Neural Multiscale Decomposition for Solving The Nonlinear Klein-Gordon Equation with Time Oscillation

Huanhuan Gao*; Zhangyong Liang

REVIEW 4 major objections 5 minor 12 references

NeuralMD solves the nonlinear Klein-Gordon equation uniformly over ε∈(0,1] by decomposing the solution into a slow envelope and a small remainder, reconstructing via a WKB expansion.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 19:29 UTC pith:AIDR4FBP

load-bearing objection Oracle-based model selection undercuts the main claim; the decomposition idea is worth engaging with, but the method as stated is not an implementable uniform solver. the 4 major comments →

arxiv 2512.00266 v7 pith:AIDR4FBP submitted 2025-11-29 math.NA cs.NA

Neural Multiscale Decomposition for Solving The Nonlinear Klein-Gordon Equation with Time Oscillation

Zhangyong Liang , Huanhuan Gao* This is my paper

classification math.NA cs.NA MSC 65M7035Q5568T07

keywords Neural multiscale decompositionNonlinear Klein-Gordon equationNonrelativistic limitSpectral biasPropagation failureWKB expansionPhysics-informed neural networksGradient correlation

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper claims that a collocation-based neural solver can handle the nonlinear Klein-Gordon equation across the full parameter range ε∈(0,1], from the relativistic regime to the nonrelativistic limit, provided the solution is first decomposed into a slowly varying envelope z and a small remainder r, with the fast O(ε²) time oscillation carried analytically by the phase factors exp(±it/ε²). A first network learns the NLSW equation for z; a second network learns the remainder r; a WKB reconstruction plus an error-based criterion decides whether r is needed. This decomposition removes the dominant cause of spectral bias, because the envelope's active temporal frequencies stay O(1) rather than O(ε⁻²), and a gated gradient-correlation correction with multiscale time perturbations is introduced to stop propagation failure that otherwise freezes the solution in time. In 1D, 2D, and 3D experiments the method reports relative errors around 0.002–0.01 at ε=0.8, 0.5, 0.1, and 0.01, with improvements of up to 99.99% over the baselines it compares against.

Core claim

The central claim is that the oscillatory NKGE solution can be represented globally in time as u(x,t)=e^{it/ε²} z(x,t) plus its complex conjugate plus r(x,t), where z solves the nonlinear Schrödinger equation with wave operator (NLSW) with well-prepared initial data and r solves a remainder equation with small initial data. As ε→0 the NLSW approximation is O(ε²) accurate, so the remainder becomes negligible; at larger ε the remainder network supplies amplitude correction. The authors argue that this decomposition makes the temporal spectrum of the trained envelope O(1) rather than O(ε⁻²), eliminating the spectral-bias breakdown, and that the gated gradient-correlation correction with random

What carries the argument

The central machinery is the global WKB/modulated-Fourier decomposition u = e^{it/ε²} z + e^{-it/ε²} z-bar + r, which splits the NKGE into an NLSW for the slow envelope z and a remainder equation for r. The key identity is the frequency-domain residual coefficients: for z the linear symbol σ_z(k,ξ)=−2k−ε²k²+|ξ|² stays O(1) in the active band |k|=O(1), while for r the symbol σ_r(k,ξ)=−ε²k²+|ξ|²+ε⁻² scales as O(ε⁻²) in the oscillatory band k∼ε⁻². This contrast explains why training on z avoids spectral bias, and why training on r alone would reintroduce it; the gated gradient-correlation correction—random time perturbations, a causal gate h(t), and multiscale time mixing—is designed to keep th

Load-bearing premise

The claim's load-bearing premise is that there is one fixed positive time-window size R>0, valid for the entire training interval, such that the networks' parameter gradients at any two times t and t+δt with |δt|≤R are never negatively correlated; the paper proves this only pointwise at each time and explicitly concedes that a single uniform R is not theoretically guaranteed and is supported only empirically.

