Pith. sign in

REVIEW 30 references

Physics-Informed Neural Networks for Discovering Periodic Orbits in the Gravitational Three-Body Problem

T0 review · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read Physics-informed neural networks recover periodic three-body orbits from sparse noisy data with no initial conditions, including families never seen in training.

desk verdict PINNs without ICs really do spit out catalog-checkable three-body periodic families from sparse noisy data, with a clean χ² split showing training data (not Glorot U vs N) shifts the family mix. read the letter →

arxiv 2607.23501 v1 pith:TTS7VLMF submitted 2026-07-26 cs.LG nlin.CDphysics.comp-ph

Nikolaos Kollias , Nikolaos Matzakos This is my paper
classification cs.LGnlin.CDphysics.comp-ph
keywords physics-informedneuralnetworksinverseproblemsunknowninitialconditionscelestialmechanicsperiodicorbitfamiliesmulti-modallosslandscapethree-bodyproblem
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

Finding periodic orbits in the chaotic three-body problem usually demands an initial guess already close to the target. This paper shows that a physics-informed neural network, given only sparse noisy observations and no initial conditions, can still recover genuine periodic orbits—and that roughly one run in four lands on a family that was not in the training data. Controlled seed ensembles and chi-squared tests separate two influences: which family a single run finds is seed-dependent, but the aggregate mix of families shifts when the training data source changes and does not shift when the weight-initialization distribution is swapped. Recovered states refine to true periodic solutions that match published catalogs, including the figure-eight choreography from Lagrange data and a Broucke–Hadjidemetriou–Hénon orbit from figure-eight data. A sympathetic reader cares because the method supplies candidate initial conditions in regions where no good guess exists to seed ordinary continuation or gradient search.

What carries the argument

The under-determined inverse PINN loss—ODE residual plus data fit, with initial conditions removed—whose multi-modal landscape lets different random seeds settle in different orbit-family basins. Enabling pieces, each validated on forward problems, are a second-order ODE residual, fixed-frequency Fourier features, percentile-based adaptive collocation, and a trainable residual scaling parameter C.

What would settle it

Re-run the same seed ensembles and classify every refined orbit by full numerical continuation into known families; if cross-family rates collapse or the chi-squared contrast between training-data sources disappears, the claim that data steers recovery of distinct genuine families fails.

Watch

Extended reading notes

Core claim

PINNs trained on sparse, noisy observations without initial conditions recover periodic orbits of the gravitational three-body problem, including families absent from the training data. Across two 100-seed ensembles, 23–25% of runs converge to such families. Changing the training data source significantly shifts the distribution of recovered families, whereas switching between Glorot Uniform and Glorot Normal does not. The recovered orbits are verifiable: refined states close as genuine periodic solutions and match catalogued families, including the figure-eight (Li–Liao I.A.1) from Lagrange data to seven digits in the scale-invariant period T*.

Load-bearing premise

That matching the scale-invariant period after numerical refinement, plus simple geometry for the classical cases, is enough to name non-classical families without independent continuation checks.

Editorial extensions

If this is right

  • Ensembles of PINNs on sparse observations can generate candidate initial conditions without a prior guess for conventional search.
  • To diversify recovered families, vary the training data source rather than the Glorot initialization variant.
  • Cross-family recovery is bidirectional: Lagrange data can yield the figure-eight, and figure-eight data can yield BHH-type orbits.
  • The niche is inverse and exploratory settings; the method is not a substitute for integrators on well-posed initial-value problems.

Reading between the lines

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

  • The same data-shaped multi-basin pattern may appear in other multi-stable oscillators; the Duffing system, flagged as an open test by the authors, is the natural next check.
  • If mode-connectivity or Hessian probes show separated basins whose relative volumes track the training data, the statistical picture would gain a geometric basis the paper leaves open.
  • Moderate observation noise may act as a regularizer that keeps several basins reachable; a controlled noise sweep would test whether the cross-family fraction is noise-dependent.
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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical PINN discoveries verified against external classical reductions and published catalogs, not against quantities defined from the same fit.

