arxiv: 2604.03020 · v1 · submitted 2026-04-03 · 🧮 math.NA · cs.NA

Recognition: 2 theorem links

· Lean Theorem

Generalized Transferable Neural Networks for Steady-State Partial Differential Equations

Tao Cheng , Lili Ju , Zhonghua Qiao , Xiaoping Zhang

Authors on Pith no claims yet

Pith reviewed 2026-05-13 18:01 UTC · model grok-4.3

classification 🧮 math.NA cs.NA

keywords neuraltransnethiddennetworkpdestransferableaccuracyadditional

0 comments

The pith

GTransNet extends single-hidden-layer TransNet by adding hidden layers with symmetry-constrained biases and variance-controlled weights to improve accuracy and stability for oscillatory steady-state PDE solutions.

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

Standard numerical methods for partial differential equations can be slow or unstable for problems with wiggly solutions. Neural networks provide an alternative by learning to approximate the solution directly. The earlier TransNet uses one hidden layer whose parameters are fixed in advance using uniform partitions of the domain; this works for smooth cases but runs into saturation and poor conditioning when the solution oscillates rapidly. GTransNet keeps the same interpretable first-layer construction but adds symmetry on the neuron biases and then stacks further hidden layers that have no biases and draw their weights from a controlled-variance distribution. The goal is to retain the original method's efficiency and transparency while gaining the expressive power of deeper networks for harder PDEs.

Core claim

We propose a generalized transferable neural network (GTransNet) for solving steady-state PDEs, which augments the original TransNet design with additional hidden layers while preserving its interpretable feature-generation mechanism. In particular, the first hidden layer of GTransNet retains TransNet's parameter sampling strategy but incorporates an additional symmetry constraint on the neuron biases, while the subsequent hidden layers omit bias terms and employ a variance-controlled sampling strategy for selecting neuron weights.

Load-bearing premise

That the symmetry constraint on the first-layer biases and the variance-controlled sampling in subsequent layers will simultaneously improve accuracy for oscillatory solutions and avoid introducing new saturation or conditioning problems that offset the gains.

Figures

Figures reproduced from arXiv: 2604.03020 by Lili Ju, Tao Cheng, Xiaoping Zhang, Zhonghua Qiao.

**Figure 1.** Figure 1: Visualization over the square (−1, 1)2 of a sample hidden-layer neuron ψm(x) in TransNet associated with B1.5(0) in two dimensions, where the dashed line indicates the corresponding partition hyperplane. From left to right: γ = 2, 6, 14 [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗

**Figure 2.** Figure 2: Distribution histograms of the hidden-layer neuron activation values in TransNet associated with B1.5(0) in two dimensions. Activation values are collected from 500 input points over the unit square (−1, 1)2 , with each point corresponding to 1000 different hidden-layer neurons, illustrating how the response pattern and concentration vary along with the shape parameter γ. From left to right: γ = 2, 6, 14. … view at source ↗

**Figure 3.** Figure 3: Numerical results of fitting the high-frequency function f(x) = sin(30πx) in the interval (−1, 1) by using a TransNet associated with B1.1(0) in one dimension. 1000 uniformly distributed collocation points are used. From Left to right: γ = 2, 14; From top to bottom: M = 200, 1000. 3 Generalized Transferable Neural Networks To overcome the limitations of TransNet in handling high-frequency and large-gradien… view at source ↗

**Figure 4.** Figure 4: Network architecture of the proposed GTransNet (16) with L hidden layers. The GTransNet-based solver for the PDE problem (1) then leads to the following loss minimization problem: Given {Γ, A, r, {Wi} L i=2} (i.e., the network parameters for all the L hidden layers), find uNN defined by (16) such that min α FLoss(uNN; {α}), (17) which is similar to (5) for the TransNet solver. For linear PDE problems, the … view at source ↗

**Figure 5.** Figure 5: Distribution histograms of the hidden-layer neuron activation values in the proposed GTransNet associated with B1.5(0) in two dimensions. Activation values are collected from 500 input points over the unit square (−1, 1)2 , with each point corresponding to 2000 different neurons in the first hidden layer and 1000 neurons in the second and third hidden layers. From left to right: γ = 2, 6, 14. First row: th… view at source ↗

