Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

DeePoly: A High-Order Accuracy Scientific Machine Learning Framework for Function Approximation and Solving PDEs

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read DeePoly claims that a two-stage strategy — coarse DNN feature extraction followed by linear polynomial refinement — reaches high-order, near-machine-precision accuracy for smooth function approximation and PDEs while keeping the method…

desk verdict DeePoly's two-stage idea is genuinely interesting and the smooth results look right, but the paper overclaims efficiency, never confronts the conditioning problem it admits, and lacks the baseline and theory needed to support its central guarantee. read the letter →

arxiv 2506.04613 v3 pith:ZBADUAWW submitted 2025-06-05 cs.AI cs.LG

classification cs.AIcs.LG MSC 65D1541A1068T07
keywords Deep-PolynomialPhysics-InformedNeuralNetworksScientificMachineLearningHigh-OrderAccuracyPartialDifferentialEquationslinearspaceoptimizationfunctionapproximationpolynomialbasisrefinement
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 proposes DeePoly, a framework for approximating functions and solving PDEs that deliberately stops short of fully training a neural network. First, a DNN is trained just well enough to extract the features of the target (the Spotter step); then the method abandons nonlinear parameter optimization and instead solves a linear least-squares problem in the space spanned by those features plus local polynomials (the Sniper step). The central claim is that this hybrid achieves high-order accuracy — in the smooth benchmarks, errors in the $10^{-13}$ to $10^{-32}$ range versus PINNs' $10^{-4}$ to $10^{-6}$ — while staying mesh-free and scheme-free and needing fewer partitions than classical methods. The paper argues that the reason is theoretical, not just empirical: on each cell a smooth function is approximated by its Taylor polynomial, so the polynomial degree, not the number of partitions, controls the convergence order, and the DNN stabilizes the polynomial rather than carrying the final accuracy.

What carries the argument

The load-bearing mechanism is the combined linear space V = ξ ⊕ P over each subdomain Ω_i. The Spotter's DNN supplies the feature set ξ from its last hidden layer after a short, deliberately low-accuracy training; the Sniper appends the local polynomial basis P = {P_0(x̄),...,P_M(x̄)} with x̄ the normalized coordinate in the cell, then solves the convex least-squares problem min ||Σ γ_n v_n(x) − t(x)||. This converts the original non-convex parameter optimization into a two-stage process where the final error is controlled by polynomial approximation theory — Taylor expansion of a smooth function on a small interval — rather than by the unpredictable quality of a DNN's local minimum.

What would settle it

Run DeePoly on the 1D smooth function t(x)=sin(2πx)+0.5 cos(4πx) with the paper's 10th-order polynomials and 2, 4, and 8 partitions, and measure the convergence order. If the observed order is not roughly 10-13 as reported across different random seeds and sampling patterns, the claimed Taylor-guaranteed convergence is not robust. Alternatively, compute the condition number of the Sniper system matrix for the 32-partition run; if it is large enough that the reported machine-precision errors depend on cancellation, the central accuracy claim is fragile.

Watch

Extended reading notes

Core claim

The central discovery is that the accuracy bottleneck of PINN-style methods is not representation but optimization: a DNN trained to moderate accuracy is enough to capture global features, and the remaining fine error can be removed by a convex linear solve in an enriched space. DeePoly builds that space as V = span{ξ_1,...,ξ_N} ⊕ span{P_0(x̄),...,P_M(x̄)} per partition, where ξ are the last-hidden-layer outputs of the coarsely trained DNN and P_i are normalized local polynomials. Solving the linear least-squares problem in V, the Sniper step, converts the hard non-convex search into a tractable projection, and Taylor expansion of the smooth target on each cell guarantees that the convergence order matches the polynomial degree. The numerical results back this up: smooth functions converge at rates 5 to 13, the Poisson examples reach errors near machine precision, and the method remains stable on discontinuous and nonlinear problems, with the admitted caveat that discontinuous cases do not reach the same high order because of linear-system conditioning.

Load-bearing premise

The high-order accuracy promise rests on the Sniper linear system being well-conditioned and on the target being smooth enough on each partition for the chosen polynomial degree; the paper itself reports that discontinuous problems lose accuracy precisely because discontinuities worsen that conditioning.

