pith. sign in

arxiv: 2605.24651 · v1 · pith:UWQPRX76new · submitted 2026-05-23 · 🧮 math.NA · cs.LG· cs.NA

WINO: A Weak-Form Physics Informed Neural Operator for Hyperelasticity on Variable Domains

Bokai Zhu , Qinghui Zhang , Timon Rabczuk This is my paper

Pith reviewed 2026-06-30 12:55 UTC · model grok-4.3

classification 🧮 math.NA cs.LGcs.NA
keywords neural operatorphysics-informedhyperelasticityunfitted finite elementsweak formlevel-set geometrydata-free training
0
0 comments X

The pith

WINO trains a neural operator for hyperelastic problems on variable domains by minimizing weak-form residuals from an unfitted finite element method, without any reference data.

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

The paper introduces a data-free neural operator called WINO that learns displacement solutions for hyperelastic materials whose domains can change shape. It encodes geometry with a level-set function and trains solely by penalizing how much the network output violates the weak form of the equilibrium equations, plus extra terms for cut cells. This removes the usual requirement to generate thousands of converged simulation results as training pairs. If the approach works, it makes operator learning practical for families of problems where both loads and shapes vary, and the resulting predictions can initialize conventional solvers to finish faster.

Core claim

WINO is a neural operator whose parameters are obtained by minimizing the squared residuals of the weak form of the hyperelastic equations discretized with phi-FEM, together with squared penalties on the auxiliary equations that arise on cut cells. For Dirichlet problems only the homogeneous part of the displacement is learned via the phi-FEM lifting; for Neumann problems the network also outputs the auxiliary fields required by the unfitted formulation. Once trained, the operator produces displacement fields whose relative error stays below 0.04 on the reported benchmarks and supplies initial guesses that reduce the iteration count of nonlinear phi-FEM solvers.

What carries the argument

phi-FEM, the unfitted finite-element technique that represents arbitrary domain geometry by a level-set function on a fixed background mesh, allowing the weak residual of the governing equations to be evaluated directly as a training loss without body-fitted remeshing or reference solutions.

If this is right

  • WINO produces solutions whose relative error is below 0.04 on all hyperelastic benchmarks without ever seeing paired reference data.
  • Total wall-clock time for obtaining solutions drops by 50 to 80 percent relative to purely data-driven neural operators.
  • Feeding WINO outputs as initial guesses into nonlinear phi-FEM solvers measurably reduces the number of iterations required for convergence.
  • The same training procedure accommodates both Dirichlet and Neumann boundary conditions on domains whose shapes are encoded by level-set functions.

Where Pith is reading between the lines

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

  • The residual-minimization strategy may transfer to other nonlinear constitutive laws once the corresponding weak form is written.
  • Embedding the operator inside an outer optimization loop over geometry parameters could accelerate shape-design tasks that repeatedly solve hyperelastic problems.

Load-bearing premise

That penalizing the squared weak-form residuals of the phi-FEM discretization is sufficient to determine an accurate mapping from domain and load parameters to displacement fields without any supervised reference data.

What would settle it

Apply the trained WINO to a hyperelastic problem whose domain shape was withheld from training, compute the relative L2 error of the predicted displacement against a converged reference solution obtained by a body-fitted solver, and check whether the error remains below 0.04; an error above that threshold on multiple such cases would falsify the accuracy claim.

Figures

Figures reproduced from arXiv: 2605.24651 by Bokai Zhu, Qinghui Zhang, Timon Rabczuk.

