Neural Network Perturbation Theory (NNPT): Learning Residual Corrections from Exact Solutions

Boris Fain; Mutian Shen; Zhenhao Chen; Zohar Nussinov

REVIEW 3 major objections 4 minor 1 cited by

Neural Network Perturbation Theory (NNPT): Learning Residual Corrections from Exact Solutions

T0 review · 3 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read The paper claims that neural networks trained on only the residual perturbation of a three-body system hit a sharp capacity barrier at chaos onset, requiring about seven times more parameters once the Jovian mass factor exceeds roughly 15.6

desk verdict The paper's headline claim—a sevenfold capacity jump at f≈15.6—is contradicted by its own abstract, and the equalized-accuracy protocol that produces it breaks down exactly in the chaotic regime. read the letter →

arxiv 2512.01558 v2 pith:4UFXKNIV submitted 2025-12-01 physics.comp-ph

Zhenhao Chen , Mutian Shen , Boris Fain , Zohar Nussinov This is my paper

classification physics.comp-ph PACS 05.45.-a07.05.Mh

keywords neuralnetworkperturbationtheoryresiduallearningthree-bodyproblemcapacitytransitionchaosonsetKAMtoriautoencoderequalized-accuracyprotocol

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

The paper introduces Neural Network Perturbation Theory (NNPT): instead of training a network to reproduce an entire trajectory, subtract the exactly solvable Keplerian orbit and train the network to predict only the residual. Applying this to a Sun–Earth–Jupiter system with Jovian mass varied from 0.05 to 30 times its physical value, the authors report that the minimal network capacity needed to hold validation error fixed jumps sharply from about 1,200 to about 8,600 parameters at a mass factor of about 15.6. At the same point, the latent dimension required to autoencode the residual jumps from about 8 to about 32. The authors interpret this as a fundamental capacity barrier at the onset of chaos, aligned with Chirikov's resonance-overlap criterion and preceding visible KAM torus destruction.

What carries the argument

The load-bearing object is the residual signal y(t)=r_full(t)−r_exact(t), which isolates the three-body correction from the analytically known Keplerian baseline. The load-bearing protocol is equalized accuracy: compute a reference threshold E* from a large 10×128 network, then shrink architecture until that threshold is met, recording the smallest architecture that succeeds. A piecewise-linear change-point fit locates f_c; a convolutional autoencoder measures the minimal latent dimension d_min of y(t) as an independent geometric diagnostic.

What would settle it

Check whether any architecture from the grid achieves validation MSE below E_thr=1.06e-2 AU^2 at f=20 and f=30; if none does, the 8,578-parameter figure is not a capacity that meets the equalized-accuracy target. Also compare the abstract's claim that f≥17 requires only 2×32 with Table II, which reports 3×64 at f≥20; both cannot describe the same protocol.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a fixed-architecture neural network's minimal capacity to achieve a fixed error is not a smooth function of a physical control parameter. In the integrable regime (f≤15) every seed hits the target with a 2×32 MLP (1,186 parameters); in the chaotic regime (f≥20) all seeds need a 3×64 MLP (8,578 parameters). The breakpoint f_c=15.6±1.0 is selected by piecewise-linear regression with Bayesian information criterion; the 68% bootstrap confidence interval spans only two mass factors. Independent autoencoder analysis gives a fourfold jump in minimal latent dimension at the same f_c. Energy drift below 2×10^-7 across all f is used to argue the

Load-bearing premise

The protocol assumes that the single global error threshold E*≈9.64e-3 AU^2 is an attainable target at every mass factor, and that the coarse architecture grid can find the true minimal capacity; if no architecture actually reaches the threshold in the chaotic regime, the reported minimal capacities are best-effort misses rather than successes.