What would settle it

Measure, for a trained NeuralMD model at a fixed ε, the gradient-correlation inner product ⟨∂θu|_t, ∂θu|_{t+δt}⟩ on a dense time grid and compute the pointwise radii R_t=‖g(t)‖/(2L) under the assumed gradient-Lipschitz constant; if min_t R_t ≤ 0, or if at any t before the gate opens the correlation becomes negative for |δt| within the chosen perturbation sizes, then Assumption 1 fails and the stated cure for propagation failure lacks its required support. A concrete run of this check at ε=0.5, T=5 would settle it.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

The method solves the NKGE in one training pass over the whole time interval, without time-marching at the O(ε²) scale; a single network pair covers ε∈(0,1].
In the nonrelativistic regime (ε small), the remainder network becomes unnecessary: solving only the NLSW and reconstructing with the WKB phase factor yields the O(ε²) accurate solution.
The spectral-bias mechanism for the original equation no longer applies to the envelope: the residual symbol for z is O(1) in the active temporal band, so training avoids the exponential gradient suppression that stalls standard neural solvers.
The causality-gated perturbation scheme keeps the remainder network stable even when the remainder is small but oscillatory, which is what lets the method work at ε=0.8 and 0.5 where ordinary PINNs freeze.
In the reported benchmarks, NeuralMD's relative errors stay near 0.002–0.01 over the whole regime, while baselines degrade dramatically at ε≤0.5, yielding the claimed improvements of up to 99.99%.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

A natural extension is to apply the same WKB/slow-envelope splitting to other oscillatory time-dependent PDEs, such as semiclassical Schrödinger systems, since the spectral-bias mechanism identified here is generic to any fast phase with an ε⁻² carrier frequency.
Because the paper's own experiments show propagation failure also occurs at ε=1 for long horizons (T=10), the gated gradient-correlation correction could be tested as a general cure for long-time causality failure in physics-informed networks, independent of oscillation.
The decision rule selects between the envelope-only and envelope-plus-remainder reconstructions by comparing both to the exact solution; replacing that oracle with an estimator of the reconstruction error would make the method fully unsupervised and applicable where no reference solution is available.
If the uniformity in Assumption 1 is the only gap between the theorem and the method, a sharpened analysis that controls min_t R_t under explicit smoothness or spectral assumptions would convert the propagation-failure cure from empirical to rigorous.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Oracle-based model selection undercuts the main claim; the decomposition idea is worth engaging with, but the method as stated is not an implementable uniform solver.

read the letter

You should know one thing: the advertised claim of a self-contained, uniformly accurate neural solver is undone by the paper's own selection step. Equation (51) and Algorithm 1 line 26 choose between uθ1 and uθ1+rθ2 by comparing both to the exact solution û. That makes the impressive Table 1 numbers post-hoc oracle selections, not the output of an implementable algorithm. The authors state the criterion plainly — no hiding — but the reported errors are not achievable without knowing the exact solution.

What is genuinely new and good: the two-stage pretraining idea — train a network on the NLSW envelope, then train a remainder network — is a sensible way to bring WKB/MTI decomposition into collocation-based methods. The benchmark suite across 1D/2D/3D and multiple ε is useful. The gated gradient-correlation correction is a real attempt to address propagation failure, and the authors are honest that Assumption 1's uniform R is not theoretically guaranteed; they say so in the text and back it only empirically. That is the right kind of transparency.

The soft spots are real and central. The oracle selection is the load-bearing flaw; the method has no residual-based or validation-based selector to replace it. Assumption 1 underpins Theorem 2, so the proof is conditional on something the authors concede is unproved. Table 1 has no error bars. Appendices D–G are empty despite promising hyperparameter, optimizer, and inverse-problem evidence. None of those are fatal on their own, but together they thin out the support.

That said, the underlying decomposition is borrowed but sound, and the combination with a two-stage PINN could plausibly work if the selection criterion is fixed. This paper deserves a serious referee — not because the current version is acceptable, but because the core idea is salvageable and the failure mode is instructive. I would not cite it in its present form, and I would want the authors to add a residual-based selection rule and rerun the experiments before trusting any of the error tables.

For peer review: send it out, but expect major revision. The referee should focus on whether the method can be made fully unsupervised, and on whether Assumption 1 can be replaced with a checkable condition. Reading group? Maybe — it is a good case study in how a small implementation choice can undermine an otherwise reasonable method.

