Discover physical concepts and equations with machine learning

Bao-Bing Li; Shao-Feng Wu; Yi Gu

REVIEW 3 major objections 6 minor 63 references

Discover physical concepts and equations with machine learning

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A single VAE-plus-neural-ODE architecture can simultaneously rediscover physical concepts and their governing equations, and in four textbook cases the learned theory matches the real one up to linear transformations.

desk verdict Well-motivated VAE+Neural ODE pipeline for rediscovering concepts and equations, but the central claim is only checked as post-hoc consistency on training data, with no held-out generalization test. read the letter →

arxiv 2412.12161 v2 pith:FEQTEWZY submitted 2024-12-11 cs.LG cond-mat.dis-nncs.AIphysics.comp-ph

Bao-Bing Li , Yi Gu , Shao-Feng Wu This is my paper · ORCID

classification cs.LGcond-mat.dis-nncs.AIphysics.comp-ph

keywords physicalconceptdiscoveryneuralordinarydifferentialequationsvariationalautoencoderSchrödingerequationPaulilatentrepresentationphysics

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 claims that one neural architecture, a variational autoencoder feeding a neural ordinary differential equation, can recover both the physical concepts and the governing equations of a system from simulated observation sequences. Across four cases drawn from the history of physics—Copernicus' heliocentrism, Newton's law of gravity, Schrödinger's wave mechanics, and Pauli's spin–magnetic formulation—the learned latent variables are said to store linear combinations of the textbook quantities, and the neural ODE is said to reproduce the correct equations. If this holds, a general-purpose learner could rediscover core physical theory from data, needing no priors specific to the example beyond a known control variable such as a potential field. The claim is quantified by relative errors between the learned latent dynamics and the expected linear combinations of the true physical quantities.

What carries the argument

The load-bearing mechanism is the VAE–Neural-ODE–decoder loop. An encoder maps observations (sometimes only the first frame, sometimes the full sequence) to a Gaussian latent distribution; the mean is sampled at $\varepsilon=0$ to give the initial latent state $h(t_0)$, avoiding the instability that noise would inject into the dynamics. A neural ODE then integrates $\mathrm{d}h/\mathrm{d}t=f(h(t),V;\zeta)$ forward with a numerical solver, where the control variable $V$ encodes the experimental condition, and the decoder reconstructs each observed frame. The loss is a reconstruction error plus a $\beta$-weighted KL divergence that pushes the latent posterior toward a factorized prior, encouraging the latent variables to specialise. A key design choice is that the number of latent dimensions is not fixed but selected by ablation under Occam's razor, so the architecture itself reports how many physical concepts the data require.

What would settle it

Retrain the Schrödinger model with the potential $V(x)$ withheld from both the encoder and the neural ODE (only the density sequences $\rho(x)$ are given, with the same fixed initial conditions as in Section 3.3); if the latent trajectories no longer correspond to $\psi$ and $\mathrm{d}\psi/\mathrm{d}x$ and the relative errors $R_h$ and $R_f$ rise sharply, the claimed simultaneous discovery is shown to depend on the supplied potential rather than on the data alone.

Watch

Extended reading notes

Core claim

The central discovery is that concept discovery and equation discovery do not have to be staged: the same network can learn both at once. The encoder compresses observed sequences into a low-dimensional latent space; a neural ODE imposes continuous, differential dynamics on those latent variables, with a control variable such as a potential $V(x)$ or initial radius $r_0$ appearing in the right-hand side; and the decoder reconstructs the observations from the evolved latent states. In all four examples the ablation-selected latent dimension matches the number of physical quantities (two for heliocentrism, gravity, and Schrödinger; four for the Pauli system), and the paper reports that the latent representations are linear combinations of the true concepts—for instance $h_j(x)=a_j\psi(x)+b_j\,\mathrm{d}\psi/\mathrm{d}x+c_j$ for Schrödinger—while the learned governing functions are the corresponding combinations of the true right-hand sides, such as $a_j\,\mathrm{d}\psi/\mathrm{d}x+b_j V(x)\psi(x)$. The reported relative errors $R_h$ and $R_f$ (Table 1) are small, with, for example, $R_f(50,4)\approx 0.03$ for the Pauli system, which the paper takes as evidence that the correct physical theories emerge in the network.

Load-bearing premise

The load-bearing premise is that the control variable—here the random potential $V(x)$ (and the magnetic offset $B$) in the quantum examples, or $r_0$ for gravity—is known and supplied to the network, because every training sample in those examples starts from the same initial wave function and differs only through that input.

Editorial extensions

If this is right

If the architecture works as claimed, a single trained model outputs both a set of latent coordinates and an ODE describing them, giving a physicist a complete candidate theory to interpret.
The model handles control variables such as potentials and initial radii, so it applies to families of experiments rather than single trajectories.
The Pauli example indicates that hidden mechanisms like spin can be extracted even when the observable pattern is degenerate (one stripe instead of two), potentially shortening discovery time for masked phenomena.
Enforcing a second-order structure (Appendix E) can break the linear-combination ambiguity, turning latent variables into the independent concepts themselves in the Newton and Schrödinger cases.
The relative-error metrics $R_h$ and $R_f$ provide a concrete, comparable measure of how close a machine-discovered theory is to the textbook one.

