Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Transformed Diffusion-Wave fPINNs: Enhancing Computing Efficiency for PINNs Solving Time-Fractional Diffusion-Wave Equations

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

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

Pith's one-line read A change of integrand makes fractional diffusion-wave PINNs cheaper without changing the Caputo operator it computes.

desk verdict The Type-II transformation is a genuine efficiency win for smooth problems, but the 'without sacrificing accuracy' claim is contradicted by the paper's own L-shaped-domain results and should be narrowed before acceptance. read the letter →

arxiv 2506.11518 v2 pith:2BQAZOFG submitted 2025-06-13 math.NA cs.NA

Jing Li , Zhengqi Zhang This is my paper
classification math.NAcs.NA MSC 65D3035R1168T0765M70
keywords fractionalphysics-informedneuralnetworksCaputoderivativediffusion-waveequationGauss-JacobiquadratureMonteCarlointegrationintegrandtransformationcomputationalefficiencymesh-freemethods
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 targets a specific computational bottleneck in fractional physics-informed neural networks (fPINNs): for Caputo time derivatives of order \(\$\alpha$\in(1,2)\), evaluating the residual at a collocation point requires the neural network's first time derivative at \(M\) shifted quadrature points, an expensive automatic-differentiation load. The authors rewrite the Caputo integral so that the quadrature instead evaluates the network's plain values at those shifted points, while remaining exactly equal to the original operator. They prove the equality, show the endpoint singularities of the new integrand are removable, and derive Monte Carlo and Gauss-Jacobi convergence rates together with explicit storage and FLOP savings. Numerical tests on one- and two-dimensional diffusion-wave problems, a Burgers equation, and an L-shaped domain show the transformed schemes match direct-scheme accuracy while using less time and GPU memory. If the claims hold, tDWfPINNs give a strict efficiency improvement over direct quadrature fPINNs for this problem class, with the savings growing with collocation count, quadrature count, network width and depth.

What carries the argument

The load-bearing object is the Type-II transformed representation of Theorem 3.1, equation (10): \($D^{{\alpha}}$_{\mathrm{II}} f(t) := \partial_t^\$\alpha$ f(t)\) expressed through the kernel \(H_f(t,\tau)=\frac{f(t)-f(t-t\tau)-t\tau f'(t)}{(t\tau)^2}\) and the boundary term \(\frac{f(t)-f(0)-t f'(t)}{t^\$\alpha$}\). The argument runs through three lemmas: Lemma 3.1 shows \(H_f\) extends continuously to \(\tau=0\) via the second-order Taylor integral remainder so the weight \(\$tau^{{1-\alpha}}$\) stays integrable on \([0,1]\); Corollary 3.2 bounds the difference of transformed operators by \(\|f''-g''\| $T^{{2-\alpha}}$/\Gamma(3-\$\alpha$)\), matching the direct form; and Theorem 3.2 converts kernel smoothness into Gauss-Jacobi convergence, algebraic \($M^{{-r}}$\) for finitely smooth kernels and geometric \(\$rho^{{-2M}}$\) for kernels analytic on a Bernstein ellipse. The same kernel regularity is what makes the Gauss-Jacobi version, GJ-II, the empirically cheapest scheme.

What would settle it

Take the time-fractional diffusion-wave equation with exact solution \(u(t,x)=t^\$\beta$ \sin(\pi x)\) for fractional order \(\$\alpha$=1.5\) and \(1<\$\beta$<2\) (so the solution is \($C^{1}$\) but not \($C^{2}$\) in time), and compare the \($L^{2}$\) error of GJ-II against GJ-I as the quadrature order \(M\) increases from 16 to 512. If the GJ-II error plateaus or grows while GJ-I continues to improve, then the transformed scheme fails to preserve accuracy outside the smooth class the paper's Appendix A.6 flags.