Referee Report

4 major / 5 minor

Summary. The paper proposes NeuralMD, a two-stage physics-informed collocation method for the nonlinear Klein-Gordon equation (NKGE) with a dimensionless parameter ε∈(0,1]. It uses a WKB/multiscale decomposition of the solution into a slowly varying envelope z satisfying an NLSW equation and a remainder r, trains separate neural networks for z and r, and reconstructs u via u = e^{it/ε²} z + c.c. + r (Eq. (34), (50)). To address propagation failure, the method introduces a gated gradient-correlation correction with random temporal perturbations and multiscale time mixing (§3.3). An error-based criterion, Eq. (51), decides whether to include the remainder. The paper reports experiments in 1D, 2D, and 3D against many PINN baselines, claiming accuracy across the whole ε regime with improvements of up to 99.99% (Table 1).

Significance. If the claims were valid, NeuralMD would be a significant contribution: a global-in-time, unsupervised collocation solver for a highly oscillatory PDE, avoiding temporal time-stepping and ε-dependent resolution. The underlying multiscale decomposition is well motivated by existing MTI/WKB literature, and the idea of training only the slow envelope plus a learned remainder is attractive. The gated gradient-correlation strategy is also a plausible proposal for mitigating propagation failure. However, the central numerical claim is currently supported by an oracle-based model-selection step that uses the exact solution, and the theoretical support rests on an unproved uniformity assumption. As a result, the paper's main conclusions are not established in the present form.

major comments (4)

[§3.2, Eq. (51); Algorithm 1 line 26] The final prediction is selected by uθ* ← min{∥uθ1 − û∥, ∥u(θ1,θ2) − û∥}, where û is the exact NKGE solution. In any genuine forward computation û is unknown, so this selection rule is not implementable. The reported errors in Table 1, including the 99.99% improvement, are therefore obtained by comparing both candidates to ground truth and picking the better one at test time. This is an internal inconsistency in the advertised 'unsupervised' solver, not a mere missing proof. The paper should either replace Eq. (51) with an implementable criterion (e.g., based on the residual of the original NKGE or a held-out validation set) and re-run all experiments, or explicitly recast the method as oracle-assisted, which would not support the stated contribution.
[§3.4, Assumption 1 and Theorem 2] Assumption 1 postulates a uniform time-region size R>0 such that the gradient inner products for zθ1 and rθ2 are nonnegative for all t and |δt|≤R. The proof given in §3.4 establishes only pointwise radii R_t = min{∥g1(t)∥/(2L1), ∥g2(t)∥/(2L2)} and explicitly states that the uniform bound R = min_t R_t > 0 'is not theoretically guaranteed' and is treated as an assumption, supported only empirically. No empirical evidence for the uniformity is actually presented. Theorem 2 and the gated gradient-correlation correction both require this uniform R. Thus the proposed cure for propagation failure is conditional on an unverified assumption, and the theoretical justification of the method's key innovation is incomplete.
[Table 1] All reported accuracy numbers in Table 1 come from single runs, with no error bars, no standard deviations, and no information about random seeds. Given the known stochasticity of neural-network training and the claimed improvements of up to 99.99%, these numbers do not establish statistical superiority. The authors should report mean ± std over multiple independent trials for NeuralMD and for the baselines, especially in the transition and nonrelativistic regimes where the differences are most dramatic.
[Appendices D–G] Appendices D–G are listed in the manuscript and are promised in the main text (hyperparameter analysis, inverse problem experiments, optimizer impact, and additional results), but they contain no content. This is not a minor formatting issue: the missing hyperparameter and inverse-problem results are part of the claimed robustness evaluation, and their absence prevents reproducibility and assessment of the method's sensitivity to the many free parameters introduced in §3.3. These sections should be completed or removed; if they are removed, the corresponding claims in the main text should be revised.

minor comments (5)

[§2.1, Eq. (8)] The stated Fourier transform of the tanh-activated network is not correct as written (missing factors and sign conventions). This is a motivational heuristic, so I do not treat it as a central error, but it should be corrected or explicitly labeled as a schematic approximation.
[§2.2] The phrase 'For small ε near 1' is confusing; presumably it means 'for ε slightly less than 1' or 'in the mildly oscillatory regime'. Please rephrase.
[Eq. (110)] The definitions of rMAE and rRMSE are nonstandard: rMAE is normally the relative L1 error without the square root, and the current formula seems to take the square root of a sum of absolute values. Please align the notation with the text.
[Eq. (2)] The periodic boundary loss L_Bd writes uθ(a, ta) − uθ(b, tb), which mixes time indices. For a periodic-in-space boundary condition, the same time should be used; please correct the typo.
[§4.2] The number of baselines is stated inconsistently ('ten PINNs architectures' versus the list of twelve models including vanilla PINNs). Please correct the count.

