Pith. sign in

REVIEW 5 major objections 7 minor 77 references

BridgeNet: A Hybrid, Physics-Informed Machine Learning Framework for Solving High-Dimensional Fokker-Planck Equations

T0 review · 5 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper's central claim is that adding convolutional layers to a physics-informed network makes Fokker-Planck solvers more accurate, faster, and more stable than standard PINNs.

desk verdict The paper's central accuracy claims collapse because the stated exact solutions for Examples 2–5 do not satisfy the Fokker–Planck equations they are meant to solve. read the letter →

arxiv 2506.04354 v4 pith:GS2LSMVV submitted 2025-06-04 physics.comp-ph cs.LGmath-phmath.APmath.MP

Elmira Mirzabeigi , Rezvan Salehi , Kourosh Parand This is my paper
classification physics.comp-phcs.LGmath-phmath.APmath.MP MSC 68T0768T2035Q8435C9935G2068W25
keywords Fokker-Planckequationsphysics-informedneuralnetworksconvolutionalhigh-dimensionalPDEshybridmachinelearningdynamiclossweightingPINNnumericalsolver
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

BridgeNet is a hybrid solver that combines convolutional layers with physics-informed constraints to solve linear and nonlinear Fokker-Planck equations. The paper's central claim is that this architecture captures local spatial structure that fully connected PINNs miss, reaching reported mean squared errors orders of magnitude smaller with faster and more stable convergence across one-dimensional and three-dimensional tests. The mechanism is a dynamic loss that penalizes the PDE residual, boundary conditions, and initial conditions, with weights $\alpha$, $\beta$, $\gamma$ updated by a greedy sensitivity search during training. A sympathetic reader would care because a cheap, stable high-dimensional Fokker-Planck solver would make stochastic modeling in finance, biology, and complex-systems physics more tractable with neural-network methods.

What carries the argument

The central object is the BridgeNet architecture: conv-nD layers with kernel size 3 and padding 1, channel depths 256 to 128 to 64, adaptive average pooling, and two dense layers taking the feature dimension from 64 to 32 to a single output. The training signal is the composite loss $L=\alpha L_{\text{residual}}+\beta L_{\text{bc}}\cdot\mathbf{1}_{\text{PDE}}+\gamma L_{\text{ic}}$, where $L_{\text{residual}}$ substitutes the network output into the Fokker-Planck operator, $L_{\text{bc}}$ and $L_{\text{ic}}$ enforce boundary and initial data, and a greedy sensitivity search adjusts $\alpha$, $\beta$, and $\gamma$ so that the loss component with the largest influence on total error gets the most weight. This mechanism carries the argument: the CNN layers supply local spatial features, and the weighted physics terms steer the solution toward the PDE manifold. The paper attributes BridgeNet's accuracy, stability, and fast convergence to this combination rather than to a convergence theorem.

What would settle it

For Example 2, substitute $u(x,t)=x e^t$ into Eq.\ (5) with $\nu(x)=x$ and $\kappa(x)=x^2/2$ on $[0,1]\times[0,0.5]$; the residual $u_t+\partial_x(\nu u)-\tfrac12\partial_{xx}(\kappa u)$ equals $1.5\,x e^t$, which is not zero. Recomputing the paper's error tables after correcting or replacing that benchmark would settle whether the claimed errors are genuine.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that a CNN feature-extraction front end, three convolutional stages with channel depths 256, 128, and 64 followed by adaptive pooling and dense layers, can serve as the interior of a physics-informed solver and outperform a standard PINN on every Fokker-Planck example tested. The authors state that BridgeNet is the first framework to integrate CNN-based spatial processing with physics-informed constraints, and that the reported MSE, MAE, and $L^\infty$ errors are significantly lower in both training and test phases, including a three-dimensional ecosystem example with Gaussian initial condition and coupled drift and diffusion. The stated reason for the improvement is that weight sharing and local receptive fields let the network learn spatial correlations that fully connected layers treat independently, while the physics-informed loss keeps predictions consistent with the governing equations.

Load-bearing premise

The accuracy claims depend wholly on the closed-form exact solutions in Examples 2 through 5 being true solutions of the stated Fokker-Planck equations; if substitution fails, the reported errors no longer measure accuracy.

Editorial extensions

If this is right

  • If the comparisons are correct, BridgeNet achieves error metrics several orders of magnitude below standard PINNs on the tested one-dimensional linear and nonlinear Fokker-Planck equations.
  • The three-dimensional results imply that convolutional locality mitigates the curse of dimensionality enough to keep training and testing MSE near $10^{-12}$ while a standard PINN stays near $10^{-6}$.
  • Adaptive weighting of residual, boundary, and initial-condition losses removes manual tuning of the physics-constraint balance, which should transfer to other PDEs.
  • The reported convergence at initialization for the tuned high-dimensional configuration would make solution recovery nearly instant once the network is trained, albeit with higher training cost and $O(n^{10})$ complexity.
  • The authors themselves list open limits: scaling beyond three dimensions, generalizing to other PDE classes, and reducing the CNN computational overhead.

