Pith. sign in

REVIEW 6 major objections 7 minor 46 references

Over-PINNs: Enhancing Physics-Informed Neural Networks via Higher-Order Partial Derivative Overdetermination of PDEs

T0 review · 6 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 Over-PINNs gain accuracy by training on the PDE plus its differentiated forms.

desk verdict A re-implementation of gradient-enhanced PINNs under a new name, with a flawed Navier-Stokes validation that decouples vorticity from velocity; the core idea is known and the evidence does not support the central claim. read the letter →

arxiv 2506.05918 v1 pith:HEWGSYCX submitted 2025-06-06 cs.LG

Wenxuan Huo , Qiang He , Gang Zhu , Weifeng Huang This is my paper
classification cs.LG MSC 68T0765M9935Q3535K55
keywords physics-informedneuralnetworksoverdeterminedPDEsystemshigher-orderderivativesautomaticdifferentiationAllen-CahnequationNavier-Stokesequationsvorticitytransportlossfunctionaugmentation
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

Over-PINNs is a training strategy for physics-informed neural networks built on a simple idea: if the network should satisfy a PDE, it should also satisfy the equations obtained by differentiating that PDE with respect to space and time. The paper argues that these higher-order auxiliary equations are compatible with the original equation for smooth solutions, so adding their residuals to the loss function 'overdetermines' the problem and narrows the set of solutions the network can settle on. In the Allen-Cahn test case the $L^2$ relative error falls from $7.520 \times 10^{-3}$ for a standard PINN to $5.932 \times 10^{-4}$ for Over-PINN, and in a two-dimensional Navier-Stokes vortex problem the method keeps vorticity predictions accurate over longer times. A sympathetic reader would take away a recipe that improves accuracy without new data, new network architecture, or major computational overhead.

What carries the argument

The machinery is the higher-order auxiliary residual: for a primal equation, Over-PINN also minimizes the squared residual of its partial derivatives, so the optimizer sees the equation and its derivative envelope simultaneously. The paper justifies compatibility in two ways: for smooth solutions in $C^k$ the differentiated equation is a necessary consequence of the original one, and for coupled systems a Frobenius commutator condition keeps constructed constraints physically consistent. This is what lets the network treat the enlarged system as overdetermined rather than contradictory.

What would settle it

After training Over-PINN on the two-dimensional vortex problem, evaluate the pointwise violation of $\omega_\theta = \partial v_\theta/\partial x - \partial u_\theta/\partial y$ at the collocation points. If the violation is large relative to the reported vorticity error, the long-time improvement is not produced by overdetermination of the velocity solution, and the central fluid-dynamics evidence fails.

Watch

Extended reading notes

Core claim

The central claim is that an overdetermined but compatible system—the original PDE together with one or more of its partial derivatives—carries more usable physical information than the PDE alone, and that a neural network trained against all residuals at once exploits that information. For a scalar equation $L[u]=f$, the paper differentiates in $x$ to get $L_x[u]=f_x$ and adds $\|L_x[u_\theta] - f_x\|^2$ to the loss; for systems it differentiates equation-wise, substitutes auxiliary variables such as vorticity, or builds linear combinations that eliminate variables. Because automatic differentiation computes these derivatives exactly, the auxiliary constraints cost little relative to the gains, and the paper reports roughly an order-of-magnitude accuracy improvement on Allen-Cahn together with slower error growth in the Navier-Stokes experiment. The paper's own framing is that the added constraints shrink the solution space from $S$ to a subset $S'$, and least-squares optimization over that subset is what improves the prediction.

Load-bearing premise

For the Navier-Stokes experiment, the argument depends on the predicted vorticity $\omega$ actually being the vorticity of the predicted velocity field, yet the loss function never enforces $\omega = \partial v/\partial x - \partial u/\partial y$, so the extra vorticity equation may be constraining an independent output rather than the velocity solution.

