Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Neural Conjugate Flows: Physics-informed architectures with flow structure

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper introduces Neural Conjugate Flows, which conjugate an invertible neural network with an affine flow, and proves they are universal approximators for flows of autonomous ODEs.

desk verdict A promising architecture and honest experiments, but the central universality theorem is unfinished and the abstract oversells the overall NCF class. read the letter →

arxiv 2411.08326 v1 pith:74A6XXSY submitted 2024-11-13 cs.LG cs.NAmath.NA

Arthur Bizzi , Lucas Nissenbaum , João M. Pereira This is my paper · ORCID
classification cs.LGcs.NAmath.NA MSC 37C1068T07
keywords neuralconjugateflowstopologicalconjugationflowoperatorsuniversalapproximationphysics-informednetworksordinarydifferentialequationsaffineinvertible
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 Conjugate Flows (NCFs), architectures defined as $\Phi^t = H^{-1} \circ \Psi^t \circ H$, where $H$ is an invertible neural network and $\Psi^t$ is an affine flow of the form $\dot{x} = Ax + b$. Because any conjugate of a flow is again a flow, an NCF is a flow operator by construction: it automatically obeys the group laws of identity, associativity, and time-reversibility, so initial conditions, causality, and trajectory uniqueness hold exactly rather than as soft-enforced losses. The paper's central claim is that affine NCFs are universal approximators for flows of autonomous ODEs: any sufficiently well-behaved vector field can be augmented in a way that its flow is topologically conjugate to a pure translation, and coupling-layer networks can approximate the conjugating homeomorphism. If this is correct, a single architecture class can in principle reproduce any smooth ODE flow while training up to five times faster than Neural ODEs, because the inner affine flow is evaluated in closed form and in parallel. The numerical experiments on neuron models (FitzHugh-Nagumo and Hodgkin-Huxley) support the more specific claim that topology-informed NCFs extrapolate beyond the training interval better than standard physics-informed networks while remaining competitive with or faster than Neural ODEs.

What carries the argument

The load-bearing construction is the conjugation identity $\Phi^t = H^{-1} \circ \Psi^t \circ H$, which turns any invertible network $H$ into a device that transports a simple, exactly solvable flow $\Psi^t$ to a potentially complicated target flow. The universality proof uses the explicit conjugacy $H(x,a) = (\Phi^{-a}x, a)$ between the augmented flow of the target ODE and a pure translation; this shows that the approximation problem reduces to approximating that $H$ with coupling layers. The augmented 'twin' input scheme is the practical mechanism that gives each coupling-layer MLP access to the full state, restoring representational power lost when the input is split by the coupling architecture.

What would settle it

For a target ODE with a closed-form flow, such as the logistic equation $\dot{x} = x(1-x)$, compute the explicit conjugacy $H(x,a) = (\Phi^{-a}x, a)$ from the appendix and measure the approximation error when $H$ is represented by coupling layers of growing width; if the flow error does not shrink, the practical universality claim fails.

Watch

Extended reading notes

Core claim

The paper's central discovery is that topological conjugation, a classical tool from dynamical systems, can be turned directly into a neural architecture. An NCF is defined as the composition of an invertible network $H$, a chosen flow $\Psi^t$, and the inverse $H^{-1}$; this composition is exactly the solution operator of some autonomous ODE. The paper proves (Theorem 1) that any Lipschitz vector field $F$ can be augmented by one extra dimension and conjugated, via the explicit homeomorphism $H(x,a) = (\Phi^{-a}x, a)$, to the constant translation flow on the augmented space. Theorem 2 then states that affine NCFs are universal approximators for flows of autonomous ODEs, because coupling layers are universal approximators for diffeomorphisms and the explicit conjugacy is a diffeomorphism. The paper also shows that topological properties of the target flow can be enforced interpretably: restricting the matrix $A$ of the affine flow to a particular Lie algebra, such as the skew-symmetric matrices for rotational $SO(n)$ behavior, constrains the learned flow to inherit that topology. This is what the paper means by calling NCFs 'topology-informed' alternatives to Neural ODEs.

