arxiv: 2604.03321 · v1 · submitted 2026-04-02 · 💻 cs.LG · cs.AI· math.AP· physics.med-ph

Recognition: no theorem link

General Explicit Network (GEN): A novel deep learning architecture for solving partial differential equations

Genwei Ma, Ping Yang, Ting Luo, Xing Zhao

Authors on Pith no claims yet

Pith reviewed 2026-05-13 22:06 UTC · model grok-4.3

classification 💻 cs.LG cs.AImath.APphysics.med-ph

keywords deep learningPDE solvingphysics-informed neural networksexplicit networksbasis functionsrobustnessextensibility

0 comments

The pith

A general explicit network solves PDEs by mapping points to functions built from known basis functions.

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

The paper introduces the General Explicit Network (GEN) as an alternative to physics-informed neural networks for solving partial differential equations. Instead of point-to-point fitting with continuous activation functions, GEN performs point-to-function solving by constructing an explicit function component from basis functions derived from prior PDE knowledge. This targets the local characteristics and resulting poor extensibility and robustness seen in standard approaches. Readers would care if it allows machine learning PDE solvers to move beyond academic limits into more reliable practical use.

Core claim

GEN implements point-to-function PDE solving, where the function component is constructed based on prior knowledge of the original PDEs through corresponding basis functions for fitting, enabling solutions with high robustness and strong extensibility.

What carries the argument

The explicit function component of GEN, assembled from PDE-informed basis functions to perform the point-to-function mapping.

If this is right

Solutions exhibit higher robustness than those from discrete point-to-point PINN fitting.
Strong extensibility follows from swapping in appropriate basis functions for new PDE problems.
The method accounts for real solution properties that continuous activations overlook.
Practical deployment of neural PDE solvers beyond research settings becomes feasible.

Where Pith is reading between the lines

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

GEN could integrate with classical numerical expansions when basis functions are already known from analysis.
The approach opens tests on whether hybrid basis-neural models reduce data requirements for training.
Extensions might adapt basis selection automatically for PDEs where initial choices are uncertain.
Comparisons on time-evolving or high-dimensional problems would check if extensibility gains hold.

Load-bearing premise

Suitable basis functions exist, can be correctly chosen from prior PDE knowledge, and are sufficient to capture solution properties without introducing bias.

What would settle it

Finding a family of PDEs where no choice of basis functions lets GEN match or exceed PINN performance on robustness and extensibility metrics, while pointwise methods succeed.

Figures

Figures reproduced from arXiv: 2604.03321 by Genwei Ma, Ping Yang, Ting Luo, Xing Zhao.

**Figure 2.** Figure 2: Comparison among solutions for the heat equation: (a) The PINN. (b-c) The SineGEN and GaussGEN developed under the proposed [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Wave equation: (a) Heatmap comparison among the numerical solution, the PINN and the two GEN results, with the colour intensities [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Burgers’ equation: (a)-(c) Heatmap comparisons between exact solutions and GENs with various numbers of basis functions. (d)-(f) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

read the original abstract

Machine learning, especially physics-informed neural networks (PINNs) and their neural network variants, has been widely used to solve problems involving partial differential equations (PDEs). The successful deployment of such methods beyond academic research remains limited. For example, PINN methods primarily consider discrete point-to-point fitting and fail to account for the potential properties of real solutions. The adoption of continuous activation functions in these approaches leads to local characteristics that align with the equation solutions while resulting in poor extensibility and robustness. A general explicit network (GEN) that implements point-to-function PDE solving is proposed in this paper. The "function" component can be constructed based on our prior knowledge of the original PDEs through corresponding basis functions for fitting. The experimental results demonstrate that this approach enables solutions with high robustness and strong extensibility to be obtained.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

GEN proposes explicit basis-function construction for point-to-function PDE solving to fix PINN locality issues, but the abstract shows no equations, selection method, or error numbers to check whether it actually works.

read the letter

The main takeaway is that GEN reframes PDE solving as building an explicit function from basis functions drawn from prior knowledge rather than pure pointwise neural fitting. That targets the local-characteristic limitation the authors correctly flag in standard PINNs with continuous activations. If the bases can be chosen well, the approach could give better extensibility on problems where some solution structure is known in advance. The paper does a clean job stating that limitation and positioning the explicit component as the fix. The experimental claim of high robustness follows directly from that framing. The soft spot is the missing substance. The abstract contains no equations for how the basis functions are constructed or fitted, no description of the selection process, and no reported error metrics or test cases. The central assumption—that suitable, unbiased bases are available and sufficient for the target PDE—receives no supporting derivation or counter-example test. For nonlinear problems without obvious closed-form bases, this could collapse to ordinary neural fitting with the same weaknesses. Without those details it is impossible to tell whether the reported robustness holds or is an artifact of favorable examples. This work is aimed at researchers already using or extending PINNs who have some comfort with basis expansions. Someone who has tried to hand-craft bases for specific PDEs would be the right reader to judge whether the construction step is practical. The idea is coherent enough on its own terms to merit a full review rather than a desk rejection; the experiments and derivations need to be examined to see if the assumption actually delivers.

