Pith. sign in

REVIEW 4 major objections 4 minor 38 references

Breaking the Precision Ceiling in Physics-Informed Neural Networks: A Hybrid Fourier-Neural Architecture for Ultra-High Accuracy

T0 review · 4 major / 4 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 hybrid Fourier-neural network solves the Euler-Bernoulli beam equation to L2 error 1.94×10⁻⁷, with exactly 10 harmonics optimal.

desk verdict The paper's headline result is unverifiable as written because the initial conditions are never given, and the Fourier basis exactly solves the homogeneous PDE, so the 1.94e-7 error may be trivial. read the letter →

arxiv 2507.20929 v1 pith:BON625D5 submitted 2025-07-28 cs.LG cond-mat.mtrl-sciphysics.comp-ph

Wei Shan Lee , Chi Kiu Althina Chau , Kei Chon Sio , Kam Ian Leong This is my paper
classification cs.LGcond-mat.mtrl-sciphysics.comp-ph MSC 68T0765M7074K10
keywords physics-informedneuralnetworksEuler-Bernoullibeamequationfourth-orderPDEshybridFourier-neuralarchitectureultra-precisionsolutionsharmonictruncationadaptivelossweightingtwo-phaseoptimization
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

This paper claims that the apparent precision ceiling of physics-informed neural networks (PINNs) on fourth-order partial differential equations is an artifact of architecture, not a fundamental limit. The authors solve the Euler-Bernoulli beam equation, the fourth-order PDE describing transverse beam vibration, with a hybrid ansatz: a truncated Fourier series with learnable coefficients captures the dominant vibration modes, and a small deep network adds residual corrections. On their test problem they report an L2 error of $1.94\times10^{-7}$, roughly 17 times better than a standard PINN and 15–500 times better than the traditional numerical methods they compare against. They also report a counterintuitive cliff: exactly 10 harmonics is optimal, and adding more harmonics degrades the error from $10^{-7}$ to about $4\times10^{-1}$. If the claim holds, machine-learning solvers become credible for fourth-order problems that demand engineering-grade precision.

What carries the argument

The load-bearing object is the hybrid ansatz of Eq. (1): a truncated Fourier series whose coefficients are trained directly as parameters, joined to a boundary-modulated neural correction $N_{\mathrm{BC}}(t,x)=\mathcal{N}(t,x)\,\sin(\pi x/L)$. The factor $\sin(k_n x)$ makes the Fourier part satisfy the zero-displacement boundary conditions identically, and analytic differentiation of the Fourier terms sidesteps the fourth-order automatic-differentiation instabilities that limit pure PINNs. Two further mechanisms carry the precision: a two-phase optimizer that runs Adam (learning rate 0.01, plateau scheduling, gradient clipping at norm 1) for up to 2000 iterations and then L-BFGS, a quasi-Newton method with strong-Wolfe line search, until the gradient norm falls below $10^{-9}$; and a sigmoid adaptive weighting $w_\alpha=\mathrm{scale}/(1+\exp(-\log_{10} L_\alpha))$ with $\mathrm{scale}=1+N/130$ that keeps PDE residual, initial, and boundary losses from dominating one another. The exponential growth of the Hessian's condition number with $N$ is the paper's proposed mechanism for the 10-harmonic cliff.

What would settle it

Run the identical architecture and training recipe on a manufactured Euler-Bernoulli solution whose energy sits above the 10th harmonic, for instance $w(t,x)=\sin(20\pi t)\,\sin(20\pi x/L)$, which the $N=10$ ansatz cannot represent: if the L2 error stays below roughly $10^{-6}$, the hybrid method's precision is general, while a collapse toward $10^{-1}$ would show the reported result is an artifact of a low-frequency test case. A cheaper check is to set the neural term to zero ($\lambda=0$) and least-squares fit only the 20 Fourier coefficients, then compare that error to $1.94\times10^{-7}$ to see whether the network contributes anything at the optimal configuration.

