capped by the network's own accuracy; the hybrid can reach arbitrarily tight tolerances.
  • The mixed-dataset result identifies dataset composition as a controllable factor: D0 outperforms both the purely specialized and the purely randomized training sets on the hardest out-of-distribution cases.

Reading between the lines

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

  • If the 64→512 transfer is real, a natural stress test is to apply the same D0 network at 256, 512, and 1024 resolutions on identical CT slices; the resolution at which iteration counts degrade would map the effective range of the fixed receptive field.
  • The same recipe should extend to 3D transcranial volumes and higher frequencies, but the paper only demonstrates 2D and 490 kHz; those settings remain open.
  • Dataset balance looks like a tunable hyperparameter: interpolating between D0 and D1 could let a user optimize for a specific anatomy or transducer geometry without changing the architecture.
  • The comparison against a standalone learned optimizer suggests the Krylov wrapper carries much of the robustness; the same network that stagnates as a solver becomes reliable as a preconditioner, pointing toward hybrid designs as a general pattern.
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 / 5 minor

Summary. The manuscript proposes a learned nonlinear preconditioner for FGMRES solving sequences of 2D Helmholtz equations in transcranial ultrasound. A depth-4 U-Net is trained on 64×64 grids with a relative-residual (physics-informed) loss using six mixed datasets (D0–D5) combining idealized skulls, random velocity fields, and several source distributions. The network is then used as a matrix-free right preconditioner in FGMRES. The central empirical claim is in Section 4.2.2 (Table 8): the D0-trained preconditioner solves four 512×512 head-CT problems to η_b = 10^{-3} in at most 191 iterations (~12 min CPU total) and to 10^{-6} in ~46 min, whereas unpreconditioned GMRES converges on only one of the four within 5120 iterations (~12.5 h) and the learned optimizer of [26] stagnates/diverges on three of the four. The paper also reports speedups on 64×64 in-distribution tests and on idealized-skull/seismology-style benchmarks, and concludes that mixed dataset design improves generalization. The relative-residual loss and the stopping criterion are coherent; the main risk identified is the unanalyzed transfer from 64×64 training to 512×512 inference.

Significance. If the results hold, the practical significance is real: a matrix-free learned preconditioner that brings FGMRES to 512×512 transcranial ultrasound problems, with honest per-slice iteration counts, wall-clock times, same-hardware comparisons, and an unsupervised physics-informed loss. The paper is incremental over the authors' earlier framework [9]; the contribution is dataset design rather than architecture. Its strengths are the systematic six-dataset comparison, the 10^{-3}/10^{-6} head-CT results, and the clearly reported failure cases (e.g., D5 on one CT slice). The significance is limited by the four-slice real-data evaluation, the lack of resolution-transfer analysis, and some overclaimed generalization statements, but the evidence is consistent and not internally contradictory.

major comments (3)
  1. [Section 4.2.2, Table 8] The transfer from 64×64 training to 512×512 inference is the load-bearing premise, but it is only asserted ('this computational grid is 64 times larger than the 64×64 grid used during training'). The depth-4 U-Net with stride-2 downsampling maps 512×512 inputs to a 32×32 bottleneck, a resolution never seen in training; the convolutional kernels and batch-normalization statistics were not adapted. No intermediate-resolution experiments (128, 256) or diagnostics of N_θ's residual reduction on fine grids are given. Since the central practical claim depends on this transfer, please add such a resolution study or, at minimum, a larger CT cohort with per-slice results and a statement of slice selection. Without this, the four-slice evidence is not sufficient to establish the claimed out-of-distribution generalization.
  2. [Table 8 / Figure 9(c)] The claim of 'strong out-of-distribution generalization' is supported by only four CT slices, with no selection criteria or statistics. The D5-trained preconditioner fails on one of the four (5120*), and iteration counts vary by factors of 2–10 across D0–D5 on the same four problems. With n=4, the ranking of datasets could easily be slice-specific. Please provide a larger, systematically sampled test set, report per-example costs, or qualify the generalization claim to the four reported examples. This is load-bearing because the abstract and conclusions make a broad generalization claim.
  3. [Abstract and Section 4.2.2] The statement that the method 'achieves arbitrary solution accuracies' is stronger than the evidence. Table 8 reports η_b = 10^{-3} and 10^{-6} for the 512×512 head-CT problems; no 10^{-9} or 10^{-12} result is provided at that scale, and the learned optimizer comparison is only made at 10^{-3}. Please either add a tighter-tolerance head-CT run or rephrase the claim as 'high prescribed accuracies (demonstrated to 10^{-6})'. This is a claim-calibration issue rather than a technical error.
minor comments (5)
  1. [Section 2.1 vs Section 4.2.2] The paper restricts the Helmholtz formulation to ω=1, but Section 4.2.2 describes the CT problem at 490 kHz. Please clarify the nondimensionalization/scaling; as written, the frequency statement is unused and confusing.
  2. [Tables 6–9] The star notation is inconsistent. In (Its_max, Its_avg, Its_min)* the star denotes divergence, but a superscript star after dataset labels (e.g., D4⋆ in Table 7, D0⋆ in Table 8, D3⋆ in Table 9) is never explained.
  3. [Figure 3] The caption says '8 × 8 mixed speed of sounds', but the axes and text indicate 64×64 training samples. Please correct this typo.
  4. [Section 3.1, Eq. (2), Eq. (8)] Equation (2) defines σ as a real damping function, but Section 3.1 and Eq. (8) describe σ as complex. Please reconcile the notation.
  5. [Reproducibility] No data or code availability statement is provided. Given the paper's empirical nature, releasing the CT preprocessing script, trained checkpoints, and dataset-generation code would materially aid reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular steps: training objective, FGMRES stopping criterion, and test data are distinct; self-citations are contextual only.

