Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

SymFlux: deep symbolic regression of Hamiltonian vector fields

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A CNN-LSTM trained on rendered images of planar Hamiltonian vector fields outputs the symbolic Hamiltonian that generated them, with 85–88% exact-formula accuracy on its test sets.

desk verdict Worth engaging for the datasets and the visual-captioning setup, but the accuracy numbers are not trustworthy until the train/test split is clarified. read the letter →

arxiv 2507.06342 v1 pith:MI5XJ22O submitted 2025-07-08 cs.LG cs.AImath.DSmath.SG

M.A. Evangelista-Alvarado , P. Suárez-Serrato This is my paper · ORCID
classification cs.LGcs.AImath.DSmath.SG MSC 68T07
keywords HamiltonianvectorfieldssymbolicregressionHamiltonizationproblemCNN-LSTMvisualrepresentationofLie–Hamiltonsystemsplanardynamicaldeeplearning
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 targets the Hamiltonization problem in reverse: given pictures of a vector field on the plane, find the Hamiltonian function $H$ whose Hamiltonian equations produce that field. It builds three synthetic databases—symbolic Hamiltonians, numerical vector fields, and rendered images (quiver, streamline, heatmap) sampled on 50 point clouds—over a finite vocabulary of polynomial and trigonometric terms. On top of these, it trains a hybrid CNN-LSTM, called SymFlux, that reads an image and outputs the symbolic formula for $H$ token by token. The paper reports 85–88% test accuracy at recovering the exact formula within that vocabulary, with the harmonic oscillator recovered exactly and the pendulum recovered up to a constant. The same experiments show the limits: Hamiltonians with logarithmic terms or $1/y$, like the Lotka–Volterra and SIS models, are not recoverable because they lie outside the training vocabulary.

What carries the argument

The load-bearing object is the map that sends a Hamiltonian $H$ to its visual representation: the Hamiltonian vector field $X_H$ defined by $\iota_{X_H}\omega_0=dH$, evaluated on 50 point clouds in $[-10,10]^2$ and rendered as a triple of images $\eta(X_H|_{P_j})=(\theta_1,\theta_2,\theta_3)$ (quiver, streamplot, heatmap). Around this map the paper builds a token vocabulary: each Hamiltonian is treated as a sentence whose words are monomials $b_j x^h y^k$ or trig terms $b_j t_l$, each encoded as a one-hot vector, and the CNN-LSTM is asked to predict the token sequence from the image. The finite basis $\mathrm{Ham}(B_i,\Delta_l)$ is what makes the regression a closed classification-style task; the Darboux theorem justifies restricting to the standard symplectic plane, and the Weierstrass approximation theorem motivates polynomial approximation.

What would settle it

Render the SIS Hamiltonian $H(x,y)=xy(1-x)+1/y$ from Section 2.1.4 with the paper's pipeline and run a trained SymFlux model: the paper reports a wrong polynomial prediction, so this case directly refutes any unrestricted claim of Hamiltonian recovery. For the restricted claim, hold out a random element of $\mathrm{Ham}(B_3,\Delta_3)$, render its three image types, and check whether the model's exact-formula match rate reaches the reported 85–88%.

Watch

Extended reading notes

Core claim

The central claim is that symbolic Hamiltonization can be learned from visual data: a CNN-LSTM trained on rendered vector-field images can output the symbolic expression of the Hamiltonian $H$ that generates the field, on the standard symplectic plane where $\omega_0 = dx \wedge dy$ and $\iota_{X_H}\omega_0 = dH$. The authors construct the function sets $\mathrm{Ham}(B_i, \Delta_l)$ and $\mathrm{Ham}(B_i^*, \Delta_l)$ from monomials $b_j x^h y^k$ with $1\le h+k\le i$ and coefficients in a discrete set $\Delta_l$, plus the trigonometric terms $\{\cos x,\cos y,\sin x,\sin y\}$, and generate images of the corresponding vector fields. Their best models reach 85–88% exact formula recovery on test splits; the 1D harmonic oscillator is recovered exactly, and the pendulum is recovered up to a constant, which the authors note yields nearly identical vector fields. On the Lotka–Volterra and SIS Hamiltonians, whose formulas contain $\ln$ and $1/y$, the model outputs a polynomial lookalike instead, because those functions were absent from the training basis.