Reading between the lines

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

The paper leaves implicit that the Schrödinger and Pauli demonstrations require the potential $V(x)$ (and the constant $B$) to be measured and fed to the network, since all training samples share the same initial conditions; a natural control experiment would withhold $V$ and check whether the discovery collapses.
Because any invertible linear redefinition of the latent variables yields the same observations, what the network finds is an equivalence class of theories, not a unique one; the second-order constraint is one way to select a representative, and other selection principles could be explored.
The dependence on a control variable suggests a sharper reading of the claim: the model rediscovers concepts and equations given the experimental knob, not from raw data alone; testing with varied initial conditions would show whether the encoder contributes beyond passing through the control information.
One testable extension would be to apply the same architecture to a system whose true concepts are not linear combinations of obvious observables, to see how far the linear-combination success generalizes.

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: this is a well-motivated extension of SciNet — swap the tacit uniform-evolution assumption for a Neural ODE in latent space — and the Pauli spin-magnetic example is a genuinely new and interesting test. But the evidence behind the claim that "the correct physical theories can emerge" is weaker than the prose suggests. I'd send it to referees, but require substantial validation before that discovery language is justified.

What the paper does well: the architecture is simple and clearly described — β-VAE encoder, deterministic initial latent state, Neural ODE evolved under control variables (potentials, initial distance), decoder. Across four historically flavored benchmarks it shows latent trajectories that look like linear mixtures of the textbook concepts, and learned ODE right-hand sides that lie close to the expected conjugate dynamics. The authors are honest about the latent redundancy, and Appendix E is a genuine attempt — successful for Newton and Schrödinger, failing for Pauli, and reported as such — to recover independent concepts under a second-order assumption. The Discussion likewise concedes the Pauli degradation and the computational limits.

The soft spots are not minor. The headline metrics R_h and R_f are post-hoc linear fits computed on the same data used for training, against expected functions constructed from the true equations (e.g., Eq. 23 uses Eq. 20). That is a consistency check, not an independent rediscovery. The bigger problem: in the Schrödinger and Pauli cases the encoder is fed the full trajectory — input sizes 50 and 100 are the entire sequence — so the sampled initial latent state can encode trajectory-specific information. Since all samples share identical true initial conditions, the model could fit the training set as a full-curve autoencoder with an ODE bottleneck. Nothing rules that out, because the paper never tests generalization to unseen potentials, which is the defining test of a discovered law. There are also no error bars, no repeated seeds, and no code or data, so the reported 1–10% relative errors are hard to evaluate.

Proportion check: the architecture is sensible, prior work is cited fairly, and the masked-single-stripe Pauli setup is a fresh demonstration worth keeping. But the central claim outruns the evidence.

Who it is for: anyone working on AI-for-science discovery and, just as much, anyone thinking about how to evaluate claims that a network "rediscovered" a physical law. It is a useful case study in how easily consistency-on-training-data reads as discovery.

Recommendation: accept for peer review, with a referee letter that insists on held-out prediction for new potentials and initial conditions, seed-level variance, and code or data release. If those cannot be produced, the discovery claim should be scaled back.

Referee Report

3 major / 6 minor

Summary. This paper proposes a variational autoencoder (VAE) combined with a Neural ODE to jointly discover latent physical concepts and governing differential equations from simulated time-series or position-series data. The architecture maps observed data (either a single initial frame or a full trajectory) to a latent initial state, evolves it through a learned Neural ODE with optional external control variables, and decodes the resulting latent trajectory to reconstruct the observations. The method is applied to four historical physics examples: Copernican heliocentrism, Newtonian gravity, Schrödinger wave mechanics, and Pauli spin-magnetic coupling. For each example, the authors use an ablation study to select the latent dimension, then show that the learned latent representations and the Neural ODE output match linear combinations of the textbook concepts and the true equations, as measured by relative errors R_h and R_f reported in Table 1. The central claim, stated in the abstract, is that the correct physical theories can emerge in the neural network.

Significance. If validated, the work would be a notable step toward simultaneous, prior-free discovery of physical concepts and equations, extending SciNet with a continuous-dynamics model. The paper's strengths include the simplicity of the architecture, the use of an ablation-based Occam's-razor criterion for the number of concepts, and a candid discussion of limitations in Appendix E, including the failure to obtain independent concepts for the Pauli system. However, the evidence as presented does not yet establish the central claim: all quantitative metrics are computed on training data, and for two of the four examples the encoder receives the full trajectory, so the model is not demonstrated to be a forward-prediction model. The paper would be significantly strengthened by held-out tests on unseen potentials and initial conditions.

major comments (3)