full rationale

The paper's chain is: train N_theta to minimize the relative residual of the discrete Helmholtz equation on 64x64 systems (Eq. 11); insert N_theta as a nonlinear preconditioner into FGMRES (Algorithm 1, Eq. 7); measure the number of FGMRES iterations/elapsed time needed to drive the same relative residual eta_b below epsilon (Eq. 12) on 64x64 test problems, idealized skulls, and 512x512 head-CT problems. The use of the relative residual in both the training loss and the stopping criterion is a coherent choice of metric, not a fitted quantity renamed as a prediction: the reported convergence is obtained by actually running FGMRES, not by reading off the training loss. All six preconditioners are trained on source/velocity distributions that do not include the head-CT data (qure.ai, [22]) or the seismology/spherical-wave tests, and the learned-optimizer baseline is an external checkpoint from [26]; the decisive comparisons are against external or held-out objects. The self-citations [9,25,29,30] supply the preconditioning framework and prior comparisons, but the numerical evidence in Tables 6-9 is generated in this paper and does not reduce to the contents of those citations. The main caveats—the unproven transfer from 64x64 training to 512x512 application and the post-hoc identification of D0 as best on the four CT slices—are empirical generalization/selection concerns, not circularity by construction. No equation in the paper makes the test-time residual equal to the training loss by definition, and no load-bearing claim is justified solely by a self-citation.

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

The paper's claim is empirical and rests on a fitted operator. Its free parameters are the dataset mixture rules, source-amplitude normalizations, network hyperparameters, and the trained weights themselves; the structural premise with the least support is discretization transfer (64x64 to 512x512), which is asserted rather than justified. No invented physical entities appear.

