Pith. sign in

REVIEW 3 major objections 4 minor 67 references

Hierarchical Physics-Embedded Learning for Partially Known Spatiotemporal Dynamics

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read Partially known PDEs can be modeled with a two-level Fourier architecture that encodes known terms and learns the rest, cutting long-horizon errors by up to 70%.

desk verdict A genuinely new architecture with strong synthetic benchmarks, but the paper's advertised theory and hydrofoil validation are missing from the manuscript, so it needs major revision before I'd trust the claims. read the letter →

arxiv 2510.25306 v3 pith:JEB5PN6W submitted 2025-10-29 cs.LG

Xizhe Wang , Xiaobin Song , Hongbo Zhao , Qingshan Jia , Qianchuan Zhao , Hao Sun , Benben Jiang This is my paper
classification cs.LG
keywords partialdifferentialequationsphysics-embeddedlearningFourierneuraloperatorhierarchicalarchitectureCahn-Hilliardequationsymbolicregressionspatiotemporaldynamics
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 tries to establish that when a governing PDE is only partially known—some constitutive relations or combinations are known, others not—that knowledge can be built directly into a two-level neural architecture rather than being penalized or appended. At the first level the model learns or hard-embeds the fundamental symbolic pieces of the equation; at the second it learns how those pieces combine. The paper proves a hierarchical error decomposition and a parameter-complexity advantage: with the right hierarchy, the number of learnable Fourier parameters needed for a fixed accuracy grows strictly more slowly than for a single-level operator. On phase-field and interface-growth systems it reports up to roughly 70% lower long-horizon extrapolation error than baselines, and it recovers closed forms of unknown concentration-dependent diffusivity and chemical potential from sparse, noisy data. A sympathetic reader would care because many real systems—battery electrodes, phase-separating materials—have exactly this partial-knowledge structure.

What carries the argument

The carrying mechanism is the two-level AFNO stack with physics-embedding channels. The first-level AFNO module outputs multiple channels, each assigned to a distinct unknown physical term; known terms are computed by separate channels that bypass learning. The second-level AFNO module learns the combination operator, such as the divergence of a mobility times the gradient of chemical potential. A Fourier attention layer—discrete Fourier transform, frequency-domain channel mixing with a block-diagonal MLP weight matrix, and inverse transform—gives a global receptive field and is proven in the paper to approximate arbitrary differential operators. A Gaussian-kernel consistency mapping forces

What would settle it

Take a trained two-channel Cahn-Hilliard model and record the two channel outputs on held-out snapshots together with the ground-truth values of D(c) and μ_hom(c). If either channel's output correlates significantly with a linear combination of the two true terms rather than with its assigned term alone—or if a linear re-mixing of the two channels fits the data equally well—the claimed structural decoupling, and with it the closed-form recovery pipeline, is not established.

Watch

Extended reading notes

Core claim

The central claim is that a hierarchical physics-embedded adaptive Fourier neural operator (AFNO) can model partially known spatiotemporal dynamics by mirroring the compositional structure of the PDE: a first level extracts or embeds fundamental physical expressions (like mobility, homogeneous chemical potential, and gradient-energy terms), and a second level learns their governing combination, with Fourier attention layers at each level capturing nonlocal and high-order couplings. The paper argues that this decomposition provably shrinks or removes error terms for embedded known components, and that when the hierarchy matches the dynamics' composition, the number of learnable Fourier parame

Load-bearing premise

The load-bearing premise is that after training, each AFNO channel assigned to a distinct unknown physical term actually represents that term and not an arbitrary linear combination that also fits the data; if channel mixing occurs, the binning and symbolic-regression steps would recover misleading closed forms even when forward predictions are accurate.

Editorial extensions

If this is right

  • Embedding known physics as architecture, not as a soft loss, meaningfully improves long-horizon extrapolation (up to ~70% error reduction reported) while preserving physical morphology, energy behavior, and spectral structure.
  • The hierarchical error decomposition implies that adding known terms strictly removes or shrinks their contribution to the total approximation error, so partial knowledge monotonically helps when embedded correctly.
  • The parameter-complexity result gives a principled reason to structure models according to the PDE's compositional form: fewer Fourier parameters suffice at a given accuracy when the hierarchy is aligned.
  • The decoupled-channel design turns the network into an equation-discovery front end, recovering closed forms of D(c) and μ_hom(c) from sparse, noisy data without prespecified symbolic libraries.
  • Reported robustness under 20% noise and 80x coarser sampling suggests the method remains usable where measurement data are heavily degraded.