Editorial extensions

If this is right

  • For scalar PDEs, appending one differentiated residual to the loss is enough to reduce relative error by an order of magnitude in the Allen-Cahn test, from 7.520e-3 to 5.932e-4.
  • The same differentiated-residual construction carries over to coupled systems, either by differentiating each equation or by deriving variable-substituted forms like the vorticity transport equation from Navier-Stokes.
  • Because the extra terms are produced by automatic differentiation, the method can be layered onto existing PINN optimizers and architectures without redesigning the network.
  • Long-time predictions benefit disproportionately: in the vortex evolution test, error grows more slowly with Over-PINN than with standard PINN.
  • The method adds no data requirement, so it improves accuracy in the low-data regime where PINNs are typically used.

Reading between the lines

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

  • If the benefit is mainly from having any compatible extra residual, an ablation that adds only one spatial derivative, or a randomly weighted combination, could reveal how much of the gain comes from overdetermination versus from simple regularization of the loss landscape.
  • For the Navier-Stokes experiment, computing max |omega_theta - (dv_theta/dx - du_theta/dy)| would test whether the vorticity residual is actually tied to the predicted velocity; without a coupling loss term, the reported long-time improvement may reflect a separate vorticity fit rather than overdetermination of the momentum equations.
  • On problems with non-smooth or noisy solutions, higher-order derivatives amplify small-scale error, so the method's accuracy advantage could reverse outside the smooth benchmark cases studied here.
  • The overdetermination view suggests a natural curriculum: train first on the primal equation, then anneal in the differentiated residuals, which might stabilize convergence on stiff 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

6 major / 7 minor

Summary. The paper proposes Over-PINNs, a modification of Physics-Informed Neural Networks (PINNs) in which automatic differentiation is used to generate additional residual terms from partial derivatives of the original PDE. These auxiliary equations are added to the loss function, and the method is demonstrated on the Allen-Cahn equation and on a two-dimensional incompressible Navier-Stokes vortex evolution problem. The authors report a substantial reduction in L2 error for Allen-Cahn (7.520e-3 to 5.932e-4) and claim improved long-time vorticity prediction for Navier-Stokes, while asserting that this is achieved without substantial additional computational cost.

Significance. If validated, the idea of adding differentiated residual terms would be a simple, architecture-agnostic enhancement that could improve PINN accuracy in many settings. The Allen-Cahn result is suggestive, and the derivation of the vorticity transport equation from the Navier-Stokes equations is correct in isolation. However, the current manuscript does not provide a sound theoretical justification or reliable experimental evidence for the central claim: the theory conflates exact solution spaces with minimizers of a non-convex loss, the Navier-Stokes experiment is ill-posed as written, and the cost claim is internally contradicted by the paper's own conclusion.