Watch

Extended reading notes

Core claim

The central discovery is that the Caputo derivative of order \(\$\alpha$\in(1,2)\) can be represented in an alternative, exactly equal form whose quadrature integrand contains \(f(t-t\tau)\) rather than \(f'(t-t\tau)\). Specifically, Theorem 3.1 proves that \(\partial_t^\$\alpha$ f(t)\) equals a boundary term involving \(f'(t)\) and \(f'(0)\) plus an integral of the second-order Taylor remainder \(H_f(t,\tau)=\frac{f(t)-f(t-t\tau)-t\tau f'(t)}{(t\tau)^2}\) against the weight \(\$tau^{{1-\alpha}}$\). Because \(H_f\) extends continuously to \(\tau=0\) with value \(-\tfrac12 f''(t)\), the apparent \((t\tau)^{-2}\) singularity is removable, the weighted integral is well defined, and the representation inherits the same stability bound as the original operator. Replacing the direct Type-I kernel \(K_f(t,\tau)=\frac{f'(t)-f'(t-t\tau)}{t\tau}\) by \(H_f\) converts each of the \(M\) shifted quadrature evaluations from a time-derivative graph to an ordinary value graph, cutting graph storage by \(8NM(d+2LH)\) bytes and training-backward FLOPs by \(NM(4A_{\mathrm{mac}}+5LH)\) for a fully connected network of depth \(L\) and width \(H\).

Load-bearing premise

The accuracy-preservation claim rests on the kernels \(K_f(t,\cdot)\) and \(H_f(t,\cdot)\) being smooth enough on \([0,1]\) that Gauss-Jacobi quadrature converges quickly; when the underlying solution loses temporal regularity the quadrature rate degrades and the numerical evidence for 'no accuracy sacrifice' weakens.

Editorial extensions

If this is right

  • For fully connected tanh networks, every shifted quadrature point in the residual can be changed from a time-derivative graph to an ordinary value graph, saving \(8NM(d+2LH)\) bytes of graph storage and \(NM(4A_{\mathrm{mac}}+5LH)\) training-backward FLOPs without changing the fractional operator.
  • Gauss-Jacobi quadrature on the transformed kernels converges faster than Monte Carlo for smooth solutions, so GJ-II typically needs far fewer quadrature nodes \(M\) for the same accuracy, making it the most favorable accuracy-cost scheme in the tested settings.
  • The transformed residual can be evaluated at arbitrary scattered collocation points, so it composes naturally with residual-based adaptive sampling (RAD) and with irregular geometries such as an L-shaped domain.
  • In the same point budget, the mesh-free GJ-II formulation with adaptive resampling outperforms standard stencil-based fPINNs on a fractional ODE test, both in time (117 s vs 343.6 s) and relative error (8.85e-3 vs 4.38e-1).
  • The claimed efficiency gain scales with problem size: the savings grow with the number of collocation points \(N\), quadrature points \(M\), input dimension \(d\), network depth \(L\), and hidden width \(H\), and are largest when the hidden width dominates.

Reading between the lines

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

  • The same integrand-transformation trick should transfer to other fractional or nonlocal operators whose kernels are difference quotients of a derivative, such as tempered or variable-order Caputo derivatives, as long as the quotient kernel admits a Taylor remainder that compensates the endpoint singularity.
  • The paper's own Table 4 shows that for \(\alpha=1.75\) on the L-shaped domain, GJ-II's error (2.04e-1) is worse than GJ-I's (3.70e-2); a reader should expect the 'no accuracy sacrifice' claim to hold only for problems whose solutions are smooth enough that the kernels \(K_f\) and \(H_f\) are well resolved by the chosen quadrature.
  • Because the Type-II raw quotient amplifies endpoint roundoff by \((t\tau)^{-2}\), practical implementations will benefit from stable endpoint evaluation (expm1-style formulas or Taylor replacement), a careful cutoff \(\delta\) for Monte Carlo, and moderate \(M\) for Gauss-Jacobi; these are implementation choices the paper analyzes but does not fully automate.
  • A natural testable extension is to build a training-time schedule that adapts the number of quadrature nodes \(M\) according to the estimated kernel smoothness, which could combine the speed of GJ-II with robustness to loss of temporal regularity.
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 a transformed representation of the Caputo fractional derivative of order α∈(1,2) for use in fractional physics-informed neural networks (fPINNs). The central object is Theorem 3.1, an integration-by-parts identity that rewrites the direct Type-I Caputo integral (involving shifted first derivatives f′(t−tτ)) as a Type-II form involving shifted function values f(t−tτ) and an apparently singular kernel (tτ)^{-2}. The authors prove that the singularity is removable for f∈C², establish endpoint regularity and a representation-level stability bound, derive Monte Carlo and Gauss–Jacobi quadrature error estimates, and provide a complexity analysis claiming reductions in graph storage and training-backward FLOPs proportional to NM(d+2LH). The numerical section compares MC-I, MC-II, GJ-I, and GJ-II on one- and two-dimensional diffusion-wave problems, a Burgers equation with adaptive sampling, a fractional ODE diagnostic, and an L-shaped-domain example, reporting substantial time and memory savings for the Type-II schemes.