Load-bearing premise

Everything rests on the assumption that the true Hamiltonian of the observed vector field is a linear combination of the finite monomial and trigonometric terms with discrete coefficients used to build the training databases; the paper's own Lotka–Volterra and SIS failures show that when this fails, the formula is not recovered.

Editorial extensions

If this is right

  • If SymFlux's accuracy holds, a researcher who has only a rendered flow field can recover the symbolic energy function that produces it, without knowing the equations of motion.
  • The three databases become reusable benchmarks and training resources for any future symbolic-regression method aimed at Hamiltonian systems.
  • The mathematical pendulum is recovered up to a coefficient change, and the paper observes that the resulting vector-field images are nearly indistinguishable, so the model can reproduce the dynamics even when the formula is not exact.
  • Accuracy rises with more visual representations per Hamiltonian; the paper reports growth from about 75% to 89% from data augmentation, so scaling the dataset is a concrete route to better recovery.
  • Out-of-vocabulary Hamiltonians, including Lotka–Volterra and SIS, are mispredicted; recovering them requires retraining on an expanded basis that includes logarithms and reciprocal terms.

Reading between the lines

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

  • A natural extension is to enlarge the vocabulary with logarithmic, exponential, rational, and higher-frequency trigonometric terms and measure how exact-match accuracy changes; the paper's own failure pattern suggests the bottleneck is vocabulary coverage, not model capacity.
  • An untested consequence is that the method is well-posed only up to transformations of $H$ that leave $X_H$ unchanged, such as adding a constant; any exact-formula accuracy metric secretly fixes a representative, and a fair benchmark would report accuracy on equivalence classes of Hamiltonians.
  • One could test the same image-to-symbol pipeline on higher-dimensional symplectic systems; Darboux's theorem guarantees local equivalence to the standard plane, but the visual encoding would need a new rendering scheme, so scalability is open.
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

4 major / 6 minor

Summary. The paper introduces SymFlux, a CNN-LSTM architecture that takes images of planar vector fields as input and outputs a symbolic Hamiltonian function from a finite vocabulary of polynomial and trigonometric monomials. The authors construct three new synthetic databases (symbolic, numerical, and visual) of Hamiltonian vector fields on the standard symplectic plane, and report symbolic regression accuracies of 73–88% on four database variants. The headline claim is that SymFlux identifies the correct Hamiltonian from its visual representation with 85–88% accuracy on test data. The paper also presents four showcase examples (harmonic oscillator, pendulum, Lotka–Volterra, SIS) and discusses where the model succeeds and where it fails.

Significance. If the empirical claims hold, the paper would be a useful contribution: it is the first to apply visual symbolic regression to the Hamiltonization problem, it releases code and synthetic datasets, and it systematically incorporates a known classification of Lie–Hamilton systems. The central contribution is the dataset-generation pipeline and the demonstration that a CNN-LSTM can map renderings of vector fields to symbolic Hamiltonians within a closed vocabulary. However, the significance is bounded by three issues: the reported accuracies are not reproducible from the text as written (split ambiguity, undefined accuracy threshold, no error bars), the method only works inside the finite training span, and—most importantly—the paper's explicit vector-field formulas appear to use the opposite sign from the stated symplectic convention. The sign issue threatens the correctness of the entire dataset-label mapping and must be resolved before the empirical claims can be accepted.