Figure 1
Figure 1. Figure 1: The deformation mapping χ under the load tN . where C = F T F is the right Cauchy-Green deformation tensor, and J = det(F) is the Jacobian deter￾minant of the deformation gradient, representing the local volume ratio of the material before and after deformation. µ and λ are the Lamé parameters that characterize the intrinsic mechanical properties of the material. In this paper, we only formulate the hypere… view at source ↗
Figure 2
Figure 2. Figure 2: φ-FEM meshes and facet sets for an elliptic domain: the exact boundary Γ (purple); facets belonging to Γ G h (green); facets on ∂Ωh (red); active interior cells Th \ T Γ h (blue); cut cells T Γ h (red fill); Cartesian background mesh T O h (overall grid). Additionally, we define Γ G h := ∂Ω Γ h \ ∂Ωh as the facets between the Th \ T Γ h and the T Γ h . To better understand φ-FEM, we illustrate the meshes a… view at source ↗
Figure 3
Figure 3. Figure 3: The architecture of the Fourier neural operator for Dirichlet boundary conditions and Neumann boundary conditions. mesh. In the Neumann setting, we denote D (m) input := (φ (m) h , t (m) h ,f (m) h ). We assume a set of paired samples is given by Ddata = {(D (m) input; u ref,(m) h )} Ns m=1. (2.27) In this supervised setting, we denote the corresponding data-driven neural operator by G data θ , which predi… view at source ↗
Figure 4
Figure 4. Figure 4: Schematic flowchart of WINO training: inputs without labeled data, FNO prediction on nodes, calculation of loss (weak form based residuals), and summation of loss over the Cartesian background mesh. the Gaussian quadrature points of each element from the shape-function gradients and the isoparametric map (2.34), so the weak and squared residuals do not rely on additional numerical differentiation of the ne… view at source ↗
Figure 5
Figure 5. Figure 5: Elliptical shape benchmark. (a) Evolution of the mean relative L 2 errors during training; (b) Relative errors after training; (c) Representative test sample, from left to right: input function fh gh, exact displacement, WINO prediction, and absolute error contour. the outer Newton residual histories, the total number of Newton iterations across all incremental loading steps, and box plots of the total GMR… view at source ↗
Figure 6
Figure 6. Figure 6: NOWS comparison for the elliptical shape case. (a) Outer Newton residual and total Newton iterations for classical φ-FEM (4 load increments) and NOWS. (b) Box plots of total GMRES iteration counts for all load increments. 3.1.2 Random shape In this part, we consider the hyperelastic problem (3.4) on the random shape domain. The level-set function φ is constructed from a sum of three Gaussian functions, φ(x… view at source ↗
Figure 7
Figure 7. Figure 7: Random shape benchmark. (a) Five test samples of the level-set functions and corresponding geometries; (b) Evolution of the mean relative L 2 errors during training; (c) Relative errors after training; (d) Box plots of the relative L 2 errors for WINO with different mesh sizes (from 16 × 16 to 128 × 128) [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: WINO performance for the random shape case: (a) In-distribution test sample; (b) Out-of￾distribution test sample. From left to right: input functions gh, fh, exact displacement field, WINO prediction, and absolute error contour. 16 × 16 to 128 × 128 (Fig. 7d). For the 16 × 16 and 32 × 32 settings, the retained Fourier modes in FNO are set to 6 and 12, respectively. Even on the coarsest 16 × 16 mesh, most s… view at source ↗
Figure 9
Figure 9. Figure 9: NOWS comparison for the random shape case. (a) Outer Newton residual and total Newton iterations for classical φ-FEM (4 load increments) and NOWS. (b) Box plots of total GMRES iteration counts for all load increments. s = 0.1 and l = 0.5. The prediction and absolute error are shown in Fig. 8b. Under this OOD setting, errors are moderately larger than in-distribution errors, while the predicted displacement… view at source ↗
Figure 10
Figure 10. Figure 10: The geometry and mesh tags for the plate with a hole case. 3.2.1 Plate with a hole In this subsection, we consider the hyperelastic problem on a square plate with an elliptic hole, which is formulated as −∇ · P(F(u)) = 0, in Ω, u = 0, on ΓD, P(F(u)) · n = tN , on ΓT , P(F(u)) · n = 0, on ΓN , (3.14) where the background domain is the square O = [0, 1]2 , and the physical domain Ω ⊂ O is a square plate wit… view at source ↗
Figure 11
Figure 11. Figure 11: The evolution of the mean relative L 2 errors during training progress (left) and the relative errors after training is displayed in the box plot (right) for the plate with a hole case [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Representative test sample for the plate with a hole case. (a) Input boundary condition tN and WINO-predicted deformed configuration; (b) x- and y-displacement components for the exact solutions (top row), WINO predictions (middle row), and pointwise absolute error contours (bottom row). tractions. As a result, we interrogate the trained operator by evaluating WINO on a monotonically increasing sequence o… view at source ↗
Figure 13
Figure 13. Figure 13: WINO generalization for the plate with a hole case. (a) WINO-predicted displacement fields for increasing nominal traction levels ty ∈ {15, 25, 35, 45}; (b) Out-of-distribution test: exact solution (left), WINO prediction (middle), and absolute error contour (right). accuracy gains. Inference cost is similar for all three models (about 5 ms per forward pass in our setup), reflecting the fact that they sha… view at source ↗
Figure 14
Figure 14. Figure 14: NOWS comparison for the plate with a hole case. (a) Outer Newton residual history for φ-FEM and NOWS. (b) Box plots of the total GMRES iteration counts. system, which is solved by restarted GMRES (restart m = 80, relative tolerance rtol = 10−5 ) in every Newton iteration in this case. In practice, the 64 × 64 discrete fields returned by WINO, including the displacement uh, the auxiliary variables yh and p… view at source ↗
Figure 15
Figure 15. Figure 15: Cook’s membrane case. (a) Geometry of the Cook’s membrane. (b) Evolution of the mean relative L 2 error during training. (c) Box plot of relative errors after training [PITH_FULL_IMAGE:figures/full_fig_p030_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Representative test sample for the Cook’s membrane case. (a) WINO-predicted deformed configuration; (b) x- and y-displacement components for the exact solutions (top row), WINO predictions (middle row), and pointwise absolute error contours (bottom row). Q9 elements on quadrilateral meshes). Owing to strong nonlinearity in this membrane benchmark, the Newton linearizations are solved with sparse LU factor… view at source ↗
Figure 17
Figure 17. Figure 17: WINO predicted displacement fields for increasing traction levels ty = {5, 10, 15}. Remark 3.3 Since this benchmark involves mixed Dirichlet and Neumann conditions, the loaded right edge is taken as the vertical segment at x = 4 between (4, 1.4) and (4, 2), while the two slanted sides are encoded by the level-set product (3.18). We use a uniform 51×51 Cartesian background on O = [0, 4]×[0, 2] and fix the … view at source ↗
Figure 18
Figure 18. Figure 18: Pressure vessel case. (a) Geometry of the pressure vessel. (b) Evolution of the mean relative L 2 error during training. (c) Box plot of relative errors after training. (d) convergence of φ-FEM: relative L 2 , H1 seminorm, and energy-norm errors versus mesh size [PITH_FULL_IMAGE:figures/full_fig_p033_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Representative test sample for the pressure vessel case. (a) WINO-predicted deformed con￾figuration; (b) x- and y-displacement components for the exact solutions (top row), WINO predictions (middle row), and pointwise absolute error contours (bottom row). Here NQ9 is the Q9 shape-function matrix. Accordingly, the network inputs and outputs are represented on a 101 × 101 grid, while the loss is computed on… view at source ↗
read the original abstract

