Pith. sign in

REVIEW 3 major objections 4 minor 27 references

DFSC claims that when the dominant linear fractional evolution is known through a Mittag-Leffler propagator, that propagator can be turned into a differentiable layer that certifies its own truncation error, leaving neural networks to learn

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 14:50 UTC pith:TQEOMAV2

load-bearing objection A well-scoped differentiable Mittag-Leffler layer with a real but narrow certified bound; the title overpromises, the body mostly doesn't. the 3 major comments →

arxiv 2607.29038 v1 pith:TQEOMAV2 submitted 2026-07-31 cs.LG

DFSC: Error-Controlled Differentiable Mittag-Leffler Propagation for Fractional Scientific Machine Learning

Ning Hu , Haitao Duan , Shuqun Li , Chuyang Hu This is my paper
classification cs.LG MSC 34A0865F6068T07
keywords fractional scientific machine learningMittag-Leffler functiondifferentiable programmingspectral layerhybrid neural modelserror certificationfractional inverse problemsscientific software
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that when the dominant linear part of a fractional dynamical system is known through a Mittag-Leffler propagator, that propagator should be a first-class differentiable layer rather than something to relearn or reintegrate. It introduces MLSL, a layer that evaluates the modal Mittag-Leffler evolution, keeps fractional orders and operator parameters trainable, and lets a neural network learn only the unresolved residual dynamics. The layer comes with an adaptive budget controller and, on the negative real axis with 0 < alpha <= 1 and beta > 0, a certified first-omitted-term bound; outside that regime results are explicitly labeled empirical. A reader should care because this offers a principled way to combine known fractional structure with machine-learned corrections, with a declared error contract instead of silent extrapolation. The paper also shows the complement: when the backbone is mismatched, the hybrid is compact but not the most accurate.

Core claim

The paper's central claim is that the Mittag-Leffler spectral layer (MLSL) implements the exact retained modal Mittag-Leffler evolution of a linear Caputo system, up to spatial truncation, special-function approximation, and floating-point error, while preserving derivatives with respect to fractional orders and operator parameters. In the declared reliability domain — z = -x <= 0, 0 < alpha <= 1, beta > 0 — Equation (8) certifies the series-truncation remainder: once term magnitudes enter the decreasing alternating regime, |E - partial sum| <= x^N / Gamma(alpha N + beta), and DFSC returns this bound only after checking sign, parameter domain, and regime entry. The associated adaptive contro

What carries the argument

The load-bearing object is the Mittag-Leffler spectral layer (MLSL), a differentiable map that applies the modal propagator E_alpha(-mu_n t^alpha) to a retained eigenbasis (or a Lanczos/Arnoldi basis) and preserves gradients with respect to alpha, beta, operator parameters, inputs, and residual weights. Around it, the paper places: a certified scalar alternating-series bound (Equation 8) on the negative real axis; a smooth blend (Equation 5) between series and asymptotic branches; an adaptive controller (Equation 7) that grows truncation depth or Krylov dimension until successive evaluations agree; and Proposition 1, which says the layer applies the exact retained modal evolution up to the s

Load-bearing premise

The load-bearing premise is that the system's fractional evolution is representable by a diagonalizable linear operator on the retained modes and that every evaluated spectral argument lies inside the validated reliability domain; if either fails, the certified evolution statement and bound do not apply.

What would settle it

Take a single eligible scalar case (0 < alpha <= 1, beta > 0, x > 0) where the alternating regime has been entered, evaluate E_alpha,beta(-x) with a high-precision reference, and check whether the true remainder exceeds x^N / Gamma(alpha N + beta) for some N; a violation would disprove Equation (8). For the layer claim, run MLSL on a non-diagonalizable Jordan-block operator and compare the output and gradients to a dense high-precision reference; divergence beyond floating-point would expose the missing projection or Krylov bound.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • For any model whose linear backbone has a Mittag-Leffler representation, training can optimize fractional orders and residual weights jointly, eliminating history reconstruction at selected query times.
  • In the negative-real regime, the returned bound lets a practitioner distinguish resolved error from truncation error, turning a black-box neural integrator into an error-reporting component.
  • Prepared Lanczos bases make repeated evaluations for different orders and times faster by roughly 5-16x, directly benefiting inverse problems and order searches.
  • Hybrid MLSL-plus-residual models can reach low error with far fewer parameters than pure networks when the backbone matches the data; when it does not, pure networks retain the advantage.
  • DFSC is positioned as a complement to general fractional solvers, with applicability limited to structurally matched cases rather than universal replacement.