Referee Report

2 major / 2 minor

Summary. The paper proposes a General Explicit Network (GEN) architecture for solving partial differential equations (PDEs). It critiques physics-informed neural networks (PINNs) for relying on discrete point-to-point fitting with continuous activations that produce local solution characteristics and limited extensibility/robustness. GEN instead performs point-to-function solving, where the function component is explicitly constructed from prior-knowledge basis functions chosen to fit the PDE. Experimental results are reported to demonstrate high robustness and strong extensibility compared to existing approaches.

Significance. If the central claims hold and the basis-function construction can be made reliable, GEN would represent a meaningful advance over PINNs by replacing implicit local fitting with an explicit, knowledge-informed function representation. This could improve robustness on problems where suitable bases exist and enable better generalization across related PDE instances. The approach also opens a route to hybrid symbolic-numeric solvers when prior knowledge is strong.

major comments (2)

[Abstract and §3] Abstract and §3 (GEN architecture): The point-to-function claim rests on the assumption that suitable basis functions are available, correctly chosen, and sufficient to capture solution properties without bias. No general procedure for basis selection or verification is described, so the method reduces to a standard neural fit when this prior knowledge is weak or misspecified (e.g., generic polynomials on a nonlinear PDE without closed form). This assumption is load-bearing for the generality and robustness claims.
[§4] §4 (Experiments): The reported results claim high robustness and strong extensibility, yet the abstract and available description supply no quantitative error metrics, baseline comparisons, or details on the PDE classes tested. Without these, it is impossible to verify whether performance gains exceed those of well-tuned PINNs or other explicit-function hybrids.

minor comments (2)

[§3] Notation for the function component and basis expansion is introduced without a clear equation or diagram showing how the network output is combined with the explicit basis term.
[Abstract] The abstract states that continuous activations lead to 'local characteristics'; a brief reference to the relevant PINN literature or a short derivation of this locality effect would help readers unfamiliar with the critique.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and have revised the manuscript to improve clarity and completeness.

read point-by-point responses

Referee: [Abstract and §3] Abstract and §3 (GEN architecture): The point-to-function claim rests on the assumption that suitable basis functions are available, correctly chosen, and sufficient to capture solution properties without bias. No general procedure for basis selection or verification is described, so the method reduces to a standard neural fit when this prior knowledge is weak or misspecified (e.g., generic polynomials on a nonlinear PDE without closed form). This assumption is load-bearing for the generality and robustness claims.

Authors: We agree that the selection of basis functions is central to GEN and that the manuscript would benefit from an explicit procedure. In the revised version we have added a dedicated subsection in §3 that outlines a systematic approach: (i) identify known analytic properties of the target PDE (linearity, separability, boundary conditions), (ii) select candidate bases (polynomials, Fourier, or problem-specific functions) accordingly, and (iii) verify sufficiency by checking residual norms on a small validation set before full training. When prior knowledge is weak we explicitly note that the architecture reverts to a more general neural fit, and we have updated the abstract to state this limitation clearly. These additions preserve the core claim while making the method’s scope transparent. revision: yes
Referee: [§4] §4 (Experiments): The reported results claim high robustness and strong extensibility, yet the abstract and available description supply no quantitative error metrics, baseline comparisons, or details on the PDE classes tested. Without these, it is impossible to verify whether performance gains exceed those of well-tuned PINNs or other explicit-function hybrids.