Circularity Check

1 steps flagged

The final 'prediction' is selected by closeness to the ground-truth solution û (Eq. 51 / Algorithm 1), so the reported accuracy numbers are not produced by an implementable unsupervised solver.

specific steps

fitted input called prediction [Section 3.2 'Error criterion', Eq. (51); Algorithm 1, line 26]
"Further, we introduce an error-based criterion to determine whether the remainder should be incorporated for amplitude compensation, as uθ∗ ← min{∥u θ∗ 1 −ˆu∥, ∥u(θ∗ 1 ,θ∗ 2 ) −ˆu∥}. (51) where, ˆu is the ground truth solution of NKGE."

The final output uθ* is defined as whichever of the two WKB reconstructions has smaller error against û, the exact NKGE solution. In any actual deployment û is unknown, so the algorithm as stated cannot produce a prediction; the 'prediction' is an oracle selection benchmarked on the target. The Table 1 rMAE/rRMSE values and the reported 'improvements up to 99.99%' are therefore partly self-referential: the reported solution is chosen by comparing candidates to the very solution used to evaluate them. The decomposition and residual losses themselves are not circular, but the final reconstruction step reduces to the ground-truth input.

full rationale

The core multiscale decomposition (WKB envelope z plus remainder r) is a rearrangement of the NKGE with the oscillating phase factored out; its error estimates (29a)-(29b) are cited from the numerical-analysis literature rather than derived from the paper's own fitted quantities, so that part of the derivation is self-contained. No load-bearing chain of self-citations is used: the MTI/NLSW convergence rests on prior external results (Bao et al., Faou-Schratz). The central circular step is the error criterion: Eq. (51) and Algorithm 1 line 26 define the final prediction as the candidate closest to the ground truth û. Because û is the unknown target, the method's benchmark numbers are not the output of an implementable unsupervised solver; they are post-hoc oracle selections. This affects the headline accuracy claims (Table 1's 'improvements up to 99.99%') but not the construction of the z and r networks themselves, hence partial circularity rather than a fully self-referential derivation. Separately, the uniform time-region size R in Assumption 1 is explicitly admitted to be 'not theoretically guaranteed' and supported only empirically, so Theorem 2 is conditional; this is a missing-proof limitation, not a circular reduction. Empty Appendices D-G similarly remove supporting details but do not create circularity. Score 6 reflects one load-bearing reduction-by-construction (ground-truth-based selection) in an otherwise externally grounded derivation.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The derivation rests on established WKB/MTI results and the authors' own prior convergence theorems, plus an explicit unproven uniformity assumption and standard smoothness/convexity conditions for the generalization bounds. No new physical entities are introduced.

free parameters (6)

Perturbation region sizes {R1,R2,R3} = §4.1: {1e-2,5e-2,9e-2}; §4.3: {0.03,0.05,0.07}
Hand-set time perturbation scales used to define multiscale neighborhoods; values are inconsistent between two sections and are not justified by analysis.
Perturbation counts {k1,k2,k3} = {3,5,7}
Number of perturbed time samples per scale, fixed by hand; Appendix G claims more perturbations improve performance, so results depend on these choices.
Gate steepness α = 5
Controls steepness of causal gate h(t); chosen a priori.
Gate shift update parameters η, Δmax, ϵ = η=1e-3, Δmax=0.1, ϵ not specified
Update of γ uses these hyperparameters; no sensitivity analysis in the main text.
Loss weights λRes, λIc, λBd = not reported
Critical for PINN training; never specified in Eqs. (5), (39), (46) or in the experiments.
Model selection using ground truth û = û (exact solution)
Eq. (51) selects between z-only and z+r reconstructions by comparing to exact solution; this is the most consequential 'parameter' and is data leakage.