Reading between the lines

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

  • Editorial extension: the reported gains mix two changes, the CNN architecture and the adaptive loss weighting, so the sharp test is an ablation that changes one at a time; the paper's ablation removes physics but does not isolate the weighting scheme.
  • Editorial extension: because the CNN input represents spatial coordinates on a structured grid-like layout, extending the method to irregular or non-Cartesian domains would require point-based or graph convolutions, and whether the local-connectivity benefit survives there is untested.
  • Editorial extension: the asymptotic complexity reported for BridgeNet is $O(n^{10})$, so in dimensions four through six the per-iteration cost may grow steeply; running the same benchmark in those dimensions would show whether the convergence-speed advantage outweighs the cost.
  • Editorial extension: a corrected benchmark with verified exact solutions or manufactured solutions would tell readers whether the reported accuracies are the method's achievements or artifacts of the chosen test problems.
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

5 major / 7 minor

Summary. The paper proposes BridgeNet, a hybrid framework that combines convolutional neural network layers with physics-informed neural network training to solve linear and nonlinear Fokker-Planck equations. The method incorporates a loss function with adaptive weights for the residual, boundary, and initial condition terms, and is evaluated on five one-dimensional examples, one three-dimensional example, and an exponential growth ODE in the appendix. The paper claims that BridgeNet achieves lower error metrics, faster convergence, and better stability than standard PINNs across these benchmarks.

Significance. If the reported results were valid, the combination of CNN-based local feature extraction with physics-informed constraints would be a useful engineering contribution, especially for structured spatial domains. The paper also includes an ablation study and a hyperparameter sensitivity analysis, which are appropriate validation steps. However, the core validation is invalid: the reference solutions for Examples 2-5 do not satisfy the stated Fokker-Planck equations, and the high-dimensional example lacks a stated reference solution. The model selection on the test set further compromises the independence of the reported accuracy claims. The manuscript therefore does not support its central claims of accuracy, stability, and superiority over PINNs.

major comments (5)
  1. [§5.1, Examples 2-5] The stated exact solutions do not satisfy the given Fokker-Planck equations. Direct substitution into Eq. (5) gives nonzero residuals: for Example 2 (ν=x, κ=x²/2, F=x e^t) the residual is 1.5 x e^t; for Example 3 (ν=-(x+1), κ=x² e^t, F=(x+1)e^t) the residual is -(x+1)e^t - (3x+1)e^{2t}; for Example 4 (ν=4F/x - x/3, κ=F, F=x² e^t) the residual is 6x² e^{2t}; for Example 5 (ν=(7/2)F, κ=xF, F=x/(t+1)) the residual is 3x/(t+1)². Because the error metrics in Tables 4-7 are computed against these invalid references, the accuracy comparisons, the PINN baselines, and the stability conclusions in Section 5.1 are uninterpretable.
  2. [§5.2, Eq. (7) and Table 8] The three-dimensional Fokker-Planck equation in Eq. (7) omits the 1/2 prefactor on the diffusion term that appears in the general form in Eq. (6). More importantly, no reference or exact solution is stated for the high-dimensional example, so the errors reported in Tables 8 and 9 have no defined ground truth. The initial condition in Eq. (8) alone does not define a benchmark solution.
  3. [§5.2.2, Table 11] The paper selects Setup E because it has the lowest test MSE and then reports that setup as the method's result. This is model selection on the test set; the test errors are therefore not independent predictions. The comparison with PINN is optimistically biased because the reported BridgeNet numbers come from the best of twelve configurations while no analogous selection procedure is described for the PINN baseline.
  4. [§5.2.2, Table 10] The claim that BridgeNet converges at initialization (0 epochs, 0.51 s) with MSE 1.5×10^-12 is internally inconsistent with the reported training time of 140.4 s and with the need for any training at all. This metric, described as 'converges at initialization', requires clarification or removal; as written it appears physically implausible for a randomly initialized neural network.
  5. [§5.1, Example 3 and Eq. (5)] The drift and diffusion coefficients in Eq. (5) are written as ν(x) and κ(x), but Example 3 sets κ(x)=x² e^t, which is explicitly time-dependent. This inconsistency, together with the invalid reference solution noted above, makes the example ill-posed as stated. The time-dependent form is introduced later in Section 4 but is not used consistently in the examples.
minor comments (7)
  1. [§5.2.2] The paragraph beginning 'Hyper-parameter Sensitivity Analysis:' is duplicated verbatim in the text.
  2. [§5.2.2] The sentence 'Training is performed using both Adam optimizer' should be 'using the Adam optimizer' or should specify which optimizers are compared; as written it is unclear.
  3. [§5.1, Example 4] The drift coefficient ν=4u/x - x/3 contains a 1/x term and is singular at x=0, but the stated domain is [0,1]×[0.5,1]; the treatment of this singularity should be discussed.
  4. [§3.1.2] The text says ReLU is used for linear Fokker-Planck equations, but the examples in Section 5.1 state Hyperbolic Tangent for several linear cases; the description of activation function selection should be reconciled with actual usage.
  5. [§5.2.2, Table 10] The Big-O complexity entries O(n^6), O(n^8), and O(n^10) are asserted without any derivation or explanation of the variable n; this should be clarified or removed.
  6. [§1] The claim that BridgeNet is 'the first framework to integrate CNN-based spatial processing with physics-informed constraints' is an overclaim; existing CNN-PINN hybrids should be cited or the claim should be softened.
  7. [§3.3.1] The phrase 'three distinct models' in the data generation section should be 'three sampling schemes' (uniform, linearly spaced, Latin hypercube) for clarity.