Authors: We acknowledge that the abstract and introductory description did not contain quantitative metrics. The full §4 already reports relative L2 errors, comparisons against standard PINNs, and results on concrete PDE classes (1-D Burgers, 2-D Poisson, wave equation). To address the concern we have inserted a concise results summary table and explicit error figures into the abstract and the opening of §4, together with a statement of the exact baseline implementations and hyper-parameter settings used. This makes the performance claims directly verifiable from the front matter. revision: yes

Circularity Check

0 steps flagged

GEN derivation remains self-contained; basis-function construction is an external assumption, not a reduction to fitted inputs

full rationale

The paper proposes GEN as a point-to-function solver whose function component is built from prior-knowledge basis functions. No equations, fitting procedures, or self-citations are shown that would make any claimed prediction equivalent to its inputs by construction. The central claim rests on an external modeling choice (availability and correctness of bases) rather than on any internal loop that renames a fit as a prediction or imports uniqueness via self-citation. Because the derivation introduces no self-referential reduction and treats basis selection as an independent modeling step, the architecture is non-circular on the supplied text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the availability of suitable basis functions derived from prior PDE knowledge; no explicit free parameters, axioms, or invented entities are named in the abstract, but the construction step implicitly assumes domain knowledge supplies an adequate function space.

pith-pipeline@v0.9.0 · 5447 in / 1114 out tokens · 27200 ms · 2026-05-13T22:06:02.550970+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

47 extracted references · 47 canonical work pages · 4 internal anchors

[1]

Machine learning solutions looking for pde problems

, 2025. Machine learning solutions looking for pde problems. Nature Machine Intelligence 7, 1

work page 2025
[2]

Neural operator: Graph kernel network for partial differential equations, in: ICLR 2020 workshop on integration of deep neural models and differential equations

Anandkumar, A., Azizzadenesheli, K., Bhattacharya, K., Kovachki, N., Li, Z., Liu, B., Stuart, A., 2020. Neural operator: Graph kernel network for partial differential equations, in: ICLR 2020 workshop on integration of deep neural models and differential equations. 10

work page 2020
[3]

Critical investigation of failure modes in physics-informed neural networks, in: AiAA SCITECH 2022 Forum, p

Basir, S., Senocak, I., 2022. Critical investigation of failure modes in physics-informed neural networks, in: AiAA SCITECH 2022 Forum, p. 2353

work page 2022
[4]

Physics informed neural network with fourier feature for natural convection problems

Bounnah, Y ., Mihoubi, M.K., Larbi, S., 2025. Physics informed neural network with fourier feature for natural convection problems. Engineering Applications of Artificial Intelligence 146, 110327

work page 2025
[5]

Envisioning better benchmarks for machine learning pde solvers

Brandstetter, J., 2025. Envisioning better benchmarks for machine learning pde solvers. Nature Machine Intel- ligence 7, 2–3

work page 2025
[6]

Promising directions of machine learning for partial differential equations

Brunton, S.L., Kutz, J.N., 2024. Promising directions of machine learning for partial differential equations. Nature Computational Science 4, 483–494

work page 2024
[7]

Physics-informed neural networks (pinns) for fluid mechanics: A review

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

work page 2021
[8]

Laplace neural operator for solving differential equations

Cao, Q., Goswami, S., Karniadakis, G.E., 2024. Laplace neural operator for solving differential equations. Nature Machine Intelligence 6, 631–640

work page 2024
[9]

Pf-pinns: Physics-informed neural networks for solving coupled allen-cahn and cahn-hilliard phase field equations

Chen, N., Lucarini, S., Ma, R., Chen, A., Cui, C., 2025. Pf-pinns: Physics-informed neural networks for solving coupled allen-cahn and cahn-hilliard phase field equations. Journal of Computational Physics , 113843

work page 2025
[10]

Experience report of physics-informed neural networks in fluid simulations: pitfalls and frustration

Chuang, P.Y ., Barba, L.A., 2022. Experience report of physics-informed neural networks in fluid simulations: pitfalls and frustration. arXiv preprint arXiv:2205.14249

work page arXiv 2022
[11]

Predictive limitations of physics-informed neural networks in vortex shedding

Chuang, P.Y ., Barba, L.A., 2023. Predictive limitations of physics-informed neural networks in vortex shedding. arXiv preprint arXiv:2306.00230

work page arXiv 2023
[12]

Scientific machine learning through physics–informed neural networks: Where we are and what’s next

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

work page 2022
[13]

Multilayer feedforward networks are universal approximators

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

work page 1989
[14]

Fl-ntk: A neural tangent kernel-based framework for federated learning analysis, in: International Conference on Machine Learning, PMLR

Huang, B., Li, X., Song, Z., Yang, X., 2021. Fl-ntk: A neural tangent kernel-based framework for federated learning analysis, in: International Conference on Machine Learning, PMLR. pp. 4423–4434

work page 2021
[15]