major comments (4)
  1. [§2.1.1, Eq. (5); also Eqs. (8), (11), (14)] With the stated convention ι_{X_H}ω = dH and ω = dx∧dy, the Hamiltonian vector field of H = ½(y² + α²x²) is X_H = y ∂_x − α²x ∂_y, not −y ∂_x + x ∂_y as given in Eq. (5). The same sign reversal appears in the pendulum formula in Eq. (8), the Lotka–Volterra formula in Eq. (11), and the SIS formula in Eq. (14), as well as in the predicted-field discussion in Table 6. Because the visual database is generated from these vector fields (via the PoissonGeometry module), either the images are actually of X_{−H} while the labels are H, or the written equations do not match the data. Please verify the sign convention used by the module, correct the formulas, and confirm that each image label corresponds to the vector field actually rendered. This is load-bearing for the validity of the training data and the accuracy claims.
  2. [§4.5 with §3.3.1 and §3.4] The training section says only that 'the datasets were split into 75% training and 25% testing.' Since each Hamiltonian class has 50 images, one per point cloud (Sec. 3.3.1, 3.4), an image-level split would place re-renderings of training Hamiltonians in the test set. The reported 85–88% test accuracies would then measure recognition of point-cloud variation rather than generalization to unseen Hamiltonian functions. Please state explicitly whether the split was performed at the level of Hamiltonian classes, and if it was not, repeat the evaluation with a class-level split and report the resulting accuracies.
  3. [§4.6.3, Definition 4.1, Table 5] Eq. (16) defines a Euclidean distance between token vectors but the paper never states the threshold used to declare a prediction correct. Table 5 reports 'Accuracy' without defining whether d = 0 (exact token-vector equality) or some tolerance was required. In addition, no error bars or number of training runs are given, so the 85–88% figures are not reproducible and may reflect a single run. Please define the accuracy criterion explicitly and include variance over repeated runs with different seeds.
  4. [Section 1 and Table 6] The abstract and introduction claim that SymFlux identifies 'the correct Hamiltonian function' from its visual representation, but Table 6 shows that three of the four showcase predictions are incorrect: the pendulum is recovered only up to a scaling factor, the Lotka–Volterra Hamiltonian is wrong, and the SIS Hamiltonian is wrong. The authors attribute those failures to terms (logarithmic, 1/y) outside the training vocabulary, which is an inherent limitation of the closed-vocabulary supervised setup described in Sec. 3.1.3. The central claim should therefore be qualified in the abstract and introduction to say 'within the span of the predefined polynomial/trigonometric basis.' Also, the statement in §4.6.3 that the original and predicted pendulum vector fields produce 'identical vector field representations' is incorrect: the y-component differs by a factor of 2, so the visual representations are not identical.
minor comments (6)
  1. [§3.1.3 vs §3.4 and Tables 3–5] Section 3.1.3 states that the work uses Ham(B₂,∆₃), Ham(B₂,∆₅), Ham(B₃,∆₃), and Ham(B*₂,∆₅), but Table 3 and the training/test tables use Ham(B*₂,∆₃) for the trigonometric set. Please reconcile this inconsistency.
  2. [§4.4] The tokenization example gives H(x, y) = ½y² + x² as the harmonic-oscillator Hamiltonian in Ham(B₂,∆₅), but Eq. (4) with α = 1 gives H(x, y) = ½(y² + x²). The example should use the same expression as the ground truth in Table 6.
  3. [Section 5] The conclusion states 'accuracies of up to 89%,' but the maximum accuracy reported in Table 5 is 88%, and the maximum in Table 4 is 87%. Please correct the number or clarify what the 89% refers to.
  4. [Tables 4 and 5] Table 4 is introduced as 'the performance of our SymFlux models over 50 epochs' without specifying whether it reports training or validation accuracy; Table 5 is said to use the test dataset. Please label the split explicitly in each table caption.
  5. [§3.3.1] The uniform density is written as p(x) = x/(b−a); it should be p(x) = 1/(b−a) on [a, b).
  6. [Figure 7 caption and figure cross-references] In the Figure 7 caption, the second vector field is expressed with ∂/∂x but should use ∂/∂y. In addition, the text near Figure 5 refers to panels '4a' and '4b' instead of '5a' and '5b.' Other typos include 'RestNet' for ResNet, 'fiels' in the Section 3.2 title, 'perespective', 'strating', and 'traning.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the supervised CNN-LSTM benchmark is self-contained, with implementation self-citations that are not load-bearing.