axioms (6)

domain assumption Validity of the global WKB/modulated Fourier expansion u=e^{it/ε²}z+c.c.+r with remainder in Eq. (34) for all ε∈(0,1] on [0,T].
Taken from MTI literature (Bao et al. 2014, Faou & Schratz 2014); the paper does not prove it for the global time interval and all regimes.
domain assumption H2/H3 convergence estimates (Eq. 29): ∥u−u_sw∥_{H1}≤C0ε² and ∥u−u_s∥_{H1}≤(C1+C2T)ε².
These are cited theorems from prior work; the paper relies on them to justify dropping or keeping the remainder.
ad hoc to paper Assumption 1: uniform R>0 such that gradient inner products remain nonnegative across time regions.
Explicitly stated as an assumption in §3.4 after Eq. (94); supports Theorem 2 and the gated correction.
domain assumption Bounded mixed derivatives ∂²z/(∂θ∂t) and ∂²r/(∂θ∂t) for the trained networks.
Used to make gradient maps Lipschitz in the proof of Assumption 1; not established for the actual training trajectory.
domain assumption Loss function L is L-Lipschitz and β-smooth, convex or bounded non-convex in θ.
The generalization bounds in Appendix A rely on these properties, but neural network losses are generally non-convex and the constants are not verified.
standard math Universal Approximation Theorem and Kolmogorov–Arnold representation theorem.
Used to justify MLP and KAN expressiveness; standard results.

pith-pipeline@v1.3.0-alltime-deepseek · 49457 in / 14413 out tokens · 124000 ms · 2026-08-03T19:29:34.737845+00:00 · methodology

0 comments

read the original abstract

In this paper, we propose a neural multiscale decomposition method (NeuralMD) for solving the nonlinear Klein-Gordon equation (NKGE) with a dimensionless parameter $\varepsilon\in(0,1]$ from the relativistic regime to the nonrelativistic limit regime. The solution of the NKGE propagates waves with wavelength at $O(1)$ and $O(\varepsilon^2)$ in space and time, respectively, which brings the oscillation in time. Existing collocation-based methods for solving this equation lead to spectral bias and propagation failure. To mitigate the spectral bias induced by high-frequency time oscillation, we employ a multiscale time integrator (MTI) to absorb the time oscillation into the phase. This decomposes the NKGE into a nonlinear Schr\"odinger equation with wave operator (NLSW) with well-prepared initial data and a remainder equation with small initial data. As $\varepsilon \to 0$, the NKGE converges to the NLSW at rate $O(\varepsilon^{2})$, and the contribution of the remainder equation becomes negligible. Furthermore, to alleviate propagation failure caused by medium-frequency time oscillation, we propose a gated gradient correlation correction strategy to enforce temporal coherence in collocation-based methods. As a result, the approximation of the remainder term is no longer affected by propagation failure. Comparative experiments with existing collocation-based methods demonstrate the superior performance of our method for solving the NKGE with various regularities of initial data over the whole regime.

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 11 linked inside Pith

[7]

Setpinns: Set-based physics-informed neural networks.arXiv preprint arXiv:2409.20206,

Mayank Nagda, Phil Ostheimer, Thomas Specht, Frank Rhein, Fabian Jirasek, Stephan Mandt, Marius Kloft, and Sophie Fellenz. Setpinns: Set-based physics-informed neural networks.arXiv preprint arXiv:2409.20206,

Pith/arXiv arXiv
[8]

Understand- ing the difficulty of training physics-informed neural networks on dynamical systems

Franz M Rohrhofer, Stefan Posch, Clemens G¨ oßnitzer, and Bernhard C Geiger. Understand- ing the difficulty of training physics-informed neural networks on dynamical systems. arXiv preprint arXiv:2203.13648, 162, 2022a. Franz M Rohrhofer, Stefan Posch, Clemens G¨ oßnitzer, and Bernhard C Geiger. On the role of fixed points of dynamical systems in training...

Pith/arXiv arXiv
[9]

Respecting causality is all you need for training physics-informed neural networks.arXiv preprint arXiv:2203.07404, 2022b

Sifan Wang, Shyam Sankaran, and Paris Perdikaris. Respecting causality is all you need for training physics-informed neural networks.arXiv preprint arXiv:2203.07404, 2022b. Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why pinns fail to train: A neural tangent kernel perspective.Journal of Computational Physics, 449:110768, 2022c. Sifan Wang, Bow...