We propose a Weak-form Physics-Informed Neural Operator (WINO), a data-free framework that combines the efficiency of neural operators with the geometric flexibility of the $\varphi$-finite element method ($\varphi$-FEM). $\varphi$-FEM is an unfitted method that accommodates geometric variations without body-fitted meshes, where the domain geometry is represented by the level-set function $\varphi$. To impose the boundary conditions, Dirichlet problems adopt the $\varphi$-FEM lifting so only the homogeneous displacement contribution is learned, whereas traction-driven Neumann problems additionally predict the auxiliary fields necessary for the unfitted weak formulation. Parameters are trained by minimizing squared weak-form residuals aligned with $\varphi$-FEM together with squared penalties on the cut-cell auxiliary equations, which removes the need for large paired datasets of converged reference solutions. After training, WINO outputs can seed the nonlinear $\varphi$-FEM solvers as neural operator warm starts (NOWS), which reduce iteration counts relative to traditional cold-started solvers. Numerical benchmarks show that WINO achieves high accuracy below 0.04 across all benchmarks, while reducing total computational time by 50--80\% compared with purely data-driven methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes WINO, a weak-form physics-informed neural operator for hyperelasticity on variable domains. It combines neural operators with the unfitted ϕ-FEM discretization (level-set geometry representation), trains data-free by minimizing squared weak-form residuals plus penalties on cut-cell auxiliary equations, and uses the resulting operator outputs as warm starts (NOWS) for nonlinear ϕ-FEM solvers. Benchmarks are claimed to reach errors below 0.04 with 50-80% total time reduction versus purely data-driven approaches.

