pith. sign in

arxiv: 2606.10686 · v2 · pith:WEKKTQ4Tnew · submitted 2026-06-09 · ⚛️ physics.comp-ph · astro-ph.IM· cs.LG

An adaptive framework for the axisymmetric pulsar magnetosphere using physics-informed Kolmogorov-Arnold networks

Spyros Rigas , Ioannis Contopoulos , Georgios Alexandridis , Antonios Nathanail This is my paper

Pith reviewed 2026-06-27 11:00 UTC · model grok-4.3

classification ⚛️ physics.comp-ph astro-ph.IMcs.LG
keywords pulsar magnetospherephysics-informed neural networksKolmogorov-Arnold networksaxisymmetric solutionsadaptive trainingcurrent sheetforce-free electrodynamics
0
0 comments X

The pith

Kolmogorov-Arnold networks with adaptive training solve the axisymmetric pulsar magnetosphere to PDE residuals of order 10^{-6} and handle stellar radii reduced by 80 percent.

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

The paper develops a refined neural-network method to compute the magnetic structure of a rotating neutron star's magnetosphere by training the network to obey the plasma force-balance equations directly. Standard networks are replaced by Kolmogorov-Arnold networks inside an automated adaptive training loop that adjusts loss weights and network parameters on the fly according to a physics-based convergence test. This combination removes manual hyperparameter searches, raises final accuracy by two orders of magnitude, shortens training to under twenty minutes in single precision, and permits solutions when the stellar radius is only one-fifth the light-cylinder radius. The same runs also supply an empirical correction to the analytic relation between the magnetic flux that opens to infinity and the radial location of the equatorial T-point.

Core claim

Physics-informed Kolmogorov-Arnold networks equipped with an automated adaptive training pipeline and a physics-based convergence criterion produce self-consistent axisymmetric pulsar magnetosphere solutions whose PDE residuals reach mean squared values of order 10^{-6} in double precision, converge in under twenty minutes in single precision, remain stable for stellar radii reduced by up to eighty percent, and yield a corrected algebraic link between open magnetic flux and the position of the equatorial T-point.

What carries the argument

Kolmogorov-Arnold networks placed inside a physics-informed neural network with domain decomposition at the separatrix and equatorial current sheet, driven by an automated adaptive training pipeline that adjusts weights according to a physics-based convergence criterion.

If this is right

  • Magnetosphere solutions become feasible for compact stars whose radius is a small fraction of the light-cylinder radius, a regime previously inaccessible to both traditional solvers and baseline PINNs.
  • Training completes in minutes rather than hours and requires no manual search over network depth, width, or loss weights.
  • Varying the open flux produces a quantitative correction to the previously used formula that relates flux to T-point position.
  • Self-consistent solutions are obtained without separate treatment of the polar cap or light-cylinder boundaries beyond the domain decomposition already employed.

Where Pith is reading between the lines

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

  • The same adaptive KAN-PINN structure could be applied to time-dependent or three-dimensional magnetosphere problems where scale separation between star and light cylinder is even more extreme.
  • Analogous automated KAN pipelines may address other thin-current-sheet plasma configurations such as solar coronal current sheets or accretion-disk coronae.
  • Release of the PulsarX library allows direct verification of the reported residuals and T-point correction on independent hardware and parameter choices.

Load-bearing premise

The physical model of infinitesimally thin separatrix and current-sheet discontinuities remains accurate enough that the neural network can satisfy the governing equations to the reported precision even at the newly accessible small stellar radii.

What would settle it

Independent high-resolution force-free or MHD simulations run at several open-flux values should reproduce the corrected algebraic relation between open flux and equatorial T-point location to within a few percent.

Figures

Figures reproduced from arXiv: 2606.10686 by Antonios Nathanail, Georgios Alexandridis, Ioannis Contopoulos, Spyros Rigas.

