Physics-informed neural networks for high-dimensional solutions and snaking bifurcations in nonlinear lattices

Fidya Almira Suheri; Hadi Susanto; Muhammad Luthfi Shahab; Rudy Kusdiantara

REVIEW 3 major objections 5 minor 48 references

Physics-informed neural networks for high-dimensional solutions and snaking bifurcations in nonlinear lattices

T0 review · 3 major / 5 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 physics-informed neural network that takes lattice indices as inputs, trained with Levenberg–Marquardt root-finding, can approximate steady states of the discrete Allen–Cahn equation in up to five dimensions, trace its snaking…

desk verdict Solid PINN-for-lattices extension with strong 1D/2D validation, but the high-dimensional accuracy claim runs ahead of the evidence. read the letter →

arxiv 2507.09782 v1 pith:BTKXOVXY submitted 2025-07-13 math.NA cs.LGcs.NAcs.NEmath.OCnlin.PS

Muhammad Luthfi Shahab , Fidya Almira Suheri , Rudy Kusdiantara , Hadi Susanto This is my paper · ORCID

classification math.NAcs.LGcs.NAcs.NEmath.OCnlin.PS MSC 65H1065P3068T0737M20

keywords physics-informedneuralnetworksnonlinearlatticesdiscreteAllen-CahnequationsnakingbifurcationsLevenberg-MarquardtalgorithmstochasticNewtonmethodlinearstabilityanalysishigh-dimensionalproblems

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 show that a physics-informed neural network can replace direct numerical solvers for three hard tasks on nonlinear lattices: finding steady states, tracing snaking bifurcation diagrams, and deciding linear stability. The proposed trick is to feed the lattice index into a shallow network and use Levenberg–Marquardt rather than gradient descent to drive the residual equations to zero, which keeps the number of unknowns equal to the network weights instead of the lattice size. In high dimensions, a stochastic mini-batch version of the same root-finding update makes a five-dimensional Allen–Cahn lattice tractable with only 181 parameters. If the claims hold, PINNs become a practical surrogate for problems where forming the full Jacobian is the bottleneck.

What carries the argument

The load-bearing object is the mapping $u(i,W)$ from normalized lattice indices to solution values, with the residual system $f_i(u(i,W),\mu)=0$ treated as a root-finding problem in weight space. The Levenberg–Marquardt update $(J^TJ+\lambda I)^{-1}J^TF$ replaces gradient-based loss minimization, so the Jacobian side length is the weight count, 33 to 181, rather than the number of lattice sites. Continuation is carried out in weight space: the predictor extrapolates $(W^{[k]},\mu^{[k]})$ and the auxiliary equation, norm-based in 1D and arclength-based in 2D, is added to the residual system. For stability, the network output is put through an absolute value to enforce the non-sign-changing eigenvector that Sturm–Liouville theory associates with the largest eigenvalue, with $\lambda$ trained as an extra parameter and $\|v\|_2=1$ as a constraint.

What would settle it

Compute the same five-dimensional problem at a second parameter value, for example $\mu=-0.3$ or $c=0.1$, using the same 181-parameter network and stochastic batches; if the full-system test MSE does not stay near $10^{-9}$ or if different random seeds produce visibly different solution slices, the claimed scalability to high dimensions is not a general property of the method.

Watch

Extended reading notes

Core claim

The paper's central claim is that a two-hidden-layer PINN with Gaussian activation, trained to solve the residual equations directly, can approximate steady states of the discrete Allen–Cahn equation with cubic and quintic nonlinearities and compute their bifurcation and stability properties. On one- and two-dimensional lattices, the network reproduces the snaking bifurcation diagrams of site- and bond-centered solutions, with absolute errors below $5\times10^{-12}$ and $5\times10^{-13}$, respectively, and the largest eigenvalue curves match the direct solver, giving the alternating stable and unstable branches. On three-, four-, and five-dimensional lattices, fixed-width networks with 161, 171, and 181 trainable parameters respectively solve systems of 2,197, 28,561, and 371,293 nonlinear equations; the five-dimensional run uses a stochastic Levenberg–Marquardt update on 1,000 sampled equations per iteration and reaches a test MSE of about $4\times10^{-9}$.