Significance. If the data-free training via the composite weak-form loss is shown to be consistent for nonlinear hyperelasticity and to generalize across held-out level-set domains, the approach would offer a useful bridge between physics-informed operators and unfitted FEM, reducing reliance on paired reference data while accelerating nonlinear solves via warm starts.

major comments (2)
  1. [Abstract] Abstract: the central claim that minimizing squared weak-form residuals aligned with ϕ-FEM (plus cut-cell penalties) yields accurate solutions without any paired reference data rests on an unverified assumption that the true displacement is a global minimizer of the composite loss for nonlinear strain-energy functions; no derivation of consistency or analysis of spurious minima is provided.
  2. [Abstract] Abstract: the reported accuracy below 0.04 is not accompanied by any indication that the benchmarks were performed on held-out variable domains (rather than training geometries), which is required to substantiate the operator's generalization claim across level-set-defined domains.
minor comments (1)
  1. [Abstract] Abstract: the phrase 'high accuracy below 0.04' is imprecise without specifying the error measure (e.g., L2 or energy norm) or the precise benchmark problems.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on the theoretical justification and the need to clarify generalization in the benchmarks. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that minimizing squared weak-form residuals aligned with ϕ-FEM (plus cut-cell penalties) yields accurate solutions without any paired reference data rests on an unverified assumption that the true displacement is a global minimizer of the composite loss for nonlinear strain-energy functions; no derivation of consistency or analysis of spurious minima is provided.

    Authors: We agree that the manuscript provides no formal derivation of consistency nor analysis of spurious minima for the nonlinear hyperelastic composite loss. The approach is motivated by the variational structure of the weak form (zero residual at the true solution) and is supported only by empirical results. We will revise the manuscript to add an explicit discussion of this assumption, its potential limitations for nonlinear problems, and references to related analyses in weak-form PINNs. revision: yes

  2. Referee: [Abstract] Abstract: the reported accuracy below 0.04 is not accompanied by any indication that the benchmarks were performed on held-out variable domains (rather than training geometries), which is required to substantiate the operator's generalization claim across level-set-defined domains.

    Authors: The full manuscript reports the accuracy below 0.04 on held-out level-set domains unseen during training, as detailed in the numerical experiments section. This detail is omitted from the abstract. We will revise the abstract to state explicitly that the reported accuracy holds on held-out variable domains. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation grounded in external weak-form physics

full rationale

The paper's core training procedure minimizes squared residuals of the weak form taken directly from the established ϕ-FEM formulation (an unfitted method using level-set geometry), together with auxiliary penalties. This loss is defined by the PDE weak form and boundary conditions, which are independent inputs not derived from or fitted to the neural operator outputs. No equations reduce the claimed solution to a self-definition, a renamed fit, or a self-citation chain; the method is benchmarked against reference solutions on variable domains. The approach follows standard physics-informed operator learning and remains self-contained against external verification.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review; no explicit free parameters or invented entities identified in the provided text.

axioms (1)
  • domain assumption The weak form of the hyperelastic equations is a valid way to enforce the physics.
    The training relies on this standard assumption in variational methods.

pith-pipeline@v0.9.1-grok · 5743 in / 1200 out tokens · 48955 ms · 2026-06-30T12:55:31.585278+00:00 · methodology

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

34 extracted references · 6 canonical work pages · 4 internal anchors

  1. [1]

    Nonlinear solid mechanics: a continuum approach for engineering science,

    G. A. Holzapfel, “Nonlinear solid mechanics: a continuum approach for engineering science,” 2002

    2002

  2. [2]

    Topics in finite elasticity: hyperelasticity of rubber, elastomers, and biological tissues— with examples,

    M. F. Beatty, “Topics in finite elasticity: hyperelasticity of rubber, elastomers, and biological tissues— with examples,” 1987

    1987

  3. [3]

    O. C. Zienkiewicz, R. L. Taylor, P. Nithiarasu, and J. Zhu, The finite element method . Elsevier, 1977, vol. 3

    1977

  4. [4]