Watch

Extended reading notes

Core claim

The central claim is that a hybrid representation of the solution — $w(t,x)=\sum_{n=1}^{N}[a_n\cos(\omega_n t)+b_n\sin(\omega_n t)]\sin(k_n x)+\lambda\mathcal{N}(t,x)$, with the Fourier coefficients $a_n,b_n$ as independent learnable parameters and $\mathcal{N}$ a tanh network of 27,905 parameters — breaks the precision ceiling that had pinned PINNs at $10^{-5}$–$10^{-6}$ relative error for fourth-order problems. On the Euler-Bernoulli equation $\partial^2 w/\partial t^2 + c^2\,\partial^4 w/\partial x^4=0$ over the domain $[0,1]\times[0,10]$, the authors measure an L2 error of $1.94\times10^{-7}$ on a dense $100\times100$ validation grid. The companion discovery is an optimal-truncation cliff: sweeping $N$ from 5 to 50 shows the error falling monotonically to a minimum at exactly $N=10$, then jumping to $4.02\times10^{-1}$ at $N=15$ and staying near $10^{-1}$ thereafter. The reported Hessian condition number grows from about $10^3$ at $N=10$ to over $10^7$ at $N=30$, which the authors read as optimization intractability overwhelming added representation capacity.

Load-bearing premise

The headline accuracy rests on the unstated choice of a test problem whose exact solution is dominated by the first 10 Fourier modes of the chosen sine basis: the initial conditions $w_0(x)$ and $v_0(x)$ are never specified, so if the true solution carried significant energy above the 10th harmonic, the truncated series could not represent it at $1.94\times10^{-7}$.

Editorial extensions

If this is right

  • If the $1.94\times10^{-7}$ result reproduces, the precision ceiling of PINNs on fourth-order PDEs is an architectural effect, and other high-order equations (Timoshenko beams, plate and biharmonic problems) become candidates for the same treatment.
  • The 10-harmonic cliff implies that model capacity can actively hurt ultra-precision convergence, so harmonic count becomes a first-order hyperparameter that must be swept rather than increased blindly.
  • Sub-30-minute training on a single GPU makes ultra-precision PINNs practical for engineering iteration loops rather than toy benchmarks.
  • The reported comparisons — about 17 times over standard PINNs and 15–500 times over the traditional numerical baselines cited — give other neural PDE solvers a concrete benchmark to beat on the Euler-Bernoulli problem.
  • The adaptive weighting scheme removes a manual-tuning burden, which the paper argues is a precondition for deploying PINNs in safety-critical settings.

Reading between the lines

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

  • The paper never states the initial conditions $w_0(x)$ and $v_0(x)$, so the most direct test of the headline claim is to rerun the identical pipeline on a manufactured solution with energy in harmonics 11–20; if the error climbs toward $10^{-1}$, the 10-harmonic optimum is a property of the chosen test case, not of the method.
  • Because the neural correction is scaled by $\lambda=10^{-8}$, the network may not be load-bearing at the optimum: a pure least-squares fit of the 20 Fourier coefficients to the same loss would isolate the contribution of the spectral ansatz from the contribution of the network.
  • An unstated corollary is that the method's generality is bounded by how well a single fixed Fourier basis matches the problem's physical modes; the paper's own limitation discussion points to Chebyshev or wavelet bases as the natural extension.
  • The sharp success at $N=10$ alongside failure at $N=15$ suggests a connectivity threshold in the loss landscape; if that is generic, other hybrid spectral-neural solvers should show similar cliffs, which is worth knowing before adopting them.
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 / 4 minor

Summary. The paper proposes a hybrid Fourier-neural architecture for the Euler-Bernoulli beam equation, combining a truncated Fourier series with a deep neural network scaled by lambda = 1e-8, trained by Adam followed by L-BFGS. The central claim is that this architecture breaks a 'precision ceiling' for fourth-order PDEs, achieving an L2 error of 1.94e-7 with exactly 10 harmonics, and that adding more harmonics causes catastrophic degradation. A harmonic-count sweep from N=5 to N=50 is reported, together with comparisons to FEM and standard PINNs.