Significance. If the accuracy-preservation claim held across the tested class, this would be a useful and moderately significant contribution to mesh-free fractional PINNs: it converts NM shifted derivative evaluations into NM shifted value evaluations, with a clean derivation, explicit complexity savings, and extensive supporting diagnostics. The theoretical core is genuinely sound and self-contained: Theorem 3.1 is an exact identity for f∈C², Lemma 3.1 correctly identifies the removable endpoint singularities, and the stability bound in Corollary 3.2 avoids the apparent (tτ)^{-2} instability at the continuous level. The paper also ships code and includes unusually detailed derivative-level rate checks, endpoint-conditioning diagnostics, and empirical complexity verification in the appendices. However, the headline claim that the transformation improves efficiency 'without sacrificing accuracy' is not supported by the paper's own experiments: Table 4 shows that at α=1.75 on the L-shaped domain, GJ-II has relative error 2.04e-1 versus 3.70e-2 for GJ-I, a fivefold degradation, and the Monte Carlo Type-II scheme is likewise worse than Type-I.

major comments (3)
  1. [Abstract; §4.6, Table 4] The abstract and introduction claim that tDWfPINNs achieve superior efficiency 'without sacrificing accuracy,' but Table 4 provides a direct counterexample within the paper itself: for the L-shaped domain with α=1.75, GJ-II has relative error 2.04e-1 versus 3.70e-2 for GJ-I, and MC-II has 1.67e-1 versus 8.70e-2 for MC-I. The efficiency gain is real, but the accuracy-preservation claim is not valid as stated. The paper should either restrict the claim to the regimes where parity actually holds, or provide an implementation fix (e.g., stable divided differences or endpoint handling for network outputs) that restores Type-II accuracy in the tested cases.
  2. [Appendix A.6; Theorem 3.2] The Gauss–Jacobi error estimates in Theorem 3.2 assume the induced kernels K_f and H_f are C^r with r≥1, and the proof relies on Jackson-type polynomial approximation. For underlying solutions that are only C¹ but not C², the Taylor remainders in (11)–(12) fail, and the Type-II kernel H_f inherits a stronger singularity than K_f. The paper's own shifted-power benchmark in Appendix A.6 with β=1.35 demonstrates exactly this degradation when the nonsmooth point lies inside the memory interval. Since many diffusion-wave solutions of interest do not have C² temporal regularity, the claim that Type-II preserves accuracy for the general diffusion-wave class is not established. The authors should either prove convergence for lower-regularity kernels or explicitly limit the claim to smooth solutions and network approximations.
  3. [Proposition 3.2; Appendix A.4; §4.6] The endpoint-conditioning analysis shows that raw Type-II quotients amplify evaluation errors by a factor scaling like δ^{-α}, which is worse than the Type-I factor δ^{1-α}. The L-shaped-domain results at α=1.75 are consistent with this mechanism, since the smallest Gauss–Jacobi node becomes very small as α→2 and large M. If the observed fivefold degradation in Table 4 is a raw-quotient artifact, the paper should demonstrate that a stable quotient evaluation (such as the expm1/Taylor replacement used for the exponential benchmark in Appendix A.2) restores parity on the PDE problem; if it is not an artifact, the accuracy-cost trade-off for Type-II must be reported honestly in the abstract and conclusions.
minor comments (6)