Load-bearing premise

The load-bearing premise is that the conjugating map $H$ used in the universality proof, which is defined from the exact flow of the target ODE, can actually be approximated well enough by coupling layers that the error in the conjugated flow stays small.

Editorial extensions

If this is right

  • NCFs inherit exact flow group properties (identity, associativity, invertibility) by construction, eliminating the need for soft initial-condition and causality penalties in physics-informed losses.
  • Affine NCFs are universal approximators for flows of autonomous ODEs, meaning any smooth autonomous system can in principle be represented with sufficient augmentation and coupling-layer capacity.
  • Topological constraints can be hard-wired into the architecture: choosing $\Psi$ from a specific Lie group imposes that group's topology on the learned flow, making oscillatory or conservative behavior interpretable and structurally guaranteed.
  • Affine flows are evaluable in closed form and in parallel, so NCFs train up to five times faster than Neural ODEs, which rely on sequential numerical integration.
  • In the neuron-model experiments, topology-informed NCFs extrapolate beyond the training time interval better than MLP-PINNs and match or improve on Neural ODEs, suggesting the architecture learns latent dynamics rather than mere interpolation.

Reading between the lines

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

  • A natural extension of the paper's argument is to continuous normalizing flows: because an NCF is a diffeomorphism at every time $t$, it can serve as a flow-based generative model whose inner affine flow is fully parallelizable, potentially making likelihood estimation much faster than Neural-ODE-based generators.
  • The paper's own failure on strongly nonlinear phenomena (a trajectory that first spirals out and then converges to a limit cycle) suggests the practical scope of affine-NCF universality is narrower than the theorem; the architecture will likely perform best when the target flow is a mild deformation of an integrable system.
  • The twin-augmentation scheme doubles the input dimension to give coupling layers full state access; a testable alternative would be to use an invertible residual architecture that sees the entire state without doubling, which could preserve flow structure at lower memory cost.
  • The fivefold speed advantage over Neural ODEs comes from closed-form parallel evaluation of the affine flow; using richer flows that still have closed-form solutions, such as linear flows with time-dependent coefficients, might extend accuracy while retaining most of the speed benefit.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces Neural Conjugate Flows (NCFs), architectures of the form Φ_t = H^{-1} ∘ Ψ_t ∘ H, where H is an invertible neural network (implemented with coupling layers) and Ψ is an affine flow. The authors show by construction that such maps are flows, claim that affine NCFs are universal approximators for flows of autonomous ODEs (Theorems 1 and 2), and argue that topological properties can be imposed through the choice of Ψ (e.g., a skew-symmetric A for oscillatory behavior). The empirical section compares MLP-PINNs, NCF, NCF-T, and Neural ODEs on FitzHugh-Nagumo and Hodgkin-Huxley models. The results support the value of NCF-T for extrapolation, but the vanilla NCF performs poorly in the first experiment, and the central universal-approximation theorem is not proved in the manuscript as written.

Significance. If the universality result were fully established, this would be a meaningful contribution to physics-informed machine learning: NCFs provide exact flow structure, a parallelizable inner flow, and a mechanism for enforcing topological constraints. The construction of Theorem 1, which shows that any Lipschitz ODE can be augmented and conjugated to a translation, is conceptually appealing. However, the paper's main advertised claim currently rests on an incomplete proof, and the experimental evidence only supports the topologically enforced variant NCF-T, not unqualified 'NCF' gains. The architecture itself is worth further study, and the paper contains reproducible-looking experimental details, but the theoretical and empirical claims need to be tightened before publication.