Circularity Check

1 steps flagged · score 6.0 of 10

BridgeNet's headline accuracy is partly circular: Setup E is chosen as the best of twelve configurations on the test set, and that same test error is then reported as the method's independent result; the exact-solution benchmarks are additionally not solutions of the stated FPEs.

  1. fitted input called prediction [Section 5.2.2, 'Hyper-parameter Sensitivity Analysis' (Table 11 discussion); echoed in Table 10 and Section 6.1]
    "Table 11 reports the effect of varying CNN depth (2–4 layers), kernel size (3×3 vs. 5×5) and learning rate (10−3 vs. 10−4) across twelve configurations (Setups A–L). Performance is evaluated in terms of training/testing MSE, training time (s) and convergence speed (epochs to reach the predefined loss threshold). Our Setup E, which employs three convolutional layers, a 3×3 kernel and a learning rate of 10−3, outperforms all alternatives: it achieves the lowest training MSE (1.5×10−12) and testing MSE (2.7×10−12)."

    The test MSE is used as the selection criterion among twelve hyperparameter configurations, and the configuration with the lowest test MSE (Setup E) is then presented as BridgeNet's result: Table 10 reports MSE 1.5×10−12 for BridgeNet, and Section 6.1 claims 'BridgeNet achieved an MSE as low as 10−12 for both training and testing datasets.' Since the same test split was used to choose the configuration and to compute the reported number, the 'prediction' is not an independent out-of-sample result; it is the argmin of the test-error criterion by construction. This is the classic selected-on-the-test-set form of fitted input called prediction.

full rationale

No load-bearing self-citation or uniqueness-imported-from-authors pattern appears: the paper's own prior works (refs. 3, 10, 63) are background citations and do not supply the core argument. The central derivation—using a CNN to minimize a PINN loss—is not circular in itself. However, the paper's central quantitative claims are partially circular in evaluation: the hyperparameter study in Table 11 selects Setup E by lowest testing MSE and then reports that same selected value as BridgeNet's accuracy, so the headline error metrics are optimistically selected statistics rather than independent predictions. Separately, the claimed 'precise solutions' in Examples 2–5 do not satisfy Eq. (5): direct substitution yields nonzero residuals (e.g., Example 2 gives residual 1.5 x e^t), and the high-dimensional example reports errors without stating an exact reference solution or the correct 1/2 prefactor. These are benchmark-validity/correctness defects that further void the reported accuracy comparisons, but they are not themselves circularity; they are why the accuracy claims cannot be accepted. On the circularity axis, the test-set selection is the one concrete reduction: a number chosen because it is smallest is then quoted as the method's measured performance.

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

The central claim rests on four assumptions, none of which is independently verified in the paper: the exact-solution benchmarks (false), the existence of a convolution-friendly grid (unspecified), the presence of a 3D reference solution (undefined), and the standard behavior of the optimizers. The free parameters are the loss weights and the chosen CNN hyperparameters, both fit during training or on the test set.