  1. [Abstract] There is a typo in the abstract: 'Diffsuion-Wave' should be 'Diffusion-Wave'; the same typo appears in the title line of the provided text.
  2. [§1, contributions list] The phrase 'mesh-free qudrature methods' contains a typo ('qudrature' should be 'quadrature').
  3. [§4.2, Table 1] The table uses 'OoM' for out-of-memory but the caption spells it 'OoM means out of memory'; this is acceptable but could be made consistent with standard 'OOM'.
  4. [§3.1.3, Eq. (33)] The symbol L is used both for the number of hidden layers in the complexity analysis and for the number of training iterations in Section 4.2; this collides and should be disambiguated (e.g., L_net and L_iter).
  5. [Appendix A.3, Table A.5] The table reports 'GJ-II raw rel. error' at M=10 as 1.40e-13, which is many orders smaller than the corresponding MC errors and also smaller than the 'stable' column at that row; the entry appears plausible because Gauss–Jacobi is exact for the exponential benchmark, but a footnote explaining why the raw value is already near machine precision would improve readability.
  6. [General] The paper cites [34, Lemma 2.10] for the starting representation in Lemma 2.1 but does not state the precise smoothness assumptions of that lemma; since Theorem 3.1 and all subsequent estimates hinge on f∈C², it would be helpful to state explicitly in Section 2 that f∈C² is assumed throughout.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central equivalence theorem is a self-contained integration-by-parts identity, and the only self-citations are non-load-bearing background references.

full rationale

The paper's central claim is Theorem 3.1, which derives the transformed Type-II representation (Eq. 5) directly from the Caputo integral via an explicit integration-by-parts argument: Eq. (8) is substituted into Eq. (7), with the Taylor remainder G_t(r) used to justify the endpoint limits. No target result is assumed in the proof, and the operator equivalence D_I^alpha f = D_II^alpha f = partial_t^alpha f is established independently of the quadrature schemes that follow. Lemma 3.1 and the quadrature error estimates use standard Taylor remainders and Gauss-Jacobi approximation theory, not fitted constants. The computational savings in Section 3.1.3 are derived by counting graph storage and backward FLOPs and are verified empirically in Appendix A.7 against predicted log-log slopes, so they are not circular. The accuracy-preservation claim is empirical, and the paper itself documents a counterexample in Table 4 (alpha=1.75, L-shaped domain: GJ-II 2.04e-1 vs GJ-I 3.70e-2) and a reduced regularity benchmark in Appendix A.6; this is a correctness limitation, not a circularity. The self-citations [27] and [31] are used only as background for adaptive sampling and for the alpha<1 Gauss-Jacobi predecessor; the present proof does not rely on any unverified result from those works, so they are not load-bearing. Overall, the derivation chain is self-contained against external mathematical facts and empirical benchmarks, with no fitted input renamed as a prediction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central mathematical claim rests on standard fractional calculus and smoothness assumptions on the network output. The computational complexity claim assumes a bias-free dense network and specific autograd graph costs. The numerical comparison uses manually chosen hyperparameters and a cutoff parameter that is analyzed but not optimized.