Load-bearing premise

The high-dimensional claims rest on the assumption that a two-hidden-layer network with ten neurons per layer can accurately represent the localized solutions and that the stochastic Levenberg–Marquardt iteration converges to the desired root, and this is demonstrated at only one parameter setting for three to five spatial dimensions.

Editorial extensions

If this is right

In one and two dimensions, the PINN-plus-continuation framework reproduces the snaking bifurcation diagrams and the alternating stable and unstable branches, with pointwise errors below $5\times10^{-12}$ and $5\times10^{-13}$.
The Jacobian size in the solver is set by the network weight count, not the lattice size: 33 parameters for the 1D system, 85 for 2D, and 161 to 181 for 3D to 5D problems with up to 371,293 equations.
Stochastic Levenberg–Marquardt, sampling 1,000 equations per iteration, gives a five-dimensional test MSE near $4\times10^{-9}$ in about 108 seconds, making steady-state computation feasible where full Jacobians exceed memory.
Enforcing positivity on the eigenvector output isolates the largest eigenvalue in agreement with direct solvers, so stability classification of snaking branches can be obtained without forming the full linearized operator.

Reading between the lines

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

Our inference: the accuracy comparison to traditional numerical methods is currently against direct Newton solves on the same lattices; an obvious next test is benchmarking against sparse direct solvers or preconditioned Newton–Krylov methods at matched accuracy, since those also reduce per-iteration cost.
Our inference: the method's scalability claim would be strengthened or bounded by sweeping $\mu$ and $c$ in 3D to 5D; success at a single parameter point does not establish that the 181-parameter network represents the full solution manifold.
Our inference: the stochastic sampling mechanism is viable precisely because every network weight affects every residual, and the paper notes this would fail for local finite-difference variables; the same idea should transfer to other globally coupled or graph-based discrete systems.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The real news is that the paper gives a working PINN formulation for discrete lattice equations, coupling Levenberg–Marquardt with pseudo-arclength continuation. In 1D and 2D the validation is genuinely strong: errors against direct solves are below 1e-12 and 1e-13, the snaking diagrams track the reference curves, and the eigenvalue module with the absolute-value output is a sensible use of Sturm–Liouville theory. The stochastic LM variant for 5D is worth taking seriously, and showing train/test residual curves is a nice touch. The sensitivity appendices for alpha, gamma, beta1, beta2 are useful and honest about tuning.

Where the paper oversells is in the abstract's claim of accuracy comparable to or better than traditional methods, especially in high-dimensional regimes. In 4D and 5D the only accuracy evidence is the MSE of the nonlinear residual at one parameter value (mu = -0.5, c = 0.05). A residual of 4e-9 in 5D is not a solution error; without a reference solution, a conditioning estimate, or a smaller-grid direct solve, you cannot tell whether the PINN found the right steady state or just something close in the residual norm. In 3D the 'True' curve in Figure 10 is never explained, which should be fixed. The continuation and stability modules, the strongest parts, are only exercised in 1D/2D. So the high-dimensional part is an extrapolation, not a demonstrated result.

Smaller issues: the stochastic Newton update in Eq. (23) is a restricted Gauss-Newton-like step, and its convergence is only empirical. Network sizes are tiny (181 parameters in 5D), which helps cost, but the expressiveness claim needs at least a heuristic argument. No code or data is provided, and the custom MATLAB implementation is described but not shared. The continuation parameters are tuned on the test problems, though the sensitivity study partly mitigates that.

The paper is written for people working on PINN-based bifurcation tracking or numerical methods for discrete nonlinear lattices. It is a real, somewhat incremental extension of the authors' earlier PDE work, with strong low-dimensional evidence. It deserves a serious referee. I would recommend major revision: rewrite the high-dimensional claims, add a genuine high-dimensional accuracy check (even on a smaller lattice where a direct solver is feasible), explain the 3D true solution, and report solution error, not just residual. With those changes this could be a useful paper for Physica D.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a physics-informed neural network (PINN) framework for steady-state solutions of discrete nonlinear lattices, for snaking bifurcation diagrams via pseudo-arclength continuation, and for linear-stability analysis via a positivity-constrained eigenvector computation. The method parameterizes lattice solutions by a small two-hidden-layer network, trains with Levenberg–Marquardt, and includes a stochastic sampling variant for five-dimensional problems. Numerical experiments on the discrete Allen–Cahn equation cover dimensions one through five. In one and two dimensions, PINN solutions, bifurcation diagrams, and eigenvalue curves are compared with direct numerical solutions and match to very high accuracy. In three to five dimensions, the evidence consists mainly of the residual MSE at a single parameter setting, with no direct reference-solution comparison reported in four and five dimensions.