Reading between the lines

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

  • The same channel-decoupling trick should generalize to PDEs with more than two unknown constitutive terms, but only if the identifiability assumption is checked: nothing in the presented analysis rules out benign mixtures of terms in the learned channels.
  • The parameter-complexity advantage could be turned into a model-selection rule: estimate the compositional structure of an unknown PDE, choose the hierarchy accordingly, and expect fewer parameters when the guess is right.
  • Because the kernel-consistency mapping enforces dependence on the input concentration alone, applying this pipeline to experimental data with hidden auxiliary variables (temperature, stress, history) will require binning on those variables too—otherwise the recovered 'functions' may be projections of higher-dimensional relations.
  • The reported low-variance premature convergence in the no-physics setting hints that physics embedding acts as a strong regularizer; an explicit latent-entanglement metric before and after embedding would test that interpretation directly.
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 / 4 minor

Summary. The paper introduces HPE-AFNO, a two-level hierarchical neural operator for partially known spatiotemporal PDEs. The first level extracts fundamental symbolic components (e.g., diffusivity, chemical potential), the second learns their governing combination; known physics can be embedded as hard-wired channels. The authors report lower RMSE than PeRCNN and FNO-2D on Cahn–Hilliard, Allen–Cahn, deterministic KPZ, and complex Ginzburg–Landau dynamics, robustness to sparse/noisy data, and symbolic recovery of D(c) and mu_hom(c) via binning plus deep symbolic regression. The version submitted for review also claims a proof of hierarchical error decomposition and a parameter-complexity advantage, but no theorem or proof appears in the manuscript.

Significance. If the theoretical claims were substantiated, the architecture would be a meaningful contribution to physics-embedded operator learning: the two-level decomposition and hard embedding of known terms are conceptually appealing and the empirical comparisons, though limited to two baselines, are consistent across four benchmark systems. The robustness sweeps and the decoupled equation-discovery pipeline are useful additions. However, the central theoretical guarantees are currently unsupported, the abstract promises an experimental hydrofoil validation that is absent from the text, and the channel-identifiability assumption behind the symbolic discovery results is not justified. The paper also does not provide code or data, limiting reproducibility.

major comments (3)
  1. [§5.1, after Eq. (3)] The sentence 'We prove that this Fourier attention layer can approximate any differential operator (see “Fourier attention layer for universal approximation of differential operators” for detailed information)' points to a section that does not exist in the main text or the supplement. No theorem statement, assumption set, or proof of the claimed hierarchical error decomposition or of the parameter-complexity advantage appears anywhere. Since the abstract's central theoretical claim is a proven quantitative comparison of parameter growth, this is a load-bearing unsupported assertion, not a presentation issue. The authors must either provide the theorem and proof or remove the claim from the abstract and framing.
  2. [Abstract vs. main text] The abstract states validation on 'experimental hydrofoil wake data' and uses it as evidence of the method's real-world applicability. The full text, including the Supplementary Information, contains no hydrofoil experiment; the closest evidence is synthetic CH/AC/dKPZ/CGL data. The full-text abstract also omits the hydrofoil claim, so the two abstracts are inconsistent. If the hydrofoil experiment belongs to a different version, it must be restored or the abstract revised; as it stands, the manuscript promises a validation that is absent.
  3. [§5.2 and Fig. 4] The method 'structurally assign[s] each AFNO channel to a distinct unknown component,' but no mechanism enforces that a channel output equals the intended physical term rather than an arbitrary linear combination of terms. The kernel-consistency constraint of Eq. (4) only enforces continuity of each channel as a function of concentration; it does not establish identifiability. Without identifiability, the concentration-binning analysis and subsequent deep symbolic regression can recover misleading closed forms even when forward predictions are accurate. Please state the identifiability condition and provide an ablation (e.g., channel-permutation or mixing tests) demonstrating that the learned channels separate D(c) and mu_hom(c) rather than a rotation of them.
minor comments (4)
  1. [§2.1 and Abstract] The abstract describes 'state-of-the-art physics encoded and neural operator baselines,' but only PeRCNN and FNO-2D are compared. If the paper claims SOTA comparisons, additional recent baselines are needed; otherwise the wording should be tempered.
  2. [References] References [53]–[55] appear in the reference list but are not cited anywhere in the main text or supplement. Either cite them or remove them.
  3. [§5.1, Eq. (4)] The Gaussian kernel width sigma is a free hyperparameter, but no sensitivity analysis is reported. Since the kernel-based mapping is the basis for the continuity guarantee, the dependence of the results on sigma should be discussed or empirically examined.
  4. [Supplementary Table 1] The 'hard thresholding fraction' is set to 1.0 while the method description emphasizes sparsity via hard thresholding. Please clarify whether this setting disables the thresholding and how the reported sparsity threshold interacts with it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found; the theoretical proof is missing, but no claimed prediction reduces to fitted inputs by construction.