full rationale

The paper's derivation chain is: choose a finite Hamiltonian vocabulary (Ham(B, Delta)), generate symbolic and numerical Hamiltonian vector fields from those Hamiltonians, render visual representations, then train a CNN-LSTM to map images to the symbolic Hamiltonian label. This is a standard supervised symbolic-regression setup: the Hamiltonian label is the ground truth used to generate the input image, and accuracy is measured on a held-out test set (Section 4.5, Table 5). No equation or fitting procedure makes the predicted output equal to the input by construction; the model is not fitting a parameter to the test data and then reporting that fit as a prediction. The custom metric in Definition 4.1 is a token-vector Euclidean distance, and though the paper does not specify a threshold for 'accuracy', this is a reporting ambiguity rather than a circular reduction. The closed-vocabulary limitation (Table 6, Section 5) is explicitly acknowledged by the authors: Lotka-Volterra and SIS fail because logarithmic and 1/y terms are outside the training basis. That is a generalizability limitation, not circularity. Self-citations to the authors' PoissonGeometry and NumericalPoissonGeometry modules are used for symbolic and numerical computation tooling, not as load-bearing theoretical premises, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the chosen architecture or basis. The only potentially concerning issue, the unstated split granularity in Section 4.5, cannot be established from the paper text as an image-level rather than class-level split, and in any case would be data leakage rather than a by-construction identity between input and output. Overall, the central accuracy claim has independent empirical content within the stated vocabulary.

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

The central claim rests on a hand-chosen polynomial and trigonometric vocabulary with discrete coefficients, on the correctness of the authors' own Poisson geometry software for generating ground truth, and on the classificational work of Ballesteros et al. No new physical entities are introduced. The reported accuracy is also hostage to an undefined correctness threshold.

free parameters (4)
  • Coefficient grid Delta_l = Delta3 = {-1,0,1}; Delta5 = {-1,-0.5,0,0.5,1}; Delta7 and Delta9 as listed in Section 3.1.3
    The discrete coefficient set defines which Hamiltonians can exist in the dataset, and therefore which symbolic outputs the model can produce; accuracy is only measured inside this grid.
  • Polynomial degree bound i = i=2 and i=3 for the four datasets actually used
    Only monomials of total degree at most i are included, so the model cannot represent Hamiltonians of higher degree.
  • Number of point clouds per Hamiltonian = 50
    Each vector field is sampled on 50 point clouds chosen by hand; this determines dataset size and the augmentation the authors say improved accuracy.
  • Visualization tuple components = 3: quiver, streamplot, heatmap
    The choice of which visualizations to concatenate into the CNN input is fixed by hand and its sufficiency is not analyzed.
assumptions (4)
  • standard math Darboux theorem: every symplectic manifold is locally symplectomorphic to (R^2n, omega0)
    Used in Section 3 to justify working only on the standard symplectic plane.
  • standard math Weierstrass approximation theorem: continuous functions on a closed interval are uniformly approximable by polynomials
    Invoked in Section 3 to motivate restricting Hamiltonian energy functions to polynomials; it is an approximation heuristic, not an exact reduction.
  • domain assumption Ballesteros et al. classification of planar Lie-Hamilton systems is correct
    Contribution (3) says the databases include all planar Lie-Hamiltonian vector fields that are not integrals, based on this cited classification; the dataset composition depends on it.
  • domain assumption The PoissonGeometry and NumericalPoissonGeometry modules (self-cited [11-13]) compute vector fields, point-cloud evaluations, and images correctly
    All ground truth labels and inputs are generated with these modules, so any bug propagates into training and evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SymFlux: deep symbolic regression of Hamiltonian vector fields." pith.science (2026). https://pith.science/paper/MI5XJ22O