free parameters (3)
  • Loss weights α, β, γ = Algorithm 1 updates; as written, all approach 0
    The total loss is linear in α, β, and γ, so each sensitivity equals the corresponding loss component, which is nonnegative; the update subtracts lr times a positive value, so the weights are pushed to their lower bound. The paper does not state a stopping rule that preserves nonzero weights. These weights control the physics constraints, so the reported solution depends on them.
  • CNN architecture (3 layers, channels 256/128/64, kernel 3, padding 1) = Setup E (Table 11)
    Chosen by hyperparameter sweep; Setup E was declared optimal based on the lowest training/test MSE, i.e., selected on the test set.
  • Learning rate schedule = e.g., 1e-4 to 1e-2 by example; Adam then LBFGS for some cases
    Per-example values are hand-picked; no sensitivity analysis for the main results.
assumptions (4)
  • domain assumption The stated analytic expressions solve the Fokker-Planck equations as written (Section 5.1, Examples 2 through 5).
    Direct substitution shows these expressions do not satisfy Eq. (5), e.g., Example 2 residual 1.5 x e^t. This assumption is load-bearing because all error metrics are computed against these expressions.
  • ad hoc to paper A structured grid exists so that conv-nD layers can be applied to the collocation points (Section 3.1.1, Figure 1).
    The paper does not specify how arbitrary sampled points (uniform, LHS) are reshaped into a tensor with spatial adjacency for convolution; without this, the CNN block is undefined.
  • domain assumption The 3D FPE (Eq. 7) with the given drift and diffusion has a known reference solution used to compute Table 8 errors.
    No reference solution is stated for the 3D case; Table 8 reports MSE, MAE, and L-infinity without defining the ground truth.
  • standard math Automatic differentiation and Adam/LBFGS optimizers compute the residual gradients correctly.
    Standard assumption for PINN training; not verified by the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BridgeNet: A Hybrid, Physics-Informed Machine Learning Framework for Solving High-Dimensional Fokker-Planck Equations." pith.science (2026). https://pith.science/paper/GS2LSMVV

@misc{pith2026250604354,
  author       = {Pith},
  title        = {Pith review of: BridgeNet: A Hybrid, Physics-Informed Machine Learning Framework for Solving High-Dimensional Fokker-Planck Equations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GS2LSMVV}},
  note         = {Machine review of arXiv:2506.04354}
}
read the original abstract

BridgeNet is a novel hybrid framework that integrates convolutional neural networks with physics-informed neural networks to efficiently solve non-linear, high-dimensional Fokker-Planck equations (FPEs). Traditional PINNs, which typically rely on fully connected architectures, often struggle to capture complex spatial hierarchies and enforce intricate boundary conditions. In contrast, BridgeNet leverages adaptive CNN layers for effective local feature extraction and incorporates a dynamically weighted loss function that rigorously enforces physical constraints. Extensive numerical experiments across various test cases demonstrate that BridgeNet not only achieves significantly lower error metrics and faster convergence compared to conventional PINN approaches but also maintains robust stability in high-dimensional settings. This work represents a substantial advancement in computational physics, offering a scalable and accurate solution methodology with promising applications in fields ranging from financial mathematics to complex system dynamics.

Figures

Figures reproduced from arXiv: 2506.04354 by the authors.

Figure 1
Figure 1. A schematic representation of a Convolutional Neural Network (CNN) integrated with physics-informed neural networks (PINNs) for solving high-dimensional Fokker-Planck equations is illustrated. The input consists of a vector of spatial (X) and temporal (t) data, where X is represented as an n-1 dimensional matrix. The data is split into training (90%) and testing (10%) datasets. The CNN block processes the input thro… view at source ↗
Figure 2
Figure 2. (Left) Residual function; (Right) Ln of the loss function for the FPEs used to [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. (Left) Residual function; (Right) Ln of the loss function for the FPEs used to [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: (Left) Residual function; (Right) Ln of the loss function for the FPEs used to [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: (Left) Residual function; (Right) Ln of the loss function for the FPEs used to [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: (Left) Residual function; (Right) Ln of the loss function for the FPEs used to [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Ln of loss function for the high-dimensional of FPEs used to compare the [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]
Figure 8
Figure 8. Figure 8: (a) Comparison of loss convergence for Standard PINN, BridgeNet, and Pure CNN. [PITH_FULL_IMAGE:figures/full_fig_p025_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 70 canonical work pages

  1. [1]

    M. M. Moayeri, M. Hemami, Solving partial differential equations by ls- svm, in: Learning with Fractional Orthogonal Kernel Classifiers in Support VectorMachines: Theory, AlgorithmsandApplications, Springer, 2023, pp. 171–198

    work page 2023

  2. [2]

    A. H. H. Rasanan, N. J. Evans, J. Rieskamp, J. A. Rad, Numerical ap- proximation of the first-passage time distribution of time-varying diffusion decision models: A mesh-free approach, Engineering Analysis with Bound- ary Elements 151 (2023) 227–243

    work page 2023

  3. [3]

    Designing of knowledge-based potentials via B-spline basis functions for native proteins detection

    E. Mirzabeigi, S. Mortezazadeh, R. Salehi, H. Naderi-Manesh, Designing of knowledge-based potentials via b-spline basis functions for native proteins detection, arXiv preprint arXiv:2305.05728 (2023). 29

    work page Pith review arXiv 2023

  4. [4]