full rationale

The paper's central empirical results are benchmarked against externally generated ground-truth synthetic data and external baselines (PeRCNN, FNO-2D), so the reported error reductions and symbolic recovery results are not fitted quantities renamed as predictions. Known-term embedding is exact by construction, but this is an architectural feature rather than a circular step: the paper does not claim to 'predict' the embedded terms. The equation-discovery pipeline trains only on the final state loss and then interprets unsupervised AFNO channel outputs via binning and deep symbolic regression; recovering D(c) and mu_hom(c) is therefore not statistically forced by the loss, though identifiability is assumed rather than proven. The abstract's proof claims—hierarchical error decomposition and a strictly slower parameter-growth advantage—are not auditable because Methods §5.1 points to a 'Fourier attention layer for universal approximation of differential operators' section that does not appear in the manuscript or supplement. Likewise, the abstract promises experimental hydrofoil wake data that are absent from the reported experiments. These are missing-support/evidence gaps, not circular reasoning: there is no quoted equation or fitted parameter that reduces to the paper's own inputs. No load-bearing self-citation was found; the only author-overlapping citation (ref. 51) is used for a standard phase-field modeling choice and is not the basis of the paper's claimed theoretical or empirical contribution.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new particles, forces, or physical entities. Its load-bearing additions are the two-level composability assumption, the channel identifiability assumption, and the unproved universal-approximation/complexity claims. The only explicit fitted/knob-like numbers in the method are architectural hyperparameters and the Gaussian kernel width, none of which are rigorously justified.

free parameters (3)
  • Gaussian kernel width sigma = not specified
    Used in Eq. (4) of Methods §5.2 to enforce continuity of the learned unknown terms; no value or tuning procedure is given.
  • AFNO sparsity threshold and hard thresholding fraction = 0.01 and 1.0
    Supplementary Table 1; these manually chosen hyperparameters control frequency-domain sparsity and attention concentration, with no sensitivity analysis.
  • Forward Euler time step delta t = 0.01 s in the refined reconstruction
    Appears in Eq. (3); the recursive prediction and extrapolation error accumulation depend on this chosen discretization, which is set by the data generation rather than optimized or ablated.
assumptions (5)
  • ad hoc to paper Compositional separability of the target PDE: F(u) = Combine(Extract(u)), with low-dimensional Extract and Combine.
    The whole architecture in Methods §5.1 and Supplementary Note 1 assumes this two-level decomposition; the claimed parameter-complexity advantage depends on the hierarchy aligning with the true compositional structure.
  • domain assumption Channel-to-term identifiability: each AFNO channel can be assigned to a distinct unknown physical term.
    Methods §5.2 states that each AFNO channel learns a distinct unknown component, but no identifiability or disentanglement guarantee is provided.
  • ad hoc to paper Fourier attention layers can universally approximate differential operators.
    Asserted in Methods §5.1 with a pointer to a section that is not present in the supplied manuscript; no proof or citation to a specific external theorem is given.
  • domain assumption Forward Euler integration with delta t = 0.01 is sufficiently accurate for autoregressive prediction.
    Eq. (3) uses a forward Euler update; the accuracy of long-horizon recursive predictions depends on this discretization being adequate for all four test systems.
  • domain assumption The regular solution (Flory-Huggins) free-energy model with chi=3 and kappa=1 describes the Cahn-Hilliard data.
    Methods §5.3 adopts this model from prior literature; the CH experiments and the equation-discovery targets are defined by this choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Physics-Embedded Learning for Partially Known Spatiotemporal Dynamics." pith.science (2026). https://pith.science/paper/JEB5PN6W

