Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Reliable and efficient steady CFD from surrogate predictions through Newton-Krylov correction

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A neural surrogate plus a short Newton–Krylov correction yields solver-converged steady RANS solutions even for transonic airfoils far outside the training distribution.

desk verdict A solid surrogate-Newton warm-start study with a genuinely OOD benchmark, but the filtered benchmark and missing code make the headline reliability claim more conditional than the 15.5x speedup suggests. read the letter →

arxiv 2608.04400 v1 pith:AKAZDLQB submitted 2026-08-05 physics.flu-dyn

classification physics.flu-dyn
keywords neuralsurrogatesNewton–KrylovmethodsJacobian-freesteadyRANSCFDout-of-distributiongeneralizationtransonicairfoiloptimizationflying-wingflows
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 tries to establish that the unreliable, out-of-distribution predictions of a neural flow surrogate can be made solver-grade by using them as initial guesses for Newton–Krylov correction inside the target CFD solver. The authors show that after a fixed budget of about 15 Newton–Krylov updates, the median normalized residual of steady RANS solutions on 20,052 out-of-distribution transonic airfoil cases drops from around $10^{-3}$ to $10^{-10}$, more than seven orders of magnitude, while lift, drag, and field errors fall sharply. They further show that the same coupling improves the agreement between online optimization objectives and converged CFD re-evaluations for supercritical airfoils, and that a three-dimensional flying-wing version converges planform-extrapolated cases. The payoff, if the claim holds, is that surrogates no longer need to be trusted as standalone predictions; a small amount of solver-driven correction buys numerical convergence and design reliability at a fraction of CFD cost.

What carries the argument

The mechanism is surrogate-initialized Jacobian-free Newton–Krylov (JFNK) correction. A Taylor expansion of the target residual about the surrogate prediction $\hat{\mathbf{U}}_s$ gives the linear correction equation $\mathbf{J}\,\delta\mathbf{U}+\mathbf{R}(\hat{\mathbf{U}}_s)\approx\mathbf{0}$; solving this repeatedly with Krylov methods and native residual evaluations drives the state to the discrete root. Two diagnostics carry the argument: the alignment between $\mathbf{J}(\mathbf{U}^\star-\hat{\mathbf{U}}_s)$ and $-\mathbf{R}(\hat{\mathbf{U}}_s)$, with benchmark medians of 0.963 and 0.929 for the two surrogate models, and the observation that median trajectories reach the $10^{-8}$ threshold within 9–11 updates. An optional staged schedule interleaves a small number of NK corrections inside the Schrödinger-bridge sampling of FSB-DiT, which mainly improves the high-residual tail.

What would settle it

Run the coupled workflow on a sequence of benchmark geometries ordered by increasing $d_5$ distance from the training set and measure the alignment $\alpha = \langle \mathbf{J}(\mathbf{U}^\star-\hat{\mathbf{U}}_s), -\mathbf{R}(\hat{\mathbf{U}}_s)\rangle / (\|\mathbf{J}(\mathbf{U}^\star-\hat{\mathbf{U}}_s)\|\.|\mathbf{R}(\hat{\mathbf{U}}_s)\|)$; if a geometry class shows low alignment, say $\alpha < 0.5$, while the median residual trajectory fails to drop below $10^{-8}$ within 15 updates, the convergence-basin assumption fails for that class.

Watch

Extended reading notes

Core claim

The central claim is that a neural surrogate and a Newton–Krylov solver are complementary: the surrogate supplies a full-field prediction that is globally organized and close to the target steady state, while the discrete residual of the target solver provides the correction that removes the remaining numerical inconsistency. Concretely, starting from the surrogate prediction $\mathbf{U}^{(0)}=\hat{\mathbf{U}}_s$, the framework iterates the Newton equation $\mathbf{J}(\mathbf{U}^{(k)})\,\Delta\mathbf{U}^{(k)} = -\mathbf{R}(\mathbf{U}^{(k)};\boldsymbol{\theta})$ with a Jacobian-free Krylov linear solve, accepting the state when the case-normalized residual ratio $r(\mathbf{U};\boldsymbol{\theta})\le 10^{-8}$. On the optimization-derived out-of-distribution benchmark, FSB-DiT with 15 updates lowers the median residual ratio from $1.29\times10^{-3}$ to $9.46\times10^{-11}$, the median field MSE from $2.60\times10^{-5}$ to $5.30\times10^{-12}$, median $|\Delta C_L|$ from $1.03\times10^{-2}$ to $1.49\times10^{-6}$, and median $|\Delta C_D|$ from $7.48\times10^{-4}$ to $9.56\times10^{-8}$.