free parameters (5)
  • training-dataset mixture probabilities = c: [0.4,0.4,0.1,0.1]; b: [0.6,0.3,0.05,0.05]
    Hand-chosen composition rules in Table 4; the paper's central finding is that this composition governs generalization, so these are the primary design degrees of freedom being tested.
  • dirac source amplitude = 60 (dirac/dirac1), 30 (dirac4)
    Amplitudes chosen so localized sources have magnitude comparable to N(0,1) over the domain (Section 3.2, block 2); a normalization that affects training.
  • U-Net hyperparameters = depth 4, 32 channels, kernel 8, stride 2, batch 32, LR 1e-3 to 1e-5, 500 epochs
    Hand-set in Tables 2-3; depth 4 is justified only as 'sufficient' (Section 3.1); the preconditioner's quality depends on these choices.
  • trained network weights theta = ~832K parameters per dataset (final checkpoints not released)
    Fitted by Adam to the residual loss (11); the central experiments measure FGMRES performance using these fitted weights, so the paper's empirical claim is a statement about these fitted values.
  • PML and frequency settings = l=4, sigma_max=2, omega=1
    Inherited from the prior setup (Table 1); the whole study operates in this regime (c in [1,2], dimensionless frequency 1), and the clinical claim (490 kHz) is only mapped to it implicitly.
assumptions (5)
  • domain assumption Fourier spectral discretization with a PML is an adequate representation of the unbounded-domain Helmholtz problem
    Section 2.1; all results are for this discretization, inherited from [25,26,29]; failure of PML tuning would corrupt all reported residuals.
  • ad hoc to paper A U-Net trained at 64x64 provides a useful approximate inverse of the Helmholtz operator at 512x512
    The 64x transfer is the enabler of Section 4.2.2's headline results; it is asserted and empirically probed on four CT slices, never derived or analyzed, and no ablation on intermediate resolutions is reported.
  • domain assumption Truncated idealized skull fields are a sufficient training proxy for real head-CT velocity fields
    Dataset construction (Section 3.2, block 1b) relies on this proxy; transfer from D0/D2 to real CT data is attributed to it, but no quantitative similarity measure is given.
  • standard math FGMRES with a nonlinear preconditioner remains well-defined and convergent as long as the preconditioned basis Z_k is full rank
    Section 2.2 invokes Saad's flexible Arnoldi relation (4); the paper correctly relies on the standard result that a nonlinear preconditioner can be used inside FGMRES.
  • domain assumption The 2D Helmholtz model at omega=1 with c in [1,2] captures the practical transcranial ultrasound regime
    Used in all experiments (Table 1 and Section 3.2); the clinically framed claim (490 kHz, 60 mm aperture transducer) is only studied through this dimensionless regime.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural operator preconditioning from mixed dataset for the Helmholtz equations: Application to transcranial ultrasound." pith.science (2026). https://pith.science/paper/XHZW6YV7

@misc{pith2026260716906,
  author       = {Pith},
  title        = {Pith review of: Neural operator preconditioning from mixed dataset for the Helmholtz equations: Application to transcranial ultrasound},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XHZW6YV7}},
  note         = {Machine review of arXiv:2607.16906}
}
read the original abstract

This work develops a neural operator preconditioned subspace method for sequences of linear systems arising from the discretization of the two-dimensional Helmholtz equation in transcranial ultrasound applications. The problem involves strongly heterogeneous, patient-dependent velocity fields that induce severe wave distortion and pose significant challenges for standard iterative solvers. Building on neural network preconditioning framework of Giraud et al. (HAL RR-9593, 2025) and the idealized skull dataset used for the learned optimizer of Stanziola et al. (JCP 441, 2021), neural operator preconditioners are trained on six mixed velocity-source datasets combining randomized source configurations and idealized skull-based velocity fields with random noise. The proposed mixed-dataset strategy aims to improve both computational efficiency and generalization across varying configurations. The neural operator is trained on a coarse grid using a physics-informed loss based on the relative residual of the discrete Helmholtz equation and is incorporated as a nonlinear preconditioner within flexible GMRES (FGMRES). Numerical experiments demonstrate that the resulting hybrid method efficiently solves practical transcranial ultrasound problems on grids 64 times larger than those used during training, whereas both classical GMRES and the learned optimizer fail to converge within comparable computational budgets. Moreover, the proposed method achieves arbitrary solution accuracies and exhibits strong out-of-distribution generalization across diverse source and velocity configurations. This work highlights the importance of dataset design in scientific machine learning and provides a practical framework for integrating matrix-free neural operator preconditioning with Krylov subspace methods for solving practical large-scale Helmholtz problems.

Figures

Figures reproduced from arXiv: 2607.16906 by the authors.