[§3.3 and §3.4, Eqs. (33)–(34), Table 1] All reported relative errors R_h and R_f are computed over the training data, with no train/test split or held-out trajectories. In the Schrödinger and Pauli examples, the encoder input is the complete observed density trajectory (Section 3.3; observation input sizes 50 and 100 in Table 2), so the latent initial state h(t0) can depend on the entire trajectory rather than on the known initial conditions. Consequently, the model can act as a full-curve autoencoder with a Neural ODE bottleneck, and low training-set errors do not imply that the learned ODE plus the known initial conditions predicts data for new potential functions. The authors should hold out a portion of the potentials, initialize the latent state from the true initial conditions (or a short prefix of the trajectory), and report the prediction error for the remaining density sequence; this test is essential for the claim that correct physical theories have been discovered.
[§3.2, Eq. (23); also §3.3 Eq. (29) and §3.4 Eq. (36)] The expected governing functions f_j are constructed by substituting the true equations (e.g., Eq. (20)) into the derivative of the post-hoc linear fit of the latent representations (Eq. (22)). The comparison between the learned f_j and these expected expressions therefore measures in-sample consistency between the neural network and the linear fit applied to the same trajectories, not an independent prediction of the governing equations. To support the discovery claim, the authors should assess whether the learned Neural ODE, initialized from the known initial state and driven by V(x) or r0, produces accurate trajectories for held-out data without refitting the linear coefficients.
[§3.3, Eq. (27)] In both quantum examples, all training samples share identical initial conditions for the wave functions and their derivatives, and the only sample-specific input is the control variable V(x), which is fed directly into the Neural ODE. The model therefore cannot be said to have discovered the Schrödinger or Pauli equations from raw observations alone: the potential is a provided input. The authors should either test the model without V(x) as input or explicitly qualify the claim to state that the discovery is conditional on knowing the potential. They should also discuss whether this is consistent with the statement in the Introduction that the method does not rely on prior knowledge specific to particular examples.

minor comments (6)

[§3.4 and Table 1] The text states that R_f(50,4) is approximately 0.03, but Table 1 reports R_f(100,4) = 0.08; the discrepancy between these numbers should be resolved.
[§2, Eq. (12) and Appendix B, Eq. (B.3)] The loss notation is inconsistent: Eq. (12) writes the KL term as β·D_KL, while Eq. (B.3) has β/2 factored into the sum; the constants should be defined consistently.
[§3.1, Eq. (15)] The expression θ_m = atan2(sin θ_m, cos θ_m) is redundant and circular; it should be replaced by a direct definition of θ_m using the preceding formulas.
[Footnote 2] The additional Mean Relative Error regularization used only in the Newton example should be justified, as it is an example-specific modification that may affect the generality of the method.
[Appendix E vs. Abstract] The abstract and conclusion claim simultaneous discovery of concepts and equations across all examples, but Appendix E reports that for the Pauli system the model does not capture four independent concepts; the main text should acknowledge this limitation in the abstract or conclusion.
[General] The manuscript does not mention code or data availability; providing the code and simulated datasets would aid reproducibility.

Circularity Check

2 steps flagged · score 5.0 of 10

Evaluation is partly circular: success metrics are in-sample linear fits of the learned variables to textbook concepts, and for the wave-mechanics examples the encoder sees the full trajectory, so the claimed 'emergence' of predictive theories is not demonstrated outside the training curves.

fitted input called prediction [Section 3.4, Eqs. (33)-(34); Table 1; see also Eqs. (22)-(23)]
"while ˆhj(xi) denotes the predicted value obtained by fitting 4000 sample points. These predicted values are linear combinations of ψ1(x), dψ1(x)/dx, ψ2(x), and dψ2(x)/dx. Similarly, fj(xi) are the spatial derivatives of different latent representations, and ˆfj(xi) are their corresponding predicted values, which are linear combinations of dψ1(x)/dx, dψ2(x)/dx, d2ψ1(x)/dx2 and d2ψ2(x)/dx2."

The metrics Rh and Rf compare each learned latent component hj and governing function fj to ĥj and f̂j built by fitting the very same 4000 sample points onto the textbook basis. The 'predicted values' are therefore in-sample linear regression fits, not outputs of the trained model on unseen conditions. A small residual shows only that the learned variables lie close to the span of the chosen textbook concepts, which is a goodness-of-fit of the linear ansatz rather than an independent predictive test. In addition, the target f̂ uses the true right-hand sides of the governing equations (e.g., Eq. (23) uses Eq. (20)), so the comparison measures agreement with a linear fit of the learned vector field onto the true dynamics rather than a prediction of future data.
other [Sections 3.3-3.4; Table 2 (observation input sizes 50 and 100)]
"The input and label data are the trajectories of all 2000 sets of density distribution ρ(x) observed in the sequential 50 positions. ... we further compute the corresponding 4,000 probability density trajectories ρ(x), which serve as both the input and label for our neural network model."