Editorial extensions

If this is right

  • Smooth steady PDEs can be solved to near machine precision by a short DNN warm-up followed by one linear solve per partition, avoiding the long training runs that characterize PINNs.
  • Mesh-free and scheme-free properties are preserved, so the accuracy gain does not require constructing meshes, upwinding, or custom discretizations.
  • Partition counts can be dramatically lower than in classic finite-element or finite-difference thinking, because the polynomial degree supplies the convergence rate.
  • The same two-stage logic carries over to time-dependent and nonlinear problems through time-stepping and pseudo-time continuation, with the DNN kept fixed between steps.
  • For discontinuous solutions the method is stable but loses its high order, pointing to conditioning rather than representation as the remaining obstacle.

Reading between the lines

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

  • The framework implicitly redefines the DNN's job: any reasonably trained network can serve as a feature extractor, so the marginal value of expensive fine-tuning is low once the Sniper step exists; this suggests cheap early-stopped networks plus polynomial refinement could replace long training in other contexts.
  • Because accuracy is carried by the polynomial space, the approach should combine naturally with adaptive or error-driven partitioning — refine only where the residual is large — which the authors list as future work but which follows directly from their convergence argument.
  • The admitted conditioning failure on discontinuities suggests a concrete test: pre-multiply the Sniper system by a weight function that down-weights residual equations near shocks and see whether smooth-region accuracy returns to high order; that would isolate the conditioning mechanism from the representational one.
  • Nothing in the argument ties the Sniper basis specifically to ordinary polynomials; any local basis with known approximation order (wavelets, piecewise Hermite) might substitute, and the features themselves could be replaced by any learned representation, so the true contribution is the two-stage division of labor.
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 / 6 minor

Summary. The paper proposes DeePoly, a two-stage framework for function approximation and PDE solving. In the Spotter stage, a DNN is trained to moderate accuracy and its last-hidden-layer features are extracted; in the Sniper stage, these features are combined with local polynomial bases on a partition of the domain, and the combined coefficient vector is obtained by solving a linear least-squares problem. The authors claim that this hybrid approach achieves high-order accuracy (inherited from polynomial approximation) while retaining the mesh-free, scheme-free flexibility of PINNs, and they present numerical experiments for smooth and discontinuous function approximation, Poisson equations, a linear convection equation, the Allen-Cahn equation, and lid-driven cavity flow. The central claims are that DeePoly significantly enhances both accuracy and efficiency and that it inherits convergence guarantees from classical polynomial theory.

Significance. If the claimed high-order accuracy and efficiency were fully supported, the two-stage Spotter/Sniper idea would be a practically valuable contribution to scientific machine learning, giving a route to much higher accuracy than vanilla PINNs on smooth problems while keeping the mesh-free character. The paper deserves credit for a clear and simple architecture, an open-source implementation (at least for function approximation), and 1D/2D smooth experiments that exhibit convergence orders consistent with the polynomial degree. However, the significance is substantially tempered by three issues: the efficiency claim is contradicted by the paper's own timing tables; the reported machine-precision errors in the smooth PDE examples may be artifacts of underdetermined least-squares interpolation rather than genuine approximation accuracy; and the theoretical section does not actually prove a convergence theorem for the DeePoly method, instead restating classical polynomial approximation results. The paper therefore reads more as a promising numerical/software exposition than as a rigorous accuracy-and-efficiency analysis.