full rationale

The paper’s load-bearing chain is experimental, not definitional. PINNs are trained on sparse noisy observations plus the Newtonian residual without initial conditions; candidate states are then least-squares refined and checked by independent numerical closure (δ_T) and by scale-invariant invariants (T*, L*) against classical central-configuration reductions (T*_Lag = 6π/2^{3/2}, T*_Eul = 5π/2) and external catalogs (Li–Liao I.A.1 figure-eight; Broucke–BHH literature). Those reference values are not fitted from the PINN outputs. Conservation diagnostics are never in the training loss. The χ² tests compare controlled experimental conditions (training-data source vs. Glorot Uniform/Normal) rather than restating labels. Methodological pieces (second-order ODE form, fixed-frequency Fourier features on forward problems with known T, percentile RAR, trainable C) are ablated on forward tasks and do not force the cross-family discovery claims. Self-citations (e.g. Matzakos & Sfyrakis) are related-work context, not uniqueness theorems or ansatz smuggling that carry the central result. The authors’ own “-like” qualifier and admission that BHH continuation was not performed are scope limits, not circular reductions. No step reduces a claimed prediction to its inputs by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

The central claim is empirical and rests on standard Newtonian planar three-body dynamics, standard PINN optimization, and catalog/invariant identification conventions—not on new physical entities. Load-bearing modeling choices include softening, fixed Fourier frequencies tied to a known period (forward fig-eight only), percentile RAR bands, trainable residual scale C, noise level, loss weights, and geometric/T* classification thresholds. Free parameters are hyperparameters and classification cutoffs, not fitted physical constants claiming universality.

free parameters (6)
  • softening ε = 10^{-9}
    Pairwise distance regularizer √(r²+ε) with ε=10^{-9} chosen to keep forces finite during training; affects residual landscape before convergence.
  • trainable residual scale C (C0) = C0=3; final C≈17 in one reported run
    Initialized at 3 and jointly optimized so residual weight decays as 1/C^4; enables data term influence in the no-IC setting.
  • loss weights wr, wd and AdamW hyperparameters = ensemble: wr=wd=1, η=1e-4, wd_decay=0.004
    Hand-chosen and changed between single-run (wr=0.1, wd=40, η=1e-3) and ensemble (wr=wd=1, η=1e-4) protocols; directly shape basin sampling.
  • modified RAR percentile band and n_add = p_low=60, p_high=95, n_add=64
    60th–95th percentile residual sampling with n_add=64; design choice controlling collocation adaptation.
  • observation noise level and sample count = N=90, 20% noise
    90 points at 20% noise define the inverse problem; authors note >50% noise failed but give no full noise ablation.
  • family classification thresholds (collinearity, equilateral, T* bins) = collinearity O(10^{-16}); equilateral O(10^{-11}); δ_T<10^{-8} discovery gate
    Machine-precision collinearity and O(10^{-11}) equilateral cuts on refined trajectories, plus T* matching for BHH/other, determine reported family counts and χ² tables.
assumptions (5)
  • domain assumption Planar Newtonian three-body equations with G=1 and equal masses are the true generative dynamics of the observations.
    Section 2; all residuals and catalogs assume this model.
  • domain assumption Scale-invariant period T*=T|E|^{3/2} (and L*) uniquely enough identifies classical families and Li–Liao/BHH membership for the reported taxonomy.
    Section 5.2 and Appendix D; authors note T* alone does not determine a unique orbit for BHH.
  • domain assumption Least-squares refinement of PINN-inferred ICs to δ_T<10^{-8} certifies a genuine periodic solution in the basin of the network output.
    Discovery definition in Section 5; standard in computational celestial mechanics but refinement can travel far (high ΔIC “other” cases).
  • standard math Pearson χ² independence tests on 4×2 family-count tables validly compare init and data effects (with figure-eight merged into other).
    Sections 5.2.3 and 5.3; expected frequencies >5 stated.
  • ad hoc to paper Fixed-frequency Fourier features using the known orbital period improve forward figure-eight PINNs primarily via harmonics, not T/3 phase encoding.
    Section 3.3 rank argument and Table 11; methodological enabler for forward validation, not for the inverse no-IC claim.
invented entities (2)
  • Percentile-band modified RAR (60th–95th residual sampling)
    purpose: Avoid standard RAR overfitting to singularity-adjacent outliers while refining collocation.
    Algorithm 1; engineering variant of existing RAR/RAD, not a new physical object.
  • Trainable residual scaling parameter C with 1/C^4 loss factor
    purpose: Automatically anneal PDE residual weight relative to data loss without external loss-balancing loops.
    Section 3.5; optimization device related to prior adaptive weighting / term scaling literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-Informed Neural Networks for Discovering Periodic Orbits in the Gravitational Three-Body Problem." pith.science (2026). https://pith.science/paper/TTS7VLMF