Editorial extensions

If this is right

In regimes where dynamics are chaotic, generic black-box surrogates will need sharply more parameters or training time; correction learning that subtracts exact baselines avoids the worst of this barrier.
Capacity metrics under equalized accuracy can serve as an early-warning diagnostic for chaos onset, detecting functional complexity before Poincaré sections show torus destruction.
NNPT's recipe — learn the residual after an exact solution — transfers to any system with a known analytic approximation, potentially cutting parameter counts by orders of magnitude.
Sequential two-stage corrections did not refine the result, meaning naive hierarchy does not circumvent the capacity barrier; better architectural priors are needed.
The autoencoder dimensionality transition gives an unsupervised way to locate the same critical point, independent of the supervised capacity sweep.

Reading between the lines

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

The internal contradiction between the abstract's non-monotonic capacity profile (peak at f=5, fall to 2×32 at f≥17) and the full text's monotonic jump (2×32 for f≤15, 3×64 for f≥20) means the reported 'minimal capacity' values are protocol-sensitive; a reader should verify which numbers correspond to the published tables.
If the capacity barrier is truly about the dynamics, the same discontinuity should appear for other fixed-capacity approximators such as polynomials or Fourier series; running the equalized-accuracy sweep with a non-neural regressor would test the claim without relying on ML training details.
One could sharpen the Chirikov connection by computing resonance-overlap widths for the specific Sun–Earth–Jupiter parameters and comparing the predicted f range to the measured breakpoint, rather than relying on the approximate f~10–16 quoted.
For practical surrogate design, this suggests a cheap diagnostic: sweep a physical parameter, hold target error fixed, and watch for a capacity step; systems with such steps may require hybrid physics-plus-ML models.

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.

Desk Editor's Note

read the letter

Quick take: don't trust the central claim yet. The abstract reports a non-monotonic capacity profile peaking at f=5 and needing only 2×32 (1,186 params) at f≥17; the full text reports a monotonic sevenfold jump to 3×64 (8,578 params) at f_c=15.6. The abstract even quotes the sequential-correction ratio as ~0.997 while the body says the second-stage residual is 13–19% larger. That is not a wording slip; it's the paper's main result changing between front matter and body.

Credit where due: the correction-learning comparison is solid and useful. With identical 2×32 networks, residual learning beats direct trajectory learning by 28–54× in validation MSE. That is a real, reproducible-style result and consistent with existing delta-learning literature (refs 6, 21), so the framing is not new but the controlled demonstration is clean. The symplectic integrator energy check is good. The autoencoder dimensionality jump is a nice idea, though it inherits the same f-region problems.

Soft spots are serious. The equalized-accuracy protocol defines E* as the median validation error of the 10×128 reference network, then asks smaller networks to meet it. Figure 3 shows that same reference network crosses the threshold at f≈20; at f=30 it is well above. Table II nevertheless reports 3×64 as the 'picked' architecture for f=20 and 30, without showing the validation MSE of the picked networks. Since the protocol says 'if no architecture meets the threshold, we report the best-performing configuration,' those points are likely best-effort failures, not capacities that achieve fixed accuracy. The sevenfold jump from 2×32 to 3×64 may just be the grid's next-smallest architecture after the threshold stops being reachable. The coarse grid (2×32, 2×48, 3×32, 3×64...) and arbitrary 10% tolerance can manufacture a step.

Also minor: Table I counts 10×128 as 149,122 parameters while the text says 168,322; the formula gives 149,122.

Bottom line: The residual-learning contribution is worth keeping, but the chaos-capacity barrier claim is not supported by the evidence as written. This deserves a serious referee only if the authors can resolve the abstract/body contradiction and show actual MSE values for the picked architectures in the chaotic regime. I would engage with a revised version; I would not cite the capacity transition in its current form.