Significance. The low-dimensional results are strong and constitute a useful demonstration: with only 33 or 85 trainable parameters, the method reproduces snaking bifurcation diagrams, solution profiles, and stability transitions of a nontrivial discrete system at high accuracy, and the sensitivity studies in Appendices A and B address the main continuation hyperparameters. If the high-dimensional results were validated against reference solutions, the stochastic Levenberg–Marquardt variant would be a meaningful scaling result for lattice problems where full Jacobians are prohibitive. As it stands, however, the paper's central abstract claim—accuracy comparable to or better than traditional methods in high-dimensional regimes—is not established by the reported evidence, because no ground-truth comparison is given for the four- and five-dimensional cases.

major comments (3)

[§5.4, §5.5, Table 2] The high-dimensional accuracy claim rests on residual MSE rather than on solution error. Table 2 reports MSEs of 2.1e-12 in four dimensions and 4.0e-9 in five dimensions; these quantities are mean squares of the nonlinear residuals f_i, not errors in the solution u. In five dimensions the RMS residual is about 6e-5, and the pointwise solution error could be substantially larger depending on the conditioning of the discrete Allen–Cahn operator at the computed state. No reference solution, error norm, or independent solver comparison is provided for 4D and 5D, and the experiments are performed at a single parameter value (mu = -0.5, c = 0.05). To support the abstract's claim, the authors should report a direct accuracy measure for at least one 4D/5D case, for example by comparison with a reference solution obtained by an independent iterative solver, by continuation from a known low-dimensional branch, or by a manufactured solution; at minimum, they should state the relation between residual MSE and solution error.
[§5.3, Fig. 10a] The curve labeled 'True' in Fig. 10a is not defined. The text says only that the figure shows a one-dimensional slice of the solution, but it does not state how the reference solution was computed or whether this reference is the same direct solver used for the one- and two-dimensional cases. Without this information the three-dimensional claim cannot be verified. If the reference is a direct 2197-by-2197 Newton solve, the authors should say so and report the maximum pointwise difference; if it is a different approximation, the label should be changed accordingly.
[§4.4, Eq. (21)] The stochastic Newton update in Eq. (21), W <- W - J_{S_k}^{-1} F_{S_k}, is not well-defined for the five-dimensional problem because J_{S_k} is rectangular (1001 by 181 in Table 2) and therefore not invertible. The stochastic Levenberg–Marquardt update in Eq. (23) is well-defined and is presumably what was implemented, but the text should not present the inverse form as the method. In addition, the authors should clarify how convergence of the stochastic iteration is assessed; reporting the residual on the full system after a fixed 1000 iterations is not by itself evidence that a root of the nonlinear system has been found.

minor comments (5)

[Eq. (29)] The quantity called the norm, ||u|| = (1/(1+sqrt(1+mu))) * sum_i u_i^2, is not a norm; it is a scaled sum of squares. Either rename it or correct the definition to avoid confusion in the continuation constraint.
[§5.3] The text '153 points' appears to be a rendering error for '15^3 points'; please correct the typography.
[Table 2] The one-dimensional MSE of 2.1e-32 seems surprisingly small relative to the solution errors around 1e-12 reported in Fig. 3; please verify the reported value and clarify what quantity is being measured.
[Figs. 5 and 9] The stability diagrams display only PINN curves, with no direct-solver overlay; the authors should state explicitly whether the stable/unstable classification was verified against direct eigenvalue computations on all branches, or only on the eigenvalue curves shown in Figs. 4 and 8.
[General] A reproducibility statement or code/data availability section would strengthen the paper; none is currently provided.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity; the method is benchmarked against direct solvers, with only a minor self-cited activation ansatz.

ansatz smuggled in via citation [Section 4.1, after Eq. (15)]
"Following the success of previous work in nonlinear bifurcation problems [20], we adopt the Gaussian activation function σ(x) = e^{−x^2/2} throughout this study."