@misc{pith2026260723501,
  author       = {Pith},
  title        = {Pith review of: Physics-Informed Neural Networks for Discovering Periodic Orbits in the Gravitational Three-Body Problem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TTS7VLMF}},
  note         = {Machine review of arXiv:2607.23501}
}
abstract

Locating periodic solutions of chaotic dynamical systems normally requires an initial guess close enough to the target orbit for numerical continuation or gradient-based search to converge. We show that Physics-Informed Neural Networks (PINNs) trained on sparse, noisy observations \emph{without} initial conditions recover periodic orbits of the gravitational three-body problem, including orbit families absent from the training data. The method rests on a second-order ODE formulation, fixed-frequency Fourier features, percentile-based adaptive refinement, and a trainable scaling parameter, each validated on forward problems. Across two 100-seed ensembles, $23$--$25\%$ of runs converge to families not present in the training data. We then ask what determines which family emerges. Two $\chi^2$ tests give a consistent answer: changing the training data source significantly shifts the distribution of recovered families ($p < 0.001$, Cram\'{e}r's $V = 0.339$), whereas switching between the two initialization distributions tested does not ($p = 0.620$, $V = 0.094$). The random seed selects which family a given run recovers; the \emph{distribution} the weights are drawn from does not shift the aggregate frequencies, but the training data does. The evidence is empirical: we do not characterize the loss landscape analytically, and PINNs remain slower than conventional integrators on well-posed initial-value problems. What the experiments establish is that the recovered orbits are verifiable rather than merely plausible: the identified ones refine to genuine periodic solutions, a network trained on Lagrange data recovers the figure-eight choreography (Li--Liao class I.A.1, matched to seven significant digits in $T^*$), and one trained on figure-eight data recovers a Broucke--Hadjidemetriou--H\'{e}non orbit closing to $\delta_T < 10^{-9}$.

Figures

Figures reproduced from arXiv: 2607.23501 by the authors.

Figure 1
Figure 1. Autonomous orbit discovery without initial conditions. Trained on the same [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. Glorot Uniform: orbit family distribution across 100 random seeds (top: color [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Glorot Uniform: periodic orbit invariant map for all 100 seeds, plotted [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Glorot Uniform with BHH training data: orbit family distribution across 100 [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Glorot Uniform with BHH training data: periodic orbit invariant map ( [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Representative BHH periodic orbits recovered by PINNs trained on noisy BHH [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Cross-family orbit discovery. (a) PINN trained on 90 noisy observations (dots) [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Least-squares refined orbit discovered by a PINN trained on noisy Lagrange data [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: Euler collinear orbits: PINN trajectories (dashed) vs. numerical reference [PITH_FULL_IMAGE:figures/full_fig_p032_9.png]
Figure 10
Figure 10. Figure 10: Evolution of residual distributions during modified RAR iterations for La [PITH_FULL_IMAGE:figures/full_fig_p033_10.png]
Figure 11
Figure 11. Figure 11: Lagrange orbits with modified RAR: (a) Loss convergence showing train and [PITH_FULL_IMAGE:figures/full_fig_p033_11.png]
Figure 12
Figure 12. Figure 12: Figure-eight orbit trajectories: (a) Standard PINN fails to capture the sym [PITH_FULL_IMAGE:figures/full_fig_p034_12.png]
Figure 13
Figure 13. Figure 13: Comparison of a standard feedforward network (a) and a PINN with hard [PITH_FULL_IMAGE:figures/full_fig_p036_13.png]
Figure 14
Figure 14. Figure 14: Data-driven PINN with trainable C: (a) Trajectories (dashed) vs. numerical solution (solid) and noisy data (dots). (b) C increases approximately linearly, acting as an implicit annealing schedule. D Glorot Normal initialization: detailed results This appendix presents…
Figure 15
Figure 15. Figure 15: Glorot Normal: orbit family distribution across 100 random seeds, with the [PITH_FULL_IMAGE:figures/full_fig_p038_15.png]
Figure 16
Figure 16. Figure 16: Glorot Normal: periodic orbit invariant map. The structure mirrors the GU [PITH_FULL_IMAGE:figures/full_fig_p038_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 5 canonical work pages