major comments (3)
  1. [§3.3.1, Tables 5 and 6] The efficiency claim in the abstract and introduction ('significantly enhances both high-order accuracy and efficiency') is not supported by the reported timings. In Table 5, DeePoly takes 18.06 s versus PINNs' 13.73 s for Case 1, and in Table 6 it takes 49.10 s versus 34.73 s for Case 2; Table 7 also shows DeePoly slightly slower (17.45 s vs. 16.83 s). A fair efficiency comparison should be time-to-accuracy (e.g., time to reach a target error), not wall-clock time at different final accuracies. As written, the data show a slowdown, not an efficiency improvement.
  2. [§3.1 and §3.3.1] The reported training and test errors for the smooth Poisson problems are likely not genuine approximation errors but instead reflect interpolation in underdetermined least-squares solves. With 10×10 partitions and 5000 training points, each cell contains roughly 50 collocation points, while the Sniper space V has 25 DNN features plus 36 polynomial basis functions (degree 5 in 2D), giving 61 unknowns per cell. The system is underdetermined, so training MSE near 1e-32 can be achieved by interpolating the cell's collocation points. The test grid 50×50 places about 25 points per cell, again comparable to the number of unknowns. The paper should report the condition number of the Sniper least-squares matrix, specify the regularization used, and show results on a genuinely independent or much finer test grid to rule out interpolation artifacts.
  3. [§4 and §3.3.2] The theoretical analysis does not provide a theorem for the DeePoly method. The claimed 'guarantee of Taylor expansion' (Figure 1 caption) is a restatement of classical polynomial approximation theory, and no proof is given that the two-stage procedure with a DNN feature space ξ achieves the polynomial convergence order, nor are the required regularity and conditioning assumptions stated. This is not merely a presentation gap: the paper's own §3.3.2 and Conclusion item 3 ascribe the poor accuracy on discontinuous problems to 'conditioning properties of the large linear system,' an explicit admission that the advertised accuracy mechanism is conditioning-limited exactly where the 'diverse problem types' claim is made. The central accuracy promise is therefore conditional on a conditioning assumption that is neither analyzed nor numerically checked.
minor comments (6)
  1. [Tables 1-4] The 'Order' column is computed between successive rows, but the definition of 'order' is not stated in any table caption; clarify whether it is the log-ratio of errors divided by log-ratio of partition counts, point counts, or cell widths.
  2. [Table 4] The 'Points' entries such as '200001' and '200002' appear to concatenate the point count with a partition count (e.g., 20000 points in 1 partition vs. 2 partitions). This is confusing and should be split into separate 'Points' and 'Sections' columns.
  3. [§5] The open-source repository currently contains only the function approximation examples; the PDE examples that support the main claims are not publicly available, which limits reproducibility of the central results.
  4. [§3.1] The 2D target function is not explicitly assigned to a variable; the parenthetical note in the text confirms this. The equation should be written as a proper definition, e.g., t2(x,y)=...
  5. [Abstract] There is a typo in the abstract: 'open-source projectDeePoly' is missing a space before and after 'project'.
  6. [Conclusion] Conclusion item 3 repeats the conditioning limitation already stated in §3.3.2; consider consolidating limitations into a single dedicated paragraph in the main text so the reader sees the scope of the guarantee before the numerical section.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the high-order accuracy claim is explicitly inherited from the polynomial basis and no fitted parameter is renamed as a prediction.

full rationale

Walking the derivation chain, the claimed high-order convergence is explicitly inherited from the polynomial basis P in the Sniper space: the Figure 1 caption states the accuracy is 'same to the order of the polynomial basis' and is guaranteed 'by Taylor expansion of smooth function.' The experimental 'Order' columns are explained by the chosen polynomial degree (e.g., 'This example employs 10th-order polynomial basis functions, which explains the approximately 10th-order convergence accuracy'), so the paper does not present a fitted parameter as a prediction or derive the order from the DNN component. No equation defines a target quantity in terms of the fitted output; the Sniper least-squares projection onto V = ξ⊕P is a standard linear approximation step, and its error bound follows from the classical approximation properties of P. The only self-citation is [19], which appears in future-work remarks about conditioning and is not load-bearing for the central claim. The conditioning and regularity gaps noted in Sections 3.3.2 and the Conclusion are correctness/robustness limitations, not circular reductions. Therefore no circular step meeting the quoted-evidence standard is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method's headline convergence order is inherited from the user-chosen polynomial degree and classical approximation theory; the DNN's contribution is asserted but not isolated. No new physical entities are introduced.

free parameters (3)
  • Polynomial degree M = 10 for 1D; [5,5] for 2D
    User-chosen; observed convergence order matches M, so the 'high-order' result is imposed by this choice rather than derived.
  • Number of partitions per dimension = 1-32 (1D); 2-10 per dimension (2D)
    User-chosen; convergence orders are measured across these chosen grids, so the reported orders depend on the selected partition schedule.
  • DNN architecture and hidden dimensions = [12,32,32,25] in examples
    Chosen by hand; no sensitivity study is provided, so the method's dependence on the Spotter network size is unknown.
assumptions (4)
  • standard math Smooth functions can be approximated with error on the order of h^(M+1) by degree-M polynomials on small subdomains (Taylor/Chebyshev theory).
    Underlies the claimed convergence order; invoked in Section 4 and Figure 1(e).
  • domain assumption Features extracted from a moderately trained DNN form a basis that stabilizes polynomial approximation and remains useful after freezing.
    No ablation or proof is given; the entire claimed benefit of the Spotter rests on this assumption.
  • domain assumption The Step 2 linear least-squares system is well-conditioned enough to realize the theoretical accuracy.
    The paper notes in Section 3.3.2 and the Conclusion that conditioning limits accuracy for discontinuous problems, so this assumption is not always valid.
  • domain assumption Target solutions have sufficient local regularity on each partition for the polynomial bases to converge at the stated order.
    Required for polynomial convergence; not verified beyond the specific test problems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeePoly: A High-Order Accuracy Scientific Machine Learning Framework for Function Approximation and Solving PDEs." pith.science (2026). https://pith.science/paper/ZBADUAWW