major comments (3)
  1. [Theorem 2 and Appendix A] Theorem 2 does not follow from the material in Appendix A as written. Theorem 1 constructs an exact conjugacy H(x,a)=(Φ^{-a}x,a) using the target flow Φ itself (Eq. 28), and the appendix verifies H^{-1} Ψ_t H = Φ̂_t by algebraic manipulation. The proof then stops. What is missing is the approximation-transfer step: if a coupling-layer ensemble H_n approximates H, one must prove that H_n^{-1} Ψ_t H_n approximates Φ̂_t uniformly for t∈[0,T], including uniform convergence of the inverses H_n^{-1} to H^{-1}. Without such a lemma, the 'universal approximator' claim of Theorem 2 is an assertion, not a proven statement. This is the load-bearing part of the central theoretical claim.
  2. [Abstract and Table 1] The abstract states that NCFs lead to computational gains in estimating and extrapolating latent ODE dynamics. Table 1, however, shows that the vanilla NCF, exactly as defined by Eq. (14), achieves Lacc = 6.7×10^{-2}, which is two orders of magnitude worse than the MLP baseline (4.9×10^{-4}). Only NCF-T, with the skew-symmetric matrix projection in Eq. (16), is competitive. The claims need to be qualified to the topologically enforced NCF-T variant, or the discrepancy must be explained.
  3. [Sections 3.2 and 3.3 vs Theorem 2] The augmentation used in the implementation (duplicating the input, Eq. 15) differs structurally from the augmentation used in the proof of Theorem 1 (adding a time coordinate a with dynamics da/dt=1, Eq. 25). The proof of Theorem 2 does not establish universality for the duplicated architecture. If the universality claim is meant to cover the implemented architecture, the authors need to state and prove it for that specific construction; otherwise, the scope of the theorem should be clarified.
minor comments (5)
  1. [Appendix A, Theorem 1 statement] The theorem states that the result holds for any positive integer m, but the proof constructs only the case m=1 with G(x,a)=1. A sentence explaining that additional dimensions can be padded with trivial dynamics would resolve this gap.
  2. [Footnote 3] The footnote contains a typo: 'they are are two distinct concepts' should read 'they are two distinct concepts'.
  3. [Section 3.2] The text calls coupling layers a 'Universal Approximator for homeomorphisms', while the cited result (Teshima et al. 2020) is specifically about diffeomorphisms. The paper should state the precise class of maps covered by the cited result, since the approximation-transfer lemma may require differentiability or C^1 convergence.
  4. [Section 3.1] The sentence 'The proof is straightforward (see Appendix A)' is not accurate because Appendix A does not contain a proof of Theorem 2; it only proves Theorem 1. This should be corrected.
  5. [Section 4.3] The statement 'NCFs both interpolate and extrapolate solutions better than Neural ODEs' should be checked against Table 2: NCF-T does extrapolate much better, but the difference in interpolation accuracy is less dramatic, and the vanilla NCF is not consistently better. The wording should be more nuanced.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the universality proof legitimately constructs the conjugacy from the target flow, and the cited coupling-layer universality is independent; the main weakness is an omitted approximation-transfer lemma, which is a rigor gap rather than a circular reduction.

full rationale

The paper's central theoretical claim is that affine Neural Conjugate Flows are universal approximators for flows of autonomous ODEs. Appendix A proves Theorem 1 by explicitly defining the conjugacy map H(x,a) = (Phi^{-a}x, a) using the flow Phi of the target system, then verifying algebraically that H^{-1} Psi^t H = Phi^t. This is a valid existential construction in the style of the flow-box theorem: it exhibits a witness H and checks the conjugacy identity; it does not assume the conclusion. The fact that H is built from the exact flow makes the theorem non-constructive in practice, but it does not make the proof circular. Theorem 2 then relies on the external, independently published universality of coupling layers (Teshima et al. 2020) to approximate H; there is no load-bearing self-citation chain. The paper does omit a proof that uniform approximation of H by coupling layers transfers to uniform approximation of the conjugated flow H^{-1} Psi^t H, i.e., a composition/topological-group continuity argument. That omission is a genuine correctness gap in the proof of Theorem 2, but it is not an instance of a prediction reducing to its inputs by construction. The empirical comparisons against MLP-PINNs and Neural ODEs are benchmarked against independent baselines, not against the paper's own fitted values, so the experimental claims are self-contained. Overall, no circular step meeting the required standard is present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The theoretical universality claim is cheap: it buys representational power by allowing H to encode the exact flow, so the ledger is dominated by a powerful H (a universal approximator of diffeomorphisms) and a global-existence assumption. The practical model adds learnable affine parameters A and b, and the topology enforcement via skew-symmetric projection is an additional modeling choice.