  1. [1]

    Breen, Christopher N

    Philip G. Breen, Christopher N. Foley, Tjarda Boekholt, and Simon Portegies Zwart. Newton versus the machine: solving the chaotic three-body problem using deep neural networks. Monthly Notices of the Royal Astronomical Society, 494: 0 2465--2470, 2020

    2020

  2. [2]

    On relative periodic solutions of the planar general three-body problem

    Roger Broucke. On relative periodic solutions of the planar general three-body problem. Celestial Mechanics, 12: 0 439--462, 1975

    1975

  3. [3]

    A remarkable periodic solution of the three-body problem in the case of equal masses

    Alain Chenciner and Richard Montgomery. A remarkable periodic solution of the three-body problem in the case of equal masses. Annals of Mathematics, 152: 0 881--901, 2000

    2000

  4. [4]

    Scientific machine learning through physics-informed neural networks: Where we are and what's next

    Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics-informed neural networks: Where we are and what's next. Journal of Scientific Computing, 92: 0 88, 2022

    2022

  5. [5]

    De motu rectilineo trium corporum se mutuo attrahentium

    Leonhard Euler. De motu rectilineo trium corporum se mutuo attrahentium. Novi Commentarii Academiae Scientiarum Petropolitanae, 11: 0 144--151, 1767

  6. [6]

    Sukumar, Ankit Srivastava, and Hadi Meidani

    Rini Jasmine Gladstone, Mohammad Amin Nabian, N. Sukumar, Ankit Srivastava, and Hadi Meidani. FO-PINNs : A first-order formulation for physics-informed neural networks. Engineering Analysis with Boundary Elements, 174: 0 106161, 2025. arXiv:2210.14320

    arXiv 2025

  7. [7]

    Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

    George Em Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics, 3: 0 422--440, 2021

    2021

  8. [8]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2015

    2015

Show all 30 references
  1. [9]

    Lagaris, Aristidis Likas, and Dimitrios I

    Isaac E. Lagaris, Aristidis Likas, and Dimitrios I. Fotiadis. Artificial neural networks for solving ordinary and partial differential equations. IEEE Transactions on Neural Networks, 9: 0 987--1000, 1998

    1998

  2. [10]

    Essai sur le probl\` e me des trois corps

    Joseph-Louis Lagrange. Essai sur le probl\` e me des trois corps. Prix de l'Acad\' e mie Royale des Sciences de Paris , 9, 1772

  3. [11]

    More than six hundred new families of Newtonian periodic planar collisionless three-body orbits

    Xiaoming Li and Shijun Liao. More than six hundred new families of Newtonian periodic planar collisionless three-body orbits. Science China Physics, Mechanics & Astronomy, 60: 0 129511, 2017. doi:10.1007/s11433-017-9078-5

    2017 doi

  4. [12]

    One family of 13315 stable periodic orbits of non-hierarchical unequal-mass triple systems

    XiaoMing Li, XiaoChen Li, and ShiJun Liao. One family of 13315 stable periodic orbits of non-hierarchical unequal-mass triple systems. Science China Physics, Mechanics & Astronomy, 64: 0 219511, 2021. doi:10.1007/s11433-020-1624-7

    2021 doi

  5. [13]

    Three-body problem---from Newton to supercomputer plus machine learning

    Shijun Liao, Xiaoming Li, and Yu Yang. Three-body problem---from Newton to supercomputer plus machine learning. New Astronomy, 96: 0 101850, 2022

    2022

  6. [14]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proceedings of the 7th International Conference on Learning Representations (ICLR), 2019