Figure 1
Figure 1. Illustration of the PML boundary condition on two-dimensional domain ([ [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of our U-Net architecture with 4 depth. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Visualize some samples from the different types of the training dataset [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Schematic representation of the training framework. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Results of the CNN-preconditioning trained with dataset [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Visualize the complex wavefield in real-imag part (upper two plots of each subplot (a) to (d)) and [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Convergence in terms of ηb = 10−12 of CNN-FGMRES and GMRES with maxIts = 512 without restrat for L examples with varying velocity c iid∼ TruncIdealSkulls from the test dataset and the fixed source term b iid∼ dirac1 with fixed grid-point locates at 60 × 32 within 2D do…
Figure 8
Figure 8. Figure 8: Visualize the complex wavefield in real-imag part (upper two plots of each subplot (a) to (d)) [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Convergence in terms of ηb = 10−3 for (a)-(c), and ηb = 10−6 for (d) of CNN-FGMRES and GMRES with maxIts = 5120 with restrat equals 10, together with the learned optimizer proposed in [26] with maxIts = 200, 5120, on four (i.e., L = 4) practical head CT scans problems.…
Figure 10
Figure 10. Figure 10: Evolution of the complex wavefield predicted by the learned optimizer [ [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Visualize four head CT scans on 2D domain [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]
Figure 12
Figure 12. Figure 12: Counterpart of Figure [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]
Figure 13
Figure 13. Figure 13: Convergence histories of the learned optimizer [ [PITH_FULL_IMAGE:figures/full_fig_p024_13.png]
Figure 14
Figure 14. Figure 14: Convergence in terms of ηb = 10−12 of CNN-FGMRES and GMRES with maxIts = 512 without restart for L = 20 simplified models of spherical waves and seismology problems. In (a)-(d), the presented CNN-FGMRES results are for the CNN-preconditioning trained with dataset D3. …
Figure 15
Figure 15. Figure 15: Visualize spherical waves at (a)-(b) and seismology examples (c)-(f) on 2D domain 64 × 64 solved by CNN-FGMRES with ηb = 10−12 and dataset D3. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 5 canonical work pages

  1. [26]

    A Helmholtz Equation Solver Using Un- supervised Learning: Application to Transcranial Ultrasound

    A. Stanziola, S. R. Arridge, B. T. Cox, and B. E. Treeby. “A Helmholtz Equation Solver Using Un- supervised Learning: Application to Transcranial Ultrasound”. In:Journal of Computational Physics 441 (2021), p. 110430.issn: 00219991.doi:10 . 1016 / j . jcp . 2021 . 110430.url:https : / / www . sciencedirect.com/science/article/abs/pii/S0021999121003259

    2021

  2. [9]

    Giraud, C

    L. Giraud, C. Kruse, P. Mycek, M. Shpakovych, and Y.-F. Xiang.Neural network preconditioning: a case study for the solution of the parametric Helmholtz equation. Technical Report 9593. Inria Centre at the University of Bordeaux, 2025.url:https://inria.hal.science/hal-05157038

    2025

  3. [1]

    Conditional Denoising Diffusion Model-BasedRobustMRImageReconstructionfromHighlyUndersampledData

    M. Alsubaie, W. Liu, L. Gu, O. C. Andronesi, S. M. Perera, and X. Li. “Conditional Denoising Diffusion Model-BasedRobustMRImageReconstructionfromHighlyUndersampledData”.In:arXiv:2510.06335 (2025).doi:10.48550/arXiv.2510.06335. arXiv:2510.06335.url:http://arxiv.org/abs/2510. 06335

    work page doi:10.48550/arxiv.2510.06335 2025

  4. [2]

    Multigrid-Augmented Deep Learning Preconditioners for the Helmholtz Equation

    Y. Azulay and E. Treister. “Multigrid-Augmented Deep Learning Preconditioners for the Helmholtz Equation”. In:SIAM Journal on Scientific Computing45.3 (2023), S127–S151.issn: 1064-8275.doi: 10.1137/21M1433514.url:https://epubs.siam.org/doi/10.1137/21M1433514

    work page doi:10.1137/21m1433514.url:https://epubs.siam.org/doi/10.1137/21m1433514 2023

  5. [3]

    A Perfectly Matched Layer for the Absorption of Electromagnetic Waves

    J.-P. Berenger. “A Perfectly Matched Layer for the Absorption of Electromagnetic Waves”. In:Journal of Computational Physics114.2 (1994), pp. 185–200.issn: 0021-9991.doi:10.1006/jcph.1994.1159. url:https://www.sciencedirect.com/science/article/pii/S0021999184711594

    arXiv 1994

  6. [4]

    Autoencoders

    C. M. Bishop and H. Bishop. “Autoencoders”. In:Deep Learning: Foundations and Concepts. Cham: Springer International Publishing, 2024, pp. 563–579.isbn: 978-3-031-45468-4.doi:10.1007/978-3- 031-45468-4_19.url:https://doi.org/10.1007/978-3-031-45468-4_19

    doi:10.1007/978-3- 2024

  7. [5]

    Message Passing Neural PDE Solvers

    J. Brandstetter, D. E. Worrall, and M. Welling. “Message Passing Neural PDE Solvers”. In:Interna- tional Conference on Learning Representations (ICLR). 2022.doi:10.48550/arXiv.2202.03376

    arXiv doi:10.48550/arxiv.2202.03376 2022

  8. [6]

    OpenFWI: Large-scale Multi-structural Benchmark Datasets for Full Waveform Inversion

    C.-Y. Deng, S.-H. Feng, H.-C. Wang, X.-T. Zhang, P. Jin, Y.-N. Feng, Q.-L. Zeng, Y.-P. Chen, and Y.-Z. Lin. “OpenFWI: Large-scale Multi-structural Benchmark Datasets for Full Waveform Inversion”. In:Advances in Neural Information Processing Systems. Ed. by S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh. Vol. 35. Curran Associates, Inc.,...