Figure 1
Figure 1. Figure 1: Evolution of the NTK eigenvalue spectra during 50,000 training iterations for the baseline MLP configuration (left) and the proposed adaptive framework (right). The color scale indicates the training progression (×104 iterations), moving from initialization (dark red) to the final iteration (blue). Vertical dashed lines indicate the expansion of the non-plateau eigenvalue regime. As illustrated in [PITH_F… view at source ↗
Figure 2
Figure 2. Figure 2: Evolution of the geometric complexity during the first 50,000 training iterations for the closed-line region network cl (left) and the open-line region network op (right). Both panels compare the standard MLP baseline (red) against the RGA KAN architecture of the adaptive framework (blue). S. Rigas et al.: Preprint submitted to Elsevier Page 10 of 25 [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Self-consistent pulsar magnetosphere solution under single precision (FP32), where 𝑧 = 𝑟 cos 𝜃 and 𝑥 = 𝑟 sin 𝜃. Left: Magnetic field lines shown as contours of constant flux. The equatorial T-point is located at 𝑥𝑇 = 0.919 𝑅LC. Right: Normalized poloidal current distribution |𝐼(Ψ)|𝑅LC∕Ψ𝑆 across the open magnetic field lines. The dotted line denotes the theoretical analytic expression for the split monopole… view at source ↗
Figure 4
Figure 4. Figure 4: Evolution of the MSE of the alignment condition’s residuals during single-precision training. Left: Close-up of the first training cycle (initial 10,000 iterations) plotted on a logarithmic horizontal axis. Right: Complete multi-cycle optimization trajectory spanning all training cycles. bottleneck in the original framework, these FP64 results offer a rigorous numerical correction to the value reported in … view at source ↗
Figure 5
Figure 5. Figure 5: Same as [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Unphysical magnetosphere solution obtained when both RBA and LRA are deactivated under single precision (FP32), where 𝑧 = 𝑟 cos 𝜃 and 𝑥 = 𝑟 sin 𝜃. Left: Magnetic field lines shown as contours of constant flux. Right: Normalized poloidal current distribution |𝐼(Ψ)|𝑅LC∕Ψ𝑆 across the open magnetic field lines. occasionally regulate the competing losses well enough to guide a favorable seed to convergence. To … view at source ↗
Figure 7
Figure 7. Figure 7: Converged magnetic field line configurations for decreasing stellar radii at polar cap angle 𝜃pc = arcsin √ 1.231𝑅∗ : 𝑅∗ = 0.2 (top-left), 𝑅∗ = 0.15 (top-middle), 𝑅∗ = 0.125 (top-right), 𝑅∗ = 0.1 (bottom-left), 𝑅∗ = 0.0625 (bottom-middle), and 𝑅∗ = 0.05 (bottom-right). Across all configurations, the maximum deviation of the equatorial T-point from the baseline location (𝑥𝑇 = 0.925 𝑅LC) is 0.009 𝑅LC. S. Rig… view at source ↗
Figure 8
Figure 8. Figure 8: Converged magnetic field line configurations for 𝑅∗ = 0.25 and varying flux ratios at polar cap opening angle 𝜃pc = 𝑚 √ 𝑅∗ : 𝑚 = 1.4 (top-left), 𝑚 = 1.35 (top-middle), 𝑚 = 1.3 (top-right), 𝑚 = 1.25 (center-left), 𝑚 = 1.2 (center-middle), 𝑚 = 1.176 (center-right), 𝑚 = 1.15 (bottom-left), 𝑚 = 1.144 (bottom-middle), and 𝑚 = 1.139 (bottom-right). S. Rigas et al.: Preprint submitted to Elsevier Page 18 of 25 [… view at source ↗
Figure 9
Figure 9. Figure 9: Magnetic flux ratio 𝜌 as a function of the equatorial T-point position 𝑥𝑇 ∕𝑅LC at polar cap opening angle 𝜃pc = 𝑚 √ 𝑅∗ , with varying 𝑚. The dashed line represents the curve 𝜌 = 1.138∕𝑥𝑇 derived from the baseline configuration for 𝑅∗ = 0.25. Scatter points indicate the average values computed across three independent seeds, with error bars corresponding to standard error. In addition to this physically sig… view at source ↗
Figure 10
Figure 10. Figure 10: Self-consistent pulsar magnetosphere solution under double precision (FP64), where 𝑧 = 𝑟 cos 𝜃 and 𝑥 = 𝑟 sin 𝜃. Left: Magnetic field lines shown as contours of constant flux. The equatorial T-point is located at 𝑥𝑇 = 0.922 𝑅LC. Right: Normalized poloidal current distribution |𝐼(Ψ)|𝑅LC∕Ψ𝑆 across the open magnetic field lines. The dotted line denotes the theoretical analytic expression for the split monopol… view at source ↗
read the original abstract