Neural tangent kernel: Convergence and generalization in neural networks

Jacot, A., Gabriel, F., Hongler, C., 2018. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems 31

work page 2018
[16]

Fourier warm start for physics-informed neural networks

Jin, G., Wong, J.C., Gupta, A., Li, S., Ong, Y .S., 2024. Fourier warm start for physics-informed neural networks. Engineering Applications of Artificial Intelligence 132, 107887

work page 2024
[17]

Solving inverse problems in physics by optimizing a discrete loss: Fast and accurate learning without neural networks

Karnakov, P., Litvinov, S., Koumoutsakos, P., 2024. Solving inverse problems in physics by optimizing a discrete loss: Fast and accurate learning without neural networks. PNAS nexus 3, pgae005

work page 2024
[18]

Physics-informed machine learning

Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L., 2021. Physics-informed machine learning. Nature Reviews Physics 3, 422–440

work page 2021
[19]

Automatic differentiation in deep learning

Ketkar, N., Moolayil, J., Ketkar, N., Moolayil, J., 2021. Automatic differentiation in deep learning. Deep Learning with python: learn best practices of deep learning models with PyTorch , 133–145

work page 2021
[20]

Neural operator: Learning maps between function spaces with applications to pdes

Kovachki, N., Li, Z., Liu, B., Azizzadenesheli, K., Bhattacharya, K., Stuart, A., Anandkumar, A., 2023. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research 24, 1–97

work page 2023
[21]

Characterizing possible failure modes in physics-informed neural networks

Krishnapriyan, A., Gholami, A., Zhe, S., Kirby, R., Mahoney, M.W., 2021. Characterizing possible failure modes in physics-informed neural networks. Advances in neural information processing systems 34, 26548–26560. 11

work page 2021
[22]

Phase-field deeponet: Physics-informed deep operator neural network for fast simulations of pattern formation governed by gradient flows of free-energy functionals

Li, W., Bazant, M.Z., Zhu, J., 2023. Phase-field deeponet: Physics-informed deep operator neural network for fast simulations of pattern formation governed by gradient flows of free-energy functionals. Computer Methods in Applied Mechanics and Engineering 416, 116299

work page 2023
[23]

Tutorials: Physics-informed machine learning methods of computing 1d phase-field models

Li, W., Fang, R., Jiao, J., Vassilakis, G.N., Zhu, J., 2024. Tutorials: Physics-informed machine learning methods of computing 1d phase-field models. APL Machine Learning 2

work page 2024
[24]

Fourier Neural Operator for Parametric Partial Differential Equations

Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., Anandkumar, A., 2020a. Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895

work page internal anchor Pith review Pith/arXiv arXiv 2010
[25]

Mul- tipole graph neural operator for parametric partial differential equations

Li, Z., Kovachki, N., Azizzadenesheli, K., Liu, B., Stuart, A., Bhattacharya, K., Anandkumar, A., 2020b. Mul- tipole graph neural operator for parametric partial differential equations. Advances in Neural Information Pro- cessing Systems 33, 6755–6766

work page
[26]

Fourier neural operator for parametric partial differential equations, in: International Conference on Learning Representations

Li, Z., Kovachki, N.B., Azizzadenesheli, K., Bhattacharya, K., Stuart, A., Anandkumar, A., et al., . Fourier neural operator for parametric partial differential equations, in: International Conference on Learning Representations

work page
[27]

KAN: Kolmogorov-Arnold Networks

Liu, Z., Wang, Y ., Vaidya, S., Ruehle, F., Halverson, J., Solja ˇci´c, M., Hou, T.Y ., Tegmark, M., 2024. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756

work page internal anchor Pith review Pith/arXiv arXiv 2024
[28]

Learning nonlinear operators via deeponet based on the universal approximation theorem of operators

Lu, L., Jin, P., Pang, G., Zhang, Z., Karniadakis, G.E., 2021. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence 3, 218–229

work page 2021
[29]

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

McGreivy, N., Hakim, A., 2024. Weak baselines and reporting biases lead to overoptimism in machine learning for fluid-related partial differential equations. Nature Machine Intelligence 6, 1256–1269

work page 2024
[30]

PyTorch: An Imperative Style, High-Performance Deep Learning Library

Paszke, A., 2019. Pytorch: An imperative style, high-performance deep learning library. arXiv preprint arXiv:1912.01703

work page internal anchor Pith review Pith/arXiv arXiv 2019
[31]

Automatic differentiation in pytorch

Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., Lerer, A., 2017. Automatic differentiation in pytorch

work page 2017
[32]