Load-bearing premise

The load-bearing premise is that a surrogate prediction, even one with a residual far above the $10^{-8}$ tolerance, lies inside the Newton convergence basin of the target solver, so that a fixed budget of about 15 Newton–Krylov updates reaches the solver-accepted root on nearly every out-of-distribution case.

Editorial extensions

If this is right

  • With a fixed 15-update NK budget, the coupled workflow reaches the solver-convergence threshold on the median out-of-distribution transonic airfoil case, so neural surrogates can be used as warm starts rather than accepted predictions.
  • In online supercritical airfoil optimization with six NK updates per evaluation, the gap between the online objective and converged-CFD verification drops from 11.8 to 1.7 drag counts for RAE2822 and from 10.0 to 0.4 for OAT15A, while the coupled workflow retains a 15.5-fold generation-level speedup over CFD.
  • Post-correction accuracy improves faster than uncorrected accuracy as training data grows: across six training-set sizes, median residuals after ten NK updates decrease by more than three orders of magnitude, from $8.46\times10^{-6}$ to $1.59\times10^{-9}$, showing that surrogate initialization quality directly controls the accuracy achieved at fixed NK budget.
  • The same coupling transfers to three-dimensional planform-extrapolated flying wings: 15 NK updates lower the median residual ratio from $1.32\times10^{-2}$ to $1.36\times10^{-9}$ and the median volume field MSE from $1.15\times10^{-4}$ to $3.09\times10^{-13}$.

Reading between the lines

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

  • One extension the paper does not test: the alignment diagnostic between $\mathbf{J}\delta\mathbf{U}$ and $-\mathbf{R}$ could serve as a cheap pre-screening criterion to decide whether to spend NK budget on a given surrogate prediction, or to feed back into training as an explicit objective.
  • If surrogate predictions generally lie in Newton basins across other stiff PDEs, the same coupling could replace pseudo-time-marching warm starts in unstructured-mesh or topology-varying solvers; the paper notes JFNK is compatible in principle with unstructured discretizations and would only require state transfer and preconditioning on the target mesh.
  • The near-linear log-log scaling of post-correction error with training-set size suggests a testable scaling law: for a fixed NK budget, the final error is controlled by initialization quality, so gains from more data and gains from more solver updates may be interchangeable engineering levers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This manuscript proposes a solver-coupled surrogate–Newton framework for steady RANS CFD. A neural surrogate (Direct DiT, FSB-DiT, or HM-FNO) predicts the full flow field; the prediction is injected as the initial iterate into ADflow's Jacobian-free Newton–Krylov solver, which then reduces the native residual to the solver threshold. The authors construct an out-of-distribution benchmark from intermediate airfoils of transonic optimization trajectories, report residual and error reductions after a fixed NK budget, analyze the alignment between the surrogate error and the linearized residual, deploy the correction in online airfoil optimization, and extend the approach to three-dimensional flying-wing configurations.

Significance. The core idea—using a learned full-field predictor only to initialize a residual-controlled Newton solver—is a sound and practical division of labor, and the paper provides direct evidence that this can convert OOD surrogate predictions into solver-converged states on a substantial benchmark. The benchmark is derived from actual optimization trajectories rather than canonical geometries, the correction uses the target solver's native residual and JFNK module, and the paper includes controlled comparisons against ANK and pseudo-time marching, training-data scaling studies, and a three-dimensional extension. If the reported numbers are robust, the framework would be a useful tool for accelerating steady aerodynamic CFD. The main reservations concern the unquantified filtering of benchmark cases, the use of a differentiable residual for the linearization diagnostic, and the lack of uncertainty quantification in the timing comparison.