    An introduction to the finite element method,

    J. N. Reddy, “An introduction to the finite element method,” in Dynamics of Earth’s Fluid System . CRC Press, 2026, pp. 199–226

    2026

  5. [5]

    Finite element implementation of incompressible, transversely isotropic hyperelasticity,

    J. A. Weiss, B. N. Maker, and S. Govindjee, “Finite element implementation of incompressible, transversely isotropic hyperelasticity,” Computer methods in applied mechanics and engineering , vol. 135, no. 1-2, pp. 107–128, 1996

    1996

  6. [6]

    J. W. Thomas, Numerical partial differential equations: finite difference methods . Springer Science & Business Media, 2013, vol. 22

    2013

  7. [7]

    DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators

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

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  8. [8]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators,

    L. Lu, P. Jin, G. Pang, Z. Zhang, and G. E. Karniadakis, “Learning nonlinear operators via deeponet based on the universal approximation theorem of operators,” Nature machine intelligence , vol. 3, no. 3, pp. 218–229, 2021

    2021

  9. [9]

    Fourier Neural Operator for Parametric Partial Differential Equations

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anand- kumar, “Fourier neural operator for parametric partial differential equations,” arXiv preprint arXiv:2010.08895, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  10. [10]

    Synergistic learning with multi-task deeponet for efficient pde problem solving,

    V. Kumar, S. Goswami, K. Kontolati, M. D. Shields, and G. E. Karniadakis, “Synergistic learning with multi-task deeponet for efficient pde problem solving,” Neural Networks , vol. 184, p. 107113, 2025. 37

    2025

  11. [11]

    Physics-informed neural operator for learning partial differential equations,

    Z. Li, H. Zheng, N. Kovachki, D. Jin, H. Chen, B. Liu, K. Azizzadenesheli, and A. Anandkumar, “Physics-informed neural operator for learning partial differential equations,” ACM/IMS Journal of Data Science , vol. 1, no. 3, pp. 1–27, 2024

    2024

  12. [12]

    Learning the solution operator of parametric partial dif- ferential equations with physics-informed deeponets,

    S. Wang, H. Wang, and P. Perdikaris, “Learning the solution operator of parametric partial dif- ferential equations with physics-informed deeponets,” Science advances, vol. 7, no. 40, p. eabi8605, 2021

    2021

  13. [13]

    Physics-informed machine learning in biomedical science and engineering,

    N. Ahmadi, Q. Cao, J. D. Humphrey, and G. E. Karniadakis, “Physics-informed machine learning in biomedical science and engineering,” Annual Review of Biomedical Engineering , vol. 28, no. 1, pp. 309–336, 2026

    2026

  14. [14]

    Variational physics-informed neural operator (vino) for solving partial differential equations,

    M. S. Eshaghi, C. Anitescu, M. Thombre, Y. Wang, X. Zhuang, and T. Rabczuk, “Variational physics-informed neural operator (vino) for solving partial differential equations,” Computer Methods in Applied Mechanics and Engineering , vol. 437, p. 117785, 2025

    2025

  15. [15]

    Finite element operator network for solving parametric pdes,

    J. Y. Lee, S. Ko, and Y. Hong, “Finite element operator network for solving parametric pdes,” arXiv preprint arXiv:2308.04690 , 2023

    work page arXiv 2023

  16. [16]

    NOWS: Neural Operator Warm Starts for Accelerating Iterative Solvers

    M. S. Eshaghi, C. Anitescu, N. Valizadeh, Y. Wang, X. Zhuang, and T. Rabczuk, “Nows: Neural operator warm starts for accelerating iterative solvers,” arXiv preprint arXiv:2511.02481 , 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Pretrain finite element method: A pretraining and warm-start framework for pdes via physics-informed neural operators,

    Y. Wang, Z. Hao, M. S. Eshaghi, C. Anitescu, X. Zhuang, T. Rabczuk, and Y. Liu, “Pretrain finite element method: A pretraining and warm-start framework for pdes via physics-informed neural operators,” arXiv preprint arXiv:2601.03086 , 2026

    work page arXiv 2026

  18. [18]