@misc{pith2026250604613,
  author       = {Pith},
  title        = {Pith review of: DeePoly: A High-Order Accuracy Scientific Machine Learning Framework for Function Approximation and Solving PDEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZBADUAWW}},
  note         = {Machine review of arXiv:2506.04613}
}
read the original abstract

Recently, machine learning methods have gained significant traction in scientific computing, particularly for solving Partial Differential Equations (PDEs). However, methods based on deep neural networks (DNNs) often lack convergence guarantees and computational efficiency compared to traditional numerical schemes. This work introduces DeePoly, a novel framework that transforms the solution paradigm from pure non-convex parameter optimization to a two-stage approach: first employing a DNN to capture complex global features, followed by linear space optimization with combined DNN-extracted features (Spotter) and polynomial basis functions (Sniper). This strategic combination leverages the complementary strengths of both methods -- DNNs excel at approximating complex global features (i.e., high-gradient features) and stabilize the polynomial approximation while polynomial bases provide high-precision local corrections with convergence guarantees. Theoretical analysis and numerical experiments demonstrate that this approach significantly enhances both high-order accuracy and efficiency across diverse problem types while maintaining mesh-free and scheme-free properties. This paper also serves as a theoretical exposition for the open-source project DeePoly.

Figures

Figures reproduced from arXiv: 2506.04613 by the authors.

Figure 1
Figure 1. Graphic illustration of DeePoly (a) The local minimas are close to the global minimum however they are separated in parameter space. (b) It is effective to find a relatively good local minima however it expensive to traverse from one local minima to another. (c) Spotter works as a features extractor to obtain similar features of the target, without the requirement of high-precision approximation (d) Sniper works loc… view at source ↗
Figure 2
Figure 2. 1D smooth function example result with 32 partitions. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. 2D smooth function example result with 10x10 partitions. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: 1D discontinuous function result. Next, we test a 2D example with complex large gradients. The target function is defined as:    f(x, y) = z(x, y) + g(x, y) z(x, y) = h tanh(s · d(x, y)) d(x, y) = y − A sin(fπx) g(x, y) = 0.4 sin(3x) cos(2y) with parameters ar…
Figure 5
Figure 5. Figure 5: 2D large gradient problem result (10x10 partitions). [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: 2D Poisson equation with sin(πx) sin(πy) source term 14 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: 2D Poisson equation with sin(4πx) sin(4πy) source term 3.3.2 Linear Convection Equation Then we consider the linear convection equation with a large gradient initial condition: ∂u ∂t + a ∂u ∂x = 0 u(x, 0) = tanh(100 ∗ (x − 0.3)) where a = 0.3 is the convection speed. T…
Figure 8
Figure 8. Figure 8: Linear convection equation with discontinuous initial condition. The method maintains [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Allen-Cahn equation result at T=1. 3.4.2 Navier-Stokes Equation - Lid-Driven Cavity Flow The incompressible Navier-Stokes equations represent one of the most fundamental and chal￾lenging systems in fluid dynamics, described by: ∂u ∂t + (u · ∇)u = −∇p + 1 Re ∇2u ∇ · u =…
Figure 10
Figure 10. Figure 10: Final converged solution of Re=100 lid-driven cavity flow showing the velocity field [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]
Figure 11
Figure 11. Figure 11: Streamline plot of the Re=100 lid-driven cavity flow, showing the main circulation [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]
Figure 12
Figure 12. Figure 12: Conceptual optimization landscape for approximating y=x with [PITH_FULL_IMAGE:figures/full_fig_p024_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CLINN: Conservation Law Informed Neural Network for Approximating Discontinuous Solutions

    math.NA 2025-09 conditional novelty 4.0 of 10

    CLINN combines implicit-form, boundedness, and Rankine-Hugoniot loss terms with adaptive refinement, cutting MSE by up to 99.2% on scalar conservation law benchmarks.