    Eslahchi, S

    M. Eslahchi, S. Esmaili, N. Namaki, R. Salehi, Application of finite differ- ence method in solving a second-and fourth-order pde blending denoising model, Mathematical Sciences (2021) 1–14

    work page 2021

  5. [5]

    A. Raj, U. Simsekli, A. Rudi, Efficient sampling of stochastic differential equations with positive semi-definite models, Advances in Neural Informa- tion Processing Systems 36 (2024)

    work page 2024

  6. [6]

    A. H. H. Rasanan, N. J. Evans, J. Rieskamp, J. A. Rad, Response time and accuracy modeling through the lens of fractional dynamics: A foundational primer, in: Computation and Modeling for Fractional Order Systems, El- sevier, 2024, pp. 1–27

    work page 2024

  7. [7]

    Marchetti, A fast surrogate cross validation algorithm for meshfree rbf collocation approaches, Applied Mathematics and Computation 481 (2024) 128943

    F. Marchetti, A fast surrogate cross validation algorithm for meshfree rbf collocation approaches, Applied Mathematics and Computation 481 (2024) 128943

    work page 2024

  8. [8]

    Risken, The fokker-planck equation (1996)

    H. Risken, The fokker-planck equation (1996)

    work page 1996

Show all 77 references
  1. [9]

    Elgin, The fokker-planck equation: methods of solution and applications, Optica Acta: International Journal of Optics 31 (11) (1984) 1206–1207

    J. Elgin, The fokker-planck equation: methods of solution and applications, Optica Acta: International Journal of Optics 31 (11) (1984) 1206–1207

    1984

  2. [10]

    A. H. Karami, S. Rezaee, E. Mirzabeigi, K. Parand, Comparison of pre- training and classification models for early detection of alzheimer’s disease using magnetic resonance imaging, in: 8th International Conference on Combinatorics Cryptography, Computer Science and Computation, 2023

    2023

  3. [11]

    S.Wojtkiewicz, L.Bergman, Numericalsolutionofhighdimensionalfokker- planck equations, in: 8th ASCE Specialty Conference on Probablistic Me- chanics and Structural Reliability, Notre Dame, IN, USA, Citeseer, 2000

    2000

  4. [12]

    Panju, K

    M. Panju, K. Parand, A. Ghodsi, Symbolically solving partial differential equations using deep learning, arXiv preprint arXiv:2011.06673 (2020)

    2020 arXiv

  5. [13]

    Omidi, B

    M. Omidi, B. Arab, A. H. Rasanan, J. Rad, K. Parand, Learning nonlinear dynamics with behavior ordinary/partial/system of the differential equa- tions: looking through the lens of orthogonal neural networks, Engineering with Computers 38 (Suppl 2) (2022) 1635–1654

    2022

  6. [14]

    Hajimohammadi, F

    Z. Hajimohammadi, F. Baharifard, A. Ghodsi, K. Parand, Fractional chebyshev deep neural network (fcdnn) for solving differential models, Chaos, Solitons & Fractals 153 (2021) 111530

    2021

  7. [15]

    Sirignano, K

    J. Sirignano, K. Spiliopoulos, Dgm: A deep learning algorithm for solving partial differential equations, Journal of computational physics 375 (2018) 1339–1364

    2018

  8. [16]

    L. Lu, R. Pestourie, S. G. Johnson, G. Romano, Multifidelity deep neural operators for efficient learning of partial differential equations with appli- cation to fast inverse design of nanoscale heat transport, Physical Review Research 4 (2) (2022) 023210. 30

    2022

  9. [17]

    Babaei, K

    M. Babaei, K. M. Mohammadi, Z. Hajimohammadi, K. Parand, Jdnn: Jacobi deep neural network for solving telegraph equation, arXiv preprint arXiv:2212.12700 (2022)

    2022 arXiv

  10. [18]

    Hajimohammadi, K

    Z. Hajimohammadi, K. Parand, A. Ghodsi, Legendre deep neural network (ldnn) and its application for approximation of nonlinear volterra fredholm hammerstein integral equations, arXiv preprint arXiv:2106.14320 (2021)

    2021 arXiv

  11. [19]

    M. M. Moayeri, J. Rad, K. Parand, Dynamical behavior of reaction– diffusion neural networks and their synchronization arising in modeling epileptic seizure: a numerical simulation study, Computers & Mathematics with Applications 80 (8) (2020) 1887–1927

    2020

  12. [20]

    A. H. Hadian-Rasanan, D. Rahmati, S. Gorgin, K. Parand, A single layer fractional orthogonal neural network for solving various types of lane– emden equation, New Astronomy 75 (2020) 101307

    2020

  13. [21]