    Level set methods: an overview and some recent results,

    S. Osher and R. P. Fedkiw, “Level set methods: an overview and some recent results,” Journal of Computational physics , vol. 169, no. 2, pp. 463–502, 2001

    2001

  19. [19]

    Level set methods and dynamic implicit surfaces,

    S. Osher, R. Fedkiw, and K. Piechor, “Level set methods and dynamic implicit surfaces,” Appl. Mech. Rev., vol. 57, no. 3, pp. B15–B15, 2004

    2004

  20. [20]

    ϕ-fem: a finite element method on domains defined by level-sets,

    M. Duprez and A. Lozinski, “ ϕ-fem: a finite element method on domains defined by level-sets,” SIAM Journal on Numerical Analysis , vol. 58, no. 2, pp. 1008–1028, 2020

    2020

  21. [21]

    φ-fem, a finite element method on domains defined by level-sets: the neumann boundary case,

    M. Duprez, V. Lleras, and A. Lozinski, “ φ-fem, a finite element method on domains defined by level-sets: the neumann boundary case,” 2020

    2020

  22. [22]

    φ-fem-fno: a new approach to train a neural operator as a fast pde solver for variable geometries,

    M. Duprez, V. Lleras, A. Lozinski, V. Vigon, and K. Vuillemot, “ φ-fem-fno: a new approach to train a neural operator as a fast pde solver for variable geometries,” Communications in Nonlinear Science and Numerical Simulation , p. 109131, 2025

    2025

  23. [23]

    Cut finite element methods,

    E. Burman, P. Hansbo, M. G. Larson, and S. Zahedi, “Cut finite element methods,” Acta Numerica, vol. 34, pp. 1–121, 2025

    2025

  24. [24]

    Bonet and R

    J. Bonet and R. D. Wood, Nonlinear continuum mechanics for finite element analysis . Cambridge university press, 1997

    1997

  25. [25]

    C. T. Kelley, Solving nonlinear equations with Newton ’s method . SIAM, 2003

    2003

  26. [26]

    Nitsol: A newton iterative solver for nonlinear systems,

    M. Pernice and H. F. Walker, “Nitsol: A newton iterative solver for nonlinear systems,” SIAM Journal on Scientific Computing , vol. 19, no. 1, pp. 302–318, 1998. 38

    1998

  27. [27]

    Function minimization by conjugate gradients,

    R. Fletcher and C. M. Reeves, “Function minimization by conjugate gradients,” The computer journal, vol. 7, no. 2, pp. 149–154, 1964

    1964

  28. [28]

    A survey of nonlinear conjugate gradient methods,

    W. W. Hager and H. Zhang, “A survey of nonlinear conjugate gradient methods,” Pacific journal of Optimization, vol. 2, no. 1, pp. 35–58, 2006

    2006

  29. [29]

    Gmres: A generalized minimal residual algorithm for solving non- symmetric linear systems,

    Y. Saad and M. H. Schultz, “Gmres: A generalized minimal residual algorithm for solving non- symmetric linear systems,” SIAM Journal on scientific and statistical computing , vol. 7, no. 3, pp. 856–869, 1986

    1986

  30. [30]

    Gmres with deflated restarting,

    R. B. Morgan, “Gmres with deflated restarting,” SIAM Journal on Scientific Computing , vol. 24, no. 1, pp. 20–37, 2002

    2002

  31. [31]

    Petsc users manual,

    S. Balay, S. Abhyankar, M. Adams, J. Brown, P. Brune, K. Buschelman, L. Dalcin, A. Dener, V. Eijkhout, W. Gropp et al. , “Petsc users manual,” 2019

    2019

  32. [32]

    SOAP: Improving and Stabilizing Shampoo using Adam

    N. Vyas, D. Morwani, R. Zhao, M. Kwun, I. Shapira, D. Brandfonbrener, L. Janson, and S. Kakade, “Soap: Improving and stabilizing shampoo using adam,” arXiv preprint arXiv:2409.11321 , 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  33. [33]

    Gaussian processes in machine learning,

    C. E. Rasmussen, “Gaussian processes in machine learning,” in Summer school on machine learning . Springer, 2003, pp. 63–71

    2003

  34. [34]

    G. H. Golub and C. F. Van Loan, Matrix computations . JHU press, 2013. 39

    2013