    2019

  7. [15]

    DeepXDE : A deep learning library for solving differential equations

    Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. DeepXDE : A deep learning library for solving differential equations. SIAM Review, 63: 0 208--228, 2021

    2021

  8. [16]

    Matzakos and Christos A

    Nikolaos M. Matzakos and Christos A. Sfyrakis. Comparing physics-informed and neural ODE approaches for modeling nonlinear biological systems: A case study based on the Morris--Lecar model. arXiv preprint arXiv:2603.26921, 2026. URL https://arxiv.org/abs/2603.26921

    2026

  9. [17]

    Meyer, Glen R

    Kenneth R. Meyer, Glen R. Hall, and Dan Offin. Introduction to Hamiltonian Dynamical Systems and the N -Body Problem , volume 90 of Applied Mathematical Sciences. Springer, New York, 2 edition, 2009. doi:10.1007/978-0-387-09724-4

    2009 doi

  10. [18]

    Central configurations

    Richard Moeckel. Central configurations. Scholarpedia, 9 0 (4): 0 10667, 2014. doi:10.4249/scholarpedia.10667

    2014 doi

  11. [19]

    Braids in classical dynamics

    Cristopher Moore. Braids in classical dynamics. Physical Review Letters, 70: 0 3675--3679, 1993

    1993

  12. [20]

    Advancing solutions for the three-body problem through physics-informed neural networks

    Manuel Santos Pereira, Lu \' s Tripa, N \'e lson Lima, Francisco Caldas, and Cl \'a udia Soares. Advancing solutions for the three-body problem through physics-informed neural networks. arXiv preprint arXiv:2503.04585, 2025. URL https://arxiv.org/abs/2503.04585

    2025 arXiv

  13. [21]

    Les M\' e thodes nouvelles de la m\' e canique c\' e leste , volume 1

    Henri Poincar\' e . Les M\' e thodes nouvelles de la m\' e canique c\' e leste , volume 1. Gauthier-Villars, Paris, 1892

  14. [22]

    On the spectral bias of neural networks

    Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In Proceedings of the 36th International Conference on Machine Learning, volume 97 of PMLR, pages 5301--5310, 2019

    2019

  15. [23]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378: 0 686--707, 2019

    2019

  16. [24]

    Veronica Saz Ulibarrena, Philipp Horn, Simon Portegies Zwart, Elena Sellentin, Barry Koren, and Maxwell X. Cai. A hybrid approach for solving the gravitational N -body problem with artificial neural networks. Journal of Computational Physics, 496: 0 112596, 2024

    2024

  17. [25]

    Physics-informed neural networks without loss balancing: A direct term scaling approach for nonlinear 1D problems

    Theodosios Theodosiou and Christoforos Rekatsinas. Physics-informed neural networks without loss balancing: A direct term scaling approach for nonlinear 1D problems. F1000Research, 14: 0 1252, 2026. doi:10.12688/f1000research.169129.2. Version 2

    2026 doi

  18. [26]

    Dmitra s inovi\' c

    Milovan S uvakov and V. Dmitra s inovi\' c . Three classes of Newtonian three-body planar periodic orbits. Physical Review Letters, 110: 0 114301, 2013

    2013

  19. [27]

    Understanding and mitigating gradient flow pathologies in physics-informed neural networks

    Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing, 43: 0 A3055--A3081, 2021 a

    2021

  20. [28]

    On the eigenvector bias of Fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks

    Sifan Wang, Hanwen Wang, and Paris Perdikaris. On the eigenvector bias of Fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 384: 0 113938, 2021 b

    2021

  21. [29]

    A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks

    Chenxi Wu, Min Zhu, Qinyang Tan, Yadhu Kartha, and Lu 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: 0 115671, 2023

    2023

  22. [30]

    Learning and discovering multiple solutions using physics-informed neural networks with random initialization and deep ensemble

    Zongren Zou, Zhicheng Wang, and George Em Karniadakis. Learning and discovering multiple solutions using physics-informed neural networks with random initialization and deep ensemble. Proceedings of the Royal Society A, 481: 0 20250205, 2025. doi:10.1098/rspa.2025.0205

    2025

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.