major comments (6)
  1. [§4.2.3] The loss term L_OE is defined in terms of the pressure p and its spatial derivatives, but the network outputs are stated to be (u,v) or (u,v,ω) (Table 3 and §4.2.3). No pressure output is described, so L_OE is not computable from the network as written. This makes the Navier-Stokes experiment ill-posed.
  2. [§4.2.3, §4.2.4] Nowhere in the loss function is the kinematic relation ω = ∂v/∂x − ∂u/∂y imposed, either as an additional residual or as an architectural constraint. Since L_IC and L_BC constrain only u and v, and L_HE is evaluated on a separately predicted ω, the vorticity transport equation is an independent PDE for a third output rather than an overdetermination constraint on the velocity field. The claimed connection between the vorticity equation and the original system is therefore not realized for the network outputs, so this experiment cannot support the central accuracy claim.
  3. [§2.3] The argument that the over-determined system improves accuracy relies on the exact solution-space inclusion S′ ⊆ S, but the training procedure minimizes a non-convex empirical loss and does not produce exact solutions. The linear least-squares analogy (Ax=b with m>n) does not transfer to this setting; no result connects the inclusion of additional residual terms to a smaller approximation error for the trained network. Thus the claimed 'mathematical proof' of accuracy improvement is not established.
  4. [Abstract and §5] The abstract states that the method achieves improvements 'without incurring substantial additional computational costs,' while the conclusion states that the high-order terms 'substantially increase computational complexity' and can lead to 'prohibitive memory requirements.' These statements directly contradict each other, and the computational-cost component of the central claim is therefore unsupported.
  5. [§4.1] The Allen-Cahn improvement is based on a single training run. No error bars, random seeds, collocation/initial/boundary point counts, or loss weighting coefficients are reported. With only one seed and no sensitivity analysis, the reported L2 drop from 7.520e-3 to 5.932e-4 cannot be distinguished from optimization noise or incidental hyperparameter effects. Since this is the only valid experimental demonstration, the empirical evidence for the central claim is weak.
  6. [§4.2.4] The text announces three comparative configurations for the traditional PINN baseline, but then defines only two loss functions (NS only and vorticity only). The third configuration (NS plus vorticity) is not defined and does not appear in Figure 7, making the comparison incomplete and the plot labels ambiguous.
minor comments (7)
  1. [§2.1] Typo 'cOver-PINNs' in the first bulleted list; it should read 'Over-PINNs.'
  2. [Throughout] The names Over-PINN, OverPINN, and Over-PINNs are used inconsistently; please unify the terminology.
  3. [§4.2.2/§4.2.4] The vorticity is denoted ω in the derivation but w in the simulation section; please make the notation consistent.
  4. [§3] The 'Differential-Algebraic Combinations and Coupled Constraints' conditions (Conditions 1 and 2) are introduced but never used in the numerical experiments; their relevance to the proposed framework should be clarified or removed.
  5. [§4.1] The loss weights λ1, λ2 in L_Total = L_original + λ L_auxiliary are not specified; the experiments appear to use λ=1 implicitly, which should be stated explicitly.
  6. [Figure 4] The 'error metrics' being compared (pointwise absolute error, L2 error, etc.) are not defined in the caption; please clarify.
  7. [References] Several references (e.g., [21], [23], [27]) are incomplete, missing journal, volume, or page information; please standardize the bibliography.

Circularity Check

1 steps flagged · score 4.0 of 10

The overdetermination rationale is self-definitional: the auxiliary equations are exact derivatives of the original PDE, so the claimed solution-space restriction is by construction no restriction; the numerical benchmarks are independent.

  1. self definitional [Section 2.2 and Section 2.3, around the differentiated-equation display and the S′⊆S argument]
    "By differentiating the original equation with respect to the spatial variable xi or the time variable t, a new partial differential equation is obtained: ... Given that the original equation holds within Ω and the solution u meets the smoothness conditions, the above new equation is necessarily valid within Ω and compatible with the original equation. ... Let the solution space of the original equation system be S, and the solution space of the new equation system with high-order equations be S′. Then S′ ⊆ S."

    The auxiliary equation is constructed as the partial derivative of the original equation: if L[u]=f, the added equation is ∂xi(L[u])=∂xi f. Every exact solution of the original equation automatically satisfies the auxiliary equation under the stated smoothness assumptions, so the pair is not an overdetermined system of independent equations. In fact the opposite inclusion also holds, so S′=S, and the claimed 'further restricted' solution space is by construction no restriction at all. The theoretical mechanism for accuracy improvement therefore reduces to the original PDE content; the only independent evidence is the empirical comparison against external reference solutions.