The pulsar magnetosphere has only recently been addressed using Physics-Informed Neural Networks (PINNs), by deploying a domain-decomposition approach and treating the separatrix and equatorial current sheet as infinitesimally thin discontinuities. However, this baseline requires extensive manual hyperparameter tuning, achieves limited final accuracy and demands several hours of training. We refine this framework by introducing domain-specific neural architectures based on Kolmogorov-Arnold networks, an automated adaptive training pipeline and a physics-based convergence criterion that eliminate the need for manual calibration. The proposed methodology delivers self-consistent axisymmetric magnetosphere solutions with mean squared errors of the PDE residuals at O(1e-6) in double precision - an improvement of two orders of magnitude over the baseline - while achieving convergence in under 20 minutes in single precision. Importantly, the method reliably resolves stellar radii reduced by up to 80% compared to the baseline, overcoming the severe spatial scale disparities that also challenge traditional solvers. Furthermore, by varying the flux that opens to infinity, we provide a correction to the equation that connects it to the equatorial T-point's position. The complete framework is released as the open-source library PulsarX.

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 / 2 minor

Summary. The manuscript introduces an adaptive PINN framework for the axisymmetric pulsar magnetosphere that replaces standard neural networks with Kolmogorov-Arnold networks, adds an automated adaptive training pipeline, and employs a physics-based convergence criterion. It reports self-consistent solutions with PDE residual MSE of O(1e-6) in double precision (two orders of magnitude better than a cited baseline), convergence in under 20 minutes in single precision, reliable solutions at stellar radii reduced by up to 80%, and a correction to the relation between open flux and the equatorial T-point position. The complete framework is released as the open-source library PulsarX.

Significance. If the reported accuracy, runtime, and scale-handling gains hold under independent verification, the work would represent a meaningful advance in applying neural solvers to relativistic astrophysical problems with disparate spatial scales. The open-source release and the claimed correction to the flux-T-point equation are concrete strengths that support reproducibility and potential impact on the field.

major comments (2)
  1. [§4.2, Table 2] §4.2 and Table 2: the two-order-of-magnitude improvement in PDE residual MSE is stated relative to a manually tuned baseline PINN, but the manuscript does not report the exact hyperparameter search budget or final residual values achieved by that baseline under identical domain decomposition and loss weighting; without these numbers the quantitative claim cannot be assessed.
  2. [§5.3, Eq. (18)] §5.3, Eq. (18): the correction to the open-flux versus T-point relation is presented as a new result obtained by varying the open flux, yet the manuscript provides neither the functional form of the original relation nor a direct side-by-side comparison of the new fit against analytic expectations or prior numerical solutions.
minor comments (2)
  1. [Figure 3, §3.1] Figure 3 caption and §3.1: the notation for the KAN spline order and grid size is introduced only in the caption; move the definitions into the main text for clarity.
  2. [§6] §6: the open-source repository link is given but the exact commit hash or release tag used for the reported results is not stated.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment and the recommendation of minor revision. The comments are constructive and we address them point-by-point below, indicating the revisions we will make to strengthen the manuscript.

read point-by-point responses

  1. Referee: [§4.2, Table 2] §4.2 and Table 2: the two-order-of-magnitude improvement in PDE residual MSE is stated relative to a manually tuned baseline PINN, but the manuscript does not report the exact hyperparameter search budget or final residual values achieved by that baseline under identical domain decomposition and loss weighting; without these numbers the quantitative claim cannot be assessed.