    J. A. Rad, K. Rashedi, K. Parand, H. Adibi, The meshfree strong form methods for solving one dimensional inverse cauchy-stefan problem, Engi- neering with Computers 33 (2017) 547–571

    2017

  14. [22]

    Rabiei, K

    K. Rabiei, K. Parand, Collocation method to solve inequality-constrained optimal control problems of arbitrary order, Engineering with Computers 36 (1) (2020) 115–125

    2020

  15. [23]

    Parand, H

    K. Parand, H. Yari, M. Delkhosh, Solving two-dimensional integral equa- tions of the second kind on non-rectangular domains with error estimate, Engineering with Computers 36 (2020) 725–739

    2020

  16. [24]

    Hajiollow, Y

    A. Hajiollow, Y. Lotfi, K. Parand, A. Hadian, K. Rashedi, J. Rad, Recover- ing a moving boundary from cauchy data in an inverse problem which arises in modeling brain tumor treatment: the (quasi) linearization idea combined with radial basis functions (rbfs) approximation, Eng...

    2021

  17. [25]

    Lotfi, K

    Y. Lotfi, K. Parand, K. Rashedi, J. A. Rad, Numerical study of tempera- ture distribution in an inverse moving boundary problem using a meshless method, Engineering with Computers 37 (1) (2021) 461–475

    2021

  18. [26]

    M. M. Moayeri, A. H. Hadian-Rasanan, S. Latifi, K. Parand, J. Rad, An efficient space-splitting method for simulating brain neurons by neuronal synchronization to control epileptic activity, Engineering with Computers 38 (1) (2022) 819–846

    2022

  19. [27]

    Baharifard, K

    F. Baharifard, K. Parand, M. Rashidi, Novel solution for heat and mass transfer of a mhd micropolar fluid flow on a moving plate with suction and injection, Engineering with Computers 38 (1) (2022) 13–30. 31

    2022

  20. [28]

    Parand, M

    K. Parand, M. Razzaghi, R. Sahleh, M. Jani, Least squares support vector regression for solving volterra integral equations, Engineering with Com- puters 38 (Suppl 1) (2022) 789–796

    2022

  21. [29]

    Asghari, A

    M. Asghari, A. H. Hadian Rasanan, S. Gorgin, D. Rahmati, K. Parand, Fpga-orthopoly: a hardware implementation of orthogonal polynomials, Engineering with Computers 39 (3) (2023) 2257–2276

    2023

  22. [30]

    Hajimohammadi, S

    Z. Hajimohammadi, S. Shekarpaz, K. Parand, The novel learning solutions to nonlinear differential models on a semi-infinite domain, Engineering with Computers 39 (3) (2023) 2169–2186

    2023

  23. [31]

    Raissi, P

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

    2019

  24. [32]

    X. Meng, G. E. Karniadakis, A composite neural network that learns from multi-fidelity data: Application to function approximation and inverse pde problems, Journal of Computational Physics 401 (2020) 109020

    2020

  25. [33]

    Y. Shin, J. Darbon, G. E. Karniadakis, On the convergence of physics informed neural networks for linear second-order elliptic and parabolic type pdes, arXiv preprint arXiv:2004.01806 (2020)

    2020 arXiv

  26. [34]

    Z. Hu, A. D. Jagtap, G. E. Karniadakis, K. Kawaguchi, When do extended physics-informed neural networks (xpinns) improve generalization?, arXiv preprint arXiv:2109.09444 (2021)

    2021 arXiv

  27. [35]

    Namaki, M

    N. Namaki, M. Eslahchi, R. Salehi, The use of physics-informed neural network approach to image restoration via nonlinear pde tools, Computers & Mathematics with Applications 152 (2023) 355–363

    2023

  28. [36]

    A. D. Jagtap, Z. Mao, N. Adams, G. E. Karniadakis, Physics-informed neural networks for inverse problems in supersonic flows, Journal of Com- putational Physics 466 (2022) 111402

    2022

  29. [37]

    L. Liu, S. Liu, H. Xie, F. Xiong, T. Yu, M. Xiao, L. Liu, H. Yong, Dis- continuity computing using physics-informed neural networks, Journal of Scientific Computing 98 (1) (2024) 22

    2024

  30. [38]

    Donnelly, A

    J. Donnelly, A. Daneshkhah, S. Abolfathi, Physics-informed neural net- works as surrogate models of hydrodynamic simulators, Science of the Total Environment 912 (2024) 168814

    2024

  31. [39]

    H. Lu, Q. Wang, W. Tang, H. Liu, Physics-informed neural networks for fully non-linear free surface wave propagation, Physics of Fluids 36 (6) (2024). 32

    2024

  32. [40]