full rationale

The numerical core is not circular: the Allen-Cahn and Navier-Stokes experiments are evaluated against exact or spectral reference solutions with identical architectures and optimizers, and no target quantity is fitted and then reported as a prediction. The one self-definitional step is the theoretical justification in Sections 2.2-2.3, where the 'higher-order auxiliary equations' are exactly the partial derivatives of the original PDE; the asserted solution-space restriction is vacuous because every solution of the original equation already satisfies the differentiated equation. The abstract's claim of 'no substantial additional computational costs' is also contradicted by the conclusion's admission that high-order terms 'substantially increase computational complexity' and may cause 'prohibitive memory requirements'; this is an internal inconsistency rather than a circularity. Separately, the Navier-Stokes implementation has a validity gap: L_HE uses a separately predicted ω without enforcing ω=∂v/∂x−∂u/∂y as a loss term or architectural constraint, so the vorticity equation is not demonstrated to be a consequence of the predicted velocity field. That gap weakens the fluid-dynamics evidence but is not a reduction-by-construction circularity. Overall score reflects the self-definitional overdetermination argument while recognizing that the empirical demonstrations are independently benchmarked.

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

The central claim rests on smoothness of the PDE solution, an unproven link from exact solution spaces to non-convex loss minimization, and, in the Navier-Stokes experiment, an unstated coupling between the output vorticity and the velocity derivatives. The loss weights and sampling counts are unreported free choices. No new physical entities are introduced.

free parameters (3)
  • Loss weighting coefficients lambda1, lambda2 (implicitly 1 in the reported losses) = not reported
    In Section 2.1 the framework is defined with lambda1 and lambda2, but the experiment losses in Sections 4.1 and 4.2.3 sum terms without stated weights or an ablation. The accuracy gain could depend on these weights, so they are a hand-chosen degree of freedom.
  • Collocation, initial, and boundary sampling counts N_c, N_i, N_b = not reported
    The losses are defined as averages over N_c, N_i, and N_b points, but the actual counts are never given. Sampling density controls how strongly different constraints are weighted and affects the reported errors.
  • Fourier embedding dimension and scaling factor = 256, 1
    Used to enforce periodic boundary conditions in both experiments. These are fixed design choices, not fitted to data, but they affect expressivity and are part of the unspecified reproduction setup.
assumptions (5)
  • domain assumption The PDE solution is smooth enough (u in C^k) for all differentiated auxiliary equations to hold pointwise.
    Stated in Section 2.2; Over-PINNs requires differentiating the PDE and the network to high order. The numerical solutions are not verified to have the needed regularity, and the Allen-Cahn initial condition is not C^1-periodic at the boundary, so this regularity assumption is load-bearing.
  • ad hoc to paper If the original PDE residual vanishes on the domain, the differentiated residual also vanishes; conversely, minimizing both residuals in the loss yields a more accurate solution.
    The first half is a standard compatibility fact, but the paper's improvement claim in Section 2.3 relies on the unproved second half: that S' subset S for exact solutions implies better minimizers of a non-convex neural-network loss.
  • domain assumption The extra network output omega in the Navier-Stokes experiment is the vorticity dv/dx - du/dy of the predicted velocity.
    No loss term enforces this relation in Section 4.2.3; L_IC and L_BC constrain only u and v, and L_HE constrains omega via the transport equation. If the relation does not hold, the vorticity constraint is decoupled from the velocity solution.
  • ad hoc to paper Conditions 1 and 2 (rank-deficient Jacobian J_K and commutation [D_alpha,d, L_i] = 0) are sufficient to construct physically admissible eliminated-variable constraints.
    Stated in Section 3 without proof; these conditions are not used in the numerical experiments and no example or theorem establishes sufficiency.
  • standard math Clairaut's theorem: mixed partial derivatives commute for sufficiently smooth fields.
    Used in Section 4.2.2 to cancel pressure and derive the vorticity transport equation; valid under smoothness, but the regularity of the network approximations is not verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Over-PINNs: Enhancing Physics-Informed Neural Networks via Higher-Order Partial Derivative Overdetermination of PDEs." pith.science (2026). https://pith.science/paper/HEWGSYCX

@misc{pith2026250605918,
  author       = {Pith},
  title        = {Pith review of: Over-PINNs: Enhancing Physics-Informed Neural Networks via Higher-Order Partial Derivative Overdetermination of PDEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HEWGSYCX}},
  note         = {Machine review of arXiv:2506.05918}
}
read the original abstract