Where Pith is reading between the lines

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

  • A natural extension is to derive analogous certified bounds for E_alpha,beta on other rays or for complex arguments within the tested radius, which would enlarge the validated domain beyond the negative real axis.
  • Because the certified bound is scalar, one could lift it through the spectral decomposition to matrix arguments only if the eigendecomposition error is also bounded; currently that link is missing, so matrix-level certificates remain open.
  • The layer design suggests a reusable template for other special-function propagators and for variable- or distributed-order models, which the paper marks experimental.
  • One could test the practical value of the reliability metadata by comparing optimization trajectories with and without strict rejection of out-of-domain arguments.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper presents DFSC, a PyTorch software ecosystem built around a differentiable Mittag-Leffler spectral layer (MLSL) for fractional linear evolution. It assumes a diagonalizable linear Caputo operator, evaluates the modal Mittag-Leffler propagator with adaptive truncation/Krylov work control, and supports trainable fractional orders and neural residual corrections. The main theoretical result is a certified alternating-series remainder bound for scalar Mittag-Leffler functions on the negative real axis, Eq. (8), together with a detailed error decomposition and propositions on spectral consistency, query complexity, and piecewise differentiability. The paper validates the method on numerical identities, inverse problems, hybrid learning, external solvers, and four real experimental domains, and it releases the package with reproducibility data.

Significance. If the claims hold, DFSC is a useful contribution to fractional scientific machine learning: it provides a differentiable, batched spectral propagator with a rigorous scalar truncation bound in a well-defined regime and separates certified from empirical components honestly. The paper's strengths include a parameter-free derivation of Eq. (8), clear scoping in Sections 3.2 and 8, and a substantial software artifact with tests, data provenance, and extensive experiments. The reported evidence is broad and the paper repeatedly states where the structured primitive fails, which increases confidence. The main risks are not mathematical correctness but possible overclaiming in the title and the absence of a diagonalizability check in the reliability metadata.

major comments (3)
  1. [§3.4, Assumption 1; §4.3] The bound in Eq. (8) is rigorous only for the scalar truncated series. The full forward map Eq. (3) additionally requires the spectral decomposition Eq. (2); Assumption 1 states diagonalizability, but the implementation and ReliabilityReport do not appear to check or warn when this assumption fails. A defective or nearly defective operator can therefore produce an incorrect output while still being reported as validated. Since this is a core component of the 'error-controlled' claim, please add an explicit diagonalizability check/warning or make clear in §4.3 and §8 that diagonalizability is user-declared and not part of the certified reliability metadata.
  2. [Title; Abstract; §3.2] The title's 'error-controlled' is stronger than the guarantees actually delivered. Section 3.2 states that the adaptive indicator η_j is not claimed to bound spectral truncation, model reduction, or systematic special-function error; Appendix A and §8 state that only the series remainder has a rigorous bound and that Krylov, projection, blend, and solver errors are empirical. To avoid overclaiming, qualify the title or abstract (e.g., 'error-aware' or 'locally certified') and state in the abstract that the certified bound applies to the scalar series truncation only.
  3. [§4.5 vs §9; Table 8] Section 4.5 says version 0.1.0 is 'not yet distributed on PyPI' and has 'no hosted API site'; §9 states it is available from PyPI and the documentation site, and Table 8 repeats the latter. This direct contradiction must be resolved because availability claims are part of the reproducibility contract.
minor comments (4)
  1. [§5.1] The 60-case certification experiment should specify how α, |z|, and truncation depths were chosen, and report the parameter values of the single uncertified case.
  2. [§3.2, Eq. (5)] The transition interval [r−, r+] is never assigned numerical values in the text. Please report the defaults used, as they affect the blend behavior and reproducibility.
  3. [§6.2, Table 3] Clarify whether the reported runtimes include basis construction or only the repeated-query amortized path, and specify the CPU/GPU configuration for the timing numbers.
  4. [Fig. 3] The statement that the adaptive gradient agrees with a fixed 180-term series 'to double-precision resolution at every probe' should explicitly note that this is an empirical scan, not a theorem, and specify the norm used.