For the Schrödinger and Pauli examples, the encoder receives the complete observed trajectory, so the initial latent state h(t0) can encode information about the entire curve, including the specific potential V(x). The Neural ODE is then trained only to reconstruct that same already-encoded trajectory. Consequently, the reported low reconstruction errors and the Rh/Rf values are in-sample autoencoding results. They do not demonstrate that the model predicts the trajectory for a new potential from the fixed initial conditions and the control variable alone, because trajectory-specific information could be absorbed into h(t0). The claim that 'the correct physical theories can emerge' is thus partly an autoencoding consistency check rather than evidence of an emergent predictive theory.

full rationale

The paper does not rely on load-bearing self-citation, imported uniqueness theorems, or ansatz-smuggling via citation; the cited prior works set up the examples but the VAE+Neural ODE combination is trained with an unsupervised reconstruction loss, so the latent representations and learned vector fields are not directly supervised with the textbook equations. The central derivation therefore has independent content. However, the evaluation of the central claim is partially circular in two concrete ways. First, the reported metrics Rh and Rf are computed by linearly fitting the learned latent variables and vector fields to the textbook concepts on the same 4000 training samples, so the 'predicted values' in Eqs. (33)-(34) are fitted values rather than forecasts; low errors quantify in-sample linearity, not predictive discovery. Second, in the wave-mechanics and spin examples the encoder input is the full observed trajectory, making the model a full-curve autoencoder with a Neural ODE bottleneck; no held-out potential is ever tested, so the claim that correct physical theories 'emerge' is not shown to generalize beyond the training curves. These issues make the strongest claim substantially weaker than advertised, but they do not make the entire derivation equivalent to its inputs by definition, because the network could have failed to produce linear textbook-compatible representations even in-sample. A score of 5 reflects partial circularity in the evaluation and prediction claims rather than complete definitional circularity.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The paper's central claim is supported mainly by architectural choices (ODEs, latent dimension, β-VAE) and a post-hoc linear interpretation, all of which are inputs to the analysis rather than outputs of discovery. The cost is that the model cannot distinguish between a genuine physical concept and an arbitrary linear reparameterization of it.

free parameters (6)

Latent space dimension = 2, 2, 2, 4 for Copernicus, Newton, Schrödinger, Pauli
Chosen by ablation on the loss curves (Figs. 2b, 3c, 4b, 5c) using a subjective threshold for 'no significant increase in loss'; the dimension is equated with the number of physical concepts.
KL weight β = 0.01, 0, 0.001, 0.0001, 0.1 (see Table 3)
Hand-tuned per example; controls the disentanglement pressure and can be zero (Newton), so the stated mechanism is not active in all demonstrations.
Prior standard deviation σ_h = 0.1
Set in Appendix B and used in the KL term; affects the magnitude of latent regularization.
Post-hoc linear fit coefficients = a_j, b_j, c_j (and d_j, e_j) fitted per latent dimension
Used to claim latent variables are linear combinations of true concepts (Eqs. 22, 28, 35); the same fitted coefficients define the expected governing functions in the success metric.
Random potential ranges and magnetic field = V in (-3,0) or [-2,-1.5]; B=1
Data-generation choices (Appendix C, D); the model's difficulty and the Pauli result depend on these ranges.
Initial quantum state = ψ(0)=1, ψ'(0)=1 for all samples
In Sections 3.3 and 3.4; shared initial conditions mean the encoder carries no sample-specific information in the quantum examples.

assumptions (6)

domain assumption The observed phenomena can be described by ordinary differential equations, and the latent dynamics can be written as an autonomous ODE.
Section 2, Eq. (18); this rules out non-ODE systems and assumes a finite-dimensional latent state.
domain assumption A neural network can represent the true governing function f with sufficient accuracy.
Used throughout; no expressivity or convergence guarantee is given.
ad hoc to paper The smallest latent dimension with no significant loss increase equals the number of physical concepts.
Section 2, 'Number of physical concepts'; this is an Occam heuristic, not a proven identifiability result.
ad hoc to paper Learned latent representations are interpretable as linear combinations of true physical concepts.
Post-hoc ansatz in Eqs. (22), (28), (35); nonlinear decodings are not considered.
ad hoc to paper The true equations can be reorganized as second-order equations when trying to recover independent concepts.
Appendix E; the authors call this a universal constraint, but it is imposed externally and fails to produce independent concepts for the Pauli case.
domain assumption The β-VAE KL term promotes disentanglement of latent variables.
Section 2 and Appendix B; empirical property from prior work, not proven for this architecture with ε=0.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Discover physical concepts and equations with machine learning." pith.science (2026). https://pith.science/paper/FEQTEWZY

@misc{pith2026241212161,
  author       = {Pith},
  title        = {Pith review of: Discover physical concepts and equations with machine learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FEQTEWZY}},
  note         = {Machine review of arXiv:2412.12161}
}

read the original abstract

Machine learning can uncover physical concepts or physical equations when prior knowledge from the other is available. However, these two aspects are often intertwined and cannot be discovered independently. We extend SciNet, which is a neural network architecture that simulates the human physical reasoning process for physics discovery, by proposing a model that combines Variational Autoencoders (VAE) with Neural Ordinary Differential Equations (Neural ODEs). This allows us to simultaneously discover physical concepts and governing equations from simulated experimental data across various physical systems. We apply the model to several examples inspired by the history of physics, including Copernicus' heliocentrism, Newton's law of gravity, Schr\"odinger's wave mechanics, and Pauli's spin-magnetic formulation. The results demonstrate that the correct physical theories can emerge in the neural network.