@misc{pith2026250706342,
  author       = {Pith},
  title        = {Pith review of: SymFlux: deep symbolic regression of Hamiltonian vector fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MI5XJ22O}},
  note         = {Machine review of arXiv:2507.06342}
}
read the original abstract

We present SymFlux, a novel deep learning framework that performs symbolic regression to identify Hamiltonian functions from their corresponding vector fields on the standard symplectic plane. SymFlux models utilize hybrid CNN-LSTM architectures to learn and output the symbolic mathematical expression of the underlying Hamiltonian. Training and validation are conducted on newly developed datasets of Hamiltonian vector fields, a key contribution of this work. Our results demonstrate the model's effectiveness in accurately recovering these symbolic expressions, advancing automated discovery in Hamiltonian mechanics.

Figures

Figures reproduced from arXiv: 2507.06342 by the authors.

Figure 1
Figure 1. A diagram of a general SymFlux system, employing CNN and LSTM submodules. This model processes an image representation of a vector field through a CNN to extract a feature vector. This vector is then input to an LSTM, which predicts the symbolic form of the Hamiltonian, m0 + . . . + mn . (1) We first curated a diverse set of analytical functions to serve as ground-truth Hamiltonian energy functions with respect to t… view at source ↗
Figure 2
Figure 2. Three distinct images representing the harmonic oscillator Hamiltonian vector field in eq. ( [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Three images associated with the mathematical pendulum’s Hamiltonian vector field [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Three images associated with the Lotka–Volterra predator-prey model’s Hamiltonian vector field [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Examples of canonical (4a) and random (4b) point clouds defined in [ [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: An illustration of how we create visual representations strating from a symbolic expression. In this [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Visual representation of the vector field: [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Continue with ORCID 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

  1. Space filling positionality and the Spiroformer

    cs.LG 2025-07 reject novelty 4.0 of 10

    A transformer that predicts vectors sampled along a polar spiral on a sphere reaches about 90 percent training accuracy but lower validation accuracy, so the proposed geometric ordering remains unvalidated.

Reference graph

Works this paper leans on

42 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    Tensorflow: a system for large-scale machine learning

    Martín Abadi et al. Tensorflow: a system for large-scale machine learning. In Proceedings of the 12th USENIX Conference on Operating Systems Design and Implementation, OSDI’16, page 265–283. USENIX Association, 2016

    work page 2016

  2. [2]

    V .I. Arnold. Mathematical methods of classical mechanics, volume 60. Springer, 1989

    1989

  3. [3]

    Geometrical aspects of the hamiltonization problem of dynamical systems

    Misael Avendaño-Camacho, Claudio César García-Mendoza, José Crispín Ruíz-Pantaleón, and Ed- uardo Velasco-Barreras. Geometrical aspects of the hamiltonization problem of dynamical systems. Symmetry, Integrability and Geometry: Methods and Applications (SIGMA), 18, 2022

    work page 2022

  4. [4]

    Ballesteros, A

    A. Ballesteros, A. Blasco, F.J. Herranz, J. de Lucas, and C. Sardón. Lie–Hamilton systems on the plane: Properties, classification and applications. Journal of Differential Equations, 258(8):2873–2907, 2015

    work page 2015

  5. [5]

    Symplectic recurrent neural networks

    Zhengdao Chen, Jianyu Zhang, Martín Arjovsky, and Léon Bottou. Symplectic recurrent neural networks. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020

    work page 2020

  6. [6]

    F. Chollet. Xception: Deep learning with depthwise separable convolutions. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1800–1807, Los Alamitos, CA, USA, July 2017. IEEE Computer Society

    work page 2017

  7. [7]

    Cooke and V .I

    R. Cooke and V .I. Arnold. Ordinary Differential Equations. Springer Textbook. Springer Berlin Heidelberg, 1992

    work page 1992

  8. [8]