Referee Report

3 major / 4 minor

Summary. The paper introduces Neural Network Perturbation Theory (NNPT), in which MLPs are trained to predict only the residual y(t)=r_full-r_Kepler after subtracting the analytic Keplerian solution, using the planar circular restricted three-body problem as a testbed. The Jovian mass factor f is varied from 0.05 to 30. A reference 10×128 network defines a global target E*=9.64×10^-3 AU^2 (median validation MSE across all f and seeds), and an equalized-accuracy protocol shrinks the architecture until E_thr=1.06×10^-2 AU^2 is met. The full text claims a sharp sevenfold minimal-capacity jump (1,186→8,578 parameters, 2×32→3×64) and a threefold training-time increase at f_c=15.6±1.0, aligned with Chirikov's resonance-overlap criterion, while autoencoder analysis claims a fourfold intrinsic-dimensionality jump (d_min: 8→32) at the same f_c. The abstract, however, reports a non-monotonic capacity profile peaking at f=5 and decreasing to 2×32 at f≥17. The paper concludes that NNPT provides a general parameter-efficient surrogate strategy and that physical chaos imposes fundamental capacity barriers on fixed-architecture networks.

Significance. The NNPT decomposition idea is attractive, and the direct comparison showing 28–54× lower MSE for correction learning versus full-trajectory learning is a useful sanity check. The use of a symplectic integrator with energy-drift monitoring is also methodologically sound. However, the central quantitative claims are not supported by the evidence as presented. The equalized-accuracy target is derived from the median error of the same reference network that is later shown to exceed the acceptance threshold in the chaotic regime; the capacity values reported for f=20 and f=30 may be best-effort failures rather than capacities that achieve the stated fixed accuracy. This, together with the direct contradiction between the abstract and the full text, means the claimed chaos-onset capacity barrier and its Chirikov alignment are not established. If the protocol were repaired and the inconsistencies resolved, the underlying research question would remain interesting, but the current manuscript does not deliver a reliable answer.

major comments (3)

[Abstract vs. §III.B and Table II] The headline result is internally contradictory. The abstract states that capacity 'peaks at f=5 in the late integrable regime (3×32, 2242 parameters), remains elevated through the transition region, then decreases in the fully chaotic regime (f≥17, requiring only 2×32 with 1186 parameters)—a 47% reduction from peak.' The full text (Fig. 4, Table II) instead reports a monotonic sevenfold jump at f_c=15.6, with f≥20 requiring 3×64 (8,578 parameters). The abstract also quotes f_c=16.6±2.8 while the body uses f_c=15.6±1.0. These are mutually incompatible descriptions of the paper's central observation; as written, the paper does not have a well-defined headline result.
[§II.H, Fig. 3, Table II] The equalized-accuracy protocol fails exactly in the regime used to define the capacity barrier. In §II.H, E* is the median validation MSE of the 10×128 reference over all (f,seed) pairs, and E_thr=1.06×10^-2 AU^2. Figure 3 shows the reference network itself has MSE above E_thr for f≈20 and larger. The protocol states that if no architecture meets the threshold, the best-performing configuration is reported. Table II lists 3×64 for f=20 and f=30 but gives no validation MSE for these choices. Therefore P(f)=8,578 may be the smallest grid architecture whose performance is closest to an unmet threshold, not the minimal capacity that actually reaches E*. The sevenfold jump is thus not evidence of a capacity barrier at fixed accuracy.
[§II.J.3, Fig. 7] The autoencoder dimensionality claim has the same threshold problem. The reconstruction criterion is E_AE_thr=10^-4 AU^2, but Fig. 7(a) reports that at f=30 even d=64 gives E_rec≈5×10^-4 AU^2, above the threshold. Consequently, the values d_min≈32–40 shown in Fig. 7(b) for f≥20 are not minimal dimensions that meet the stated reconstruction criterion; they are best-effort values at the largest tested latent dimension. The claimed fourfold dimensionality jump (8→32) at f_c=15.6 is therefore not supported by the data presented.

minor comments (4)