Partial differential equations (PDEs) serve as the cornerstone of mathematical physics. In recent years, Physics-Informed Neural Networks (PINNs) have significantly reduced the dependence on large datasets by embedding physical laws directly into the training of neural networks. However, when dealing with complex problems, the accuracy of PINNs still has room for improvement. To address this issue, we introduce the Over-PINNs framework, which leverages automatic differentiation (AD) to generate higher-order auxiliary equations that impose additional physical constraints. These equations are incorporated as extra loss terms in the training process, effectively enhancing the model's ability to capture physical information through an "overdetermined" approach. Numerical results illustrate that this method exhibits strong versatility in solving various types of PDEs. It achieves a significant improvement in solution accuracy without incurring substantial additional computational costs.

Figures

Figures reproduced from arXiv: 2506.05918 by the authors.

Figure 1
Figure 1. Architecture of the physics-informed neural network (PINN) framework. The network directly incorporates [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The distribution of the exact solution, the predicted solution, and the absolute error when using the traditional [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. The distribution of the exact solution, the predicted solution, and the absolute error when using the OverPINN [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Error distribution comparison: OverPINN vs. traditional PINN for Allen-Cahn equation solutions [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Initial vorticity distribution (t = 0 s) The numerical solution employs a hybrid Crank-Nicolson/RK4 scheme with spectral discretization: • Viscous terms: Implicit Crank-Nicolson (O(∆t 2 )) • Convective terms: Explicit 4th-order Runge-Kutta • Spectral operations: RFFT f…
Figure 6
Figure 6. Figure 6: Evolution of the vorticity field at integer seconds (numerical solution). [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Comparison of prediction results: (a) Traditional PINN using vorticity transport equation loss, showing exact [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages

  1. [1]

    Brunton and J

    Steven L. Brunton and J. Nathan Kutz. Promising directions of machine learning for partial differential equations. Nature Computational Science

    work page

  2. [2]

    Machine learning solutions looking for pde problems.Nature Machine Intelligence, 7(1):1–1, jan 01 2025

    work page 2025

  3. [3]

    Envisioning better benchmarks for machine learning pde solvers.Nature Machine Intelligence, 7(1):2–3, jan 2025

    Johannes Brandstetter. Envisioning better benchmarks for machine learning pde solvers.Nature Machine Intelligence, 7(1):2–3, jan 2025

    work page 2025

  4. [4]

    Leveque.Numerical Methods for Conservation Laws

    Randall J. Leveque.Numerical Methods for Conservation Laws. Numerical Methods for Conservation Laws, 1992

    work page 1992

  5. [5]

    Comparative performance analysis of numerical discretization methods for electro- chemical model of lithium-ion batteries.Journal of Power Sources, 650:237365, 09 2025

    Feng Guo and Luis Couto. Comparative performance analysis of numerical discretization methods for electro- chemical model of lithium-ion batteries.Journal of Power Sources, 650:237365, 09 2025

    work page 2025

  6. [6]

    Meerschaert and Charles Tadjeran

    Mark M. Meerschaert and Charles Tadjeran. Finite difference approximations for fractional advection–dispersion flow equations.Journal of Computational and Applied Mathematics, 172(1):65–77, 2004

    work page 2004

  7. [7]

    Alikhanov

    Anatoly A. Alikhanov. A new difference scheme for the time fractional diffusion equation.Journal of Computa- tional Physics, 280:424–438, 2015

    work page 2015

  8. [8]

    Guang-Hua Gao, Hai-Wei Sun, and Zhi-Zhong Sun. Stability and convergence of finite difference schemes for a class of time-fractional sub-diffusion equations based on certain superconvergence.Journal of Computational Physics, 280:510–528, 2015

    work page 2015

Show all 46 references
  1. [9]

    Moghaddam and J.A.T

    B.P. Moghaddam and J.A.T. Machado. A stable three-level explicit spline finite difference scheme for a class of nonlinear time variable order fractional partial differential equations.Computers & Mathematics with Applications, 73(6):1262–1269, 2017. Advances in Fractional Diff...