major comments (4)
  1. [Results, benchmark construction] The sentence 'Cases failing the prescribed CFD quality criteria were excluded' introduces a filter whose size and properties are never quantified. Because the benchmark is the main evidence for the claim that a fixed 15-update NK budget turns OOD surrogate predictions into solver-accepted roots, and because every retained case necessarily has a from-scratch CFD reference that converged to the quality criteria, the headline convergence statistics are computed on a possibly favorable subset. Please report the number of excluded cases, the criteria in full, the geometric/residual characteristics of the excluded population, and, if possible, the NK trajectories on those cases. At minimum, state this selection as a limitation in the Discussion.
  2. [Methods, 'Two-stage surrogate training' / 'Jacobian-free Newton–Krylov correction'; Fig. 4a] Fig. 4a presents alignment of JΔU with −R_pred and the normalized closure as evidence that surrogate errors are correctable by the Newton equation. However, Methods states that the differentiable residual evaluator supplies 'the local-linearization diagnostic in Fig. 4', while numerical correction always uses the native ADflow residual. The target Newton operator and the diagnostic operator are therefore different. Please either report the alignment and closure with the native ADflow residual and Jacobian-vector products, or justify why the differentiable-residual diagnostic is representative of the native solver's Newton basin.
  3. [Table 1] Table 1 shows that for Direct DiT with terminal correction the 90th-percentile residual ratio is 4.91×10−6, which is above the solver-convergence threshold 10−8; only FSB-DiT with staged correction has a 90th percentile below threshold. The abstract's 'reliable' claim and the Discussion's 'reliable and efficient steady CFD' are therefore not supported for 10% of the Direct DiT benchmark cases. Please report case-level success rates, the distribution of failures, and discuss the trade-off between surrogate family and reliability.
  4. [Table 2] The 15.5-fold generation-level speedup rests on single runtime values: 69.90 s for FSB-DiT+NK (1 GPU + 72 CPU cores) versus 1080.15 s for CFD (72 CPU cores), with no repetition or error bars. Since the speedup is a central practical claim and the configurations differ in hardware and in the fact that NK uses the surrogate's GPU inference, please provide matched-hardware timings, repeated measurements, and the variance of the comparison.
minor comments (5)
  1. [Results, benchmark construction and Methods] The 'prescribed CFD quality criteria' are never stated anywhere in the manuscript; please define the residual and force-stability thresholds that determine whether a reference case is retained.
  2. [Fig. 4b] In Fig. 4b the residual centre lines denote medians while the solver-time centre lines denote arithmetic means; mixing two statistics in one panel without explicit annotation makes the figure hard to interpret.
  3. [Results, mechanistic analysis] The statement that staged correction reduced the 95th percentile of the residual distribution by 66.8% is ambiguous for a log-distributed quantity; please report the reduction on a log10 scale and state the baseline and post-correction values.
  4. [Abstract and Table 2] The abstract reports a '15.5-fold generation-level speedup' without noting that Table 2 gives a single-generation measurement under one timing protocol; the headline number should be qualified accordingly.
  5. [Table 1] Table 1 would be much easier to read if the column structure explicitly grouped mean/median/90th percentile for each metric; as printed, the repeated triplets are easy to misread.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the surrogate supplies only the initial iterate; the corrected state is the target solver's own fixed point, and all headline metrics are measured outcomes of the native residual equations.

full rationale

The derivation chain is not circular. The surrogate prediction is explicitly defined as an initial guess (U(0)=Ûs), and every NK update solves J(U(k))ΔU(k)=-R(U(k);θ) with the native ADflow residual and JFNK module; the final state is accepted only when r(U;θ)≤10^-8, a criterion defined by the solver residual, not by any quantity fitted during surrogate training. Residual reduction, field MSE, |ΔCL| and |ΔCD| are measured after the fact against the same solver's converged state, so they are outcomes rather than fitted parameters. The training-time residual loss uses a differentiable surrogate residual evaluator, but the paper states that 'numerical correction always uses the native ADflow residual,' making the correction evaluation independent of the training objective. The alignment diagnostic in Fig. 4a uses ΔU=U*−Ûs; it is a post-hoc Taylor-consistency check, not an output that is subsequently reported as a prediction. There are no load-bearing self-citations: ADflow and the JFNK references are external, and no uniqueness theorem or prior work by these authors is invoked to force the framework. The one in-scope caveat is benchmark construction: 'Cases failing the prescribed CFD quality criteria were excluded' (Section 'Benchmark construction from aerodynamic optimization trajectories'). That is a potential selection-bias or external-validity limitation, but it does not make the NK residual reductions definitional or self-referential; the measured convergence on the retained cases is still an empirical result about a fixed 15-update budget. Overall circularity score: 0.