Significance. If the claimed accuracy were demonstrated on a well-posed, fully specified test problem, the result would be of practical interest for neural PDE solvers. The paper gives a detailed training pipeline, a code availability statement, and a systematic harmonic sweep, which are commendable. However, the mathematical structure of the ansatz makes the central claim largely tautological: each Fourier mode is an exact solution of the homogeneous beam equation, so the PDE residual is identically zero for the Fourier block regardless of coefficients, and the neural correction is scaled by 1e-8. The optimization therefore reduces to fitting initial conditions that are never specified. The reported accuracy and the 'exactly 10 harmonics' law are consequently not established as a general result.

major comments (4)
  1. [§2.3–2.4, Eqs. (1), (6), (8), Table 2] For k_n = nπ/L and ω_n = k_n^2 c, each Fourier term sin(k_n x)cos(ω_n t) and sin(k_n x)sin(ω_n t) satisfies Eq. (6) exactly, so the Fourier block contributes identically zero to the PDE residual L_pde in Eq. (8) for any coefficients. With the neural correction scaled by λ = 10^{-8} in Algorithm 1, the PDE loss is essentially determined by the tiny neural term, and the optimization reduces to fitting the initial-condition losses (9)–(10). Because the initial data w_0(x) and v_0(x) are never specified, the reported error 1.94×10^{-7} is unfalsifiable: if the omitted data lie in span{sin(nπx/L) : n ≤ 10}, the result is expected by construction, whereas any high-frequency content cannot be represented at all. Please state the initial data and repeat the experiments for initial data containing modes beyond N=10.
  2. [Table 3, N=15 row] For smooth band-limited data, adding five more sine modes cannot make a least-squares or spectral fit six orders of magnitude worse; the jump from 1.94×10^{-7} at N=10 to 4.02×10^{-1} at N=15 is evidence of optimization failure rather than a representational limit. The paper's claim of exponential growth of the Hessian condition number is asserted but not measured. Report training curves and final initial-condition losses for each N, the converged Fourier coefficients, and diagnostics such as gradient norms and condition-number estimates; otherwise 'exactly 10 harmonics' cannot be distinguished from a local-minimum artifact.
  3. [§3.1 and Figs. 7–8] All numerical claims appear to come from a single training run; no random seeds, repeated trials, or confidence intervals are reported. Given the order-of-magnitude accuracy claims and the sharp threshold at N=10, at least 3–5 independent runs per configuration are required before the 'catastrophic degradation' and '17-fold improvement' can be supported statistically.
  4. [§3 and §4 comparison paragraphs] The baseline comparisons are internally inconsistent and under-specified: the paper variously claims 15–500×, 108-fold, and 15–30× improvement over traditional numerical methods, and the FEM/PINN baselines are not described (mesh, element order, architecture, training budget, or initial data). A controlled comparison on identical initial and boundary data is necessary for the central claim that the hybrid method 'breaks the precision ceiling'; as written, the improvement factor is not a well-defined quantity.
minor comments (4)
  1. [Algorithm 1, Phase 2 break condition] The condition 'if ∥∇L∥ < 10^{-9} or loss increases' can stop L-BFGS prematurely; report how often this condition triggered and the loss at termination.
  2. [Eq. (12)] Equation (12) uses w_alpha and L_alpha without defining alpha; define the index set and the correspondence with Eq. (7).
  3. [Figures 1–15] Figures 1–15 are referenced but not included in the arXiv text, so the visual evidence (error heatmaps, training curves) cannot be inspected by the reader.
  4. [Abstract and Conclusion] The abstract and conclusion give inconsistent baseline improvement factors (15–500× vs. 108-fold vs. 15–30×); harmonize these numbers.

Circularity Check

3 steps flagged · score 7.0 of 10

The headline 'precision breakthrough' reduces to a Fourier fit of unstated initial conditions: every Fourier term exactly solves the homogeneous beam PDE, the neural correction is scaled to 1e-8, and w0/v0 are never specified.

  1. self definitional [Section 2.3 Eq. 1; Section 2.4 Eq. 6; Table 2 (omega_n = k_n^2 c); Algorithm 1 line 3]
    "w(t, x) = sum_{n=1}^N [a_n cos(omega_n t) + b_n sin(omega_n t)] sin(k_n x) + lambda * N(t, x) (1) ... partial^2 w/partial t^2 + c^2 partial^4 w/partial x^4 = 0 (6) ... Set lambda = 10^{-8} (scaling factor) ... omega_n = k_n^2 c"

    The Fourier block is an exact solution of Eq. (6) by construction: for omega_n = k_n^2 c, (partial_t^2 + c^2 partial_x^4)[sin(k_n x)(a_n cos omega_n t + b_n sin omega_n t)] = 0. Therefore L_pde in Eq. (8) is identically zero for the Fourier part and O(lambda^2) for the network; with lambda = 10^{-8}, the PDE constraint is effectively empty. Training reduces to fitting the initial-condition losses (9)-(10). If w_0 and v_0 are linear combinations of the first 10 sine modes, the exact solution lies exactly in the ansatz span, so the reported 1.94e-7 is an in-sample fit of the chosen test function, not an independent prediction.