free parameters (2)
  • Affine flow matrix A = learned; skew-symmetric projection for NCF-T; A=0 in the universality proof
    The affine flow Ψ_t has a learnable matrix A that is trained by gradient descent. The proof of universality uses A=0, so A is not needed for the theorem but is essential in practice; for NCF-T it is projected to skew-symmetric to enforce oscillatory topology.
  • Affine flow offset b = learned
    The affine flow includes a learnable offset vector b. The initialization scheme sets it to satisfy F(x0)=A x0 + b, so its value influences convergence and is fitted during training.
assumptions (4)
  • domain assumption Coupling layers are universal approximators for diffeomorphisms
    Used to claim that H can approximate any homeomorphism; cited from Teshima et al. (2020) and not proved in the paper. The target H constructed in the universality proof is not shown to be in the approximable class with explicit error bounds.
  • domain assumption The target ODE admits a global flow (solutions exist for all times)
    The conjugation construction evaluates Φ^{-a} for unbounded a and relies on the flow group property for all real times. Theorem 2 states 'where global existence is assured', and the examples satisfy this, but the theorem as phrased requires it.
  • ad hoc to paper Approximating H by a coupling layer implies approximation of H^{-1} Ψ H
    Needed for Theorem 2 to be an approximation theorem. The paper does not state or prove a continuity result for the map H ↦ H^{-1} Ψ H, so it is an unstated assumption that approximation error in H transfers to the conjugated flow.
  • standard math Picard-Lindelöf existence and the group property of flows
    Basic ODE theory used to define Φ, its group property Φ^{t+τ}=Φ^t∘Φ^τ, and invertibility; these are standard and not proved in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Conjugate Flows: Physics-informed architectures with flow structure." pith.science (2026). https://pith.science/paper/74A6XXSY

@misc{pith2026241108326,
  author       = {Pith},
  title        = {Pith review of: Neural Conjugate Flows: Physics-informed architectures with flow structure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/74A6XXSY}},
  note         = {Machine review of arXiv:2411.08326}
}
read the original abstract

We introduce Neural Conjugate Flows (NCF), a class of neural network architectures equipped with exact flow structure. By leveraging topological conjugation, we prove that these networks are not only naturally isomorphic to a continuous group, but are also universal approximators for flows of ordinary differential equation (ODEs). Furthermore, topological properties of these flows can be enforced by the architecture in an interpretable manner. We demonstrate in numerical experiments how this topological group structure leads to concrete computational gains over other physics informed neural networks in estimating and extrapolating latent dynamics of ODEs, while training up to five times faster than other flow-based architectures.

Figures

Figures reproduced from arXiv: 2411.08326 by the authors.

Figure 1
Figure 1. Spurious convergence of an MLP to equilibrium, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Deforming a harmonic oscillator’s orbit to match [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. The NCF pipeline: Change variables to the conjugate manifold, iterate, then change back. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: A standard coupling layer ensemble in R 2 . Inverse evaluation may be done by reversing the arrows and replacing + for −. Notice that each MLP only ‘sees’ half the input, leading to reduced representation power. 3.3 The flow Ψ Our architecture is best suited for system…
Figure 7
Figure 7. Figure 7: Our augmented Affine Neural Conjugate, in full. The following operations are applied in order, from left to right: input [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Reference values for the membrane potential and [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Reconstructed and extrapolated membrane poten [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 23 canonical work pages

  1. [1]

    S.; Januschowski, T.; and G \"u nnemann, S

    Bilo s , M.; Sommer, J.; Rangapuram, S. S.; Januschowski, T.; and G \"u nnemann, S. 2021. Neural flows: Efficient alternative to neural ODEs. Advances in neural information processing systems, 34: 21325--21337

    work page 2021

  2. [2]

    J.; Brunton, S

    Bramburger, J. J.; Brunton, S. L.; and Kutz, J. N. 2021. Deep learning of conjugate mappings. Physica D: Nonlinear Phenomena, 427: 133008

    work page 2021

  3. [3]

    Chen, R. T. Q.; Rubanova, Y.; Bettencourt, J.; and Duvenaud, D. 2018. Neural ordinary differential equations. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS'18, 6572–6583. Red Hook, NY, USA: Curran Associates Inc

    work page 2018

  4. [4]