The only cited justification for the central nonlinear ansatz of the network, the Gaussian activation in Eq. (15), is reference [20], which is prior work by two of the same authors. The present paper does not derive or independently test this activation choice; it imports it from the authors' own prior work, matching the ansatz-smuggled-in-via-citation pattern. This is a genuine but minor self-citation: the central accuracy claims do not reduce to the citation because the paper validates the resulting PINN solutions against direct numerical solvers in one and two dimensions, so the activation choice is not load-bearing for the final conclusions.

full rationale

The central derivation chain is self-contained with respect to independent benchmarks. In one and two dimensions, PINN steady states, snaking bifurcation diagrams, and largest eigenvalues are compared directly against ground-truth solutions obtained by traditional numerical methods; the paper reports absolute errors below 5e-12 in one dimension and below 5e-13 in two dimensions, so the accuracy claim is externally supported rather than reduced to a fit. The continuation procedure solves the original nonlinear system F(u,mu)=0 together with an auxiliary arclength or norm constraint, and the plotted diagrams track actual solutions of the governing equations rather than fitted values. The stability analysis relies on classical Sturm-Liouville theory and the positivity of the principal eigenvector, not on any parameter fitted to the target eigenvalue. In the 3D-5D cases, the reported MSE is the residual of the discrete Allen-Cahn system defined in Eq. (14), which is the training objective itself; presenting residual MSE as 'accuracy' is a validation weakness rather than a circular reduction, and in the five-dimensional case a full-system test residual is additionally reported. The only self-citation concern is the Gaussian activation function, which is adopted from the authors' own prior work without independent derivation in this paper; because the method is subsequently validated against direct solvers, this is a minor self-citation and does not make the central derivation circular.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the network's representational capacity, the convergence of the optimization, and standard spectral theory. These are domain assumptions of the numerical method rather than physical constants. The continuation and sampling hyperparameters are tuned by hand on the test problem, so they are listed as free parameters. No new entities such as particles, forces, or dimensions are introduced.

free parameters (6)

Continuation weight alpha = 10
Chosen by sensitivity analysis in Appendix A to balance the residual of the nonlinear system against the continuation constraint; tuned on the same 1D Allen-Cahn test problem.
Norm step gamma (1D) = 10/1000
Controls the step size in solution norm for norm-based continuation; tuned in Appendix A to resolve snaking turning points accurately.
Arclength weights beta1, beta2, delta (2D) = beta1=1000/40, beta2=100, delta=1
Set the step sizes in solution norm and parameter for arclength-based continuation; tuned in Appendix B on the 2D Allen-Cahn problem.
Stochastic batch size s(S_k) (5D) = 1000
Number of randomly sampled equations per Levenberg-Marquardt iteration; Table 1 shows the effect of batch size on test MSE and time, and 1000 is chosen as a compromise.
Training iterations = 1000
All experiments run for a fixed 1000 iterations; convergence is judged by the MSE at iteration 1000 rather than by a tolerance-based stopping rule.
Network hidden layer sizes = (4,4) in 1D, (7,7) in 2D, (10,10) in 3D-5D
Architecture hyperparameters chosen ad hoc; no systematic search is reported, and the table in Section 3 explores hidden size only for the 3D case.