free parameters (3)
  • Endpoint cutoff ε (or δ) = ε=10^{-7} for MC-I/MC-II in §4; ε_I=10^{-16}, ε_II=10^{-7} in Appendix A.2
    Caps (tτ) denominators near τ=0 to avoid floating-point blowup; bias O(δ^{2−α}) is analyzed (Prop. 3.3) but no optimality criterion is given.
  • Loss weights λ_bd, λ_init = λ=1.0 for 1D problems; λ_init=100 and λ_bd=1 (implicit) for Burgers; λ_bd=1 for others
    Set by hand; no grid search or sensitivity study is reported.
  • Training hyperparameters (epochs, LR, iterations) = e.g., 5000/10000 epochs, Adam lr 1e-3/1e-4/1e-5, L=3/10/20/50 iterations
    Chosen based on earlier diagnostics; not derived from theory; affect reported errors and timings.
assumptions (5)
  • standard math Caputo derivative definition and Gamma function properties are used throughout.
    Definition in §1 and used in Theorem 3.1 and corollaries.
  • domain assumption f∈C^2([0,T]) for the transformation to hold (Theorem 3.1); the neural network output is assumed to satisfy this.
    The proof of Eq. (5) uses integration by parts and Taylor remainders requiring two derivatives; tanh networks are smooth, so this holds for u_θ.
  • standard math Gauss-Jacobi nodes and weights for weight τ^{1-α} on [0,1] can be computed accurately (reference [35]).
    Used in Corollary 3.3 and all GJ experiments; no implementation error analysis is provided for the node computation.
  • ad hoc to paper Complexity analysis assumes a bias-free fully connected network and specific autograd graph storage (Section 3.1.3).
    FLOP/storage formulas (33)-(40) omit biases; the actual experimental network may include biases, and the constants c_σ are not specified.
  • domain assumption Reference solutions used for error computation (Mittag-Leffler solutions and the numerical reference on the L-shaped domain) are sufficiently accurate.
    Errors are computed against these references; any error in them would propagate to the reported relative errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transformed Diffusion-Wave fPINNs: Enhancing Computing Efficiency for PINNs Solving Time-Fractional Diffusion-Wave Equations." pith.science (2026). https://pith.science/paper/2BQAZOFG

@misc{pith2026250611518,
  author       = {Pith},
  title        = {Pith review of: Transformed Diffusion-Wave fPINNs: Enhancing Computing Efficiency for PINNs Solving Time-Fractional Diffusion-Wave Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2BQAZOFG}},
  note         = {Machine review of arXiv:2506.11518}
}
abstract

We propose transformed Diffsuion-Wave fractional Physics-Informed Neural Networks (tDWfPINNs) for efficiently solving time-fractional diffusion-wave equations with fractional order $\alpha\in(1,2)$. Conventional numerical methods for these equations often compromise the mesh-free advantage of Physics-Informed Neural Networks (PINNs) or impose high computational costs when computing fractional derivatives. The proposed method avoids first-order derivative calculations at quadrature points by introducing an integrand transformation technique, significantly reducing computational costs associated with fractional derivative evaluation while preserving accuracy. We conduct a comprehensive comparative analysis applying this integrand transformation in conjunction with both Monte Carlo integration and Gauss-Jacobi quadrature schemes across various time-fractional PDEs. Our results demonstrate that tDWfPINNs achieve superior computational efficiency without sacrificing accuracy. Furthermore, we incorporate the proposed approach into adaptive sampling approaches such as the residual-based adaptive distribution (RAD) for the time-fractional Burgers equation with order $\alpha\in(1,2)$, which exhibits complex solution dynamics. The experiments show that the Gauss-Jacobi method typically outperforms the Monte Carlo approach; however, careful consideration is required when selecting the number of quadrature points. Overall, the proposed tDWfPINNs offer a significant advancement in the numerical solution of time-fractional diffusion-wave equations, providing an accurate and scalable mesh-free alternative for challenging fractional models.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 25 canonical work pages

  1. [1]

    Mainardi, Fractional relaxation-oscillation and fractional diffusion-wave phenomena, Chaos, Solitons & Frac- tals 7 (9) (1996) 1461–1477.doi:10.1016/0960-0779(95)00125-5

    F. Mainardi, Fractional relaxation-oscillation and fractional diffusion-wave phenomena, Chaos, Solitons & Frac- tals 7 (9) (1996) 1461–1477.doi:10.1016/0960-0779(95)00125-5

    doi:10.1016/0960-0779(95)00125-5 1996

  2. [2]