Circularity Check

0 steps flagged

No significant circularity: Eq. (8) is an independent theorem and the propagation formulas are definitional; caveats are scope limitations, not circular reasoning.

full rationale

The paper's central derivation is self-contained rather than circular. Equation (8) is not fitted: it is a parameter-free alternating-series bound justified in Appendix A by gamma-ratio monotonicity (r_k = x Gamma(alpha k + beta)/Gamma(alpha k + alpha + beta)) and the alternating-series theorem, with the implementation only checking sign, parameter domain, and regime entry before assigning rigorous status. The MLSL forward map (3) is the definitional modal solution of the linear Caputo system under Assumption 1's diagonalizability premise; Proposition 1 is a consistency statement decomposing errors (Eq. 10) into projection, model reduction, and numerical evaluation, and the paper explicitly says only the restricted series remainder has a rigorous bound and that Krylov, projection, hybrid, and solver errors are diagnosed empirically (Sec. 3.2, Sec. 8, App. A). The adaptive work control (Eq. 7) is explicitly labeled an empirical indicator, not a claimed rigorous bound. The inverse and hybrid experiments fit trainable orders on manufactured/real data and then evaluate recovery or held-out transfer; that is the intended application, not a fitted quantity renamed as a prediction. The only self-reference ([11]) points to the released software artifact and is not used to justify any load-bearing theorem; no uniqueness result or ansatz is imported from the authors' prior work. The limitations concerning defective operators and absent end-to-end error bounds are honest scope restrictions relevant to correctness risk, but they do not make any step of the derivation equivalent to its inputs.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central formula Eq (3) is definitional (the exact solution of the linear Caputo system when diagonalizable), so the mathematical pull is small: three standard-analysis ingredients plus two scoping assumptions. The certified-bound machinery is parameter-free; the main hidden degrees of freedom are implementation-level (blend routing interval, adaptive work schedule), both hand-chosen and unpublished. No new physical entities are postulated.

free parameters (2)
  • Series/asymptotic blend transition interval [r-, r+]
    Hand-chosen routing thresholds for the Eq (5) blend; no values or selection procedure are published, and thresholds depend on detached policy variables so gradients do not pass through the routing decision (Sec 3.2).
  • Adaptive work schedule w1 < ... < wJ
    Nested evaluation schedule for truncation depth or Krylov dimension (Sec 3.2). The selected candidate, and therefore the selected-path gradient, depends on this schedule; no schedule is published in the paper.
axioms (5)
  • standard math Gamma-ratio monotonicity: psi(y+alpha)-psi(y) > 0 for y>0, alpha>0, making r_k = x*Gamma(alpha k + beta)/Gamma(alpha k + alpha + beta) nonincreasing
    Core of the alternating-series remainder proof for Eq (8), Appendix A; standard digamma monotonicity.
  • standard math Alternating-series theorem: a decreasing alternating tail is bounded in magnitude by its first omitted term
    Invoked in Eq (8) and Appendix A after the regime-entry check a_N <= a_{N-1}.
  • standard math Scalar Caputo mode solution: CD^alpha_t u_hat + mu u_hat = 0 implies u_hat(t) = E_alpha(-mu t^alpha) u_hat(0)
    Textbook Mittag-Leffler solution of the scalar fractional relaxation equation, used in Sec 3.1 and Appendix A.
  • domain assumption A_beta is diagonalizable on the retained subspace with arguments in the declared reliability domain
    Assumption 1 (Sec 3.4). If false, Eq (3), the certified remainder, and the differentiability claims do not apply.
  • domain assumption Validated regime restrictions: z = -x <= 0, 0 < alpha <= 1, beta > 0 for the certificate; complex support |z| <= 4; real paths for nonpositive spectral arguments
    Declared scope controlling where Eq (8) is certified versus labeled empirical (Sec 3.2, 5.1, 8).