Figures

Figures reproduced from arXiv: 2412.12161 by the authors.

**Figure 2.** Heliocentric solar system. (a) The φe and φm denote the heliocentric angles of Earth and Mars, while θs and θm represent the geocentric angles of the Sun and Mars. These angles are measured against a background of fixed stars. Both the Earth and Mars move in uniform circular motion around the Sun. (b) Impact of latent dimensions on loss, showing a two-dimensional space as optimal. (c) The latent representations capt… view at source ↗

**Figure 3.** The law of gravity. (a) Different types of motion trajectories determined by r0. (b) Distances r(t) from the central object with different r0. (c) Impact of latent dimensions on loss, showing that a twodimensional space is optimal. (d) The latent representations store linear combinations of r and dr/dt at the 20th (1), 50th (2), and 80th (3) moments. (e) The governing functions learned by Neural ODEs are linear com… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Wave function and Schr¨odinger equation. (a) Random potential and corresponding density distribution. (b) Impact of latent dimensions on loss, showing a two-dimensional space as optimal. (c) The latent representations store linear combinations of ψ(x) and dψ(x)/dx at t…

**Figure 5.** Figure 5: Spin and Pauli equation. (a) Silver atoms in uniform (pink) and non-uniform (green) magnetic fields. (b) Random potential and corresponding density distribution. (c) Impact of latent dimensions on loss, showing a four-dimensional space as optimal. (d) Relative error me…

**Figure 6.** Figure 6: Law of gravity with the second-order assumption. (a) Two learned latent representations at the 20th (1), 50th (2), and 80th (3) moments. (b) The governing functions learned by Neural ODEs at the 20th (1), 50th (2), and 80th (3) moments. Here h1 and h2 are the two laten…

**Figure 7.** Figure 7: The Schr¨odinger system with the second-order assumption. (a) Two learned latent representations at the 10th (1), 25th (2), and 40th (3) positions. (b) The governing functions learned by Neural ODEs at the 10th (1), 25th (2), and 40th (3) positions. By comparing [PIT…

**Figure 8.** Figure 8: The Pauli system with the second-order assumption. (a) Relative error metric between the learned latent representations and the linear combinations of the four physical concepts. (b) Relative error metric between the learned governing functions and the linear combinati…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 40 canonical work pages

[1]

Einstein, On the method of theoretical physics , Philos

A. Einstein, On the method of theoretical physics , Philos. Sci. 1, 163 (1934)

work page 1934
[2]

Maillet, Heisenberg spin chains: from quantum groups to neutron scattering experiments, in Quantum Spaces, 161 (2007)

J.-M. Maillet, Heisenberg spin chains: from quantum groups to neutron scattering experiments, in Quantum Spaces, 161 (2007)

work page 2007
[3]

Wang et al

H. Wang et al. , Scientific discovery in the age of artificial intelligence , Nature 620, 47 (2023)

work page 2023
[4]

Carleo et al

G. Carleo et al. , Machine learning and the physical sciences , Rev. Mod. Phys. 91, 045002 (2019). [arXiv:1903.10563]

arXiv 2019
[5]

Wu and M

T. Wu and M. Tegmark, Toward an artificial intelligence physicist for unsupervised learning, Phys. Rev. E 100, 033311 (2019). [arXiv:1810.10525]

arXiv 2019
[6]

Cornelio al., Combining data and theory for derivable scientific discovery with AI-Descartes, Nat

C. Cornelio al., Combining data and theory for derivable scientific discovery with AI-Descartes, Nat. Commun. 14, 1777 (2023). [arXiv:2109.01634]

arXiv 2023
[7]

Krenn et al., On scientific understanding with artificial intelligence , Nat

M. Krenn et al., On scientific understanding with artificial intelligence , Nat. Rev. Phys. 4, 761 (2022). [arXiv:2204.01467]

arXiv 2022
[8]

Evolving Scientific Discovery by Unifying Data and Background Knowledge with AI Hilbert

R. Cory-Wright et al., Evolving scientific discovery by unifying data and background knowledge with AI Hilbert , Nat. Commun. 15, 5922 (2024). [arXiv:2308.09474]

work page Pith review arXiv 2024

Show all 63 references

[9]

Iten et al., Discovering physical concepts with neural networks , Phys

R. Iten et al., Discovering physical concepts with neural networks , Phys. Rev. Lett. 124, 010508 (2020). [arXiv:1807.10300] 30

2020 arXiv
[10]

C. Wang, H. Zhai, and Y.-Z. You, Emergent Schr¨ odinger equation in an introspective machine learning architecture, Sci. Bull. 64, 1228 (2019). [arXiv:1901.11103]

2019 arXiv
[11]

A. Seif, M. Hafezi, and C. Jarzynski, Machine learning the thermodynamic arrow of time, Nat. Phys. 17, 105 (2021). [arXiv:1909.12380]