**Figure 6.** Figure 6: Visualization of some typical second and third hidden-layer neurons, ψ (2) m (x) and ψ (3) m (x), in the proposed GTransNet. From left to right: γ = 2, 6, 14. Top row: ψ (2) m (x) with W2 ∼ N (0, 1); Middle row: ψ (2) m (x) with W2 ∼ N (0, σ2 2) and δ = 0.5; Bottom row: ψ (3) m (x) with W3 ∼ N (0, σ2 3) and δ = 0.5. the distributions exhibit large variance, and most neuron values are concentrated near the … view at source ↗

**Figure 7.** Figure 7: Plots of relative L 2 errors of numerical solutions produced by the TransNet method and the GTransNet method with two and three hidden layers in Section 4.1, where N denotes the number of last hidden-layer neurons. Top-left: (S1); Top-right: (S2); Bottom row: (S3). 500 600 700 800 N 0.10 0.15 0.20 0.25 0.30 Average runtimes S1. 2D Poisson GTransNet (with 2 hidden layers) GTransNet (with 3 hidden layers) Tr… view at source ↗

**Figure 8.** Figure 8: Average running times (in seconds) of the TransNet method and the GTransNet method with two and three hidden layers in Section 4.1, where N denotes the number of neurons in the last hidden layer. From left to right: (S1), (S2) and (S3). 15 [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗

**Figure 9.** Figure 9: Plots of the relative L 2 errors of numerical solutions produced by the GTransNet method (top row) and the TransNet method (bottom row) with different values of the last hidden-layer neurons N and the shape parameter γ for the three cases of the Poisson equation in Section 4.2.1. From left to right: Cases 1 (2D), Case 2 (2D), and Case 3 (3D). 16 [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗

**Figure 10.** Figure 10: Numerical solutions (top row) and the corresponding absolute point-wise errors (bottom row) produced by the GTransNet method for the three cases of the Poisson equation in Section 4.2.1. Left column: Case 1 (2D), N = 4000, γ = 8; Middle column: Case 2 (2D), N = 4000, γ = 10; Right column: Case 3 (3D), N = 7000, γ = 4, at the two cross sections x = 0 and y = 0 [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗

**Figure 11.** Figure 11: Plots of the relative L 2 errors of numerical solutions produced by the GTransNet method (top row) and the TransNet method (bottom row) with different values of the last hidden-layer neurons N and the shape parameter γ for the 2D Helmholtz equation in Section 4.2.2. From left to right: ν = 4000, 6000, and 8000 Hz [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗

**Figure 12.** Figure 12: Numerical solutions (top row) and the corresponding absolute point-wise errors (bottom row) produced by the GTransNet method for the 2D Helmholtz equation in Section 4.2.2. Left column: ν = 4000 Hz, N = 4000, γ = 6; middle column: ν = 6000 Hz, N = 4000, γ = 10; right column: ν = 8000 Hz, N = 4000, γ = 10. of neurons N in the last hidden layer. The results exhibit behaviors similar to those observed for th… view at source ↗

**Figure 13.** Figure 13: Plots of the relative L 2 errors of numerical solutions produced by the GTransNet method (top row) and the TransNet method (bottom row) with different values of the last hidden-layer neurons N and the shape parameter γ for the 2D multiscale elliptic problem in Section 4.2.3. From left to right: ε = 0.5, 0.2, and 0.1 [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗

**Figure 14.** Figure 14: Numerical solutions uNN − up (top row) and the corresponding absolute point-wise errors of numerical solutions (bottom row) produced by the GTransNet method for the 2D multiscale elliptic problem (43) in Section 4.2.3. Left column: ε = 0.5, N = 4000, γ = 6; Middle column: ε = 0.2, N = 4000, γ = 8; Right column: ε = 0.1, N = 4000, γ = 8. with the periodic boundary condition. The exact solution is given by … view at source ↗

**Figure 15.** Figure 15: Plots of the relative L 2 errors of numerical solutions produced by the GTransNet method (left) and the TransNet method (right) method with different values of the last hidden-layer neurons N and the shape parameter γ for the 3D steady-state Allen-Cahn equation in Section 4.2.4 [PITH_FULL_IMAGE:figures/full_fig_p021_15.png] view at source ↗