    2022

Show all 30 references
  1. [7]

    Numerical Solution of Mixed-Dimensional PDEs Using a Neural Preconditioner

    N. Dimola, N. R. Franco, and P. Zunino. “Numerical Solution of Mixed-Dimensional PDEs Using a Neural Preconditioner”. In: arXiv:2505.08491 (2025).doi:10 . 48550 / arXiv . 2505 . 08491.url: https://arxiv.org/pdf/2505.08491

    2025 doi arXiv

  2. [8]

    Scientific machine learning

    F. Dietrich and W. Schilders. “Scientific machine learning”. In:Math Semesterber72 (2025), pp. 89– 115.doi:10.1007/s00591-025-00399-4

    2025 doi

  3. [10]

    N.J.Higham.Accuracy and stability of numerical algorithms.SIAM,2002.doi:10.1137/1.9780898718027

    2002 doi

  4. [11]

    DeepONet Based Preconditioning Strategies for Solving Para- metric Linear Systems of Equations

    A. Kopaničáková and G. Karniadakis. “DeepONet Based Preconditioning Strategies for Solving Para- metric Linear Systems of Equations”. In:SIAM Journal on Scientific Computing47.1 (2025), pp. C151– C181.issn: 1095-7197.doi:10.1137/24m162861x.url:http://dx.doi.org/10.1137/24M162861X

    2025 doi

  5. [12]

    Neural Operator: Learning Maps Between Function Spaces with Applications to PDEs

    N. B. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. M. Stuart, and A. Anandkumar. “Neural Operator: Learning Maps Between Function Spaces with Applications to PDEs”. In:Journal of Machine Learning Research24.89 (2023), pp. 1–97.url:http : / / jmlr . org / pa...

    2023

  6. [13]

    Neural Preconditioning Operator for Efficient PDE Solves

    Z. Li, D. Xiao, Z. Lai, and W. Wang. “Neural Preconditioning Operator for Efficient PDE Solves”. In: arXiv:2502.01337 (2025).doi:10.48550/arXiv.2505.01337.url:https://arxiv.org/pdf/2502. 01337

    2025 doi arXiv

  7. [14]

    Fourier Neural Operator for Parametric Partial Differential Equations

    Z. Li, N. B. Kovachki, K. Azizzadenesheli, B. liu, K. Bhattacharya, A. Stuart, and A. Anandkumar. “Fourier Neural Operator for Parametric Partial Differential Equations”. In:International Conference on Learning Representations. 2021.url:https://openreview.net/forum?id=c8P9NQVtmnO

    2021

  8. [15]

    AI-Augmented Adaptive Digital Twin Modeling for Brain Tumor Evolution Prediction and Treatment Scheduling

    W. Liu, M. Trimboli, and X. Li. “AI-Augmented Adaptive Digital Twin Modeling for Brain Tumor Evolution Prediction and Treatment Scheduling”. In: arXiv:2607.13877 (2026).doi:10.48550/arXiv. 2607.13877. arXiv:2607.13877.url:http://arxiv.org/abs/2607.13877. 22

    2026 doi arXiv

  9. [16]

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators

    L. Lu, P. Jin, G. Pang, Z. Zhang, and G. Karniadakis. “Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators”. In:Nature Machine Intelligence3 (2021), pp. 218–229.doi:10.1038/s42256-021-00302-5

    2021 doi

  10. [17]

    Neural Operator: Learning Maps Between Function Spaces With Applications to PDEs

    N. Kovachki, Z.-Y. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, and A. Anandkumar. “Neural Operator: Learning Maps Between Function Spaces With Applications to PDEs”. In:Journal of Machine Learning Research(2023). http://jmlr.org/papers/v24/21-1524.html, pp. 1–97

    2023

  11. [18]

    Pathak, S

    J. Pathak, S. Subramanian, P. Harrington, S. Raja, A. Chattopadhyay, M. Mardani, T. Kurth, D. Hall, Z. Li, K. Azizzadenesheli, P. Hassanzadeh, K. Kashinath, and A. Anandkumar.FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operator...