assumptions (5)

domain assumption A shallow two-hidden-layer network with Gaussian activation can approximate the solutions of the discrete Allen-Cahn equation to the reported accuracy.
The method succeeds only if the chosen architecture lies within the universal approximation regime for these specific discrete solutions; the paper provides empirical MSE but no error bounds.
domain assumption The Levenberg-Marquardt algorithm, including the stochastic variant, converges to a root of the nonlinear system F(u,mu)=0.
The paper reports successful runs but gives no convergence proof; failure of the optimization would invalidate all computed solutions and diagrams.
standard math The largest eigenvalue of the linearized discrete operator has a non-sign-changing eigenvector (Sturm-Liouville / Perron-Frobenius property).
Used in Section 4.6 to justify that enforcing positivity via the absolute value output yields the largest eigenvalue for linear stability analysis.
domain assumption The pseudo-arclength continuation constraint (Eq. 9) with the weight extrapolation predictor (Eq. 18) can track the solution branch through turning points.
The constraint and predictor are heuristic adaptations of standard continuation; success depends on branch smoothness and on the chosen tuning parameters.
domain assumption The solutions of interest are symmetric about the lattice midpoint, allowing the input transformation of Eq. (17).
Assumed from the literature [18] and used to reduce the network input domain; nonsymmetric branches would not be represented by the transformed network.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Physics-informed neural networks for high-dimensional solutions and snaking bifurcations in nonlinear lattices." pith.science (2026). https://pith.science/paper/BTKXOVXY

@misc{pith2026250709782,
  author       = {Pith},
  title        = {Pith review of: Physics-informed neural networks for high-dimensional solutions and snaking bifurcations in nonlinear lattices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BTKXOVXY}},
  note         = {Machine review of arXiv:2507.09782}
}

read the original abstract

This paper introduces a framework based on physics-informed neural networks (PINNs) for addressing key challenges in nonlinear lattices, including solution approximation, bifurcation diagram construction, and linear stability analysis. We first employ PINNs to approximate solutions of nonlinear systems arising from lattice models, using the Levenberg-Marquardt algorithm to optimize network weights for greater accuracy. To enhance computational efficiency in high-dimensional settings, we integrate a stochastic sampling strategy. We then extend the method by coupling PINNs with a continuation approach to compute snaking bifurcation diagrams, incorporating an auxiliary equation to effectively track successive solution branches. For linear stability analysis, we adapt PINNs to compute eigenvectors, introducing output constraints to enforce positivity, in line with Sturm-Liouville theory. Numerical experiments are conducted on the discrete Allen-Cahn equation with cubic and quintic nonlinearities in one to five spatial dimensions. The results demonstrate that the proposed approach achieves accuracy comparable to, or better than, traditional numerical methods, especially in high-dimensional regimes where computational resources are a limiting factor. These findings highlight the potential of neural networks as scalable and efficient tools for the study of complex nonlinear lattice systems.

Figures

Figures reproduced from arXiv: 2507.09782 by the authors.

**Figure 2.** Snaking bifurcation diagrams of the one-dimensional discrete Allen–Cahn equa [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

**Figure 3.** (a), (c), (e): True and PINN solutions corresponding to the marked points in [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: Largest eigenvalues versus ∥u∥ for the one-dimensional discrete Allen–Cahn equation: (a) c = 0.05, (b) c = 0.15. 0 2 4 6 8 10 -1 -0.8 -0.6 -0.4 -0.2 SC PINN BC PINN (a) 0 2 4 6 8 10 -1 -0.8 -0.6 -0.4 -0.2 SC PINN BC PINN (b) [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]

**Figure 5.** Figure 5: Bifurcation diagrams with linear stability indicated: (a) [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]

**Figure 6.** Figure 6: Snaking bifurcation diagrams of the two-dimensional discrete Allen–Cahn equa [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

**Figure 7.** Figure 7: (a), (c), (e): PINN solutions corresponding to black circles in Figure [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

**Figure 8.** Figure 8: Largest eigenvalue versus ∥u∥ for the two-dimensional discrete Allen–Cahn equation: (a) c = 0.05, (b) c = 0.15. SC: site-centered, BC: bond-centered. 0 5 10 15 20 25 30 35 40 -1 -0.8 -0.6 -0.4 -0.2 SC PINN BC PINN (a) 0 5 10 15 20 25 30 35 40 -1 -0.8 -0.6 -0.4 -0.2 SC…

**Figure 9.** Figure 9: Snaking bifurcation diagrams with linear stability indicated for (a) [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

**Figure 10.** Figure 10: (a) True and PINN solutions for the three-dimensional discrete Allen–Cahn [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]