    2017

  2. [10]

    Vadivel, Nallappan Gunasekaran, Haitao Zhu, Jinde Cao, and Xiaodi Li

    Sekar Elango, Ayyadurai Tamilselvan, R. Vadivel, Nallappan Gunasekaran, Haitao Zhu, Jinde Cao, and Xiaodi Li. Finite difference scheme for singularly perturbed reaction diffusion problem of partial delay differential equation with nonlocal boundary condition.Advances in Differ...

    2021

  3. [11]

    Godunov and I

    Sergei K. Godunov and I. Bohachevsky. Finite difference method for numerical computation of discontinuous solutions of the equations of fluid dynamics.Matematiˇ ceskij sbornik, 47(89)(3):271–306, 1959. hal-01620642

    1959

  4. [12]

    Mfem: A modular finite element methods library

    Robert Anderson, Julian Andrej, Andrew Barker, Jamie Bramwell, Jean-Sylvain Camier, Jakub Cerveny, Veselin Dobrev, Yohann Dudouit, Aaron Fisher, Tzanio Kolev, Will Pazner, Mark Stowell, Vladimir Tomov, Ido Akker- man, Johann Dahm, David Medina, and Stefano Zampini. Mfem: A mod...

    2021

  5. [13]

    High-order finite element methods for time-fractional partial differential equations

    Yingjun Jiang and Jingtang Ma. High-order finite element methods for time-fractional partial differential equations. Journal of Computational and Applied Mathematics, 235(11):3285–3290, 2011

    2011

  6. [14]

    Gunzburger, Clayton G

    Max D. Gunzburger, Clayton G. Webster, and Guannan Zhang. Stochastic finite element methods for partial differential equations with random input data.Acta Numerica, 23:521–650, 2014

    2014

  7. [15]

    The local discontinuous galerkin finite element methods for caputo-type partial differential equations: Mathematical analysis.Applied Numerical Mathematics, 150:587–606, 2020

    Changpin Li and Zhen Wang. The local discontinuous galerkin finite element methods for caputo-type partial differential equations: Mathematical analysis.Applied Numerical Mathematics, 150:587–606, 2020

    2020

  8. [16]

    O. C. Zienkiewicz, R. L. Taylor, P. Nithiarasu, and J. Z. Zhu.The Finite Element Method, volume 3. Elsevier, 1977

    1977

  9. [17]

    Finite volume methods

    Robert Eymard, Thierry Gallouët, and Raphaèle Herbin. Finite volume methods. InHandbook of numerical analysis, Vol. VII, volume VII ofHandb. Numer. Anal., pages 713–1020. North-Holland, Amsterdam, 2000

    2000

  10. [18]

    Spectral solutions for the time-fractional heat differential equation through a novel unified sequence of chebyshev polynomials.AIMS MATHEMATICS, 9(1):2137–2166, 2024

    Waleed Mohamed Abd-Elhameed and Hany Mostafa Ahmed. Spectral solutions for the time-fractional heat differential equation through a novel unified sequence of chebyshev polynomials.AIMS MATHEMATICS, 9(1):2137–2166, 2024. 17

    2024

  11. [19]

    M. D. Feit, J. A. Fleck, Jr., and A. Steiger. Solution of the Schrödinger equation by a spectral method.J. Comput. Phys., 47(3):412–433, 1982

    1982

  12. [20]

    Spectral methods in fluid dynamics (c

    John Boyd. Spectral methods in fluid dynamics (c. canuto, m. y. hussaini, a. quarteroni, and t. a. zang).Siam Review - SIAM REV, 30, 12 1988

    1988

  13. [21]