  2. fitted input called prediction [Section 2.4 Eqs. 9-10; Section 3 Table 3]
    "L_ic = 1/N_ic * sum_i |w(0, x_i) - w_0(x_i)|^2 (9); L_ict = 1/N_ic * sum_i |partial w/partial t (0, x_i) - v_0(x_i)|^2 (10) ... attaining an L2 error of 1.94 x 10^{-7}"

    Because the PDE term is satisfied by construction, the only informative terms in Eq. (7) are the initial-condition losses, yet the paper never states w_0(x) or v_0(x) and never writes the 'exact analytical solution' against which Table 3 is measured. The reported error is therefore not reproducible and cannot be checked against a non-bandlimited case. Calling the result a 'prediction' or a '17-fold improvement' treats the fitted Fourier coefficients of an unstated target as a forecast; the benchmark is definitionally favorable if the target contains only the first 10 modes.

1 more flagged steps
  1. other [Abstract and Section 3.1, Table 3; Section 4 Conclusions]
    "a systematic harmonic optimization study revealed a counter-intuitive discovery: exactly 10 harmonics yield optimal performance, with accuracy catastrophically degrading from 10^{-7} to 10^{-1} beyond this threshold ... the optimal count remains problem-dependent"

    The N=10 'discovery' is exactly what would result if the unstated exact solution is a sum of the first 10 sine modes: the truncated series then represents the solution exactly, and any extra harmonics are unnecessary. The paper provides no mechanism for the catastrophic jump at N=15 and no evidence that the target contains modes beyond 10. Since w_0/v_0 are omitted, the reader cannot distinguish a general precision phenomenon from a tautology of problem selection. The paper's own statement that 'the optimal count remains problem-dependent' concedes this is a property of the chosen problem, not a fundamental law.

full rationale

The central derivation chain is internally self-contained but circular in its test design. The ansatz in Eq. (1) is a truncated Fourier series plus a neural term scaled by lambda = 1e-8, and each Fourier term exactly solves the homogeneous Euler-Bernoulli equation (6) because omega_n = k_n^2 c. Consequently, the PDE residual loss (8) contributes nothing, and the optimization is effectively a least-squares fit of the Fourier coefficients to initial-condition data (9)-(10). The paper never specifies w0(x) or v0(x), so the reader cannot tell whether the exact solution is simply a superposition of the first 10 sine modes; if it is, the claimed 1.94e-7 error is guaranteed by construction rather than being a demonstrated breakthrough. The 'exactly 10 harmonics' finding is the same problem-selection effect, and the paper itself concedes the optimal count is problem-dependent. The two-phase Adam/L-BFGS schedule, adaptive weighting, and GPU implementation are independent engineering contributions, but they do not validate the headline precision claim against any externally specified benchmark. No load-bearing self-citation chain or imported uniqueness theorem was found; the circularity is in the mathematical setup and the omitted test definition. Score 7 reflects that the central claim reduces by construction on the unstated test problem, while some peripheral methodology remains non-circular.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities. Its empirical claims rest on several fitted constants and on the unstated choice of a test problem that appears to align with the Fourier basis.

free parameters (4)
  • Harmonic truncation count N = 10
    Optimal value determined by the authors' own harmonic sweep (Table 3). Central to the paper's main 'discovery'.
  • Adaptive weight scale factor = 1.0 + N/130
    Eq. 12; the constant 130 was 'empirically determined' from ablation studies (Section 2.4).
  • Neural correction scaling lambda = 1e-8
    Set in Algorithm 1 line 3; no justification given for this value.
  • Network widths = 128-128-64-32-16-8
    Stated in Section 2.3 without ablation; a design choice.
assumptions (3)
  • domain assumption The exact solution of the test problem is well represented by 10 Fourier modes in the sin(n pi x/L) basis.
    Invoked implicitly in Section 2.1 and 3; validity depends on the unspecified initial conditions.
  • domain assumption The two-phase Adam/L-BFGS optimization with the described settings converges to the global minimum for N<=10.
    Needed for the claim that 10 harmonics is optimal; no convergence guarantee is provided.
  • ad hoc to paper The Hessian condition number grows exponentially with harmonic count N.
    Asserted in Section 3.1 to explain the performance cliff, but no measurement or derivation is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Breaking the Precision Ceiling in Physics-Informed Neural Networks: A Hybrid Fourier-Neural Architecture for Ultra-High Accuracy." pith.science (2026). https://pith.science/paper/BON625D5