    Babaei, A

    M. Babaei, A. A. Aghaei, Z. Kazemi, M. Jamshidi, R. Ghaderi, K. Parand, Solving a class of thomas–fermi equations: A new solution concept based on physics-informed machine learning, Mathematics and Computers in Simu- lation (2024)

    2024

  33. [41]

    H. D. Mazraeh, K. Parand, Gepinn: An innovative hybrid method for a symbolic solution to the lane-emden type equation based on grammatical evolution and physics-informed neural networks, Astronomy and Comput- ing (2024) 100846

    2024

  34. [42]

    Zhang, Z

    Y. Zhang, Z. Jin, L. Wang, K. Zheng, W. Jia, Stochastic dynamics of air- craft ground taxiing via improved physics-informed neural networks, Non- linear Dynamics 112 (5) (2024) 3163–3178

    2024

  35. [43]

    Lahariya, F

    M. Lahariya, F. Karami, C. Develder, G. Crevecoeur, Physics-informed recurrent neural networks for the identification of a generic energy buffer system, in: 2021 IEEE 10th Data Driven Control and Learning Systems Conference (DDCLS), IEEE, 2021, pp. 1044–1049

    2021

  36. [44]

    Haitsiukevich, A

    K. Haitsiukevich, A. Ilin, Improved training of physics-informed neural networks with model ensembles, in: 2023 International Joint Conference on Neural Networks (IJCNN), IEEE, 2023, pp. 1–8

    2023

  37. [45]

    Rafiq, G

    M. Rafiq, G. Rafiq, G. S. Choi, Dsfa-pinn: Deep spectral feature aggrega- tion physics informed neural network, IEEE Access 10 (2022) 22247–22259

    2022

  38. [46]

    Zobeiry, K

    N. Zobeiry, K. D. Humfeld, A physics-informed machine learning approach for solving heat transfer equation in advanced manufacturing and engi- neering applications, Engineering Applications of Artificial Intelligence 101 (2021) 104232

    2021

  39. [47]

    R.Pan, F.Xiao, M.Shen, ro-pinn: Areducedorderphysics-informedneural network for solving the macroscopic model of pedestrian flows, Transporta- tion Research Part C: Emerging Technologies 163 (2024) 104658

    2024

  40. [48]

    Lee, Anti-derivatives approximator for enhancing physics-informed neu- ral networks, Computer Methods in Applied Mechanics and Engineering 426 (2024) 117000

    J. Lee, Anti-derivatives approximator for enhancing physics-informed neu- ral networks, Computer Methods in Applied Mechanics and Engineering 426 (2024) 117000

    2024

  41. [49]

    Eshkofti, S

    K. Eshkofti, S. M. Hosseini, A new modified deep learning technique based on physics-informed neural networks (pinns) for the shock-induced coupled thermoelasticity analysis in a porous material, Journal of Thermal Stresses 47 (6) (2024) 798–825

    2024

  42. [50]

    Xiang, W

    Z. Xiang, W. Peng, X. Liu, W. Yao, Self-adaptive loss balanced physics- informed neural networks, Neurocomputing 496 (2022) 11–34

    2022

  43. [51]

    A. F. Psaros, K. Kawaguchi, G. E. Karniadakis, Meta-learning pinn loss functions, Journal of computational physics 458 (2022) 111121. 33

    2022

  44. [52]

    Habib, U

    A. Habib, U. Yildirim, Developing a physics-informed and physics- penalized neural network model for preliminary design of multi-stage fric- tion pendulum bearings, Engineering Applications of Artificial Intelligence 113 (2022) 104953

    2022

  45. [53]

    M. Yang, J. T. Foster, Multi-output physics-informed neural networks for forward and inverse pde problems with uncertainties, Computer Methods in Applied Mechanics and Engineering 402 (2022) 115041

    2022

  46. [54]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, Advances in neural information processing systems 25 (2012)

    2012

  47. [55]

    Palleschi, F

    V. Palleschi, F. Sarri, G. Marcozzi, M. Torquati, Numerical solution of the fokker-planck equation: A fast and accurate algorithm, Physics Letters A 146 (7-8) (1990) 378–386

    1990

  48. [56]

    Palleschi, M

    V. Palleschi, M. De Rosa, Numerical solution of the fokker-planck equation. ii. multidimensional case, Physics Letters A 163 (5-6) (1992) 381–391

    1992

  49. [57]

    Vanaja, Numerical solution of a simple fokker-planck equation, Applied Numerical Mathematics 9 (6) (1992) 533–540

    V. Vanaja, Numerical solution of a simple fokker-planck equation, Applied Numerical Mathematics 9 (6) (1992) 533–540

    1992

  50. [58]

    M. P. Zorzano, H. Mais, L. Vazquez, Numerical solution of two dimensional fokker—planck equations, Applied mathematics and computation 98 (2-3) (1999) 109–117