    Cranmer, M.; Greydanus, S.; Hoyer, S.; Battaglia, P.; Spergel, D.; and Ho, S. 2020. Lagrangian neural networks. arXiv preprint arXiv:2003.04630

    arXiv 2020

  5. [5]

    S.; Giampaolo, F.; Rozza, G.; Raissi, M.; and Piccialli, F

    Cuomo, S.; Di Cola, V. S.; Giampaolo, F.; Rozza, G.; Raissi, M.; and Piccialli, F. 2022. Scientific machine learning through physics--informed neural networks: Where we are and what’s next. Journal of Scientific Computing, 92(3): 88

    work page 2022

  6. [6]

    u r Angewandte Mathematik, Eidgen \

    De Ryck, T.; Mishra, S.; and Molinaro, R. 2022. Weak physics informed neural networks for approximating entropy solutions of hyperbolic conservation laws. In Seminar f \"u r Angewandte Mathematik, Eidgen \"o ssische Technische Hochschule, Z \"u rich, Switzerland, Rep , volume 35, 2022

    work page 2022

  7. [7]

    Dinh, L.; Sohl-Dickstein, J.; and Bengio, S. 2017. Density estimation using Real NVP . In International Conference on Learning Representations

    2017

  8. [8]

    o rr, C.; and K \

    Draxler, F.; Wahl, S.; Schn \"o rr, C.; and K \"o the, U. 2024. On the universality of coupling-based normalizing flows. arXiv preprint arXiv:2402.06578

    arXiv 2024

Show all 40 references
  1. [9]

    Dupont, E.; Doucet, A.; and Teh, Y. W. 2019. Augmented Neural ODEs. arXiv:1904.01681

    2019 arXiv

  2. [10]

    Finlay, C.; Jacobsen, J.-H.; Nurbekyan, L.; and Oberman, A. 2020. How to Train Your Neural ODE : the World of J acobian and Kinetic Regularization. In Proceedings of the 37th International Conference on Machine Learning, 3154--3164. PMLR

    2020

  3. [11]

    FitzHugh, R. 1961. Impulses and Physiological States in Theoretical Models of Nerve Membrane. Biophysical Journal, 1(6): 445--466

    1961

  4. [12]

    Guckenheimer, J.; and Oliva, R. A. 2002. Chaos in the Hodgkin--Huxley model. SIAM Journal on Applied Dynamical Systems, 1(1): 105--114

    2002

  5. [13]

    M.; Ravier, R.; Farsiu, S.; and Tarokh, V

    Hasan, A.; Pereira, J. M.; Ravier, R.; Farsiu, S.; and Tarokh, V. 2020. Learning Partial Differential Equations From Data Using Neural Networks. In 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 3962--3966

    2020

  6. [14]

    L.; and Huxley, A

    Hodgkin, A. L.; and Huxley, A. F. 1952. A quantitative description of membrane current and its application to conduction and excitation in nerve. Journal of Physiology, 117: 500--544

    1952

  7. [15]

    Humphreys, J. E. 1972. Introduction to Lie Algebras and Representation Theory. Springer New York. ISBN 9781461263982

    1972

  8. [16]

    Johnston, D.; and Wu, S. M.-S. 1994. Foundations of cellular neurophysiology. MIT press

    1994

  9. [17]

    Kidger, P. 2022. On Neural Differential Equations. arXiv:2202.02435

    2022 arXiv

  10. [18]

    P.; and Ba, J

    Kingma, D. P.; and Ba, J. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  11. [19]

    P.; and Welling, M

    Kingma, D. P.; and Welling, M. 2022. Auto-Encoding Variational Bayes. arXiv:1312.6114

    2022 arXiv

  12. [20]

    Kumar, S. K. 2017. On weight initialization in deep neural networks. arXiv preprint arXiv:1704.08863

    2017 arXiv

  13. [21]

    Lai, Z.; Mylonas, C.; Nagarajaiah, S.; and Chatzi, E. 2021. Structural identification with physics-informed neural ordinary differential equations. Journal of Sound and Vibration, 508: 116196

    2021

  14. [22]