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

The central claims rest on a small number of hand-chosen numerical settings (NK budget, residual threshold, bridge steps, loss weights) and several domain assumptions: the target solver's residual is the ground truth, the mesh topology is fixed, d5 captures geometric shift, and the excluded cases would not change the conclusions. No invented physical entities are introduced.

free parameters (5)
  • NK update budget = 15 (benchmark), 6 (online optimization)
    The convergence and speedup results depend on the number of outer Newton updates allowed; a larger budget improves accuracy but increases cost, and the paper selects these values by hand.
  • Residual convergence threshold = 1e-8
    The residual L2 ratio threshold defining solver acceptance; all residual-ratio statistics are measured against this value.
  • FSB-DiT bridge steps = 5
    The number of deterministic bridge coordinates used during FSB-DiT inference; affects surrogate prediction quality and staged correction timing.
  • Stage-2 loss weights (lambda_Cp, lambda_fpl, lambda_res) = not reported
    The weighting of pressure, perceptual, and residual losses in fine-tuning the surrogate; chosen by hand and affects surrogate accuracy, which is central to the initialization quality.
  • Volume-weighting exponent alpha = not reported
    Controls inverse-volume weighting in the field reconstruction loss; influences near-wall accuracy.
assumptions (5)
  • domain assumption ADflow's discrete residual and Jacobian are the ground-truth solver; a state with residual ratio <= 1e-8 is accepted as the steady solution.
    The central claim equates solver acceptance with correctness; the paper uses ADflow's native residual and JFNK module.
  • domain assumption Fixed mesh topology and structured grids are used across all parameterized geometries; the surrogate outputs complete cell-centered fields on this fixed topology.
    The paper explicitly states 'The evidence in this study is limited to steady aerodynamic CFD with a fixed mesh topology across parameterized geometries' in the Method overview.
  • domain assumption The d5 nearest-neighbour distance is a meaningful measure of geometric distribution shift for the OOD benchmark.
    The benchmark separates OOD from validation based on d5 percentiles; if d5 does not capture the relevant difficulty, the OOD conclusions weaken.
  • ad hoc to paper Cases failing the prescribed CFD quality criteria can be excluded from the benchmark without biasing the OOD evaluation.
    The benchmark construction excludes cases that do not meet quality criteria; this is a post-hoc selection that could remove the hardest cases.
  • domain assumption A surrogate prediction with high alignment between J delta U and -R_pred lies in the Newton convergence basin for a fixed NK budget.
    The mechanistic analysis in Fig. 4 relies on this alignment to explain why NK converges; it is tested empirically but not guaranteed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reliable and efficient steady CFD from surrogate predictions through Newton-Krylov correction." pith.science (2026). https://pith.science/paper/AKAZDLQB

@misc{pith2026260804400,
  author       = {Pith},
  title        = {Pith review of: Reliable and efficient steady CFD from surrogate predictions through Newton-Krylov correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKAZDLQB}},
  note         = {Machine review of arXiv:2608.04400}
}
read the original abstract

Neural surrogates offer a promising route to accelerating computationally expensive simulations governed by partial differential equations across science and industry. Their practical deployment, however, is limited by unreliable predictions under out-of-distribution (OOD) conditions. We develop a solver-coupled surrogate-Newton framework that uses surrogate predictions as high-quality initial guesses for Newton-Krylov iterations, thereby combining rapid global flow-field prediction with high-accuracy numerical convergence at the terminal stage. On an OOD benchmark comprising geometries sampled from actual transonic airfoil optimization trajectories, the framework lowers the median residual L_2 ratio by over seven orders of magnitude while substantially reducing field and aerodynamic errors. In practical supercritical airfoil optimization, it improves online prediction reliability while achieving a 15.5-fold generation-level speedup over CFD. We further test the framework's extension to three dimensions using a flying-wing dataset. Together, these studies demonstrate the potential of surrogate-Newton coupling to deliver accurate, efficient and scalable steady CFD across industrial workflows.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 17 canonical work pages