    1999

  51. [59]

    Dehghan, M

    M. Dehghan, M. Tatari, The use of he’s variational iteration method for solving a fokker–planck equation, Physica Scripta 74 (3) (2006) 310

    2006

  52. [60]

    M.Tatari, M.Dehghan, M.Razzaghi, Applicationoftheadomiandecompo- sition method for the fokker–planck equation, Mathematical and Computer Modelling 45 (5-6) (2007) 639–650

    2007

  53. [61]

    Lakestani, M

    M. Lakestani, M. Dehghan, Numerical solution of fokker-planck equation using the cubic b-spline scaling functions, Numerical Methods for Partial Differential Equations: An International Journal 25 (2) (2009) 418–429

    2009

  54. [62]

    Kazem, J

    S. Kazem, J. Rad, K. Parand, Radial basis functions methods for solv- ing fokker–planck equation, Engineering Analysis with Boundary Elements 36 (2) (2012) 181–189

    2012

  55. [63]

    Parand, S

    K. Parand, S. Latifi, M. Moayeri, M. Delkhosh, Generalized lagrange ja- cobi gauss-lobatto (gljgl) collocation method for solving linear and nonlin- ear fokker-planck equations, Communications in Theoretical Physics 69 (5) (2018) 519

    2018

  56. [64]

    Czarnetzki, L

    U. Czarnetzki, L. L. Alves, Describing local and non-local electron heating by the fokker–planck equation, Reviews of Modern Plasma Physics 6 (1) (2022) 31. 34

    2022

  57. [65]

    Tabandeh, N

    A. Tabandeh, N. Sharma, L. Iannacone, P. Gardoni, Numerical solution of the fokker–planck equation using physics-based mixture models, Computer Methods in Applied Mechanics and Engineering 399 (2022) 115424

    2022

  58. [66]

    S. Liu, W. Li, H. Zha, H. Zhou, Neural parametric fokker–planck equation, SIAM Journal on Numerical Analysis 60 (3) (2022) 1385–1449

    2022

  59. [67]

    N. M. Boffi, E. Vanden-Eijnden, Probability flow solution of the fokker– planck equation, Machine Learning: Science and Technology 4 (3) (2023) 035012

    2023

  60. [68]

    L. Chen, Z. Yuan, J. Qian, J. Sun, Random vibration of hysteretic systems under poisson white noise excitations, Applied Mathematics and Mechanics 44 (2) (2023) 207–220

    2023

  61. [69]

    Aszhar, I

    J. Aszhar, I. Setiawan, R. Medriati, Method for accelerating equilibrium in perfectly damped brownian motion with harmonic potential, Kasuari: Physics Education Journal (KPEJ) 7 (1) (2024) 31–43

    2024

  62. [70]

    Q. Wang, L. Zu, Dynamical analysis of a delayed stochastic lotka–volterra competitive model in polluted aquatic environments, Qualitative Theory of Dynamical Systems 23 (2) (2024) 80

    2024

  63. [71]

    T. Wang, Z. Hu, K. Kawaguchi, Z. Zhang, G. E. Karniadakis, Tensor neu- ral networks for high-dimensional fokker-planck equations, arXiv preprint arXiv:2404.05615 (2024)

    2024 arXiv

  64. [72]

    Neena, D

    A. Neena, D. P. Clemence-Mkhope, A. Awasthi, Nonstandard finite differ- ence schemes for linear and non-linear fokker–planck equations, Journal of Engineering Mathematics 145 (1) (2024) 11

    2024

  65. [73]

    Zhang, M

    G. Zhang, M. Gunzburger, W. Zhao, A sparse-grid method for multi- dimensional backward stochastic differential equations, Journal of Com- putational Mathematics (2013) 221–248

    2013

  66. [74]

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

    2018

  67. [75]

    Dewar, A root-shoot partitioning model based on carbon-nitrogen-water interactions and munch phloem flow, Functional Ecology (1993) 356–368

    R. Dewar, A root-shoot partitioning model based on carbon-nitrogen-water interactions and munch phloem flow, Functional Ecology (1993) 356–368

    1993

  68. [76]

    Saadatmandi, M

    A. Saadatmandi, M. Dehghan, M.-R. Azizi, The sinc–legendre collocation method for a class of fractional convection–diffusion equations with variable coefficients, Communications in Nonlinear Science and Numerical Simula- tion 17 (11) (2012) 4125–4136

    2012

  69. [77]

    X. Li, D. Zhao, X. Yang, Experimental and theoretical bifurcation study of a nonlinear standing-wave thermoacoustic system, Energy 135 (2017) 553–562. 35 Appendix A. Additional Examples In this appendix, we present a simple example of an ordinary differential equation (ODE) to...

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.