pith-pipeline@v1.3.0-daily-deepseek · 14806 in / 20153 out tokens · 188561 ms · 2026-08-03T14:50:40.503129+00:00 · methodology

0 comments
read the original abstract

Fractional scientific machine learning requires numerical operators that can be differentiated, batched, accelerated, and composed with neural networks. When the dominant linear fractional evolution is known through a Mittag-Leffler propagator, repeatedly reconstructing that response with a history solver or relearning it from data is unnecessary. We present DFSC, a PyTorch environment organized around the Mittag-Leffler Spectral Layer (MLSL). The layer separates known fractional propagation from data-driven corrections, so neural modules learn only unresolved dynamics while fractional orders and residual-network parameters are optimized jointly. Its adaptive algorithm increases special-function truncation depth or Lanczos dimension until successive differentiable evaluations satisfy a requested tolerance. In the negative-real alternating-series regime, DFSC additionally returns a certified first-omitted-term bound; outside that regime it explicitly labels estimates as empirical. DFSC supports dense, sparse, matrix-free, self-adjoint, generalized, and controlled complex operator paths; trainable fractional orders; direct inverse problems; residual neural composition; and CPU/GPU execution. The certified series bound covers all 59 eligible reference cases, with median bound/error effectivity 1.246 for resolved errors. Reusing a prepared batched Lanczos basis gives identical fixed-path values and reduces repeated-query time by 4.61--7.11 times on CPU and 13.07--16.22 times on an RTX 5070, excluding one-time preparation. A 27-case inverse matrix finds full-rank local curvature throughout, while remaining explicitly model-conditional. External solver and mixed real-data results support DFSC as an error-aware optional primitive for matched fractional structure, rather than a general replacement for fractional solvers or neural models.

Figures

Figures reproduced from arXiv: 2607.29038 by Chuyang Hu, Haitao Duan, Ning Hu, Shuqun Li.

Figure 1
Figure 1. Figure 1: DFSC architecture. A differentiable Mittag–Leffler core supports learning workflows [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MLSL in a hybrid neural computation graph. The known propagator supplies a [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Adaptive-gradient diagnostics. Left: the selected-path gradient and a fixed 180-term [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Adaptive Lanczos calibration on an RTX 5070 Laptop GPU. Tighter requested [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Measured direct-query MLSL and bundled L1-reference runtimes. The benchmark [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Multi-seed inverse recovery of trainable fractional orders under the tested manufactured [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Complementary evidence for the structured primitive. Left: sample efficiency and [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Held-out or extrapolation error across four reported real-task groups. No model [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

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

27 extracted references · 3 canonical work pages

  1. [1]

    Anomalous diffusion experimental data set, 2026

    Anomalous Diffusion Database. Anomalous diffusion experimental data set, 2026. URL https://anomdiffdb.github.io/DB/. Accessed 24 July 2026

    2026

  2. [2]

    Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind

    Atilim Gunes Baydin, Barak A. Pearlmutter, Alexey Andreyevich Radul, and Jeffrey Mark Siskind. Automatic Differentiation in machine learning: A survey.Journal of Machine Learning Research, 18(153):1–43, 2018

    2018

  3. [3]

    Springer, Berlin, 2010

    Kai Diethelm.The Analysis of Fractional Differential Equations: An Application-Oriented Exposition Using Differential Operators of Caputo Type. Springer, Berlin, 2010. doi: 10.1007/978-3-642-14574-2

    doi:10.1007/978-3-642-14574-2 2010

  4. [4]

    pycaputo: Evaluate fractional integrals and solve fractional odes, 2026

    Alexandru Fikl. pycaputo: Evaluate fractional integrals and solve fractional odes, 2026. URLhttps://pypi.org/project/pycaputo/. Version 0.10.2, software package