**Figure 11.** Figure 11: (a) PINN solution for the four-dimensional Allen–Cahn equation shown as [PITH_FULL_IMAGE:figures/full_fig_p021_11.png]

**Figure 12.** Figure 12: (a) PINN solution for the five-dimensional Allen–Cahn equation shown as [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 43 canonical work pages

[1]

Cuomo, V

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

work page 2022
[2]

Hornik, M

K. Hornik, M. Stinchcombe, H. White, Multilayer feedforward networks are universal approximators, Neural Networks 2 (5) (1989) 359–366. 25

work page 1989
[3]

Dissanayake, N

M. Dissanayake, N. Phan-Thien, Neural-network-based approximations for solving partial differential equations, Communications in Numerical Methods in Engineering 10 (3) (1994) 195–201

work page 1994
[4]

I. E. Lagaris, A. Likas, D. I. Fotiadis, Artificial neural networks for solving ordinary and partial differential equations, IEEE Transactions on Neural Networks 9 (5) (1998) 987–1000

1998
[5]

Raissi, P

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

2019
[6]

S. Cai, Z. Mao, Z. Wang, M. Yin, G. E. Karniadakis, Physics-informed neural networks (PINNs) for fluid mechanics: A review, Acta Mechanica Sinica 37 (12) (2021) 1727–1738

2021
[7]

Y. Chen, L. Lu, G. E. Karniadakis, L. Dal Negro, Physics-informed neural networks for inverse problems in nano-optics and metamaterials, Optics Express 28 (8) (2020) 11618–11633

work page 2020
[8]

J. Han, A. Jentzen, E. Weinan, Solving high-dimensional partial dif- ferential equations using deep learning, Proceedings of the National Academy of Sciences 115 (34) (2018) 8505–8510

work page 2018

Show all 48 references

[9]

E. R. Putri, M. L. Shahab, M. Iqbal, I. Mukhlash, A. Hakam, L. Mar- dianto, H. Susanto, A deep-genetic algorithm (deep-GA) approach for high-dimensional nonlinear parabolic partial differential equations, Computers & Mathematics with Applications 154 (2024) 120–127

2024
[10]

W. Zhu, W. Khademi, E. G. Charalampidis, P. G. Kevrekidis, Neural networks enforcing physical symmetries in nonlinear dynamical lattices: The case example of the Ablowitz-Ladik model, Physica D: Nonlinear Phenomena 434 (2022) 133264

2022
[11]

Z. Lin, Y. Chen, Pseudo grid-based physics-informed convolutional- recurrent network solving the integrable nonlinear lattice equations, Physica D: Nonlinear Phenomena 468 (2024) 134304

2024
[12]

R. Geng, J. Zu, Y. Gao, H.-K. Zhang, Separable graph Hamiltonian network: A graph deep learning model for lattice systems, Physical Review Research 6 (1) (2024) 013176. 26

2024
[13]

Zhou, X.-Y

J.-C. Zhou, X.-Y. Wen, M.-J. Guo, Symmetric difference data enhance- ment physics-informed neural network for solving discrete nonlinear lat- tice equations, Communications in Theoretical Physics (2024)

2024
[14]

Saqlain, W

S. Saqlain, W. Zhu, E. G. Charalampidis, P. G. Kevrekidis, Discovering governing equations in discrete systems using PINNs, Communications in Nonlinear Science and Numerical Simulation 126 (2023) 107498

2023
[15]

Opala, S

A. Opala, S. Ghosh, T. C. Liew, M. Matuszewski, Neuromorphic com- puting in Ginzburg-Landau polariton-lattice systems, Physical Review Applied 11 (6) (2019) 064029

2019
[16]

Stokes, J

J. Stokes, J. R. Moreno, E. A. Pnevmatikakis, G. Carleo, Phases of two- dimensional spinless lattice fermions with first-quantized deep neural- network quantum states, Physical Review B 102 (20) (2020) 205122

2020
[17]

Y. V. Kartashov, B. A. Malomed, L. Torner, Solitons in nonlinear lat- tices, Reviews of Modern Physics 83 (1) (2011) 247–305

2011
[18]