@misc{pith2026250720929,
  author       = {Pith},
  title        = {Pith review of: Breaking the Precision Ceiling in Physics-Informed Neural Networks: A Hybrid Fourier-Neural Architecture for Ultra-High Accuracy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BON625D5}},
  note         = {Machine review of arXiv:2507.20929}
}
abstract

Physics-informed neural networks (PINNs) have plateaued at errors of $10^{-3}$-$10^{-4}$ for fourth-order partial differential equations, creating a perceived precision ceiling that limits their adoption in engineering applications. We break through this barrier with a hybrid Fourier-neural architecture for the Euler-Bernoulli beam equation, achieving unprecedented L2 error of $1.94 \times 10^{-7}$-a 17-fold improvement over standard PINNs and \(15-500\times\) better than traditional numerical methods. Our approach synergistically combines a truncated Fourier series capturing dominant modal behavior with a deep neural network providing adaptive residual corrections. A systematic harmonic optimization study revealed a counter-intuitive discovery: exactly 10 harmonics yield optimal performance, with accuracy catastrophically degrading from $10^{-7}$ to $10^{-1}$ beyond this threshold. The two-phase optimization strategy (Adam followed by L-BFGS) and adaptive weight balancing enable stable ultra-precision convergence. GPU-accelerated implementation achieves sub-30-minute training despite fourth-order derivative complexity. By addressing 12 critical gaps in existing approaches-from architectural rigidity to optimization landscapes-this work demonstrates that ultra-precision is achievable through proper design, opening new paradigms for scientific computing where machine learning can match or exceed traditional numerical methods.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 15 canonical work pages

  1. [1]

    Journal of Computational Physics 378, 686–707 (2019) https://doi.org/10.1016/j.jcp

    Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics-informed neural networks: A deep learn- ing framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 378, 686–707 (2019) https://doi.org/10.1016/j.jcp. 2018.10.045

    doi:10.1016/j.jcp 2019

  2. [2]

    arXiv preprint arXiv:1711.10566 (2017)

    Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics informed deep learning (part ii): Data- driven discovery of nonlinear partial differential equations. arXiv preprint arXiv:1711.10566 (2017)

    arXiv 2017

  3. [3]

    Nature Reviews Physics 3(6), 422–440 (2021) https://doi.org/10

    Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L.: Physics- informed machine learning. Nature Reviews Physics 3(6), 422–440 (2021) https://doi.org/10. 1038/s42254-021-00314-5

    work page 2021

  4. [4]

    Journal of Scientific Computing 92(3), 88 (2022) https://doi.org/10.1007/s10915-022-01939-z

    Cuomo, S., Di Cola, V.S., Giampaolo, F., Rozza, G., Raissi, M., Piccialli, F.: Scientific machine learning through physics–informed neural networks: Where we are and what’s next. Journal of Scientific Computing 92(3), 88 (2022) https://doi.org/10.1007/s10915-022-01939-z

    doi:10.1007/s10915-022-01939-z 2022

  5. [5]

    Nature Communications 12(1), 6136 (2021) https://doi.org/10.1038/s41467-021-26434-1

    Chen, Z., Liu, Y., Sun, H.: Physics-informed learning of governing equations from scarce data. Nature Communications 12(1), 6136 (2021) https://doi.org/10.1038/s41467-021-26434-1

    doi:10.1038/s41467-021-26434-1 2021

  6. [6]

    SIAM Journal on Scientific Computing 41(4), 2603–2626 (2019) https://doi.org/10.1137/18M1229845