    Mainardi, Fractional calculus and waves in linear viscoelasticity, second edition Edition, World Scientific, New Jersey, 2022, literaturverzeichnis: Seite 433-581

    F. Mainardi, Fractional calculus and waves in linear viscoelasticity, second edition Edition, World Scientific, New Jersey, 2022, literaturverzeichnis: Seite 433-581

    work page 2022

  3. [3]

    Swain, S

    S. Swain, S. Swain, B. Panda, M. C. Tripathy, Modeling and optimal analysis of lung cancer cell growth and apoptosis with fractional-order dynamics, Computers in Biology and Medicine 188 (2025) 109837.doi:10. 1016/j.compbiomed.2025.109837

    arXiv 2025

  4. [4]

    Bayesian Inference for a Time-Fractional HIV Model with Nonlinear Diffusion

    M. BenSalah, S. Tatar, S. Ulusoy, Bayesian inference for a time-fractional hiv model with nonlinear diffusion (Mar. 2025).arXiv:2503.23638,doi:10.48550/ARXIV.2503.23638

    work page Pith review arXiv doi:10.48550/arxiv.2503.23638 2025

  5. [5]

    Metzler, J

    R. Metzler, J. Klafter, The random walk’s guide to anomalous diffusion: a fractional dynamics approach, Physics Reports 339 (1) (2000) 1–77.doi:10.1016/s0370-1573(00)00070-3

    doi:10.1016/s0370-1573(00)00070-3 2000

  6. [6]

    Metzler, J.-H

    R. Metzler, J.-H. Jeon, A. G. Cherstvy, E. Barkai, Anomalous diffusion models and their properties: non- stationarity, non-ergodicity, and ageing at the centenary of single particle tracking, Phys. Chem. Chem. Phys. 16 (44) (2014) 24128–24164.doi:10.1039/c4cp03465a

    doi:10.1039/c4cp03465a 2014

  7. [7]

    Sakamoto, M

    K. Sakamoto, M. Yamamoto, Initial value/boundary value problems for fractional diffusion-wave equations and applications to some inverse problems, Journal of Mathematical Analysis and Applications 382 (1) (2011) 426– 447.doi:10.1016/j.jmaa.2011.04.058

    doi:10.1016/j.jmaa.2011.04.058 2011

  8. [8]

    Jiang, F

    H. Jiang, F. Liu, I. Turner, K. Burrage, Analytical solutions for the multi-term time-fractional diffusion- wave/diffusion equations in a finite domain, Computers & Mathematics with Applications 64 (10) (2012) 3377– 3388.doi:10.1016/j.camwa.2012.02.042. 44

    work page doi:10.1016/j.camwa.2012.02.042 2012

Show all 40 references
  1. [9]

    O. P. Agrawal, Solution for a fractional diffusion-wave equation defined in a bounded domain, Nonlinear Dy- namics 29 (1/4) (2002) 145–155.doi:10.1023/a:1016539022492

    2002 doi

  2. [10]

    R. Du, W. Cao, Z. Sun, A compact difference scheme for the fractional diffusion-wave equation, Applied Math- ematical Modelling 34 (10) (2010) 2998–3007.doi:10.1016/j.apm.2010.01.008

    2010 doi

  3. [11]

    Sun, Fractional differential equations, De Gruyter, Berlin, 2020

    Z.-z. Sun, Fractional differential equations, De Gruyter, Berlin, 2020

    2020

  4. [12]

    Cuesta, C

    E. Cuesta, C. Lubich, C. Palencia, Convolution quadrature time discretization of fractional diffusion-wave equa- tions, Mathematics of Computation 75 (254) (2006) 673–696.doi:10.1090/s0025-5718-06-01788-1

    2006 doi

  5. [13]

    B. Jin, B. Li, Z. Zhou, Correction of high-order bdf convolution quadrature for fractional evolution equations, SIAM Journal on Scientific Computing 39 (6) (2017) A3129–A3152.doi:10.1137/17m1118816