Pith/arXiv arXiv
[10]

Propinn: Demystifying propagation failures in physics-informed neural networks.arXiv preprint arXiv:2502.00803,

51 Zhangyong Liang, Zhiping Mao, and Xiaofei Zhao Haixu Wu, Yuezhou Ma, Hang Zhou, Huikun Weng, Jianmin Wang, and Mingsheng Long. Propinn: Demystifying propagation failures in physics-informed neural networks.arXiv preprint arXiv:2502.00803,

Pith/arXiv arXiv
[1989]

Kan: Kolmogorov-arnold networks.arXiv preprint arXiv:2404.19756,

Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljaˇ ci´ c, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks.arXiv preprint arXiv:2404.19756,

Pith/arXiv arXiv
[1990]

Diederik P Kingma

45 Zhangyong Liang, Zhiping Mao, and Xiaofei Zhao 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 u(t, x) x = 32.0 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 2.8 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 1.5 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 u(t, x) x = 0.0 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 1.2 0 2 4 t 0.50 0.25 0.00 0.25 0.50 x = 31.8 Ground Truth Predi...

Pith/arXiv arXiv
[2009]

Quadratic residual networks: A new class of neural networks for solving forward and inverse problems in physics involving pdes

Jie Bu and Anuj Karpatne. Quadratic residual networks: A new class of neural networks for solving forward and inverse problems in physics involving pdes. InProceedings of the 2021 SIAM International Conference on Data Mining (SDM), pages 675–683. SIAM,

2021
[2013]

Ghazal Farhani, Alexander Kazachek, and Boyu Wang

44 Neural Multiscale Decomposition 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 u(t, x) x = 32.0 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 2.8 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 1.5 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 u(t, x) x = 0.0 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 1.2 0 2 4 t 0.50 0.25 0.00 0.25 0.50 x = 31.8 Ground Truth Prediction Figure 20...

Pith/arXiv arXiv
[2021]

Multi-scale deep neural networks for solving high dimensional pdes.arXiv preprint arXiv:1910.11710,

Wei Cai and Zhi-Qin John Xu. Multi-scale deep neural networks for solving high dimensional pdes.arXiv preprint arXiv:1910.11710,

Pith/arXiv arXiv 1910
[2022]

Pinnsformer: A transformer-based framework for physics-informed neural networks.arXiv preprint arXiv:2307.11833,

Zhiyuan Zhao, Xueying Ding, and B Aditya Prakash. Pinnsformer: A transformer-based framework for physics-informed neural networks.arXiv preprint arXiv:2307.11833,

Pith/arXiv arXiv
[2024]

Multi-scale deep neural network (mscalednn) for solving poisson-boltzmann equation in complex domains.arXiv preprint arXiv:2007.11207,

Ziqi Liu, Wei Cai, and Zhi-Qin John Xu. Multi-scale deep neural network (mscalednn) for solving poisson-boltzmann equation in complex domains.arXiv preprint arXiv:2007.11207,

Pith/arXiv arXiv 2007
[2025]

Sub-sequential physics-informed learning with state space model.arXiv preprint arXiv:2502.00318, 2025a

Chenhui Xu, Dancheng Liu, Yuting Hu, Jiajie Li, Ruiyang Qin, Qingxiao Zheng, and Jinjun Xiong. Sub-sequential physics-informed learning with state space model.arXiv preprint arXiv:2502.00318, 2025a. Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks.arXiv preprint ar...

Pith/arXiv arXiv 1901

[1] [7]

Setpinns: Set-based physics-informed neural networks.arXiv preprint arXiv:2409.20206,

Mayank Nagda, Phil Ostheimer, Thomas Specht, Frank Rhein, Fabian Jirasek, Stephan Mandt, Marius Kloft, and Sophie Fellenz. Setpinns: Set-based physics-informed neural networks.arXiv preprint arXiv:2409.20206,

Pith/arXiv arXiv

[2] [8]

Understand- ing the difficulty of training physics-informed neural networks on dynamical systems

Franz M Rohrhofer, Stefan Posch, Clemens G¨ oßnitzer, and Bernhard C Geiger. Understand- ing the difficulty of training physics-informed neural networks on dynamical systems. arXiv preprint arXiv:2203.13648, 162, 2022a. Franz M Rohrhofer, Stefan Posch, Clemens G¨ oßnitzer, and Bernhard C Geiger. On the role of fixed points of dynamical systems in training...