    Pang, G., Lu, L., Karniadakis, G.E.: fpinns: Fractional physics-informed neural networks. SIAM Journal on Scientific Computing 41(4), 2603–2626 (2019) https://doi.org/10.1137/18M1229845

    doi:10.1137/18m1229845 2019

  7. [7]

    IEEE Transactions on Neural Networks and Learning Systems (2023) https://doi.org/10.1109/TNNLS.2023.3310585

    Kapoor, T., Wang, H., N´ u˜ nez, A., Dollevoet, R.: Physics-informed neural networks for solving forward and inverse problems in complex beam systems. IEEE Transactions on Neural Networks and Learning Systems (2023) https://doi.org/10.1109/TNNLS.2023.3310585

    arXiv 2023

  8. [8]

    Mechanical Systems and Signal Processing 200, 110575 (2023) https://doi.org/10.1016/j.ymssp

    Luo, K., Kong, X., Wang, X., Jiang, T., Frøseth, G.T., Rønnquist, A.: Cable vibration mea- surement based on broad-band phase-based motion magnification and line tracking algorithm. Mechanical Systems and Signal Processing 200, 110575 (2023) https://doi.org/10.1016/j.ymssp. 2023.110575

    arXiv 2023

Show all 38 references
  1. [9]

    Engineering Applications of Artificial Intelligence 133, 108085 (2024) https://doi.org/10.1016/j.engappai.2024.108085

    Kapoor, T., Wang, H., N´ u˜ nez, A., Dollevoet, R.: Transfer learning for improved generalizability in causal physics-informed neural networks for beam simulations. Engineering Applications of Artificial Intelligence 133, 108085 (2024) https://doi.org/10.1016/j.engappai.2024.108085

    2024

  2. [10]

    Journal of Engineering Mechanics 148(2), 04021139 (2022) https://doi.org/10.1061/(ASCE)EM.1943-7889.0002062

    Vahab, M., Haghighat, E., Khaleghi, M., Khalili, N.: A physics-informed neural network approach to solution and identification of biharmonic equations of elasticity. Journal of Engineering Mechanics 148(2), 04021139 (2022) https://doi.org/10.1061/(ASCE)EM.1943-7889.0002062

    2022

  3. [11]

    arXiv preprint arXiv:2210.00518 (2022)

    Mukhametzhanov, M.S.: High precision differentiation techniques for data-driven solution of nonlinear pdes by physics-informed neural networks. arXiv preprint arXiv:2210.00518 (2022)

    2022 arXiv

  4. [12]

    IEEE Transactions on Artificial Intelligence 5(6), 2547–2557 (2024) https:// doi.org/10.1109/TAI.2022.3192362

    Wong, J.C., Ooi, C., Gupta, A., Ong, Y.-S.: Learning in sinusoidal spaces with physics-informed 13 neural networks. IEEE Transactions on Artificial Intelligence 5(6), 2547–2557 (2024) https:// doi.org/10.1109/TAI.2022.3192362

    2024

  5. [13]

    Com- puter Methods in Applied Mechanics and Engineering 365, 113028 (2020) https://doi.org/10

    Jagtap, A.D., Kawaguchi, K., Karniadakis, G.E.: Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems. Com- puter Methods in Applied Mechanics and Engineering 365, 113028 (2020) https://doi.org/10....

    2020

  6. [14]

    SIAM Review 63(1), 208–228 (2021) https://doi.org/10.1137/19M1274067

    Lu, L., Meng, X., Mao, Z., Karniadakis, G.E.: Deepxde: A deep learning library for solving dif- ferential equations. SIAM Review 63(1), 208–228 (2021) https://doi.org/10.1137/19M1274067

    2021 doi

  7. [15]

    Nature Computational Science 4(7), 483–494 (2024) https://doi.org/10.1038/ s43588-024-00643-2

    Brunton, S.L., Kutz, J.N.: Promising directions of machine learning for partial differen- tial equations. Nature Computational Science 4(7), 483–494 (2024) https://doi.org/10.1038/ s43588-024-00643-2

    2024

  8. [16]

    Physics of Fluids 36(10), 101301 (2024) https://doi.org/10.1063/5.0226562

    Zhao, C., Zhang, F., Lou, W., Wang, X., Yang, J.: A comprehensive review of advances in physics-informed neural networks and their applications in complex fluid dynamics. Physics of Fluids 36(10), 101301 (2024) https://doi.org/10.1063/5.0226562