[Throughout] Several numerical values are inconsistent and need reconciliation: f_c=16.6±2.8 (abstract) vs. 15.6±1.0 (body); energy drift bound 2×10^-4 (abstract) vs. 2×10^-7 (body); 10×128 parameter count 168,322 (§II.E and Fig. 3) vs. 149,122 (Table I and Eq. (8)). The abstract's ||y2||/||y1||≈0.997 also disagrees with §III.C, which reports the second-stage residual is 13–19% larger.
[§II.H, Table I] The architecture grid is not strictly increasing in size: 2×48 has 2,546 parameters while 3×32 has 2,242. Section II.H says architectures are tested 'in order of increasing size,' but the order between these two is ambiguous. This could affect which architecture is picked and should be clarified.
[Fig. 3 caption] The caption marks 'Transition fc = 20.0', while §III.B reports f_c=15.6. Please clarify which quantity is being indicated.
[Data and Code Availability] The reproducibility statement says data and code 'can be obtained from the corresponding author upon request.' For a computational study whose central claims rely on specific training runs and threshold choices, a public repository would materially strengthen verifiability.

Circularity Check

2 steps flagged · score 6.0 of 10

Equalized-accuracy protocol's target is the reference network's own median error, and in the chaotic regime the protocol reports best-performing failures as 'minimal capacities'; the sevenfold jump is therefore partly forced by construction.

fitted input called prediction [Section II.H (Equalized-Accuracy Protocol), applied in Fig. 4 and Table II; cf. Fig. 3]
"If no architecture meets the threshold, we report the best-performing configuration."

Under the protocol, P(f) is called 'the minimal parameter count achieving E_thr' (Section II.H). But Fig. 3 states the reference 10×128 network's MSE 'crosses the acceptance threshold E_thr = 1.06×10−2 AU2 at f≈20'; therefore for f=20 and f=30 no architecture on the grid can meet the threshold, and the '3×64' entries in Table II are best-performing failures, not capacities that achieve the target. The reported sevenfold jump (1,186→8,578) therefore compares an architecture that actually reaches E_thr with one that merely fails least badly; the 'minimal capacity' in the chaotic regime is an artifact of the grid + fallback rule, not a measured capacity barrier at fixed accuracy.
self definitional [Section II.H, Eq. (10)]
"E⋆ = median{ValMSE(f, s) : f∈ F, s∈ S} = 9.64×10−3 AU2"

The 'fixed accuracy' target is defined as the median validation error of the same reference 10×128 network whose capacity is being probed. Thus 'minimal capacity required to reach E⋆' means 'capacity required to match the reference network's own average error,' not an externally fixed physical accuracy. The threshold is therefore self-referential: the model whose capacity is being measured sets the yardstick, so the resulting P(f) and its breakpoint are defined relative to that network's performance rather than discovered from an independent accuracy requirement.

full rationale

The paper's central claim—a sharp sevenfold capacity jump at f_c ≈ 15.6 aligned with Chirikov's criterion—rests on the equalized-accuracy protocol of Section II.H. That protocol defines the target E⋆ as the median validation MSE of the reference 10×128 network itself (Eq. 10), making the 'minimal capacity to reach E⋆' a measure of matching that network's own average error. More seriously, the protocol's fallback rule ('If no architecture meets the threshold, we report the best-performing configuration') is activated in the chaotic regime: Figure 3 and Section III.A state that even the reference 10×128 crosses E_thr at f≈20. Consequently, Table II's 3×64 entries for f=20 and f=30 are best-performing failures, not architectures that achieve the target accuracy. The reported jump from 2×32 (1,186 params) to 3×64 (8,578 params) is therefore not a demonstrated capacity barrier at fixed accuracy; it is the difference between an architecture that succeeds and the largest grid architecture that fails least badly. The autoencoder analysis and the correction-learning advantage over direct learning are independent empirical contributions, and the energy-conservation check is a legitimate numerical validation; but the headline 'capacity transition' is substantially forced by the self-referential threshold and the grid/fallback protocol. Internal inconsistency between the abstract ('capacity ... decreases in the fully chaotic regime ... 2×32 with 1186 parameters') and the full text/Table II ('3×64 ... 8,578 parameters' for f≥20) further undermines the robustness of the reported transition. No load-bearing self-citation chain was found; references to the authors' earlier work ([6], [22]) are supporting examples rather than derivation inputs. Overall, the central predictive claim partially reduces to protocol construction, warranting a score of 6.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The central claims rest on a small set of empirical and protocol choices: the threshold E* is self-referential, the architecture grid is coarse, the autoencoder threshold is arbitrary, and the Chirikov alignment is post-hoc. No new physical entities are postulated.