    Authors: We agree that additional detail on the baseline would strengthen the comparison. The reported baseline residuals are taken directly from the values published in the cited reference for the standard PINN implementation. In the revised manuscript we will explicitly list the hyperparameter settings used for that baseline (as described in the reference) and confirm that domain decomposition and loss weighting match those employed in our adaptive framework. We will also note that a full re-execution of the baseline under identical conditions was not performed due to the extensive manual tuning required; this limitation will be stated transparently. revision: yes

  2. Referee: [§5.3, Eq. (18)] §5.3, Eq. (18): the correction to the open-flux versus T-point relation is presented as a new result obtained by varying the open flux, yet the manuscript provides neither the functional form of the original relation nor a direct side-by-side comparison of the new fit against analytic expectations or prior numerical solutions.

    Authors: We accept that including the original relation and a direct comparison would improve clarity. The original functional form is the one given by the baseline literature (specifically the relation derived in the cited prior work). In the revised manuscript we will state the original equation explicitly, provide the new fitted form, and add a side-by-side table (or figure) comparing both against the analytic expectation and available prior numerical solutions. This will make the correction and its improvement evident. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on empirical performance of implemented solver

full rationale

The paper introduces a KAN-based adaptive PINN framework for axisymmetric pulsar magnetosphere PDEs and reports quantitative metrics (O(1e-6) residuals, <20 min convergence, 80% smaller stellar radii) versus a cited baseline. These are obtained by running the solver on the target equations; no derivation step reduces a claimed prediction or result to a fitted parameter or self-citation by construction. The correction to the flux-T-point relation is extracted from the numerical solutions themselves. Self-citation of the baseline is present but not load-bearing for the new method's performance claims, which remain externally verifiable via the released code.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on standard domain assumptions from force-free electrodynamics and PINN methodology for PDE solving; no free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption The axisymmetric pulsar magnetosphere is governed by PDEs that can be solved via domain decomposition with infinitesimally thin separatrix and equatorial current sheet discontinuities.
    This is the baseline setup being refined and is standard in the cited prior PINN work for this problem.

pith-pipeline@v0.9.1-grok · 5751 in / 1534 out tokens · 45849 ms · 2026-06-27T11:00:57.055777+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

32 extracted references · 30 canonical work pages

  1. [1]

    Computer Methods in Applied Mechanics and Engineering 421, 116805

    Residual-based attention in physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering 421, 116805. doi:10.1016/j.cma.2024.116805. Bogovalov, S.,

    work page doi:10.1016/j.cma.2024.116805 2024

  2. [2]

    M., & Levin, Y

    JAX: composable transformations of Python+NumPy programs. URL:http://github.com/jax-ml/jax. Bransgrove,A.,Beloborodov,A.M.,Levin,Y.,2023. Radioemissionandelectricgapsinpulsarmagnetospheres. TheAstrophysicalJournalLetters 958, L9. doi:https://doi.org/10.3847/2041-8213/ad0556. Cerutti, B., Philippov, A., Parfrey, K., Spitkovsky, A.,

    work page doi:10.3847/2041-8213/ad0556 2023

  3. [3]

    Monthly Notices of the Royal Astronomical Society 448, 606–619

    Particle acceleration in axisymmetric pulsar current sheets. Monthly Notices of the Royal Astronomical Society 448, 606–619. doi:10.1093/mnras/stv042. Contopoulos,I.,2007. Theroleofreconnectioninthepulsarmagnetosphere. Astronomy&Astrophysics466,301–307. doi: 10.1051/0004-6361: 20065973. Contopoulos, I., Dimitropoulos, I., Ntotsikas, D., Gourgouliatos, K.N...

    work page doi:10.1093/mnras/stv042 2007

  4. [4]

    Contopoulos,I.,Kazanas,D.,Fendt,C.,1999.Theaxisymmetricpulsarmagnetosphere.TheAstrophysicalJournal511,351.doi: 10.1086/306652

    doi:10.3390/ universe10040178. Contopoulos,I.,Kazanas,D.,Fendt,C.,1999.Theaxisymmetricpulsarmagnetosphere.TheAstrophysicalJournal511,351.doi: 10.1086/306652. Contopoulos, I., Ntotsikas, D., Gourgouliatos, K.N., 2024b. On the pulsar y-point. Monthly Notices of the Royal Astronomical Society: Letters 527, L127–L131. doi:10.1093/mnrasl/slad153. Dherin,B.,Mun...

    work page doi:10.1086/306652 1999

  5. [5]