    2017 doi

  6. [14]

    H. Zhu, C. Xu, A fast high order method for the time-fractional diffusion equation, SIAM Journal on Numerical Analysis 57 (6) (2019) 2829–2849.doi:10.1137/18m1231225

    2019 doi

  7. [15]

    Stynes, E

    M. Stynes, E. O’Riordan, J. L. Gracia, Error analysis of a finite difference method on graded meshes for a time- fractional diffusion equation, SIAM Journal on Numerical Analysis 55 (2) (2017) 1057–1079.doi:10.1137/ 16m1082329

    2017

  8. [16]

    S. Chen, J. Shen, Z. Zhang, Z. Zhou, A spectrally accurate approximation to subdiffusion equations using the log orthogonal functions, SIAM Journal on Scientific Computing 42 (2) (2020) A849–A877.doi:10.1137/ 19m1281927

    2020

  9. [17]

    Zayernouri, G

    M. Zayernouri, G. E. Karniadakis, Fractional sturm–liouville eigen-problems: Theory and numerical approxi- mation, Journal of Computational Physics 252 (2013) 495–517.doi:10.1016/j.jcp.2013.06.031

    2013 doi

  10. [18]

    Raissi, P

    M. Raissi, P. Perdikaris, G. Karniadakis, Physics-informed neural networks: A deep learning framework for solv- ing forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics 378 (2019) 686–707.doi:10.1016/j.jcp.2018.10.045

    2019 doi

  11. [19]

    S. Cai, Z. Mao, Z. Wang, M. Yin, G. E. Karniadakis, Physics-informed neural networks (pinns) for fluid me- chanics: a review, Acta Mechanica Sinica 37 (12) (2021) 1727–1738.doi:10.1007/s10409-021-01148-1

    2021 doi

  12. [20]

    Cuomo, V

    S. Cuomo, V . S. Di Cola, F. Giampaolo, G. Rozza, M. Raissi, F. Piccialli, Scientific machine learning through physics–informed neural networks: Where we are and what’s next, Journal of Scientific Computing 92 (3) (Jul. 2022).doi:10.1007/s10915-022-01939-z

    2022 doi

  13. [21]

    G. Pang, L. Lu, G. E. Karniadakis, fpinns: Fractional physics-informed neural networks, SIAM Journal on Scientific Computing 41 (4) (2019) A2603–A2626.doi:10.1137/18m1229845

    2019 doi

  14. [22]

    Lu, Z.-p

    B. Lu, Z.-p. Hao, C. Moya, G. Lin, Fpinn-deeponet: An operator learning framework for multi-term time- fractional mixed diffusion-wave equations (2025).doi:10.2139/ssrn.5080263

    2025 doi

  15. [23]

    Y . Vats, M. Mehra, D. Oelz, A. K. Singh, A new perspective for scientific modelling: Sparse reconstruction- based approach for learning time-space fractional differential equations, Journal of Computational and Nonlinear Dynamics 19 (12) (Sep. 2024).doi:10.1115/1.4066330

    2024 doi

  16. [24]

    H. Ren, X. Meng, R. Liu, J. Hou, Y . Yu, A class of improved fractional physics informed neural networks, Neurocomputing 562 (2023) 126890.doi:10.1016/j.neucom.2023.126890

    2023

  17. [25]

    S. Wang, X. Yu, P. Perdikaris, When and why pinns fail to train: A neural tangent kernel perspective, Journal of Computational Physics 449 (2022) 110768.doi:10.1016/j.jcp.2021.110768

    2022

  18. [26]

    C. Wu, M. Zhu, Q. Tan, Y . Kartha, L. Lu, A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks, Computer Methods in Applied Mechanics and Engineering 403 (2023) 115671.doi:10.1016/j.cma.2022.115671. 45

    2023

  19. [27]