@misc{pith2026251025306,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Physics-Embedded Learning for Partially Known Spatiotemporal Dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEB5PN6W}},
  note         = {Machine review of arXiv:2510.25306}
}
read the original abstract

Partial physical knowledge--governing structures known, constitutive relations or their combinations not--pervades spatiotemporal systems. Existing scientific machine learning paradigms learn evolution largely from data, impose equations as soft constraints, or hard-code physical terms into network updates; none exploits knowledge of this form. Here we introduce the hierarchical physics-embedded adaptive Fourier neural operator, encoding such knowledge as computational architecture rather than penalizing or appending it: a first level learns or embeds fundamental physical expressions as intermediate representations, and a second level learns or embeds their governing combination, with adaptive Fourier layers capturing nonlocal, high-order couplings at each level. We prove a hierarchical error decomposition--embedding known components removes or shrinks their terms, and a parameter-complexity advantage: when the hierarchy aligns with the compositional structure of the dynamics, the number of learnable Fourier parameters sufficient for a prescribed accuracy grows strictly more slowly than for a single-level operator. Across canonical phase-field systems and experimental hydrofoil wake data, our method reduces long horizon extrapolation errors by up to ~70% relative to state-of-the-art physics encoded and neural operator baselines, while preserving physically meaningful morphology, energetic consistency, and spectral structure, and maintaining robust performance under sparse and noisy observations. The separated intermediate representations further enable symbolic recovery of unknown constitutive relations in partially specified PDEs. These results establish hierarchical physics embedding as a theoretically grounded route to prediction and discovery when governing laws are neither fully known nor absent, but partially known and compositionally organized.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 1 linked inside Pith

  1. [1]

    Brunton, S. L. & Kutz, J. N. Promising directions of machine learning for partial differential equations. Nat. Comput. Sci. 4, 483–494 (2024)

    2024

  2. [2]

    Pestourie, R. et al. Physics -enhanced deep surrogates for partial differential equations. Nat. Mach. Intell. 5, 1458–1465 (2023)

    2023

  3. [3]

    Meigel, F. J. et al. Dispersive transport dynamics in porous media emerge from local correlations. Nat. Commun. 13, 5885 (2022)

    2022

  4. [4]

    Otsuji, M. et al. A mass conserved reaction -diffusion system captures properties of cell polarity. PLoS Comput. Biol. 3, e108 (2007)

    2007

  5. [5]

    Zhao, H. et al. Learning heterogeneous reaction kinetics from X -ray videos pixel by pixel. Nature 621, 289–294 (2023)

    2023

  6. [6]

    & Frey, E

    Halatek, J. & Frey, E. Rethinking pattern formation in reaction –diffusion systems. Nat. Phys. 14, 507–514 (2018)

    2018

  7. [7]

    Burkart, T. et al. Control of protein -based pattern formation via guiding cues. Nat. Rev. Phys. 4, 511–527 (2022)

    2022

  8. [8]

    P-SV wave propagation in heterogeneous media: velocity-stress finite-difference method

    Virieux, J. P-SV wave propagation in heterogeneous media: velocity-stress finite-difference method. Geophysics 51, 889–901 (1986)

    1986

Show all 67 references
  1. [9]

    & Vese, A

    Chan, T. & Vese, A. Active contours without edges. IEEE Trans. Image Process. 10, 266–277 (2001)

    2001

  2. [10]

    B., Brantingham, P

    Short, M. B., Brantingham, P. J., Bertozzi, A. L. & Tita, G. E. Dissipation and displacement of hotspots in reaction-diffusion models of crime. Proc. Natl Acad. Sci. 107, 3961–3965 (2010)