    D.; and Karniadakis, G

    Mao, Z.; Jagtap, A. D.; and Karniadakis, G. E. 2020. Physics-informed neural networks for high-speed flows. Computer Methods in Applied Mechanics and Engineering, 360: 112789

    2020

  15. [23]

    S.; and Protopapas, P

    Mattheakis, M.; Sondak, D.; Dogra, A. S.; and Protopapas, P. 2022. Hamiltonian neural networks for solving equations of motion. Physical Review E, 105(6): 065305

    2022

  16. [24]

    A.; and Mesbah, A

    O'Leary, J.; Paulson, J. A.; and Mesbah, A. 2022. Stochastic physics-informed neural ordinary differential equations. Journal of Computational Physics, 468: 111466

    2022

  17. [25]

    J.; Mohamed, S.; and Lakshminarayanan, B

    Papamakarios, G.; Nalisnick, E.; Rezende, D. J.; Mohamed, S.; and Lakshminarayanan, B. 2021. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research, 22(57): 1--64

    2021

  18. [26]

    G.; Manickam, I.; Trask, N

    Patel, R. G.; Manickam, I.; Trask, N. A.; Wood, M. A.; Lee, M.; Tomas, I.; and Cyr, E. C. 2022. Thermodynamically consistent physics-informed neural networks for hyperbolic systems. Journal of Computational Physics, 449: 110754

    2022

  19. [27]

    Poli, M.; Massaroli, S.; Yamashita, A.; Asama, H.; Park, J.; and Ermon, S. 2021. TorchDyn: implicit models and neural numerical methods in PyTorch. In Neural Information Processing Systems, Workshop on Physical Reasoning and Inductive Biases for the Real World, volume 2

    2021

  20. [28]

    Raissi, M.; Perdikaris, P.; and Karniadakis, G. E. 2019. 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

    2019

  21. [29]

    Rao, C.; Sun, H.; and Liu, Y. 2021. Physics-informed deep learning for computational elastodynamics without labeled data. Journal of Engineering Mechanics, 147(8): 04021043

    2021

  22. [30]

    Ronneberger, O.; Fischer, P.; and Brox, T. 2015. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, 234--241. Springer

    2015

  23. [31]

    Tancik, M.; Srinivasan, P.; Mildenhall, B.; Fridovich-Keil, S.; Raghavan, N.; Singhal, U.; Ramamoorthi, R.; Barron, J.; and Ng, R. 2020. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, ...

    2020

  24. [32]

    Teshima, T.; Ishikawa, I.; Tojo, K.; Oono, K.; Ikeda, M.; and Sugiyama, M. 2020. Coupling-based invertible neural networks are universal diffeomorphism approximators. Advances in Neural Information Processing Systems, 33: 3362--3373

    2020

  25. [33]

    Viana, M.; and Espinar, J. M. 2021. Differential equations: a dynamical systems approach to theory and practice, volume 212. American Mathematical Society

    2021

  26. [34]

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

    2022 arXiv

  27. [35]

    Wang, S.; Sankaran, S.; Wang, H.; and Perdikaris, P. 2023. An expert's guide to training physics-informed neural networks. arXiv preprint arXiv:2308.08468

    2023 arXiv

  28. [36]

    Yu, J.; Lu, L.; Meng, X.; and Karniadakis, G. E. 2022. Gradient-enhanced physics-informed neural networks for forward and inverse PDE problems. Computer Methods in Applied Mechanics and Engineering, 393: 114823

    2022

  29. [37]

    Zhang, Z.-Y.; Zhang, H.; Zhang, L.-S.; and Guo, L.-L. 2023. Enforcing continuous symmetries in physics-informed neural network for solving forward and inverse problems of partial differential equations. Journal of Computational Physics, 492: 112415

    2023

  30. [38]

    V.; and Ramos, F

    Zhi, W.; Lai, T.; Ott, L.; Bonilla, E. V.; and Ramos, F. 2022. Learning Efficient and Robust Ordinary Differential Equations via Invertible Neural Networks. In Chaudhuri, K.; Jegelka, S.; Song, L.; Szepesvari, C.; Niu, G.; and Sabato, S., eds., Proceedings of the 39th Internat...

    2022

  31. [39]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  32. [40]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.