2021 arXiv
[12]

Desai, B

K. Desai, B. Nachman, and J. Thaler, Symmetry discovery with deep learning , Phys. Rev. D 105, 096031 (2022). [arXiv:2112.05722]

2022 arXiv
[13]

S. E. Otto, N. Zolman, J. N. Kutz, and S. L. Brunton, A unified framework to enforce, discover, and promote symmetry in machine learning , arXiv:2311.00212

arXiv
[14]

Lemos et al., Rediscovering orbital mechanics with machine learning, Mach

P. Lemos et al., Rediscovering orbital mechanics with machine learning, Mach. Learn.: Sci. Technol. 4, 045002 (2023). [arXiv:2202.02306]

2023 arXiv
[15]

Liu and M

Z. Liu and M. Tegmark, Machine learning conservation laws from trajectories , Phys. Rev. Lett. 126, 180604 (2021). [arXiv:2011.04698]

2021 arXiv
[16]

Alet et al., Noether networks: Meta-learning useful conserved quantities , in Advances in Neural Information Processing Systems 35 (NeurIPS 2021)

F. Alet et al., Noether networks: Meta-learning useful conserved quantities , in Advances in Neural Information Processing Systems 35 (NeurIPS 2021) . [arXiv:2112.03321]

2021 arXiv
[17]

Z. Liu, V. Madhavan, and M. Tegmark, Machine learning conservation laws from differential equations, Phys. Rev. E 106, 045307 (2022). [arXiv:2203.12610]

2022 arXiv
[18]

Einstein, Science and Religion , Nature 146, 605 (1940)

A. Einstein, Science and Religion , Nature 146, 605 (1940)

1940
[19]

Schr¨ odinger,An undulatory theory of the mechanics of atoms and molecules , Phys

E. Schr¨ odinger,An undulatory theory of the mechanics of atoms and molecules , Phys. Rev. 28, 1049 (1926)

1926
[20]

R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud, Neural ordi- nary differential equations, in Advances in Neural Information Processing Systems 31 (NeurIPS 2018) . [arXiv:1806.07366]

2018 arXiv
[21]

M. Choi, D. Flam-Shepherd, T. H. Kyaw, and ´A. Aspuru-Guzik, Learning quantum dynamics with latent neural ordinary differential equations, Phys. Rev. A 105, 042403 (2022). [arXiv:2110.10721]

2022 arXiv
[22]

Z. Lai et al., Neural modal ordinary differential equations: Integrating physics-based modeling with neural ordinary differential equations for modeling high-dimensional monitored structures, Data-Centric Engineering 3, e34 (2022). [arXiv:2207.07883] 31

2022 arXiv
[23]

Sholokhov, Y

A. Sholokhov, Y. Liu, H. Mansour, and S. Nabi, Physics-informed neural ODE (PIN- ODE): Embedding physics into models using collocation points , Sci. Rep. 13, 10166 (2023)

2023
[24]

B. C. Daniels and I. Nemenman, Automated adaptive inference of phenomenological dynamical models, Nat. Commun. 6, 8133 (2015). [arXiv:1404.6283]

2015 arXiv
[25]

D. P. Kingma and M. Welling, Auto-encoding variational Bayes , ICLR, (2014). [arXiv:1312.6114]

2014 arXiv
[26]

Higgins et al., Beta-V AE: Learning basic visual concepts with a constrained varia- tional framework, ICLR, (2017)

I. Higgins et al., Beta-V AE: Learning basic visual concepts with a constrained varia- tional framework, ICLR, (2017)

2017
[27]

Frohnert and E

F. Frohnert and E. van Nieuwenburg, Explainable representation learning of small quantum states , Mach. Learn.: Sci. Technol. 5, 015001 (2024). [arXiv:2306.05694]

2024 arXiv
[28]

Fern´ andez-Fern´ andez et al., Learning minimal representations of stochas- tic processes with variational autoencoders , Phys

G. Fern´ andez-Fern´ andez et al., Learning minimal representations of stochas- tic processes with variational autoencoders , Phys. Rev. E 110, L012102 (2024). [arXiv:2307.11608]

2024
[29]

K. Cho et al., Learning phrase representations using RNN encoder-decoder for statisti- cal machine translation, in Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, 1724 (2014). [arXiv:1406.1078]

2014 arXiv
[30]

S. L. Brunton, J. L. Proctor, and J. N. Kutz, Discovering governing equations from data by sparse identification of nonlinear dynamical systems , Proc. Natl. Acad. Sci. 113, 3932 (2016). [arXiv:1509.03580]

2016 arXiv
[31]

S. L. Brunton, M. Budiˇ si´ c, E. Kaiser, and J. N. Kutz, Modern Koopman theory for dynamical systems, SIAM Rev 64, 229 (2022). [arXiv:2102.12086]

2022 arXiv
[32]