    Astronomy & Astrophysics 705, A140

    The 3d pulsar magnetosphere with machine learning: First results. Astronomy & Astrophysics 705, A140. doi:10.1051/0004-6361/202554175. Dimitropoulos,I.,Contopoulos,I.,Mpisketzis,V.,Chaniadakis,E.,2024. Thepulsarmagnetospherewithmachinelearning:methodology. Monthly Notices of the Royal Astronomical Society 528, 3141–3152. doi:10.1093/mnras/stae192. Endean, V.G.,

    work page doi:10.1051/0004-6361/202554175 2024

  6. [6]

    Astrophysical Journal 187, 359–360

    Lorentz force-free pulsar rotating fields. Astrophysical Journal 187, 359–360. doi:10.1086/152639. Faroughi, S.A., Mostajeran, F.,

    work page doi:10.1086/152639

  7. [7]

    Neural tangent kernel analysis to probe convergence in physics-informed neural solvers: PIKANs vs. PINNs. Computers & Mathematics with Applications 215, 155–191. doi:10.1016/j.camwa.2026.05.003. Goldreich, P., Julian, W.H.,

    work page doi:10.1016/j.camwa.2026.05.003 2026

  8. [8]

    H., 1969, @doi [ ] 10.1086/150119 , https://ui.adsabs.harvard.edu/abs/1969ApJ...157..869G 157, 869

    doi:10.1086/150119. Hewish, A., Bell, S.J., Pilkington, J.D.H., Scott, P.F., Collins, R.A.,

    work page doi:10.1086/150119

  9. [9]

    J., Pilkington J

    Observation of a rapidly pulsating radio source. Nature 217, 709–713. doi:10.1038/217709a0. Jacot, A., Gabriel, F., Hongler, C.,

    work page doi:10.1038/217709a0

  10. [10]

    8580–8589

    Neural tangent kernel: convergence and generalization in neural networks, in: Proceedings of the 32nd International Conference on Neural Information Processing Systems, p. 8580–8589. Jagtap,A.D.,Mao,Z.,Adams,N.,Karniadakis,G.E.,2022. Physics-informedneuralnetworksforinverseproblemsinsupersonicflows. Journalof Computational Physics 466, 111402. doi:10.1016...

    work page doi:10.1016/j.jcp.2022.111402 2022

  11. [11]

    Kalapotharakos,C.,Contopoulos,I.,2009

    doi:10.3847/1538-4357/aab550. Kalapotharakos,C.,Contopoulos,I.,2009. Three-dimensionalnumericalsimulationsofthepulsarmagnetosphere:preliminaryresults. Astronomy & Astrophysics 496, 495–502. doi:10.1051/0004-6361:200810281. Kingma, D.P., Ba, J.,

    work page doi:10.3847/1538-4357/aab550 2009

  12. [12]

    2005, MNRAS, 361, 919, doi: 10.1111/j.1365-2966.2005.09216.x

    Simulations of the axisymmetric magnetospheres of neutron stars. Monthly Notices of the Royal Astronomical Society 367, 19–31. doi:10.1111/j.1365-2966.2005.09932.x. Krishnapriyan,A.,Gholami,A.,Zhe,S.,Kirby,R.,Mahoney,M.,2021. Characterizingpossiblefailuremodesinphysics-informedneuralnetworks, in: Advances in Neural Information Processing Systems, pp. 2654...

    work page doi:10.1111/j.1365-2966.2005.09932.x 2005

  13. [13]

    Neural Networks 136, 112–125

    A dual-dimer method for training physics-constrained neural networks with minimax architecture. Neural Networks 136, 112–125. doi:10.1016/j.neunet.2020.12.028. Liu, Z., Tegmark, M., Ma, P., Matusik, W., Wang, Y., 2025a. Kolmogorov-Arnold networks meet science. Physical Review X 15, 041051. doi:10.1103/4t7t-v19l. Liu,Z.,Wang,Y.,Vaidya,S.,Ruehle,F.,Halverso...

    work page doi:10.1016/j.neunet.2020.12.028 2020

  14. [14]

    Physical Review D 107, 064025

    Solving the teukolsky equation with physics-informed neural networks. Physical Review D 107, 064025. doi:10.1103/PhysRevD.107.064025. Mestel, L.,

    work page doi:10.1103/physrevd.107.064025

  15. [15]