    2010

  3. [11]

    Smith, G. D. Numerical Solution of Partial Differential Equations: Finite Difference Methods (Oxford Univ. Press, 1985). 28

    1985

  4. [12]

    Felippa, C. A. Introduction to finite element methods. Univ. of Colorado, 885 (2004)

    2004

  5. [13]

    Bazant, M. Z. Theory of chemical kinet ics and charge transfer based on nonequilibrium thermodynamics. Acc. Chem. Res. 46, 1144–1160 (2013)

    2013

  6. [14]

    Pope, S. B. Turbulent flows (Cambridge Univ. Press, 2000)

    2000

  7. [15]

    Davidson, P. A. Turbulence: An Introduction for Scientists and Engineers (Oxford Univ. Press, 2015)

    2015

  8. [16]

    Groot, S. R. D. & Mazur, P. Non-Equilibrium Thermodynamics (Courier Corporation, 2013)

    2013

  9. [17]

    Non-equilibrium statistical mechanics (Courier Dover Publications, 2017)

    Prigogine, I. Non-equilibrium statistical mechanics (Courier Dover Publications, 2017)

    2017

  10. [18]

    Cross, M. C. & Hohenberg, P. C. Pattern formation o utside of equilibrium. Rev. Mod. Phys. 65, 851–1112 (1993)

    1993

  11. [19]

    Lectures on phase transitions and the renormalization group (CRC Press, 2018)

    Goldenfeld, N. Lectures on phase transitions and the renormalization group (CRC Press, 2018)

    2018

  12. [20]

    Cahn, J. W. & Hilliard, J. E. Free energy of a nonuniform system. I. Interfacial free energy. J. Chem. Phys. 28, 258–267 (1958)

    1958

  13. [21]

    Cahn, J. W. Phase separation by spinodal decomposition in isotropic systems. J. Chem. Phys. 42, 93–99 (1965)

    1965

  14. [22]

    Aranson, I. S. & Kramer, L. The world of the complex Ginzburg–Landau equation. Rev. Mod. Phys. 74, 99–143 (2002)

    2002

  15. [23]

    Schmid, P. J. Dynamic mode decomposition of numerical and experimental data. J. Fluid Mech. 656, 5–28 (2010)

    2010

  16. [24]

    N., Brunton, S

    Kutz, J. N., Brunton, S. L., Brunton, B. W. & Proctor, J. L. Dynamic Mode Decomposition: Data- Driven Modeling of Complex Systems (SIAM, 2016)

    2016

  17. [25]

    Schmid, P. J. Dynamic mode decomposition and its variants. Annu. Rev. Fluid Mech. 54, 225-254 (2022)

    2022

  18. [26]

    & Weinan, E

    Han, J., Je ntzen, A. & Weinan, E. Solving high -dimensional partial differential equations using 29 deep learning. Proc. Natl Acad. Sci. 115, 8505–8510 (2018)

    2018

  19. [27]

    & Dong, B

    Long, Z., Lu, Y ., Ma, X. & Dong, B. PDE -Net: learning PDEs from data. In International Conference on Machine Learning 3208–3216 (2018)

    2018

  20. [28]

    & Brunton, S

    Vinuesa, R. & Brunton, S. L. Enhancing computational fluid dynamics with machine learning. Nat. Comput. Sci. 2, 358–366 (2022)

    2022

  21. [29]

    R., Siboni, N

    Mianroodi, J. R., Siboni, N. H. & Raabe, D. Teaching solid mechanics to Artificial Intelligence— a fast solver for heterogeneous materials. Npj Comput. Mater. 7, 99 (2021)

    2021

  22. [30]

    L., Proctor, J

    Brunton, S. L., Proctor, J. L. & Kutz, J. N. Discovering governing equations from data by spa rse identification of nonlinear dynamical systems. Proc. Natl Acad. Sci. 113, 3932–3937 (2016)

    2016

  23. [31]

    & Karniadakis, G

    Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378, 686–707 (2019)

    2019

  24. [32]