Reference graph

Works this paper leans on

31 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    Random matrices and complexity of spin glasses.Communications on Pure and Applied Mathematics, 66(2):165–201, 2013

    Antonio Auffinger, Gérard Ben Arous, and Jiří Čern` y. Random matrices and complexity of spin glasses.Communications on Pure and Applied Mathematics, 66(2):165–201, 2013

  2. [2]

    SIAM, 2018

    Léon Bottou, Frank E Curtis, and Jorge Nocedal.Optimization methods for large-scale machine learning, volume 60. SIAM, 2018

  3. [3]

    Statistics of critical points of gaussian fields on large- dimensional spaces.Physical review letters, 98(15):150201, 2007

    Alan J Bray and David S Dean. Statistics of critical points of gaussian fields on large- dimensional spaces.Physical review letters, 98(15):150201, 2007

  4. [4]

    Optimization of random feature method in the high-precision regime.Communications on Applied Mathematics and Computation, 6(2):1490–1517, 2024

    Jingrun Chen, Weinan E, and Yifei Sun. Optimization of random feature method in the high-precision regime.Communications on Applied Mathematics and Computation, 6(2):1490–1517, 2024

  5. [5]

    The loss surfaces of multilayer networks

    Anna Choromanska, Mikael Henaff, Michael Mathieu, Gérard Ben Arous, and Yann LeCun. The loss surfaces of multilayer networks. InArtificial intelligence and statistics, pages 192–

  6. [6]

    Approximation by superpositions of a sigmoidal function.Mathematics of control, signals and systems, 2(4):303–314, 1989

    George Cybenko. Approximation by superpositions of a sigmoidal function.Mathematics of control, signals and systems, 2(4):303–314, 1989

  7. [7]

    Hierarchical extreme learning machine for solving partial differential equations.Available at SSRN 4775113

    Daiwei Dong, Jiaqing Kou, Wei Suo, and Weiwei Zhang. Hierarchical extreme learning machine for solving partial differential equations.Available at SSRN 4775113

  8. [8]

    High-re solutions for incompressible flow using the navier- stokes equations and a multigrid method.Journal of Computational Physics, 48(3):387–411, 1982

    U Ghia, K N Ghia, and C T Shin. High-re solutions for incompressible flow using the navier- stokes equations and a multigrid method.Journal of Computational Physics, 48(3):387–411, 1982

Show all 31 references
  1. [9]

    Multilayer feedforward networks are universal approximators.Neural networks, 2(5):359–366, 1989

    Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators.Neural networks, 2(5):359–366, 1989

  2. [10]

    Trends in extreme learning machines: A review.Neural Networks, 61:32–48, 2015

    Gao Huang, Guang-Bin Huang, Shiji Song, and Keyou You. Trends in extreme learning machines: A review.Neural Networks, 61:32–48, 2015

  3. [11]

    Extreme learning machine: theory and applications.Neurocomputing, 70(1-3):489–501, 2006

    Guang-Bin Huang, Qin-Yu Zhu, and Chee-Kheong Siew. Extreme learning machine: theory and applications.Neurocomputing, 70(1-3):489–501, 2006

  4. [12]

    Ameya D Jagtap and George Em Karniadakis. Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations.Communications in Computational Physics, 28(5):2002–2041, 2021

  5. [13]

    Adaptive activation functions accelerate convergence in deep and physics-informed neural networks.Journal of Computational Physics, 404:109136, 2020

    Ameya D Jagtap, Kenji Kawaguchi, and George Em Karniadakis. Adaptive activation functions accelerate convergence in deep and physics-informed neural networks.Journal of Computational Physics, 404:109136, 2020

  6. [14]

    Nature Reviews Physics, 2021

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang.Physics-informed machine learning. Nature Reviews Physics, 2021. 26

  7. [15]

    Deeplearningwithoutpoorlocalminima.Advances in neural information processing systems, 29, 2016

    KenjiKawaguchi. Deeplearningwithoutpoorlocalminima.Advances in neural information processing systems, 29, 2016

  8. [16]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

  9. [17]

    Deep domain decomposition method: Elliptic problems.arXiv preprint arXiv:2004.13900, 2020