    Hauck, and Stanley Osher

    Hayden Schaeffer, Russel Caflisch, Cory D. Hauck, and Stanley Osher. Sparse dynamics for partial differential equations.Proc Natl Acad Sci U S A, 110(17):6634–6639, 2013

    2013

  14. [22]

    Highly accurate protein structure prediction with alphafold.Nature, pages 1–11, 2021

    John Jumper, Richard Evans, Alexander Pritzel, Tim Green, and Demis Hassabis. Highly accurate protein structure prediction with alphafold.Nature, pages 1–11, 2021

    2021

  15. [23]

    Deep learning for reduced order modelling and efficient temporal evolution of fluid simulations

    Pranshu Pant, Ruchit Doshi, Pranav Bahl, and Amir Barati Farimani. Deep learning for reduced order modelling and efficient temporal evolution of fluid simulations. 2021

    2021

  16. [24]

    Read, Jacob A

    Xiaowei Jia, Jared Willard, Anuj Karpatne, Jordan S. Read, Jacob A. Zwart, Michael Steinbach, and Vipin Kumar. Physics-guided machine learning for scientific discovery: An application in simulating lake temperature profiles. ACM/IMS Transactions on Data Science, (3), 2021

    2021

  17. [25]

    Integrating scientific knowledge with machine learning for engineering and environmental systems

    Jared Willard, Xiaowei Jia, Shaoming Xu, Michael Steinbach, and Vipin Kumar. Integrating scientific knowledge with machine learning for engineering and environmental systems. 2020

    2020

  18. [26]

    Meaningless comparisons lead to false optimism in medical machine learning.Plos One, 12(9), 2017

    De Masi Orianna, Kording Konrad, Recht Benjamin, and Jan Yih-Kuen. Meaningless comparisons lead to false optimism in medical machine learning.Plos One, 12(9), 2017

    2017

  19. [27]

    Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations

    Nick Mcgreivy and Ammar Hakim. Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations. 2024

    2024

  20. [28]

    Wujek and Patrick Hall

    B. Wujek and Patrick Hall. Best practices for machine learning applications. 2016

    2016

  21. [29]

    Raissi, P

    M. Raissi, P. Perdikaris, and 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, 378:686–707, 2019

    2019

  22. [30]

    Encoding physics to learn reaction–diffusion processes.Nature Machine Intelligence, 5(7):765–779, jul 2023

    Chengping Rao, Pu Ren, Qi Wang, Oral Buyukozturk, Hao Sun, and Yang Liu. Encoding physics to learn reaction–diffusion processes.Nature Machine Intelligence, 5(7):765–779, jul 2023

    2023

  23. [31]

    Karniadakis

    Ehsan Kharazmi, Zhongqiang Zhang, and George E.M. Karniadakis. hp-vpinns: Variational physics-informed neural networks with domain decomposition.Computer Methods in Applied Mechanics and Engineering, 374:113547, 2021

    2021

  24. [32]

    Enhancing convergence speed with feature enforcing physics-informed neural networks using boundary conditions as prior knowledge.Scientific Reports, 14(1):23836, oct 11 2024

    Jahani-Nasab M and Bijarchi MA. Enhancing convergence speed with feature enforcing physics-informed neural networks using boundary conditions as prior knowledge.Scientific Reports, 14(1):23836, oct 11 2024

    2024

  25. [33]

    McClenny and Ulisses M

    Levi D. McClenny and Ulisses M. Braga-Neto. Self-adaptive physics-informed neural networks.Journal of Computational Physics, 474:111722, 2023

    2023

  26. [34]

    and Jia Zhao

    Colby Wight, L. and Jia Zhao. Solving allen-cahn and cahn-hilliard equations using the adaptive physics informed neural networks.Communications in Computational Physics, 29(3):930–954, 2021

    2021

  27. [35]

    E-pinn: A fast physics-informed neural network based on explicit time-domain method for dynamic response prediction of nonlinear structures.Engineering Structures, 321:118900, 2024