    Karniadakis, G. E. et al. Physics-informed machine learning. Nat. Rev. Phys. 3, 422–440 (2021)

    2021

  25. [33]

    Cuomo, S. et al. Scientific machine learning through physics-informed neural networks: where we are and what’s next. J. Sci. Comput. 92, 88 (2022)

    2022

  26. [34]

    & Karniadakis, G

    Cai, S., Mao, Z., Wang, Z., Yin, M. & Karniadakis, G. E. Physics -informed neural networks (PINNs) for fluid mechanics: a review. Acta Mech. Sin. 37, 1727–1738 (2021)

    2021

  27. [35]

    & Luo , T

    Zhou, J., Li, R. & Luo , T. Physics-informed neural networks for solving time -dependent mode- resolved phonon Boltzmann transport equation. npj Comput. Mater. 9, 212 (2023)

    2023

  28. [36]

    Wang, F. et al. Physics -informed neural network for lithium -ion battery degradation stable modeling and prognosis. Nat. Commun. 15, 4332 (2024)

    2024

  29. [37]

    Ren, P., Rao, C., Liu, Y ., Wang, J. X. & Sun, H. PhyCRNet: Physics -informed convolutional - recurrent network for solving spatiotemporal PDEs. Comput. Methods Appl. Mech. Eng. 389, 114399 (2022). 30

    2022

  30. [38]

    Rao, C. et al. Encoding physics to learn reaction-diffusion processes. Nat. Mach. Intell. 5, 765-779 (2023)

    2023

  31. [39]

    & Karniadakis, G

    Lu, L., Jin, P., Pang, G., Zhang, Z. & Karniadakis, G. E. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nat. Mach. Intell. 3, 218–229 (2021)

    2021

  32. [41]

    Guibas, J. et al. Adaptive Fourier neural operators: efficient token mixers for transformers. In International Conference on Learning Representations, 1–15 (2022)

    2022

  33. [42]

    Xiong, W. et al. Koopman neural operator as a mesh -free solver of non -linear partial differential equations. J. Comput. Phys. 113194 (2024)

    2024

  34. [43]

    & Karniadakis, G

    Cao, Q., Goswami, S. & Karniadakis, G. E. Laplace neural operator for solving differential equations. Nat. Mach. Intell. 6, 631–640, (2024)

    2024

  35. [44]

    & Karniadakis, G

    Goswami, S., Bora, A., Yu, Y . & Karniadakis, G. E. Physics -informed deep neural operator networks. In Machine Learning in Modeling and Simulation: Methods and Applications , 219 –254 (Springer International Publishing, Cham, 2023)

    2023

  36. [45]

    Li, Z. et al. Physics-informed neural operator for learning partial differential equations. ACM/JMS J. Data Sci. 1, 1–27 (2024)

    2024

  37. [46]

    Chen, L. Q. Phase-field models for microstructure evolution. Annu. Rev. Mater. Res. 32, 113–140 (2002)

    2002

  38. [47]

    Phase-field models in materials science

    Steinbach, I. Phase-field models in materials science. Modell. Simul. Mater. Sci. Eng. 17, 073001 (2009)

    2009

  39. [49]

    A., Schulte, A., Allen, T

    Ahmed, K., Yablinsky, C. A., Schulte, A., Allen, T. & El -Azab, A. Phase field modeling of the 31 effect of porosity on grain growth kinetics in polycrystalline ceramics. Modell. Simul. Mater. Sci. Eng. 21, 065005 (2013)

    2013

  40. [50]

    & Zhang, Y .-C

    Kardar, M., Parisi, G. & Zhang, Y .-C. Dynamic scaling of growing interfaces. Phys. Rev. Lett. 56, 889–892 (1986)

    1986

  41. [51]

    D., Braatz, R

    Zhao, H., Storey, B. D., Braatz, R. D. & Bazant, M. Z. Learning the physics of pattern formation from images. Phys. Rev. Lett. 124, 060201(2020)

    2020

  42. [52]

    Petersen, B. K. et al. Deep symbol ic regression: recovering mathematical expressions from data via risk -seeking policy gradients. In International Conference on Learning Representations , 1 –26 (2020)

    2020

  43. [53]