Gabor-filtered fourier neural operator for solving partial differential equations

Qi, K., Sun, J., 2024. Gabor-filtered fourier neural operator for solving partial differential equations. Computers & Fluids 274, 106239

work page 2024
[33]

Physics-informed neural networks: A deep learning frame- work for solving forward and inverse problems involving nonlinear partial differential equations

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

work page 2019
[34]

Applications of physics informed neural operators

Rosofsky, S.G., Al Majed, H., Huerta, E., 2023. Applications of physics informed neural operators. Machine Learning: Science and Technology 4, 025022

work page 2023
[35]

On the use of fourier features-physics informed neural networks (ff-pinn) for forward and inverse fluid mechanics problems

Sallam, O., Fürth, M., 2023. On the use of fourier features-physics informed neural networks (ff-pinn) for forward and inverse fluid mechanics problems. Proceedings of the Institution of Mechanical Engineers, Part M: Journal of Engineering for the Maritime Environment 237, 846–866

work page 2023
[36]

A review of physics-informed machine learning in fluid mechanics

Sharma, P., Chung, W.T., Akoush, B., Ihme, M., 2023. A review of physics-informed machine learning in fluid mechanics. Energies 16, 2343

work page 2023
[37]

Simulating seismic multifrequency wavefields with the fourier feature physics- informed neural network

Song, C., Wang, Y ., 2023. Simulating seismic multifrequency wavefields with the fourier feature physics- informed neural network. Geophysical Journal International 232, 1503–1514

work page 2023
[38]

Physics-based deep learning

Thuerey, N., Holl, P., Mueller, M., Schnell, P., Trost, F., Um, K., 2021. Physics-based deep learning. arXiv preprint arXiv:2109.05237

work page arXiv 2021
[39]

Enhancing computational fluid dynamics with machine learning

Vinuesa, R., Brunton, S.L., 2022. Enhancing computational fluid dynamics with machine learning. Nature Computational Science 2, 358–366. 12

work page 2022
[40]

On the eigenvector bias of fourier feature networks: From regression to solving multi-scale pdes with physics-informed neural networks

Wang, S., Wang, H., Perdikaris, P., 2021. On the eigenvector bias of fourier feature networks: From regression to solving multi-scale pdes with physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering 384, 113938

work page 2021
[41]

When and why pinns fail to train: A neural tangent kernel perspective

Wang, S., Yu, X., Perdikaris, P., 2022. When and why pinns fail to train: A neural tangent kernel perspective. Journal of Computational Physics 449, 110768

work page 2022
[42]

Nas-pinn: Neural architecture search-guided physics-informed neural network for solving pdes

Wang, Y ., Zhong, L., 2024. Nas-pinn: Neural architecture search-guided physics-informed neural network for solving pdes. Journal of Computational Physics 496, 112603

work page 2024
[43]

Small-data-driven fast seismic simulations for complex media using physics-informed fourier neural operators

Wei, W., Fu, L.Y ., 2022. Small-data-driven fast seismic simulations for complex media using physics-informed fourier neural operators. Geophysics 87, T435–T446

work page 2022
[44]

Solving allen-cahn and cahn-hilliard equations using the adaptive physics informed neural networks

Wight, C.L., Zhao, J., 2020. Solving allen-cahn and cahn-hilliard equations using the adaptive physics informed neural networks. arXiv preprint arXiv:2007.04542

work page arXiv 2020
[45]

A comprehensive study of non-adaptive and residual- based adaptive sampling for physics-informed neural networks

Wu, C., Zhu, M., Tan, Q., Kartha, Y ., Lu, L., 2023. A comprehensive study of non-adaptive and residual- based adaptive sampling for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering 403, 115671

work page 2023
[46]

Sinc Kolmogorov-Arnold Network and Its Applications on Physics-informed Neural Networks

Yu, T., Qiu, J., Yang, J., Oseledets, I., 2024. Sinc kolmogorov-arnold network and its applications on physics- informed neural networks. arXiv preprint arXiv:2410.04096

work page internal anchor Pith review Pith/arXiv arXiv 2024
[47]

Fourier neural operator for solving subsurface oil/water two-phase flow partial differential equation

Zhang, K., Zuo, Y ., Zhao, H., Ma, X., Gu, J., Wang, J., Yang, Y ., Yao, C., Yao, J., 2022. Fourier neural operator for solving subsurface oil/water two-phase flow partial differential equation. Spe Journal 27, 1815–1830. 13

work page 2022