    2024 doi

  9. [17]

    arXiv preprint arXiv:2010.08895 (2021) https://doi.org/10.48550/arXiv.2010.08895

    Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandku- mar, A.: Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895 (2021) https://doi.org/10.48550/arXiv.2010.08895

    2021 doi arXiv

  10. [18]

    Communications in Computational Physics 28(5), 2002–2041 (2020) https://doi.org/10.4208/cicp.OA-2020-0164

    Jagtap, A.D., Karniadakis, G.E.: Extended physics-informed neural networks (xpinns): A gen- eralized space-time domain decomposition based deep learning framework for nonlinear partial differential equations. Communications in Computational Physics 28(5), 2002–2041 (2020) http...

    2020 doi

  11. [19]

    Computer Methods in Applied Mechanics and Engineering 374, 113547 (2021) https://doi.org/10.1016/j.cma.2020.113547

    Kharazmi, E., Zhang, Z., Karniadakis, G.E.: hp-vpinns: Variational physics-informed neural networks with domain decomposition. Computer Methods in Applied Mechanics and Engineering 374, 113547 (2021) https://doi.org/10.1016/j.cma.2020.113547

    2021

  12. [20]

    Journal of Computational Physics 496, 112603 (2024) https://doi.org/10

    Wang, Y., Zhong, L.: Nas-pinn: Neural architecture search-guided physics-informed neural net- work for solving pdes. Journal of Computational Physics 496, 112603 (2024) https://doi.org/10. 1016/j.jcp.2023.112603

    2024

  13. [21]

    arXiv preprint arXiv:2409.18426 (2024) https://doi.org/10.48550/arXiv.2409.18426

    Hwang, Y., Lim, D.-Y.: Dual cone gradient descent for training physics-informed neural networks. arXiv preprint arXiv:2409.18426 (2024) https://doi.org/10.48550/arXiv.2409.18426

    2024 doi arXiv

  14. [22]

    Journal of Computational Physics 457, 111053 (2022) https://doi.org/10.1016/j.jcp.2022.111053

    Lin, S., Chen, Y.: A two-stage physics-informed neural network method based on conserved quantities and applications in localized wave solutions. Journal of Computational Physics 457, 111053 (2022) https://doi.org/10.1016/j.jcp.2022.111053

    2022

  15. [23]

    Journal of Computational Physics 375, 1339–1364 (2018) https://doi.org/10.1016/j

    Sirignano, J., Spiliopoulos, K.: Dgm: A deep learning algorithm for solving partial differential equations. Journal of Computational Physics 375, 1339–1364 (2018) https://doi.org/10.1016/j. jcp.2018.08.029

    2018 doi

  16. [24]

    Computer Methods in Applied Mechanics and Engineering 397, 115141 (2022) https://doi.org/10.1016/j.cma.2022.115141

    Haghighat, E., Amini, D., Juanes, R.: Physics-informed neural network simulation of multiphase poroelasticity using stress-split sequential training. Computer Methods in Applied Mechanics and Engineering 397, 115141 (2022) https://doi.org/10.1016/j.cma.2022.115141

    2022

  17. [25]

    Computer Methods in Applied Mechanics and Engineering 424, 116883 (2024) https://doi.org/10.1016/j.cma.2024.116883

    Hu, Z., Shi, Z., Karniadakis, G.E., Kawaguchi, K.: Hutchinson trace estimation for high- dimensional and high-order physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering 424, 116883 (2024) https://doi.org/10.1016/j.cma.2024.116883

    2024

  18. [26]

    SIAM Journal on Scientific Computing 43(5), 3055–3081 (2021) https://doi.org/10.1137/20M1318043

    Wang, S., Teng, Y., Perdikaris, P.: Understanding and mitigating gradient flow pathologies in physics-informed neural networks. SIAM Journal on Scientific Computing 43(5), 3055–3081 (2021) https://doi.org/10.1137/20M1318043

    2021 doi

  19. [27]

    arXiv preprint arXiv:2109.01050 (2021) https: //doi.org/10.48550/arXiv.2109.01050