Taylor, J

C. Taylor, J. H. Dawes, Snaking and isolas of localised states in bistable discrete lattices, Physics Letters A 375 (1) (2010) 14–22

2010
[19]

Kusdiantara, H

R. Kusdiantara, H. Susanto, Homoclinic snaking in the discrete Swift- Hohenberg equation, Physical Review E 96 (6) (2017) 062214

2017
[20]

M. L. Shahab, H. Susanto, Neural networks for bifurcation and linear stability analysis of steady states in partial differential equations, Ap- plied Mathematics and Computation 483 (2024) 128985

2024
[21]

M. L. Shahab, H. Susanto, Corrigendum to ”Neural networks for bifur- cation and linear stability analysis of steady states in partial differential equations” [Appl. Math. Comput. 483 (2024) 128985], Applied Mathe- matics and Computation 495 (2025) 129319

2024
[22]

Fabiani, F

G. Fabiani, F. Calabr` o, L. Russo, C. Siettos, Numerical solution and bifurcation analysis of nonlinear partial differential equations with ex- treme learning machines, Journal of Scientific Computing 89 (2) (2021) 1–35

2021
[23]

Galaris, G

E. Galaris, G. Fabiani, I. Gallos, I. Kevrekidis, C. Siettos, Numerical bi- furcation analysis of PDEs from Lattice Boltzmann model simulations: a parsimonious machine learning approach, Journal of Scientific Com- puting 92 (2) (2022) 34. 27

2022
[24]

D. D. Joseph, T. S. Lundgren, Quasilinear Dirichlet problems driven by positive sources, Archive for Rational Mechanics and Analysis 49 (1973) 241–269

1973
[25]

E. J. Allen, J. A. Burns, D. S. Gilliam, Numerical approximations of the dynamical system generated by Burgers’ equation with Neumann- Dirichlet boundary conditions, ESAIM: Mathematical Modelling and Numerical Analysis 47 (5) (2013) 1465–1492

2013
[26]

J. R. Graef, C. Qian, B. Yang, A three point boundary value problem for nonlinear fourth order differential equations, Journal of Mathematical Analysis and Applications 287 (1) (2003) 217–233

2003
[27]

Liao, Homotopy analysis method in non-linear differential equations (2012)

S. Liao, Homotopy analysis method in non-linear differential equations (2012)

2012
[28]

E. K. Chong, S. H. Zak, An Introduction to Optimization, Vol. 75, John Wiley & Sons, 2013

2013
[29]

Levenberg, A method for the solution of certain non-linear problems in least squares, Quarterly of Applied Mathematics 2 (2) (1944) 164–168

K. Levenberg, A method for the solution of certain non-linear problems in least squares, Quarterly of Applied Mathematics 2 (2) (1944) 164–168

1944
[30]

D. W. Marquardt, An algorithm for least-squares estimation of non- linear parameters, Journal of the Society for Industrial and Applied Mathematics 11 (2) (1963) 431–441

1963
[31]

D. C. Liu, J. Nocedal, On the limited memory BFGS method for large scale optimization, Mathematical Programming 45 (1) (1989) 503–528

1989
[32]

D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014)

2014 arXiv
[33]

Moseley, A

B. Moseley, A. Markham, T. Nissen-Meyer, Finite basis physics- informed neural networks (FBPINNs): a scalable domain decomposition approach for solving differential equations, Advances in Computational Mathematics 49 (4) (2023) 62

2023
[34]

A. D. Jagtap, G. E. Karniadakis, Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equa- tions, Communications in Computational Physics 28 (5) (2020). 28

2020
[35]

Heinlein, A

A. Heinlein, A. Klawonn, M. Lanser, J. Weber, Combining machine learning and domain decomposition methods for the solution of par- tial differential equations-a review, GAMM-Mitteilungen 44 (1) (2021) e202100001

2021
[36]

M. L. Shahab, H. Susanto, H. Hatzikirou, A finite difference method with symmetry properties for the high-dimensional Bratu equation, Applied Mathematics and Computation 489 (2025) 129136

2025
[37]

Pedregosa, G

F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al., Scikit-learn: Machine learning in Python, The Journal of Machine Learning Research 12 (2011) 2825–2830