free parameters (4)

Acceptance threshold factor ε = 0.10
Chosen arbitrarily to define E_thr = E*(1+ε); directly determines which architectures are 'accepted' and hence the capacity step.
Reference median target E* = 9.64e-3 AU²
Computed as median validation MSE of the 10×128 reference network over all f; the protocol's target is defined by this fit, so the resulting 'capacity' is tied to a self-referential quantity.
Autoencoder reconstruction threshold E_AE_thr = 1e-4 AU²
Chosen without justification; the deduced minimal latent dimension d_min(f) depends on this absolute threshold, which may not be scale-invariant across f.
Architecture grid = 2×32, 3×32, 2×48, 3×64, 4×64, 6×96, 10×128
Discrete hand-selected grid; the coarse spacing between 2×32 and 3×64 means the reported 'sevenfold jump' could be an artifact of missing intermediate sizes.

assumptions (7)

domain assumption Sun-Earth-Jupiter planar circular restricted three-body problem adequately represents the physics of interest
Used throughout; ignores inclination and eccentricity of Jupiter, among other effects.
domain assumption Velocity Verlet with Δt=8×10^-3 yr and T=20 yr yields converged trajectories; energy drift <2×10^-7 is sufficient
Section II.B; used to generate all training data.
domain assumption The Keplerian two-body solution with fixed initial orbital elements is the correct exact baseline to subtract
Section II.C; the claimed decomposition r_full = r_Kepler + y.
ad hoc to paper A single global median validation MSE of one reference network is a valid equalized-accuracy target
Section II.H; load-bearing for the capacity metric.
domain assumption Chirikov's resonance-overlap criterion predicts chaos onset for f~10-16 in this system
Section III.E; the match to f_c=15.6 is asserted, not derived, and the predicted range is broad.
ad hoc to paper The discrete architecture grid adequately samples the capacity continuum
Section II.F; the jump between 2×32 and 3×64 is an artifact of the grid spacing.
ad hoc to paper Autoencoder reconstruction error below 1e-4 AU² indicates true latent dimensionality
Section II.J.3; threshold is arbitrary and may conflate signal amplitude with intrinsic dimension.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Neural Network Perturbation Theory (NNPT): Learning Residual Corrections from Exact Solutions." pith.science (2026). https://pith.science/paper/4UFXKNIV

@misc{pith2026251201558,
  author       = {Pith},
  title        = {Pith review of: Neural Network Perturbation Theory (NNPT): Learning Residual Corrections from Exact Solutions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4UFXKNIV}},
  note         = {Machine review of arXiv:2512.01558}
}