    Discovering symbolic models from deep learning with inductive biases

    Miles Cranmer, Alvaro Sanchez-Gonzalez, Peter Battaglia, Rui Xu, Kyle Cranmer, David Spergel, and Shirley Ho. Discovering symbolic models from deep learning with inductive biases. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS’20. Curran Associates Inc., 2020

    work page 2020

Show all 42 references
  1. [9]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009

    2009

  2. [10]

    A generalization of a SIS epidemic model with fluctuations

    Ogul Esen, Eduardo Fernández-Saiz, Cristina Sardón, and Marcin Zajac. A generalization of a SIS epidemic model with fluctuations. Mathematical Methods in the Applied Sciences, 45(7):3718–3731, 2022. SymFlux: deep symbolic regression of Hamiltonian vector fields 25

    2022

  3. [11]

    Examples of symbolic and numerical computation in Poisson geometry

    Miguel Evangelista-Alvarado, José Crispín Ruíz-Pantaleón, and Pablo Suárez-Serrato. Examples of symbolic and numerical computation in Poisson geometry. In Frank Nielsen and Frédéric Barbaresco, editors, Geometric Science of Information, pages 200–208. Springer International Pu...

    2021

  4. [12]

    On computa- tional Poisson geometry I: Symbolic foundations

    Miguel Evangelista-Alvarado, José Crispín Ruíz-Pantaleón, and Pablo Suárez-Serrato. On computa- tional Poisson geometry I: Symbolic foundations. Journal of Geometric Mechanics, 13(4):607–628, 2021

    2021

  5. [13]

    On computa- tional Poisson geometry II: Numerical methods

    Miguel Evangelista-Alvarado, José Crispín Ruíz-Pantaleón, and Pablo Suárez-Serrato. On computa- tional Poisson geometry II: Numerical methods. Journal of Computational Dynamics, 8(3):273–307, 2021

    2021

  6. [14]

    Datasheets for datasets

    Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, and Kate Crawford. Datasheets for datasets. Commun. ACM, 64(12):86–92, Nov. 2021

    2021

  7. [15]

    Hamiltonian neural networks

    Samuel Greydanus, Misko Dzamba, and Jason Yosinski. Hamiltonian neural networks. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019

    2019

  8. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR ’16, pages 770–778. IEEE, June 2016

    2016

  9. [17]

    Long short-term memory

    Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735– 1780, 11 1997

    1997

  10. [18]

    The construction of a poisson structure out of a symmetry and a conservation law of a dynamical system

    Sergio A Hojman. The construction of a poisson structure out of a symmetry and a conservation law of a dynamical system. Journal of Physics A: Mathematical and General, 29(3):667, feb 1996

    1996

  11. [19]

    J. D. Hunter. Matplotlib: A 2d graphics environment. Computing in Science & Engineering, 9(3):90– 95, 2007

    2007

  12. [20]

    Étude comparative de la distribution florale dans une portion des alpes et des jura

    Paul Jaccard. Étude comparative de la distribution florale dans une portion des alpes et des jura. Bulletin de la Société Vaudoise des Sciences Naturelles, 37:547–579, 1901

    1901

  13. [21]

    Adam: A method for stochastic optimization

    Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diega, CA, USA, 2015

    2015

  14. [22]

    John R. Koza. Genetic programming as a means for programming computers by natural selection. Proceedings of the National Academy of Sciences, 4(2):87–112, 1994

    1994

  15. [23]

    Laramee, Helwig Hauser, Helmut Doleisch, Benjamin Vrolijk, Frits H

    Robert S. Laramee, Helwig Hauser, Helmut Doleisch, Benjamin Vrolijk, Frits H. Post, and Daniel Weiskopf. The state of the art in flow visualization: Dense and texture-based techniques. Computer Graphics Forum, 23(2):203–221, 2004

    2004

  16. [24]

    Lecun, L

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998

    1998

  17. [25]

    Self-learning machines based on hamiltonian echo backpropagation

    Víctor López-Pastor and Florian Marquardt. Self-learning machines based on hamiltonian echo backpropagation. Phys. Rev. X, 13:031020, Aug 2023