    2026

  5. [6]

    Numerical evaluation of two and three parameter Mittag-Leffler functions

    Roberto Garrappa. Numerical evaluation of two and three parameter Mittag-Leffler functions. SIAM Journal on Numerical Analysis, 53(3):1350–1369, 2015. doi: 10.1137/140971191

    doi:10.1137/140971191 2015

  6. [8]

    Kilbas, Francesco Mainardi, and Sergei V

    Rudolf Gorenflo, Anatoly A. Kilbas, Francesco Mainardi, and Sergei V. Rogosin.Mittag- Leffler Functions, Related Topics and Applications. Springer, Berlin, 2014. doi: 10.1007/ 978-3-662-43930-2

    2014

  7. [9]

    Weiss, Elias Nehme, Michael Levin, Michael Chein, Eran Perlson, Yael Roichman, and Yoav Shechtman

    Naor Granik, Liron E. Weiss, Elias Nehme, Michael Levin, Michael Chein, Eran Perlson, Yael Roichman, and Yoav Shechtman. Single-particle diffusion characterization by deep learning.Biophysical Journal, 117(2):185–192, 2019. doi: 10.1016/j.bpj.2019.06.015

    work page doi:10.1016/j.bpj.2019.06.015 2019

  8. [10]

    Ling Guo, Hao Wu, Xiaochen Yu, and Tao 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: 115523, 2022. doi: 10.1016/j.cma.2022.115523

    arXiv 2022

  9. [11]

    DFSC: Error-controlled differentiable mittag–leffler propagation for fractional scientific machine learning, 2026

    Ning Hu and Haitao Duan. DFSC: Error-controlled differentiable mittag–leffler propagation for fractional scientific machine learning, 2026. URL https://doi.org/10.5281/zenodo. 21588834. MIT-licensed software. 18

    doi:10.5281/zenodo 2026

  10. [12]

    Shah, and Will Tebbutt

    Mike Innes, Alan Edelman, Keno Fischer, Chris Rackauckas, Elliot Saba, Viral B. Shah, and Will Tebbutt. A Differentiable Programming system to bridge machine learning and scientific computing, 2019. arXiv preprint

    2019

  11. [13]

    An analysis of the l1 scheme for the subdiffusion equation with nonsmooth data.IMA Journal of Numerical Analysis, 36(1):197–221, 2016

    Bangti Jin, Raytcho Lazarov, and Zhi Zhou. An analysis of the l1 scheme for the subdiffusion equation with nonsmooth data.IMA Journal of Numerical Analysis, 36(1):197–221, 2016. doi: 10.1093/imanum/dru063

    work page doi:10.1093/imanum/dru063 2016

  12. [14]

    Kilbas, Hari M

    Anatoly A. Kilbas, Hari M. Srivastava, and Juan J. Trujillo.Theory and Applications of Fractional Differential Equations, volume 204 ofNorth-Holland Mathematics Studies. Elsevier, Amsterdam, 2006

    2006

  13. [15]

    A library for learning Neural Operators, 2024

    Jean Kossaifi, Nikola Kovachki, Zongyi Li, David Pitt, Miguel Liu-Schiaffini, Robert Joseph George, Boris Bonev, Kamyar Azizzadenesheli, Julius Berner, and Anima Anandkumar. A library for learning Neural Operators, 2024. arXiv preprint

    2024

  14. [16]

    Neural Operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89): 1–97, 2023

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural Operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89): 1–97, 2023

    2023

  15. [17]

    Fourier Neural Operator for parametric partial differential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier Neural Operator for parametric partial differential equations. InInternational Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=c8P9NQVtmnO

    2021

  16. [18]

    Meerschaert, Mark Ainsworth, and George Em Karniadakis

    Anna Lischke, Guofei Pang, Mamikon Gulian, Fangying Song, Christian Glusa, Xiaoning Zheng, Zhiping Mao, Wei Cai, Mark M. Meerschaert, Mark Ainsworth, and George Em Karniadakis. What is the fractional laplacian?Journal of Computational Physics, 404: 109009, 2020. doi: 10.1016/j.jcp.2019.109009

    arXiv 2020

  17. [19]