    2022 arXiv

  12. [19]

    Meta-learning PINN loss functions

    A. F. Psaros, K. Kawaguchi, and G. E. Karniadakis. “Meta-learning PINN loss functions”. In:Jour- nal of Computational Physics458 (2022), p. 111121.issn: 0021-9991.doi:https://doi.org/10. 1016 / j . jcp . 2022 . 111121.url:https : / / www . sciencedirect . com / science / artic...

    2022

  13. [20]

    On the Compatibility of a Given Solution With the Data of a Linear System

    J. L. Rigal and J. Gaches. “On the Compatibility of a Given Solution With the Data of a Linear System”. In:Journal of the ACM14.3 (1967), pp. 543–548.doi:10 . 1145 / 321406 . 321416.url: https://doi.org/10.1145/321406.321416

    1967

  14. [21]

    U-Net: Convolutional Networks for Biomedical Image Seg- mentation

    O. Ronneberger, P. Fischer, and T. Brox. “U-Net: Convolutional Networks for Biomedical Image Seg- mentation”. In: arXiv:1505.04597 (2015).doi:10.48550/arXiv.1505.04597. arXiv:1505.04597.url: http://arxiv.org/abs/1505.04597

    2015 doi arXiv

  15. [22]

    Development and validation of deep learning algorithms for detection of critical findings in head CT scans

    S. Chilamkurthy, R. Ghosh, S. Tanamala, M. Biviji, N. G. Campeau, V. K. Venugopal, V. Mahajan, P. Rao, and P. Warier. “Development and validation of deep learning algorithms for detection of critical findings in head CT scans”. In: arXiv:1803.05854 (2018).doi:10.48550/arXiv.18...

    2018 doi arXiv

  16. [23]

    A flexible inner-outer preconditioned GMRES algorithm

    Y. Saad. “A flexible inner-outer preconditioned GMRES algorithm”. In:SIAM Journal Scientific Com- puting14 (1993), pp. 461–469

    1993

  17. [24]

    GMRES: a generalized minimal residual algorithm for solving nonsymmetric linear systems

    Y. Saad and M. Schultz. “GMRES: a generalized minimal residual algorithm for solving nonsymmetric linear systems.” In:SIAM Journal on Scientific and Statistical Computing7 (1986), pp. 856–869

    1986

  18. [25]

    Shpakovych.Neural Network Preconditioning of Large Linear Systems

    M. Shpakovych.Neural Network Preconditioning of Large Linear Systems. Technical Report 0518. Inria Centre at the University of Bordeaux, 2023.url:https://inria.hal.science/hal-04254315

    2023

  19. [27]

    Efficient preconditioning for iterative methods with graph neural networks

    V. Trifonov, A. Rudikov, O. Iliev, Y. M. Laevsky, and I. Oseledets. “Efficient preconditioning for iterative methods with graph neural networks”. In:AI4X 2025 International Conference. 2025.url: https://openreview.net/pdf?id=PjDJonBo9T

    2025

  20. [28]

    J. H. Wilkinson.Rounding Errors in Algebraic Processes. Notes on Applied Science 32. Also published by Prentice-Hall, Englewood Cliffs, NJ, USA, 1964. Reprinted by Dover Publications, New York, 1994. London, UK: HMSO, 1963

    1964

  21. [29]

    Solution of Large Linear Systems with a Massive Number of Right-Hand Sides and Ma- chine Learning

    Y.-F. Xiang. “Solution of Large Linear Systems with a Massive Number of Right-Hand Sides and Ma- chine Learning”. PhD thesis. University of Bordeaux, 2022.url:https://theses.fr/2022BORD0383

    2022

  22. [30]

    Xiang.Unsupervised convolution neural operator preconditioning for the solution of some hetero- geneous fluid PDEs

    Y.-F. Xiang.Unsupervised convolution neural operator preconditioning for the solution of some hetero- geneous fluid PDEs. Technical Report. Inria Centre at the University of Bordeaux and Inria Center of the University of Lorraine, 2025.url:https://inria.hal.science/hal-0488693...

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.