  1. [1]

    Forrester, A. I. J. & Keane, A. J. Recent advances in surrogate-based optimization. Prog. Aerosp. Sci. 45, 50–79 (2009). https://doi.org/10.1016/j.paerosci.2008.11.001

  2. [2]

    A., Kenway, G

    Mader, C. A., Kenway, G. K. W., Yildirim, A. & Martins, J. R. R. A. ADflow: an open -source computational fluid dynamics solver for aerodynamic and multidisciplinary optimization. J. Aerosp. Inf. Syst. 17, 508–527 (2020). https://doi.org/10.2514/1.I010796 26

  3. [3]

    & Karniadakis, G

    Lu, L., Jin, P., Pang, G., Zhang, Z. & Karniadakis, G. E. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nat. Mach. Intell. 3, 218–229 (2021). https://doi.org/10.1038/s42256-021-00302-5

  4. [4]

    Li, Z. et al. Fourier neural operator for parametric partial differential equations. In Proc. International Conference on Learning Representations (ICLR, 2021). https://openreview.net/forum?id=c8P9NQVtmnO

  5. [5]

    Cheng, Z. et al. Reference neural operators: learning the smooth dependence of solutions of PDEs on geometric deformations. In Proc. 41st International Conference on Machine Learning, PMLR 235, 8060– 8076 (2024). https://proceedings.mlr.press/v235/cheng24c.html

  6. [6]

    Rabeh, A. et al. Benchmarking scientific machine-learning approaches for flow prediction around complex geometries. Commun. Eng. 4, 182 (2025). https://doi.org/10.1038/s44172-025-00513-3

  7. [7]

    Zhou, X.-H. et al. Neural operator-based super-fidelity: a warm-start approach for accelerating steady-state simulations. J. Comput. Phys. 529, 113871 (2025). https://doi.org/10.1016/j.jcp.2025.113871

  8. [8]

    & Karniadakis, G

    Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial different ial equations. J. Comput. Phys. 378, 686–707 (2019). https://doi.org/10.1016/j.jcp.2018.10.045

Show all 29 references
  1. [9]

    Li, Z. et al. Physics-informed neural operator for learning partial differential equations. ACM/IMS J. Data Sci. 1, 1–27 (2024). https://doi.org/10.1145/3648506

  2. [10]

    & Perdikaris, P

    Huang, X. & Perdikaris, P. PhysicsCorrect: a training-free approach for stable neural PDE simulations. Proc. AAAI Conf. Artif. Intell. 40, 22057–22065 (2026). https://doi.org/10.1609/aaai.v40i26.39360

  3. [11]

    Krishnapriyan, A., Gholami, A., Zhe, S., Kirby, R. M. & Mahoney, M. W. Characterizing possible failure modes in physics-informed neural networks. Adv. Neural Inf. Process. Syst. 34, 26548–26560 (2021). https://proceedings.neurips.cc/paper/2021/hash/df438e5206f31600e6ae4af72f27...

  4. [12]

    Zhang, E. et al. Blending neural operators and relaxation methods in PDE numerical solvers. Nat. Mach. Intell. 6, 1303–1313 (2024). https://doi.org/10.1038/s42256-024-00910-x

  5. [13]

    Lee, Y. et al. Fast meta-solvers for 3D complex-shape scatterers using neural operators trained on a non- scattering problem. Comput. Methods Appl. Mech. Eng. 446, 118231 (2025). https://doi.org/10.1016/j.cma.2025.118231

  6. [14]

    R., Holl, P

    Um, K., Brand, R., Fei, Y. R., Holl, P. & Thuerey, N. Solver-in-the-loop: learning from differentiable physics to interact with iterative PDE-solvers. Adv. Neural Inf. Process. Syst. 33, 6111–6122 (2020). https://proceedings.neurips.cc/paper/2020/hash/43e4e6a6f341e00671e123714...