2011
[38]

Abadi, P

M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat, G. Irving, M. Isard, et al., TensorFlow: a system for large- scale machine learning, in: 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16), 2016, pp. 265–283

2016
[39]

R. H. Byrd, S. L. Hansen, J. Nocedal, Y. Singer, A stochastic quasi- Newton method for large-scale optimization, SIAM Journal on Opti- mization 26 (2) (2016) 1008–1031

2016
[40]

Kovalev, K

D. Kovalev, K. Mishchenko, P. Richt´ arik, Stochastic Newton and cubic Newton methods with simple local linear-quadratic rates, arXiv preprint arXiv:1912.01597 (2019)

2019 arXiv
[41]

Burke, E

J. Burke, E. Knobloch, Localized states in the generalized Swift- Hohenberg equation, Physical Review E—Statistical, Nonlinear, and Soft Matter Physics 73 (5) (2006) 056211

2006
[42]

Kusdiantara, H

R. Kusdiantara, H. Susanto, F. Adriano, N. Karjanto, Analysis of mul- tistability in discrete quantum droplets and bubbles, Chaos, Solitons & Fractals 187 (2024) 115410

2024
[43]

Kusdiantara, F

R. Kusdiantara, F. T. Akbar, N. Nuraini, B. E. Gunara, H. Susanto, Snakes on lieb lattice, Journal of Nonlinear Science 32 (4) (2022) 59

2022
[44]

Kusdiantara, H

R. Kusdiantara, H. Susanto, Snakes in square, honeycomb and triangu- lar lattices, Nonlinearity 32 (12) (2019) 5170

2019
[45]

H. D. Mazraeh, K. Parand, Approximate symbolic solutions to differen- tial equations using a novel combination of Monte Carlo tree search and physics-informed neural networks approach, Engineering with Comput- ers (2025) 1–29. 29

2025
[46]

H. D. Mazraeh, K. Parand, An innovative combination of deep Q- networks and context-free grammars for symbolic solutions to differ- ential equations, Engineering Applications of Artificial Intelligence 142 (2025) 109733

2025
[47]

H. D. Mazraeh, K. Parand, GEPINN: An innovative hybrid method for a symbolic solution to the Lane–Emden type equation based on gram- matical evolution and physics-informed neural networks, Astronomy and Computing 48 (2024) 100846

2024
[48]

H. D. Mazraeh, K. Parand, A three-stage framework combining neural networks and Monte Carlo tree search for approximating analytical solu- tions to the Thomas–Fermi equation, Journal of Computational Science 87 (2025) 102582. Appendix A. Effect of Varying α and γ The parameter...

2025

Pith tools

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

[1] [1]

Cuomo, V

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

work page 2022

[2] [2]

Hornik, M

K. Hornik, M. Stinchcombe, H. White, Multilayer feedforward networks are universal approximators, Neural Networks 2 (5) (1989) 359–366. 25

work page 1989

[3] [3]

Dissanayake, N

M. Dissanayake, N. Phan-Thien, Neural-network-based approximations for solving partial differential equations, Communications in Numerical Methods in Engineering 10 (3) (1994) 195–201

work page 1994

[4] [4]

I. E. Lagaris, A. Likas, D. I. Fotiadis, Artificial neural networks for solving ordinary and partial differential equations, IEEE Transactions on Neural Networks 9 (5) (1998) 987–1000

1998

[5] [5]

Raissi, P

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

2019

[6] [6]

S. Cai, Z. Mao, Z. Wang, M. Yin, G. E. Karniadakis, Physics-informed neural networks (PINNs) for fluid mechanics: A review, Acta Mechanica Sinica 37 (12) (2021) 1727–1738

2021

[7] [7]

Y. Chen, L. Lu, G. E. Karniadakis, L. Dal Negro, Physics-informed neural networks for inverse problems in nano-optics and metamaterials, Optics Express 28 (8) (2020) 11618–11633

work page 2020

[8] [8]

J. Han, A. Jentzen, E. Weinan, Solving high-dimensional partial dif- ferential equations using deep learning, Proceedings of the National Academy of Sciences 115 (34) (2018) 8505–8510

work page 2018