Hashimoto, H.-Y

K. Hashimoto, H.-Y. Hu, and Y.-Z. You, Neural ordinary differential equation and holographic quantum chromodynamics, Mach. Learn.: Sci. Technol. 2, 035011 (2021). [arXiv:2006.00712]

2021 arXiv
[33]

Rrapaj, A

E. Rrapaj, A. V. Patwardhan, E. Armstrong, and G. M. Fuller, Inference of neutrino flavor evolution through data assimilation and neural differential equations, Phys. Rev. D 103, 043006 (2021). [arXiv:2010.00695]

2021 arXiv
[34]

Chen et al., Forecasting the outcome of spintronic experiments with neural ordinary differential equations, Nat

X. Chen et al., Forecasting the outcome of spintronic experiments with neural ordinary differential equations, Nat. Commun. 13, 1016 (2022). 32

2022
[35]

Chen and Y

L. Chen and Y. Wu, Learning quantum dissipation by the neural ordinary differential equation, Phys. Rev. A 106, 022201 (2022). [arXiv:2207.09056]

2022 arXiv
[36]

Zhao et al., Optical neural ordinary differential equations, Opt

Y. Zhao et al., Optical neural ordinary differential equations, Opt. Lett. 48, 628 (2023). [arXiv:2209.12898]

2023 arXiv
[37]

Q. Li, T. Wang, V. Roychowdhury, and M. K. Jawed, Metalearning generalizable dynamics from trajectories, Phys. Rev. Lett. 131, 067301 (2023). [arXiv:2301.00957]

2023 arXiv
[38]

Liu and W

Z. Liu and W. M. Stacey, Application of Neural Ordinary Differential Equations for Tokamak Plasma Dynamics Analysis , ICLR, (2024). [arXiv:2403.01635]

2024 arXiv
[39]

Gu, Y.-K

Z.-F. Gu, Y.-K. Yan, and S.-F. Wu, Neural ODEs for holographic transport models without translation symmetry , Eur. Phys. J. C 85, 63 (2025). [arXiv:2401.09946]

2025 arXiv
[40]

Zhi et al., Learning efficient and robust ordinary differential equations via invert- ible neural networks , ICLR, (2022)

W. Zhi et al., Learning efficient and robust ordinary differential equations via invert- ible neural networks , ICLR, (2022). [arXiv:2107.01650]

2022 arXiv
[41]

J. E. Lye et al., Bose-Einstein condensate in a random potential , Phys. Rev. Lett. 95, 070401 (2005)

2005
[42]

classically non-describable two-valuedness

D. Giulini, Electron spin or “classically non-describable two-valuedness” , Stud. Hist. Philos. Sci. B 39, 557 (2008). [arXiv:0710.3128]

2008 arXiv
[43]

Pauli, Zur Quantenmechanik des magnetischen Elektrons , Z

W. Pauli, Zur Quantenmechanik des magnetischen Elektrons , Z. Physik 43, 601 (1927)

1927
[44]

R. A. Horn and C. R. Johnson, Matrix Analysis. (Cambridge University Press, 1990)

1990
[45]

W.-j. Hsu, D. Dutta, and A. Helmy, Structural Analysis of User Association Patterns in Wireless LAN, in IEEE Transactions on Mobile Computing 11, 1734 (2012). [arXiv: cs/0606002]

2012 arXiv
[46]

Morrill et al., Neural CDEs for long time series via the log-ODE method , ICLR, (2021)

J. Morrill et al., Neural CDEs for long time series via the log-ODE method , ICLR, (2021). [arXiv:2009.08295v1]

2021 arXiv
[47]

Iakovlev, C

V. Iakovlev, C. Yildiz, M. Heinonen, and H. L¨ ahdesm¨ aki,Latent neural ODEs with sparse Bayesian multiple shooting , ICLR, (2023). [arXiv:2210.03466]

2023 arXiv
[48]

Finlay, J.-H

C. Finlay, J.-H. Jacobsen, L. Nurbekyan, and A. Oberman, How to train your neural ODE: The world of Jacobian and kinetic regularization , in Proceedings of the 37th International Conference on Machine Learning , 3154 (2020). [arXiv:2002.02798] 33

2020 arXiv
[49]

Daulbaev et al., Interpolation technique to speed up gradients propagation in neural ODEs, in Advances in Neural Information Processing Systems 34 (NeurIPS 2020)

T. Daulbaev et al., Interpolation technique to speed up gradients propagation in neural ODEs, in Advances in Neural Information Processing Systems 34 (NeurIPS 2020) . [arXiv:2003.05271]

2020 arXiv
[50]

Ghosh et al., STEER: Simple temporal regularization for neural ODE , in Advances in Neural Information Processing Systems 34 (NeurIPS 2020)

A. Ghosh et al., STEER: Simple temporal regularization for neural ODE , in Advances in Neural Information Processing Systems 34 (NeurIPS 2020) . [arXiv:2006.10711]

2020 arXiv
[51]

Kelly, J

J. Kelly, J. Bettencourt, M. J. Johnson, and D. K. Duvenaud, Learning differential equations that are easy to solve, in Advances in Neural Information Processing Systems 34 (NeurIPS 2020) . [arXiv:2007.04504]