    Zhang, J

    Z. Zhang, J. Li, B. Liu, Annealed adaptive importance sampling method in pinns for solving high dimensional partial differential equations, Journal of Computational Physics 521 (2025) 113561.doi:10.1016/j.jcp. 2024.113561

    2025

  20. [28]

    Zhang, D

    T. Zhang, D. Zhang, S. Shi, Z. Guo, Spectral-fpinns: spectral method based fractional physics-informed neural networks for solving fractional partial differential equations, Nonlinear Dynamics 113 (11) (2025) 12565–12588. doi:10.1007/s11071-025-11067-6

    2025 doi

  21. [29]

    Z. Chen, C. QIU, Z. Li, Laplace-based physics-informed neural networks for solving nonlinear coefficient in- verse problem for acoustic equation with fractional derivative (2025).doi:10.2139/ssrn.5142405

    2025 doi

  22. [30]

    L. Guo, H. Wu, X. Yu, T. Zhou, Monte carlo fpinns: Deep learning method for forward and inverse problems involving high dimensional fractional partial differential equations, Computer Methods in Applied Mechanics and Engineering 400 (2022) 115523.doi:10.1016/j.cma.2022.115523

    2022

  23. [31]

    Z. Hu, K. Kawaguchi, Z. Zhang, G. E. Karniadakis, Tackling the curse of dimensionality in fractional and tempered fractional pdes with physics-informed neural networks, Computer Methods in Applied Mechanics and Engineering 432 (2024) 117448.doi:10.1016/j.cma.2024.117448

    2024

  24. [32]

    Z. Lin, Q. Ma, H. Xie, X. Yin, Solving time-fractional partial integro-differential equations using tensor neural network (2025).doi:10.48550/ARXIV.2504.01440

    2025 doi

  25. [33]

    A. A. Kilbas, Theory and applications of fractional differential equations, no. 204 in North-Holland mathematics studies, Elsevier, Boston, 2006, includes bibliographical references and index

    2006

  26. [34]

    Jin, Fractional Differential Equations: An Approach via Fractional Derivatives, Springer International Pub- lishing, 2021.doi:10.1007/978-3-030-76043-4

    B. Jin, Fractional Differential Equations: An Approach via Fractional Derivatives, Springer International Pub- lishing, 2021.doi:10.1007/978-3-030-76043-4

    2021 doi

  27. [35]

    D. W. Brzezinski, Computation of gauss-jacobi quadrature nodes and weights with arbitrary precision, in: 2018 Federated Conference on Computer Science and Information Systems (FedCSIS), 2018, pp. 297–306

    2018

  28. [36]

    Chernov, C

    A. Chernov, C. Schwab, Exponential convergence of gauss–jacobi quadratures for singular integrals over simplices in arbitrary dimension, SIAM Journal on Numerical Analysis 50 (3) (2012) 1433–1455.doi: 10.1137/100812574

    2012 doi

  29. [37]

    Tadmor, The exponential accuracy of fourier and chebyshev differencing methods, SIAM Journal on Numer- ical Analysis 23 (1) (1986) 1–10.doi:10.1137/0723001

    E. Tadmor, The exponential accuracy of fourier and chebyshev differencing methods, SIAM Journal on Numer- ical Analysis 23 (1) (1986) 1–10.doi:10.1137/0723001

    1986 doi

  30. [38]

    Podlubny, Fractional differential equations, no

    I. Podlubny, Fractional differential equations, no. v. 198 in Mathematics in science and engineering, Academic Press, San Diego, 1999

    1999

  31. [39]

    S. Wang, S. Sankaran, P. Perdikaris, Respecting causality is all you need for training physics-informed neural networks (Mar. 2022).arXiv:2203.07404

    2022 arXiv

  32. [40]

    S. Wang, Y . Teng, P. Perdikaris, Understanding and mitigating gradient flow pathologies in physics-informed neural networks, SIAM Journal on Scientific Computing 43 (5) (2021) A3055–A3081.doi:10.1137/ 20m1318043. 46

    2021

Pith tools

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