  7. [15]

    & Goswami, S

    Wang, W., Hakimzadeh, M., Ruan, H. & Goswami, S. Time-marching neural operator–FE coupling: AI- accelerated physics modeling. Comput. Methods Appl. Mech. Eng. 446, 118319 (2025). https://doi.org/10.1016/j.cma.2025.118319

  8. [16]

    Ouyang, W., Shin, Y., Liu, S.-W. & Lu, L. NOEM: efficient and scalable finite element method enabled by reusable neural operators. Nat. Comput. Sci. 6, 417–429 (2026). https://doi.org/10.1038/s43588-026-00974- 2

  9. [17]

    Eshaghi, M. S. et al. NOWS: neural operator warm starts for accelerating iterative solvers. Comput. Methods Appl. Mech. Eng. 458, 118989 (2026). https://doi.org/10.1016/j.cma.2026.118989

  10. [18]

    & Karniadakis, G

    Oh, J., Lee, Y., Darbon, J. & Karniadakis, G. E. Spectrally safe neural operator warm -starts for large-scale Newton solvers. Preprint at https://arxiv.org/abs/2606.21828 (2026)

  11. [19]

    Wang, Y. et al. Pretrain finite element method: a pretraining and warm-start framework for PDEs via physics-informed neural operators. J. Mech. Phys. Solids 214, 106682 (2026). https://doi.org/10.1016/j.jmps.2026.106682

  12. [20]

    & Papadopoulos, V

    Nikolopoulos, S., Kalogeris, I., Stavroulakis, G. & Papadopoulos, V. AI-enhanced iterative solvers for accelerating the solution of large-scale parametrized systems. Int. J. Numer. Methods Eng. 125, e7372 (2024). https://doi.org/10.1002/nme.7372

  13. [21]

    Knoll, D. A. & Keyes, D. E. Jacobian-free Newton–Krylov methods: a survey of approaches and applications. J. Comput. Phys. 193, 357–397 (2004). https://doi.org/10.1016/j.jcp.2003.08.010

  14. [22]

    & Karniadakis, G

    Kopaničáková, A. & Karniadakis, G. E. DeepONet based preconditioning strategies for solving para metric linear systems of equations. SIAM J. Sci. Comput. 47, C151–C181 (2025). https://doi.org/10.1137/24M162861X 27

  15. [23]

    Y., Du, T

    Li, Y., Chen, P. Y., Du, T. & Matusik, W. Learning preconditioners for conjugate gradient PDE solvers. In Proc. 40th International Conference on Machine Learning, PMLR 202, 19425–19439 (2023). https://proceedings.mlr.press/v202/li23e.html

  16. [24]

    Yildirim, A., Kenway, G. K. W., Mader, C. A. & Martins, J. R. R. A. A Jacobian-free approximate Newton–Krylov startup strategy for RANS simulations. J. Comput. Phys. 397, 108741 (2019). https://doi.org/10.1016/j.jcp.2019.06.018

  17. [25]

    Taghikhani, K. et al. Neural-initialized Newton: accelerating nonlinear finite elements via operator learning. Int. J. Numer. Methods Eng. 127, e70380 (2026). https://doi.org/10.1002/nme.70380

  18. [26]

    Chisholm, T. T. & Zingg, D. W. A Jacobian-free Newton–Krylov algorithm for compressible turbulent fluid flows. J. Comput. Phys. 228, 3490–3507 (2009). https://doi.org/10.1016/j.jcp.2009.02.004

  19. [27]

    & Heinlein, A

    Zandbergen, A., van Noorden, T. & Heinlein, A. Improving pseudo-time stepping convergence for CFD simulations with neural networks. Comput. Math. Appl. 196, 64–83 (2025). https://doi.org/10.1016/j.camwa.2025.07.006

  20. [28]

    & Zhang, W

    Wang, X., Liu, Y., Yang, X. & Zhang, W. A data-driven convergence booster for accelerating and stabilizing pseudo time-stepping. J. Comput. Phys. 555, 114781 (2026). https://doi.org/10.1016/j.jcp.2026.114781

  21. [29]

    & Long, M

    Wu, H., Luo, H., Wang, H., Wang, J. & Long, M. Transolver: a fast transformer solver for PDEs on general geometries. In Proc. 41st International Conference on Machine Learning , PMLR 235, 53681–53705 (2024). https://proceedings.mlr.press/v235/wu24r.html

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.