    Dingcheng Li, Kailiang Xu, Jerry M Harris, and Eric Darve. Deep domain decomposition method: Elliptic problems.arXiv preprint arXiv:2004.13900, 2020

  10. [18]

    KAN: Kolmogorov-Arnold networks.arXiv preprint arXiv:2404.19756, 2024

    Guido F Liu, Changshuo Zhu, Ling Li, Shanghua Zhao, Albert Gu, Jimmy Gao, Liping Liu, Sungik Ahn, Tuo Zhao, Anuran Makur, et al. KAN: Kolmogorov-Arnold networks.arXiv preprint arXiv:2404.19756, 2024

  11. [19]

    Discontinuity computing using physics-informed neural networks.Journal of Scientific Computing, 98(1):22, 2024

    Li Liu, Shengping Liu, Hui Xie, Fansheng Xiong, Tengchao Yu, Mengjuan Xiao, Lufeng Liu, and Heng Yong. Discontinuity computing using physics-informed neural networks.Journal of Scientific Computing, 98(1):22, 2024

  12. [20]

    Lower bounds for approximation by mlp neural networks

    Vitaly Maiorov and Allan Pinkus. Lower bounds for approximation by mlp neural networks. Neurocomputing, 25(1-3):81–91, 1999

  13. [21]

    Cambridge University Press, 1999

    Allan Pinkus.Approximation theory of the MLP model in neural networks. Cambridge University Press, 1999

  14. [22]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural net- works: A deep learning framework for solving forward and inverse problems involving non- linear partial differential equations.Journal of Computational Physics, 378:686–707, 2019

  15. [23]

    Khemraj Shukla, Juan Diego Toscano, Zhicheng Wang, Zongren Zou, and George Em Karniadakis. A comprehensive and fair comparison between mlp and kan representations for differential equations and operator networks.Computer Methods in Applied Mechanics and Engineering, 431:117290, 2024

  16. [24]

    Optimization for deep learning: theory and algorithms.arXiv preprint arXiv:1912.08957, 2020

    Ruoyu Sun, Mikhail Khodak, and Nate Kushman. Optimization for deep learning: theory and algorithms.arXiv preprint arXiv:1912.08957, 2020

  17. [25]

    Fourierfeatures let networks learn high frequency functions in low dimensional domains

    Matthew Tancik, Pratul P Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Ragha- van, UtkarshSinghal, RaviRamamoorthi, JonathanTBarron, andRenNg. Fourierfeatures let networks learn high frequency functions in low dimensional domains. InAdvances in Neural Information Pro...

  18. [26]

    Gradient align- ment in physics-informed neural networks: A second-order optimization perspective.arXiv preprint arXiv:2502.00604, 2025

    Sifan Wang, Ananyae Kumar Bhartari, Bowen Li, and Paris Perdikaris. Gradient align- ment in physics-informed neural networks: A second-order optimization perspective.arXiv preprint arXiv:2502.00604, 2025

  19. [27]

    Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM Journal on Scientific Computing, 43(5):A3055–A3081, 2021

    Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM Journal on Scientific Computing, 43(5):A3055–A3081, 2021

  20. [28]

    When and why PINNs fail to train: A neural tangent kernel perspective.Journal of Computational Physics, 449:110768, 2022

    Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why PINNs fail to train: A neural tangent kernel perspective.Journal of Computational Physics, 449:110768, 2022

  21. [29]

    Multi-stage neural networks: Function approximator of machine precision.Journal of Computational Physics, 504:112865, 2024

    Yongji Wang and Ching-Yao Lai. Multi-stage neural networks: Function approximator of machine precision.Journal of Computational Physics, 504:112865, 2024

  22. [30]

    Eigenvector bases for neural networks

    Yuzhe Wang, Mingming Sun, and Björn Eriksson. Eigenvector bases for neural networks. InInternational Conference on Learning Representations, 2020. 27

  23. [31]

    Towards understand- ing the condensation of neural networks at initial training.Advances in Neural Information Processing Systems, 35:2184–2196, 2022

    Hanxu Zhou, Zhou Qixuan, Tao Luo, Yaoyu Zhang, and Zhi-Qin Xu. Towards understand- ing the condensation of neural networks at initial training.Advances in Neural Information Processing Systems, 35:2184–2196, 2022. 28

Pith tools

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