    2023

  18. [26]

    A.J. Lotka. Elements of Physical Biology. Williams & Wilkins, 1925

    1925

  19. [27]

    Interpretable scientific discovery with symbolic regression: A review

    Nour Makke and Sanjay Chawla. Interpretable scientific discovery with symbolic regression: A review. Artificial Intelligence Review, 57(2):1–41, 2023

    2023

  20. [28]

    Marsden and T

    J.E. Marsden and T. Ratiu. Introduction to Mechanics and Symmetry: A Basic Exposition of Classical Mechanical Systems. Texts in Applied Mathematics. Springer New York, 2002. 26 M. Evangelista-Alvarado and P . Suárez-Serrato

    2002

  21. [29]

    McDuff and D

    D. McDuff and D. Salamon. Introduction to Symplectic Topology. Oxford mathematical monographs. Clarendon Press, 1998

    1998

  22. [30]

    Sympy: symbolic computing in python

    Aaron Meurer et al. Sympy: symbolic computing in python. PeerJ Computer Science, 3, January 2017

    2017

  23. [31]

    Kerastuner

    Tom O’Malley, Elie Bursztein, James Long, François Chollet, Haifeng Jin, Luca Invernizzi, et al. Kerastuner. https://github.com/keras-team/keras-tuner , 2019

    2019

  24. [32]

    The Hamiltonization problem from a global viewpoint

    V olker Perlick. The Hamiltonization problem from a global viewpoint. Journal of Mathematical Physics, 33(2):599–606, 02 1992

    1992

  25. [33]

    Rew and G

    R. Rew and G. Davis. Netcdf: an interface for scientific data access. IEEE Computer Graphics and Applications, 10(4):76–82, 1990

    1990

  26. [34]

    The Weierstrass approximation theorem

    Eugene Schenkman. The Weierstrass approximation theorem. The American Mathematical Monthly, 79(1):65–66, 1972

    1972

  27. [35]

    Distilling free-form natural laws from experimental data

    Michael Schmidt and Hod Lipson. Distilling free-form natural laws from experimental data. Science, 324(5923):81–85, 2009

    2009

  28. [36]

    Khoshgoftaar

    Connor Shorten and Taghi M. Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of Big Data, 6:60, 2019

    2019

  29. [37]

    AI Feynman: A physics-inspired method for symbolic regression

    Silviu-Marian Udrescu and Max Tegmark. AI Feynman: A physics-inspired method for symbolic regression. Science Advances, 6(16):eaay2631, 2020

    2020

  30. [38]

    Symformer: End-to-end symbolic regression using transformer-based architecture

    Martin Vastl, Jonáš Kulhánek, Jiˇrí Kubalík, Erik Derner, and Robert Babuška. Symformer: End-to-end symbolic regression using transformer-based architecture. IEEE Access, 12:37840–37849, 2024

    2024

  31. [39]

    Show and tell: A neural image caption generator

    Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156–3164, 2015

    2015

  32. [40]

    Binary codes capable of correcting deletions, insertions, and reversals

    Levenshtein Vladimir I. Binary codes capable of correcting deletions, insertions, and reversals. Soviet Physics Doklady, 10(8):707–710, 1966. Originally published in Russian in Doklady Akademii Nauk SSSR, 163(4), 845–848, 1965

    1966

  33. [41]

    E. T. Whittaker and Sir William McCrae. A Treatise on the Analytical Dynamics of Particles and Rigid Bodies. Cambridge Mathematical Library. Cambridge University Press, 1988

    1988

  34. [42]

    Catanzaro, Matthew Zabka, and Bei Wang

    Youjia Zhou, Janis Lazovskis, Michael J. Catanzaro, Matthew Zabka, and Bei Wang. Mvf designer: Design and visualization of morse vector fields, 2019. Universidad Nacional Rosario Castellanos (UNRC), Mexico City, Mexico. Max-Planck Institute for Mathematics, Bonn, and Electrica...

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.