read the original abstract

Many complex physical systems naturally decompose into an exactly solvable component augmented by a perturbative correction. Rather than directly employing neural networks to analyze complex physical systems, we introduce Neural Network Perturbation Theory (NNPT)--a correction learning approach that predicts residual perturbations after analytically subtracting known exact solutions. Using the gravitational three-body problem as testbed, we vary Jovian mass from f=0.05 to 30 times its physical value while holding network architecture fixed. An equalized-accuracy protocol with 1% tolerance reveals an unexpected non-monotonic capacity profile: capacity peaks at f=5 in the late integrable regime (3x32, 2242 parameters), remains elevated through the transition region (f~15-17), then decreases in the fully chaotic regime (f>=17, requiring only 2x32 with 1186 parameters)--a 47% reduction from peak. With symplectic integrator energy conservation below 2x10^{-4}, this counterintuitive phenomenon reflects genuine physical structure rather than numerical artifacts. Sequential correction experiments show negligible refinement (||y2||/||y1||~0.997), confirming single-stage networks capture dominant perturbative features without hierarchical decomposition. The capacity transition at f_c=16.6+-2.8 aligns with Chirikov's resonance-overlap criterion. Intermediate-complexity regimes impose maximal capacity requirements, while fully chaotic dynamics undergo ergodic smoothing--trajectory-specific fluctuations become irreducible noise, leaving only statistically smooth corrections requiring fewer parameters.

Figures

Figures reproduced from arXiv: 2512.01558 by the authors.

**Figure 3.** FIG. 3. Baseline network performance using fixed 10 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 4.** FIG. 4. Sharp capacity and training transitions at chaos onset. (a) Minimal network capacity vs. Jovian mass factor under [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Figures from the paper (5 more)

**Figure 5.** Figure 5: FIG. 5. Validation of correction learning framework using fixed 2 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: FIG. 6. Sequential correction results for 2 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: FIG. 7. Autoencoder reconstruction quality across mass factors and latent dimensions. (a) Reconstruction error [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: FIG. 8. Two-dimensional latent space representations ( [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

**Figure 9.** Figure 9: FIG. 9. Neural network capacity transitions anticipate KAM torus breakdown. (a) Poincar´e sections (radial distance vs. radial [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Dynamical and Optimization Trade-offs of Levi--Civita Coordinates for Learned Close-Encounter Dynamics
physics.comp-ph 2026-07 conditional novelty 6.0 of 10

Levi-Civita coordinates improve rollout survival of learned close-encounter models but worsen raw-basis regression conditioning, while accurate neural residual learning remains unresolved.

Reference graph

Works this paper leans on

34 extracted references · 1 linked inside Pith · cited by 1 Pith paper

[1]

This provides a controlled comparison: identical architecture, identical training protocol, but learning the full signal rather than the correction

Direct Learning Baseline We train networks to predict the complete three-body trajectoryr (3-body) E (t) directly from normalized time in- put, without subtracting the Keplerian baseline. This provides a controlled comparison: identical architecture, identical training protocol, but learning the full signal rather than the correction
[2]

The total parameter count for sequential 2×2×32 corrections (2,372 parameters) is compared against single-stage 3×64 networks (8,578 parameters)

Sequential Perturbative Corrections Motivated by higher-order perturbation theory, we ex- plore a two-stage sequential approach: Stage 1:y 1(t) =r (3-body) E (t)−r (Kepler) E (t), NN1 trained to predicty 1(t); (12) Stage 2:y 2(t) =y 1(t)−NN 1(t), NN2 trained to predicty 2(t).(13) If the second-stage residual∥y 2∥is substantially smaller than∥y 1∥, this wo...
[3]

Autoencoder Latent Space Analysis To investigate whether correction signalsy(t) contain compressible low-dimensional structure—and whether such structure changes at the capacity transition—we employ convolutional autoencoders. The encoder maps the 2D correction trajectory (y x(t), yy(t)) sampled at N= 2500 time points to ad-dimensional latent repre- senta...

2025
[4]

P. G. Breen, C. N. Foley, T. Boekholt, and S. P. Zwart, Monthly Notices of the Royal Astronomical Society494, 2465 (2020)

2020
[5]

Greydanus, M

S. Greydanus, M. Dzamba, and J. Yosinski, Advances in Neural Information Processing Systems32, 15379 (2019)

2019
[6]

Cranmer, S

M. Cranmer, S. Greydanus, S. Hoyer, P. Battaglia, D. Spergel, and S. Ho, inICLR 2020 Workshop on Inte- gration of Deep Neural Models and Differential Equations (2020)

2020
[7]