Pith/arXiv arXiv

[3] [9]

Respecting causality is all you need for training physics-informed neural networks.arXiv preprint arXiv:2203.07404, 2022b

Sifan Wang, Shyam Sankaran, and Paris Perdikaris. Respecting causality is all you need for training physics-informed neural networks.arXiv preprint arXiv:2203.07404, 2022b. Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why pinns fail to train: A neural tangent kernel perspective.Journal of Computational Physics, 449:110768, 2022c. Sifan Wang, Bow...

Pith/arXiv arXiv

[4] [10]

Propinn: Demystifying propagation failures in physics-informed neural networks.arXiv preprint arXiv:2502.00803,

51 Zhangyong Liang, Zhiping Mao, and Xiaofei Zhao Haixu Wu, Yuezhou Ma, Hang Zhou, Huikun Weng, Jianmin Wang, and Mingsheng Long. Propinn: Demystifying propagation failures in physics-informed neural networks.arXiv preprint arXiv:2502.00803,

Pith/arXiv arXiv

[5] [1989]

Kan: Kolmogorov-arnold networks.arXiv preprint arXiv:2404.19756,

Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljaˇ ci´ c, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks.arXiv preprint arXiv:2404.19756,

Pith/arXiv arXiv

[6] [1990]

Diederik P Kingma

45 Zhangyong Liang, Zhiping Mao, and Xiaofei Zhao 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 u(t, x) x = 32.0 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 2.8 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 1.5 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 u(t, x) x = 0.0 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 1.2 0 2 4 t 0.50 0.25 0.00 0.25 0.50 x = 31.8 Ground Truth Predi...

Pith/arXiv arXiv

[7] [2009]

Quadratic residual networks: A new class of neural networks for solving forward and inverse problems in physics involving pdes

Jie Bu and Anuj Karpatne. Quadratic residual networks: A new class of neural networks for solving forward and inverse problems in physics involving pdes. InProceedings of the 2021 SIAM International Conference on Data Mining (SDM), pages 675–683. SIAM,

2021

[8] [2013]

Ghazal Farhani, Alexander Kazachek, and Boyu Wang

44 Neural Multiscale Decomposition 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 u(t, x) x = 32.0 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 2.8 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 1.5 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 u(t, x) x = 0.0 0 1 2 3 4 5 t 0.4 0.2 0.0 0.2 0.4 0.6 x = 1.2 0 2 4 t 0.50 0.25 0.00 0.25 0.50 x = 31.8 Ground Truth Prediction Figure 20...

Pith/arXiv arXiv

[9] [2021]

Multi-scale deep neural networks for solving high dimensional pdes.arXiv preprint arXiv:1910.11710,

Wei Cai and Zhi-Qin John Xu. Multi-scale deep neural networks for solving high dimensional pdes.arXiv preprint arXiv:1910.11710,

Pith/arXiv arXiv 1910

[10] [2022]

Pinnsformer: A transformer-based framework for physics-informed neural networks.arXiv preprint arXiv:2307.11833,

Zhiyuan Zhao, Xueying Ding, and B Aditya Prakash. Pinnsformer: A transformer-based framework for physics-informed neural networks.arXiv preprint arXiv:2307.11833,

Pith/arXiv arXiv

[11] [2024]

Multi-scale deep neural network (mscalednn) for solving poisson-boltzmann equation in complex domains.arXiv preprint arXiv:2007.11207,

Ziqi Liu, Wei Cai, and Zhi-Qin John Xu. Multi-scale deep neural network (mscalednn) for solving poisson-boltzmann equation in complex domains.arXiv preprint arXiv:2007.11207,

Pith/arXiv arXiv 2007

[12] [2025]

Sub-sequential physics-informed learning with state space model.arXiv preprint arXiv:2502.00318, 2025a

Chenhui Xu, Dancheng Liu, Yuting Hu, Jiajie Li, Ruiyang Qin, Qingxiao Zheng, and Jinjun Xiong. Sub-sequential physics-informed learning with state space model.arXiv preprint arXiv:2502.00318, 2025a. Zhi-Qin John Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks.arXiv preprint ar...

Pith/arXiv arXiv 1901