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3:218–229, 2021

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3:218–229, 2021. doi: 10.1038/s42256-021-00302-5

    doi:10.1038/s42256-021-00302-5 2021

  18. [20]

    DeepXDE: A deep learning library for solving differential equations.SIAM Review, 63(1):208–228, 2021

    Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. DeepXDE: A deep learning library for solving differential equations.SIAM Review, 63(1):208–228, 2021. doi: 10.1137/19M1274067

    doi:10.1137/19m1274067 2021

  19. [21]

    Discretized fractional calculus.SIAM Journal on Mathematical Analysis, 17(3):704–719, 1986

    Christian Lubich. Discretized fractional calculus.SIAM Journal on Mathematical Analysis, 17(3):704–719, 1986. doi: 10.1137/0517050

    doi:10.1137/0517050 1986

  20. [22]

    fPINNs: Fractional physics-informed neural networks.SIAM Journal on Scientific Computing, 41(4):A2603–A2626, 2019

    Guofei Pang, Lu Lu, and George Em Karniadakis. fPINNs: Fractional physics-informed neural networks.SIAM Journal on Scientific Computing, 41(4):A2603–A2626, 2019. doi: 10.1137/18M1229845

    doi:10.1137/18m1229845 2019

  21. [23]

    Academic Press, San Diego, CA, 1999

    Igor Podlubny.Fractional Differential Equations: An Introduction to Fractional Deriva- tives, Fractional Differential Equations, to Methods of Their Solution and Some of Their Applications, volume 198 ofMathematics in Science and Engineering. Academic Press, San Diego, CA, 1999

    1999

  22. [24]

    FractionalDiffEq.jl: High performance fractional differential equation solver in julia, 2025

    Qingyu Qu and Wei Ruan. FractionalDiffEq.jl: High performance fractional differential equation solver in julia, 2025

    2025

  23. [25]

    DifferentialEquations.jl – a performant and feature- rich ecosystem for solving differential equations in julia.Journal of Open Research Software, 5(1):15, 2017

    Christopher Rackauckas and Qing Nie. DifferentialEquations.jl – a performant and feature- rich ecosystem for solving differential equations in julia.Journal of Open Research Software, 5(1):15, 2017. doi: 10.5334/jors.151. 19

    doi:10.5334/jors.151 2017

  24. [26]

    Universal Differential Equations for scientific machine learning, 2020

    Christopher Rackauckas, Yingbo Ma, Julius Martensen, Collin Warner, Kirill Zubov, Rohit Supekar, Dominic Skinner, Ali Ramadhan, and Alan Edelman. Universal Differential Equations for scientific machine learning, 2020. arXiv preprint

    2020

  25. [27]

    Solving the inverse source problem of the fractional poisson equation by MC-fPINNs, 2024

    Rui Sheng, Peiying Wu, Jerry Zhijian Yang, and Cheng Yuan. Solving the inverse source problem of the fractional poisson equation by MC-fPINNs, 2024. arXiv preprint

    2024

  26. [28]

    Effects of viscoelastic properties on geomembrane mechanics and deformations, 2025

    Samuel Luke Vorlet. Effects of viscoelastic properties on geomembrane mechanics and deformations, 2025. URL https://doi.org/10.5281/zenodo.14329961. CC BY 4.0 dataset

    work page doi:10.5281/zenodo.14329961 2025

  27. [29]

    Optimising neural fractional differential equations for performance and efficiency

    Bernd Zimmering, Cec ´ ılia Coelho, and Oliver Niggemann. Optimising neural fractional differential equations for performance and efficiency. InProceedings of the 1st ECAI Workshop on Machine Learning Meets Differential Equations: From Theory to Applications, volume 255 ofProceedings of Machine Learning Research, pages 1–22, 2024. URL https: //proceedings...

    2024