C. M. Bender and S. A. Orszag,Advanced Mathematical Methods for Scientists and Engineers: Asymptotic Meth- ods and Perturbation Theory(Springer-Verlag, 1999)

1999
[8]

L. D. Landau and E. M. Lifshitz,Quantum Mechanics, Non-relativistic Theory, 2nd ed., Course of Theoretical Physics, Vol. 3 (Pergamon Press, 1965)

1965

Show all 34 references

[9]

Illarionov, S

A. Illarionov, S. Sakipov, L. Pereyaslavets, I. V. Kurnikov, G. Kamath, O. Butin, E. Voronina, I. Ivah- nenko, I. Leontyev, G. Nawrocki, M. Darkhovskiy, M. Ol- evanov, Y. K. Cherniavskyi, C. Lock, S. Greenslade, S. K. Sankaranarayanan, M. G. Kurnikova, J. Potoff, R. D. Ko- rnb...

2023
[10]

Poincar´ e, Acta Mathematica13, 1 (1890)

H. Poincar´ e, Acta Mathematica13, 1 (1890)
[11]

Laskar, Nature338, 237 (1989)

J. Laskar, Nature338, 237 (1989)

1989
[12]

C. D. Murray and S. F. Dermott,Solar System Dynamics (Cambridge University Press, Cambridge, 1999)

1999
[13]

J. P. Sethna,Entropy, Order Parameters, and Complex- ity(Oxford University Press, Oxford, 2006)

2006
[14]

Sanchez-Gonzalez, J

A. Sanchez-Gonzalez, J. Godwin, T. Pfaff, R. Ying, J. Leskovec, and P. Battaglia, inProceedings of the 37th International Conference on Machine Learning, ICML 2020 (2020)

2020
[15]

P. Jin, Z. Zhang, A. Zhu, Y. Tang, and G. E. Karniadakis, Neural Networks132, 166 (2020)

2020
[16]

Batzner, A

S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nature Communications13, 2453 (2022)

2022
[17]

Kondor, Proceedings of the National Academy of Sci- ences122(2025), in press

R. Kondor, Proceedings of the National Academy of Sci- ences122(2025), in press

2025
[18]

Raissi, P

M. Raissi, P. Perdikaris, and G. E. Karniadakis, Journal of Computational Physics378, 686 (2019)

2019
[19]

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, Nature Reviews Physics3, 422 (2021)

2021
[20]

B. V. Chirikov, Physics Reports52, 263 (1979)

1979
[21]

Hairer, C

E. Hairer, C. Lubich, and G. Wanner,Geometric Numer- ical Integration: Structure-Preserving Algorithms for Or- dinary Differential Equations, 2nd ed. (Springer, Berlin, 2006)

2006
[22]

Elfwing, E

S. Elfwing, E. Uchibe, and K. Doya, Neural Networks 107, 3 (2018)

2018
[23]

D. P. Kingma and J. Ba, inProceedings of the 3rd Inter- national Conference on Learning Representations, ICLR 2015 (2015) arXiv:1412.6980

2015 arXiv
[24]

R. T. McGibbon, A. G. Taube, A. G. Donchev, K. Siva, F. Hern´ andez, C. Hargus, K. Law, J. L. Klepeis, and D. E. Shaw, The Journal of Chemical Physics147, 161725 (2017)

2017
[25]

I. V. Kurnikov, L. Pereyaslavets, G. Kamath, S. N. Sakipov, E. Voronina, O. Butin, A. Illarionov, I. Leon- tyev, G. Nawrocki, M. Darkhovskiy, M. Olevanov, I. Ivahnenko, Y. Chen, C. B. Lock, M. Levitt, R. D. Kornberg, and B. Fain, Journal of Chemical Theory and Computation20, 1...