2020 arXiv
[52]

Hey, that’s not an ODE

P. Kidger, R. T. Q. Chen, and T. J. Lyons, “Hey, that’s not an ODE”: Faster ODE adjoints via seminorms , in Proceedings of the 38th International Conference on Machine Learning, 5443 (2021). [arXiv:2009.09457]

2021 arXiv
[53]

Xia et al., Heavy ball neural ordinary differential equations , in Advances in Neural Information Processing Systems 35 (NeurIPS 2021)

H. Xia et al., Heavy ball neural ordinary differential equations , in Advances in Neural Information Processing Systems 35 (NeurIPS 2021) . [arXiv:2110.04840]

2021 arXiv
[54]

H. H. N. Nguyen, T. Nguyen, H. Vo, S. Osher, and T. Vo, Improving neural ordi- nary differential equations with Nesterov’s accelerated gradient method , in Advances in Neural Information Processing Systems 36 (NeurIPS 2022)

2022
[55]

Zhao et al., Accelerating Neural ODEs: A variational formulation-based approach , ICLR, (2025)

H. Zhao et al., Accelerating Neural ODEs: A variational formulation-based approach , ICLR, (2025)

2025
[56]

Course and P

K. Course and P. B. Nair, State estimation of a physical system with unknown gov- erning equations, Nature 622, 261 (2023)

2023
[57]

H. P. Nautrup et al., Operationally meaningful representations of physical systems in neural networks, Mach. Learn.: Sci. Technol. 3, 045025 (2022). [arXiv:2001.00593]

2022 arXiv
[58]

Shojaeeet al., LLM-SR: Scientific equation discovery via programming with large language models, ICLR, (2024)

P. Shojaeeet al., LLM-SR: Scientific equation discovery via programming with large language models, ICLR, (2024). [arXiv:2404.18400]

2024 arXiv
[59]

Wilczek, The Dirac Equation , International Journal of Modern Physics A , (2004)

F. Wilczek, The Dirac Equation , International Journal of Modern Physics A , (2004)

2004
[60]

K. He, X. Zhang, S. Ren, and J. Sun, Delving deep into rectifiers: Surpassing human- level performance on ImageNet classification , in Proceedings of the 2015 IEEE Inter- national Conference on Computer Vision , 1026 (2015). [arXiv:1502.01852] 34

2015 arXiv
[61]

Tsitouras, Runge–Kutta pairs of order 5(4) satisfying only the first column simpli- fying assumption , Comput

C. Tsitouras, Runge–Kutta pairs of order 5(4) satisfying only the first column simpli- fying assumption , Comput. Math. Appl. 62, 770 (2011)

2011
[62]

L. D. Landau and E. M. Lifshitz, Quantum Mechanics: Non-Relativistic Theory , 3rd ed. (Butterworth-Heinemann, Oxford, 1977)

1977
[63]

Basdevant, Lectures on Quantum Mechanics

J.-L. Basdevant, Lectures on Quantum Mechanics . (New York, USA: Springer Sci- ence+Business Media, LLC, 2007). 35

2007

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Einstein, On the method of theoretical physics , Philos

A. Einstein, On the method of theoretical physics , Philos. Sci. 1, 163 (1934)

work page 1934

[2] [2]

Maillet, Heisenberg spin chains: from quantum groups to neutron scattering experiments, in Quantum Spaces, 161 (2007)

J.-M. Maillet, Heisenberg spin chains: from quantum groups to neutron scattering experiments, in Quantum Spaces, 161 (2007)

work page 2007

[3] [3]

Wang et al

H. Wang et al. , Scientific discovery in the age of artificial intelligence , Nature 620, 47 (2023)

work page 2023

[4] [4]

Carleo et al

G. Carleo et al. , Machine learning and the physical sciences , Rev. Mod. Phys. 91, 045002 (2019). [arXiv:1903.10563]

arXiv 2019

[5] [5]

Wu and M

T. Wu and M. Tegmark, Toward an artificial intelligence physicist for unsupervised learning, Phys. Rev. E 100, 033311 (2019). [arXiv:1810.10525]

arXiv 2019

[6] [6]

Cornelio al., Combining data and theory for derivable scientific discovery with AI-Descartes, Nat

C. Cornelio al., Combining data and theory for derivable scientific discovery with AI-Descartes, Nat. Commun. 14, 1777 (2023). [arXiv:2109.01634]

arXiv 2023

[7] [7]

Krenn et al., On scientific understanding with artificial intelligence , Nat

M. Krenn et al., On scientific understanding with artificial intelligence , Nat. Rev. Phys. 4, 761 (2022). [arXiv:2204.01467]

arXiv 2022

[8] [8]

Evolving Scientific Discovery by Unifying Data and Background Knowledge with AI Hilbert

R. Cory-Wright et al., Evolving scientific discovery by unifying data and background knowledge with AI Hilbert , Nat. Commun. 15, 5922 (2024). [arXiv:2308.09474]

work page Pith review arXiv 2024