    Mémoires of the Société Royale des Sciences de Leège 8, 79–91

    Effects of magnetic fields. Mémoires of the Société Royale des Sciences de Leège 8, 79–91. Muslimov,A.G., Harding,A.K.,2009. Pair-starvedpulsar magnetospheres. The AstrophysicalJournal692,

    2009

  16. [16]

    Philippov,A.A.,Spitkovsky,A.,2018

    doi:10.1088/0004-637X/692/ 1/140. Philippov,A.A.,Spitkovsky,A.,2018. Ab-initiopulsarmagnetosphere:Particleaccelerationinobliquerotatorsandhigh-energyemissionmodeling. The Astrophysical Journal 855,

    work page doi:10.1088/0004-637x/692/ 2018

  17. [17]

    Pétri,J.,2012

    doi:10.3847/1538-4357/aaabbc. Pétri,J.,2012. Thepulsarforce-freemagnetospherelinkedtoitsstripedwind:time-dependentpseudo-spectralsimulations. MonthlyNoticesofthe Royal Astronomical Society 424, 605–619. doi:10.1111/j.1365-2966.2012.21238.x. Raissi, M., Perdikaris, P., Karniadakis, G.E.,

    work page doi:10.3847/1538-4357/aaabbc 2012

  18. [18]

    Raissi, P

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 378, 686–707. doi:10.1016/j.jcp.2018.10

    work page doi:10.1016/j.jcp.2018.10 2018

  19. [19]

    Physical Review Fluids 6, 073301

    Learning unknown physics of non-newtonian fluids. Physical Review Fluids 6, 073301. doi:10.1103/PhysRevFluids.6.073301. Rigas, S., Anagnostopoulos, F., Papachristou, M., Alexandridis, G., 2026a. Training deep physics-informed Kolmogorov–Arnold networks. Computer Methods in Applied Mechanics and Engineering 452, 118761. doi:10.1016/j.cma.2026.118761. Rigas...

    work page doi:10.1103/physrevfluids.6.073301 2026

  20. [20]

    Rigas,S.,Papachristou,M.,Papadopoulos,T.,Anagnostopoulos,F.,Alexandridis,G.,2024

    doi:10.21105/joss.07830. Rigas,S.,Papachristou,M.,Papadopoulos,T.,Anagnostopoulos,F.,Alexandridis,G.,2024. Adaptivetrainingofgrid-dependentphysics-informed Kolmogorov–Arnold networks. IEEE Access 12, 176982–176998. doi:10.1109/ACCESS.2024.3504962. Rigas, S., Verma, D., Alexandridis, G., Wang, Y., 2026b. Initialization schemes for Kolmogorov–Arnold network...

    work page doi:10.21105/joss.07830 2024

  21. [21]

    Neural Networks 162, 472–489

    Deep learning-accelerated computational framework based on physics informed neural network for the solution of linear elasticity. Neural Networks 162, 472–489. doi:10.1016/j.neunet.2023.03.014. Shukla, K., Toscano, J.D., Wang, Z., Zou, Z., Karniadakis, G.E.,

    work page doi:10.1016/j.neunet.2023.03.014 2023

  22. [22]

    Computer Methods in Applied Mechanics and Engineering 431, 117290

    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. doi:10.1016/j.cma.2024.117290. Soudais, A., Cerutti, B., Contopoulos, I.,

    work page doi:10.1016/j.cma.2024.117290 2024

  23. [23]

    Astronomy & Astrophysics 690, A170

    Scaling up global kinetic models of pulsar magnetospheres using a hybrid force-free-pic numerical approach. Astronomy & Astrophysics 690, A170. doi:10.1051/0004-6361/202450238. Spitkovsky,A.,2006. Time-dependentforce-freepulsarmagnetospheres:Axisymmetricandobliquerotators. TheAstrophysicalJournal648,L51. doi:10.1086/507518. Stefanou, P., Urbán, J.F., Pons, J.A.,

    work page doi:10.1051/0004-6361/202450238 2006

  24. [24]