2024
[26]

P. W. Battaglia, R. Pascanu, M. Lai, D. J. Rezende, and K. Kavukcuoglu, inAdvances in Neural Information Processing Systems, Vol. 29 (2016) pp. 4502–4510

2016
[27]

V. S. Ulibarrena, P. Horn, S. P. Zwart, E. Sellentin, B. Koren, and M. X. Cai, Journal of Computational Physics496, 112565 (2024)

2024
[28]

Reichstein, G

M. Reichstein, G. Camps-Valls, B. Stevens, M. Jung, J. Denzler, N. Carvalhais, and Prabhat, Nature566, 195 (2019)

2019
[29]

K. T. Sch¨ utt, F. Arbabzadah, S. Chmiela, K. R. M¨ uller, and A. Tkatchenko, Nature Communications8, 13890 (2017)

2017
[30]

Lecar, F

M. Lecar, F. A. Franklin, M. J. Holman, and N. W. Mur- ray, Annual Review of Astronomy and Astrophysics39, 581 (2001)

2001
[31]

Kevorkian and J

J. Kevorkian and J. D. Cole,Perturbation Methods in Applied Mathematics, Applied Mathematical Sciences (Springer-Verlag, 1981)

1981
[32]

M. H. Holmes,Introduction to Perturbation Methods, 2nd ed. (Springer, 2012)

2012
[33]

F. M. Fernandez,Introduction to Perturbation Theory in Quantum Mechanics(CRC Press, 2020)

2020
[34]

Talagrand,What Is a Quantum Field Theory?(Cam- bridge University Press, 2022)

M. Talagrand,What Is a Quantum Field Theory?(Cam- bridge University Press, 2022)

2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

This provides a controlled comparison: identical architecture, identical training protocol, but learning the full signal rather than the correction

Direct Learning Baseline We train networks to predict the complete three-body trajectoryr (3-body) E (t) directly from normalized time in- put, without subtracting the Keplerian baseline. This provides a controlled comparison: identical architecture, identical training protocol, but learning the full signal rather than the correction

[2] [2]

The total parameter count for sequential 2×2×32 corrections (2,372 parameters) is compared against single-stage 3×64 networks (8,578 parameters)

Sequential Perturbative Corrections Motivated by higher-order perturbation theory, we ex- plore a two-stage sequential approach: Stage 1:y 1(t) =r (3-body) E (t)−r (Kepler) E (t), NN1 trained to predicty 1(t); (12) Stage 2:y 2(t) =y 1(t)−NN 1(t), NN2 trained to predicty 2(t).(13) If the second-stage residual∥y 2∥is substantially smaller than∥y 1∥, this wo...

[3] [3]

Autoencoder Latent Space Analysis To investigate whether correction signalsy(t) contain compressible low-dimensional structure—and whether such structure changes at the capacity transition—we employ convolutional autoencoders. The encoder maps the 2D correction trajectory (y x(t), yy(t)) sampled at N= 2500 time points to ad-dimensional latent repre- senta...

2025

[4] [4]

P. G. Breen, C. N. Foley, T. Boekholt, and S. P. Zwart, Monthly Notices of the Royal Astronomical Society494, 2465 (2020)

2020

[5] [5]

Greydanus, M

S. Greydanus, M. Dzamba, and J. Yosinski, Advances in Neural Information Processing Systems32, 15379 (2019)

2019

[6] [6]

Cranmer, S

M. Cranmer, S. Greydanus, S. Hoyer, P. Battaglia, D. Spergel, and S. Ho, inICLR 2020 Workshop on Inte- gration of Deep Neural Models and Differential Equations (2020)

2020

[7] [7]

C. M. Bender and S. A. Orszag,Advanced Mathematical Methods for Scientists and Engineers: Asymptotic Meth- ods and Perturbation Theory(Springer-Verlag, 1999)

1999

[8] [8]

L. D. Landau and E. M. Lifshitz,Quantum Mechanics, Non-relativistic Theory, 2nd ed., Course of Theoretical Physics, Vol. 3 (Pergamon Press, 1965)

1965