    Krishnapriyan, A.S., Gholami, A., Zhe, S., Kirby, R.M., Mahoney, M.W.: Characterizing possible 14 failure modes in physics-informed neural networks. arXiv preprint arXiv:2109.01050 (2021) https: //doi.org/10.48550/arXiv.2109.01050

    2021 doi arXiv

  20. [28]

    Journal of Computational Physics 474, 111722 (2023) https://doi.org/10.1016/j.jcp.2022.111722

    McClenny, L., Braga-Neto, U.: Self-adaptive physics-informed neural networks. Journal of Computational Physics 474, 111722 (2023) https://doi.org/10.1016/j.jcp.2022.111722

    2023

  21. [29]

    Journal of Computational Physics 473, 111768 (2023) https://doi.org/10.1016/j.jcp.2022.111768

    Arzani, A., Cassel, K.W., D’Souza, R.M.: Theory-guided physics-informed neural networks for boundary layer problems with singular perturbation. Journal of Computational Physics 473, 111768 (2023) https://doi.org/10.1016/j.jcp.2022.111768

    2023

  22. [30]

    arXiv preprint arXiv:2306.15969 (2023) https://doi.org/10.48550/arXiv.2306.15969

    Cho, J., Nam, S., Yang, H., Yun, S.-B., Hong, Y., Park, E.: Separable physics-informed neural networks. arXiv preprint arXiv:2306.15969 (2023) https://doi.org/10.48550/arXiv.2306.15969

    2023 doi arXiv

  23. [31]

    Journal of Computational Physics 493, 112464 (2023) https: //doi.org/10.1016/j.jcp.2023.112464

    Penwarden, M., Jagtap, A.D., Zhe, S., Karniadakis, G.E., Kirby, R.M.: A unified scalable framework for causal sweeping strategies for physics-informed neural networks (pinns) and their temporal decompositions. Journal of Computational Physics 493, 112464 (2023) https: //doi.or...

    2023

  24. [32]

    Journal of Sound and Vibration 225(5), 935–988 (1999) https://doi.org/10.1006/ jsvi.1999.2257

    Han, S.M., Benaroya, H., Wei, T.: Dynamics of transversely vibrating beams using four engineer- ing theories. Journal of Sound and Vibration 225(5), 935–988 (1999) https://doi.org/10.1006/ jsvi.1999.2257

    1999

  25. [33]

    arXiv preprint arXiv:2009.04544 (2020)

    McClenny, L., Braga-Neto, U.: Self-adaptive physics-informed neural networks using a soft attention mechanism. arXiv preprint arXiv:2009.04544 (2020)

    2020 arXiv

  26. [34]

    In: 3rd Interna- tional Conference on Learning Representations, ICLR 2015, San Diego, CA, USA (2015)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: 3rd Interna- tional Conference on Learning Representations, ICLR 2015, San Diego, CA, USA (2015). http://arxiv.org/abs/1412.6980

    2015 arXiv

  27. [35]

    Mathematical programming 45(1-3), 503–528 (1989)

    Liu, D.C., Nocedal, J.: On the limited memory bfgs method for large scale optimization. Mathematical programming 45(1-3), 503–528 (1989)

    1989

  28. [36]

    Journal of Computational Physics 477, 111902 (2023)

    Psaros, A.F., Meng, X., Zou, Z., Guo, L., Karniadakis, G.E.: Uncertainty quantification in sci- entific machine learning: Methods, metrics, and comparisons. Journal of Computational Physics 477, 111902 (2023)

    2023

  29. [37]

    https://arxiv.org/abs/2410.19843

    Wang, Y., Bai, J., Lin, Z., Wang, Q., Anitescu, C., Sun, J., Eshaghi, M.S., Gu, Y., Feng, X.- Q., Zhuang, X., Rabczuk, T., Liu, Y.: Artificial intelligence for partial differential equations in computational mechanics: A review (2024). https://arxiv.org/abs/2410.19843

    2024

  30. [38]

    GitHub (2025) 15 Fig

    Lee, W.S., Chau, C.K.A., Sio, K.C., Leong, K.I.: PINN-Euler-Bernoulli-Beam: Ultra-precision physics-informed neural network implementation. GitHub (2025) 15 Fig. 1: Conceptual overview of the ultra-precision physics-informed neural network approach for solving the Euler-Bernou...

    2025

Pith tools

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