    Monthly Notices of the Royal Astronomical Society 526, 1504–1511

    Solving the pulsar equation using physics-informed neural networks. Monthly Notices of the Royal Astronomical Society 526, 1504–1511. doi:10.1093/mnras/stad2840. Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Singhal, U., Ramamoorthi, R., Barron, J., Ng, R.,

    work page doi:10.1093/mnras/stad2840

  25. [25]

    Monthly Notices of the Royal Astronomical Society: Letters 435, L1–L5

    Time-dependent 3d magnetohydrodynamic pulsar magnetospheres: oblique rotators. Monthly Notices of the Royal Astronomical Society: Letters 435, L1–L5. doi:10.1093/mnrasl/slt076. Timokhin,A.N.,2006. Ontheforce-freemagnetosphereofanalignedrotator. MonthlyNoticesoftheRoyalAstronomicalSociety368,1055–1072. doi:10.1111/j.1365-2966.2006.10192.x. Toscano, J.D., O...

    work page doi:10.1093/mnrasl/slt076 2006

  26. [26]

    Toscano, J.D., Wang, L.L., Karniadakis, G.E., 2025b

    doi:10.1007/ s44379-025-00015-1. Toscano, J.D., Wang, L.L., Karniadakis, G.E., 2025b. KKANs: Kurkova-Kolmogorov-Arnold networks and their learning dynamics. Neural Networks 191, 107831. doi:10.1016/j.neunet.2025.107831. Tseneklidou,D., Torres-Forné,A.,Cerdá-Durán,P.,

    work page doi:10.1016/j.neunet.2025.107831 2025

  27. [27]

    doi:10.1140/epjc/s10052-025-14942-z. S. Rigas et al.:Preprint submitted to Elsevier Page 22 of 25 An adaptive framework for the axisymmetric pulsar magnetosphere using physics-informed KANs Urbán,J.F.,Stefanou,P.,Dehman,C.,Pons,J.A.,2023. Modellingforce-freeneutronstarmagnetospheresusingphysics-informedneuralnetworks. Monthly Notices of the Royal Astronom...

    work page doi:10.1140/epjc/s10052-025-14942-z 2023

  28. [28]

    Wang, S., Li, B., Chen, Y., Perdikaris, P.,

    doi:10.1086/378849. Wang, S., Li, B., Chen, Y., Perdikaris, P.,

    work page doi:10.1086/378849

  29. [29]

    Understanding and Mitigating Gradient Flow Pathologies in Physics-Informed Neural Networks , volume =

    Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing 43, A3055–A3081. doi:10.1137/20M1318043. Wang, S., Yu, X., Perdikaris, P.,

    work page doi:10.1137/20m1318043

  30. [30]

    When and why PINNs fail to train: A neural tangent kernel perspective

    When and why pinns fail to train: A neural tangent kernel perspective. Journal of Computational Physics 449, 110768. doi:10.1016/j.jcp.2021.110768. Wang, Y., Siegel, J.W., Liu, Z., Hou, T.Y.,

    work page doi:10.1016/j.jcp.2021.110768 2021

  31. [31]

    URL:https://openreview.net/forum?id=ydlDRUuGm9

    On the expressiveness and spectral bias of KANs, in: The Thirteenth International Conference on Learning Representations. URL:https://openreview.net/forum?id=ydlDRUuGm9. Xu,K.,Tartakovsky,A.M.,Burghardt,J.,Darve,E.,2021. Learningviscoelasticitymodelsfromindirectdatausingdeepneuralnetworks. Computer Methods in Applied Mechanics and Engineering 387, 114124....

    work page doi:10.1016/j.cma.2021.114124 2021

  32. [32]

    + 2𝑁 + 2𝐷𝑠 + 𝑑H (𝑑I𝐷𝑠 + 𝑑O𝐷 + 1). (48) For the purposes of this work, we utilize RGA KANs with 2-dimensional inputs (the spherical coordinates𝑟, 𝜃) forthetworegionalPINNs:onefortheclosed-lineregionwith1-dimensionaloutput,usedtoapproximate Ψcl viaEq. (21), and one for the open-line region with 2-dimensional output, used to approximateΨop and 𝐼 via Eqs. (22...

    2024