    B., Loukas, A

    Cordonnier, J. B., Loukas, A. & Jaggi, M. On the relationship between self-attention and convolutional layers. In International Conference on Learning Representations, 1–18 (2020)

    2020

  44. [54]

    Hao, Z. et al. Dpot: Auto -regressive denoising operator transformer for large -scale PDE pre - training. In International Conference on Machine Learning, 1–20 (2024)

    2024

  45. [55]

    & Kutyniok, G

    Alberti, S., Dern, N., Thesing, L. & Kutyniok, G. Sumformer: universal approximation for efficient transformers. In Proceedings of 2nd Annual Workshop on Topological, Algebraic and Geometric Learning Workshops, 72-86 (2023)

    2023

  46. [56]

    Genton, M. G. Classes of kernels for machine learning: a statistics perspective. J. Mach. Learn. Res. 2, 299–312 (2001)

    2001

  47. [57]

    H., Brunton, S

    Rudy, S. H., Brunton, S. L., Proctor, J. L., & Kutz, J. N. Data-driven discovery of partial differential equations. Sci. Adv. 3, e1602614 (2017)

    2017

  48. [58]

    & Pappu, R

    Brangwynne, C., Tompa, P. & Pappu, R. Polymer physics of intracellular phase transitions. Nat. Phys. 11, 899–904 (2015)

    2015

  49. [59]

    Bazant, M. Z. Thermodynamic stability of driven open systems and control of phase separation by electro-autocatalysis. Faraday Discuss. 199, 423–463 (2017). Supplementary information Hierarchical Physics-Embedded Learning for Spatiotemporal Dynamical Systems Xizhe Wang1*, Xiao...

    2017

  50. [60]

    Furthermore, in the implementation of AFNO, the Fourier attention layer introduces some tricks

    A Fourier attention layer in AFNO consists of three parts: (1) employ the discre te Fourier transform to mix different tokens: 𝒁[𝑖, 𝑗, ∶] = DFT(𝑼̅)[𝑖, 𝑗, ∶] where 𝑼̅ ∈ ℝ𝐻′×𝑊′×𝑑 is obtained through patch and position embedding from 𝑼 ∈ ℝ𝐻×𝑊 (2) multiply the weight matrix 𝑾𝑓 ∈ ℂ...

  51. [61]

    Physics-encoded Recurrent Convolut ional Neural Network (PeRCNN)

    as baseline methods. Physics-encoded Recurrent Convolut ional Neural Network (PeRCNN). PeRCNN implements a recurrent architecture with shared parameters, implemented through the forward Euler scheme. Unlike HPE-AFNO, PeRCNN introduces a novel convolutional block, referred to a...

  52. [62]

    Guibas, J. et al. Adaptive Fourier neural operators: efficient token mixers for transformers. Preprint at arXiv:2111.13587 (2021)

    2021 arXiv

  53. [63]

    Rao, C. et al. Encoding physics to learn reaction -diffusion processes. Nature Mach. Intell. 5, 765– 779 (2023)

    2023

  54. [64]

    Li, Z. et al. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 1–16 (2021)

    2021

  55. [65]

    Takamoto, M. et al. PDEBench: an extensive benchmark for scientific machine learning. Adv. Neural Inf. Process. Syst. 35, 1596–1611 (2022)

    2022

  56. [66]

    L., Proctor, J

    Brunton, S. L., Proctor, J. L. & Kutz, J. N. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proc. Natl Acad. Sci. USA 113, 3932-3937 (2016)

    2016

  57. [67]

    H., Brunton, S

    Rudy, S. H., Brunton, S. L., Proctor, J. L. & Kutz, J. N. Data-driven discovery of partial differential equations. Sci. Adv. 3, e1602614 (2017)

    2017

  58. [68]

    Petersen, B. K. et al. Deep symbolic regression: recovering mathematical expressions from data via risk-seeking policy gradients. In International Conference on Learning Representations , 1–26 (2020)

    2020

  59. [69]

    Augusto, D. A. & Barbosa, H. J. Symbolic regression via genetic programming. In Proceedings 6th Brazilian Symp. Neural Networks, 173–178 ( 2000)

    2000

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.