    Cheng Su, Jingwei Liang, and Zengsheng He. E-pinn: A fast physics-informed neural network based on explicit time-domain method for dynamic response prediction of nonlinear structures.Engineering Structures, 321:118900, 2024

    2024

  28. [36]

    Pi-lstm: Physics-informed long short-term memory network for structural response modeling.Engineering Structures, 292:116500, 2023

    Fangyu Liu, Junlin Li, and Linbing Wang. Pi-lstm: Physics-informed long short-term memory network for structural response modeling.Engineering Structures, 292:116500, 2023

    2023

  29. [37]

    Physics-informed multi-lstm networks for metamodeling of nonlinear structures.Computer Methods in Applied Mechanics and Engineering, 369:113226, 2020

    Ruiyang Zhang, Yang Liu, and Hao Sun. Physics-informed multi-lstm networks for metamodeling of nonlinear structures.Computer Methods in Applied Mechanics and Engineering, 369:113226, 2020

    2020

  30. [38]

    Ppinn: Parareal physics-informed neural network for time-dependent pdes.Computer Methods in Applied Mechanics and Engineering, 370:113250, 2020

    Xuhui Meng, Zhen Li, Dongkun Zhang, and George Em Karniadakis. Ppinn: Parareal physics-informed neural network for time-dependent pdes.Computer Methods in Applied Mechanics and Engineering, 370:113250, 2020

    2020

  31. [39]

    and Em Karniadakis, George

    Ameya Jagtap, D. and Em Karniadakis, George. Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations.Communications in Computational Physics, 28(5):2002–2041, 2020

    2002

  32. [40]

    Jagtap, Kenji Kawaguchi, and George Em Karniadakis

    Ameya D. Jagtap, Kenji Kawaguchi, and George Em Karniadakis. Adaptive activation functions accelerate convergence in deep and physics-informed neural networks.Journal of Computational Physics, 2020

    2020

  33. [41]

    Deep learning method based on physics informed neural network with resnet block for solving fluid flow problems.Water, (4), 2021

    Chen Cheng and Guang Tao Zhang. Deep learning method based on physics informed neural network with resnet block for solving fluid flow problems.Water, (4), 2021. 18

    2021

  34. [42]

    Understanding and mitigating gradient pathologies in physics- informed neural networks.arXiv preprint arXiv:2001.04536, 2020

    Sifan Wang, Yujun Teng, and Paris Perdikaris. Understanding and mitigating gradient pathologies in physics- informed neural networks.arXiv preprint arXiv:2001.04536, 2020

    2001 arXiv

  35. [43]

    Multi-scale deep neural network (mscalednn) for solving poisson- boltzmann equation in complex domains.Communications in Computational Physics, 28(5):1970–2001, 2020

    Liu Ziqi, Cai Wei, and John Zhi-Qin, Xu. Multi-scale deep neural network (mscalednn) for solving poisson- boltzmann equation in complex domains.Communications in Computational Physics, 28(5):1970–2001, 2020

    1970

  36. [44]

    Automatic differentiation in machine learning: A survey.Journal of Machine Learning Research, 18:1–43, 04 2018

    Atilim Baydin, Barak Pearlmutter, Alexey Radul, and Jeffrey Siskind. Automatic differentiation in machine learning: A survey.Journal of Machine Learning Research, 18:1–43, 04 2018

    2018

  37. [45]

    Deepxde: A deep learning library for solving differential equations.SIAM Review, 63(1):208–228, 2021

    Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. Deepxde: A deep learning library for solving differential equations.SIAM Review, 63(1):208–228, 2021

    2021

  38. [46]

    Compatibility conditions for systems of iterative functional equations with non-trivial contact sets.Results in Mathematics, 76(2):68, mar 17 2021

    Jorge Buescu and Cristina Serpa. Compatibility conditions for systems of iterative functional equations with non-trivial contact sets.Results in Mathematics, 76(2):68, mar 17 2021. 19

    2021

Pith tools

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