**Figure 16.** Figure 16: Numerical solution (left) and the corresponding absolute point-wise error (right) at the three cross sections x = 0.5, y = 0.5 and z = 0.5 produced by the GTransNet method with N = 7000 and γ = 6 for the 3D steady-state Allen-Cahn equation in Section 4.2.4. In this example, GTransNet again produces much better numerical solutions than TransNet [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗

read the original abstract

Deep learning has emerged as a compelling framework for scientific and engineering computing, motivating growing interest in neural network-based solvers for partial differential equations (PDEs). Within this landscape, network architectures with deterministic feature construction have become an appealing approach, offering both high accuracy and computational efficiency in practice. Among them, the transferable neural network (TransNet) is a special class of shallow neural networks (i.e., single-hidden-layer architectures), whose hidden-layer parameters are predetermined according to the principle of uniformly distributed partition hyperplanes. Although TransNet has demonstrated strong performance in solving PDEs with relatively smooth solutions, its accuracy and stability may deteriorate in the presence of highly oscillatory solution structures, where activation saturation and system conditioning issues become limiting factors. In this paper, we propose a generalized transferable neural network (GTransNet) for solving steady-state PDEs, which augments the original TransNet design with additional hidden layers while preserving its interpretable feature-generation mechanism. In particular, the first hidden layer of GTransNet retains TransNet's parameter sampling strategy but incorporates an additional symmetry constraint on the neuron biases, while the subsequent hidden layers omit bias terms and employ a variance-controlled sampling strategy for selecting neuron weights.

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

GTransNet adds symmetry on first-layer biases and variance-controlled weights in extra layers to TransNet, but supplies no tests showing these changes actually help with oscillatory PDEs.

read the letter

The core move here is straightforward: start from TransNet's deterministic sampling of hidden-layer parameters via uniform partition hyperplanes, then stack additional layers while forcing symmetry on the first-layer biases and dropping biases from later layers in favor of variance-controlled weight selection. That combination is presented as new relative to the cited TransNet work, and the construction itself is explicit enough that someone could implement it from the description without fitting extra parameters to the PDE solution. The intent is to keep the interpretable feature generation while trying to reduce saturation and conditioning trouble on highly oscillatory steady-state problems. That part of the design is clear and internally consistent. What is missing is any demonstration that the added constraints deliver the claimed gains. The text gives the architectural rules but no error tables, no condition-number measurements, no comparison runs against plain TransNet on oscillatory test cases, and no approximation bounds. Without those, the claim that accuracy and stability improve rests on the modeling assumption rather than evidence, so it is impossible to tell whether the symmetry and variance steps create offsetting problems. The paper is aimed at researchers already working on neural solvers that use fixed or semi-fixed feature maps instead of fully trained deep networks. A reader looking for concrete architecture tweaks in that niche could extract the construction rules and try them, but anyone needing validated performance on oscillatory problems will find little to use yet. The thinking is direct and the literature engagement is honest, so the manuscript is worth sending to referees who can check whether the full version contains the missing experiments or analysis. If those turn out to be absent or weak, the work would need major revision before publication.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes a generalized transferable neural network (GTransNet) for solving steady-state PDEs. It extends the single-hidden-layer TransNet architecture by adding multiple hidden layers while retaining the interpretable parameter-sampling mechanism: the first hidden layer uses TransNet's uniform partition hyperplane sampling but adds a symmetry constraint on neuron biases, and subsequent layers are bias-free with variance-controlled weight sampling. The central claim is that these modifications simultaneously improve accuracy and stability for highly oscillatory solutions without introducing new saturation or conditioning problems.

Significance. If the architectural modifications can be shown to deliver the claimed gains, GTransNet would offer a deterministic, interpretable alternative to standard deep networks for oscillatory steady-state PDEs, potentially improving both accuracy and computational efficiency in scientific computing applications where feature construction must remain transparent.

major comments (2)

[Abstract] Abstract: the assertion that accuracy and stability improve for highly oscillatory solutions rests on the design of the symmetry constraint and variance-controlled sampling, yet the manuscript contains no numerical experiments, error tables, condition-number measurements, or approximation bounds comparing GTransNet to TransNet on any test problem; this is load-bearing for the central claim.
[Method] The description of the first-layer symmetry constraint and subsequent-layer variance sampling (detailed after the abstract) is presented as simultaneously raising accuracy and avoiding saturation/conditioning degradation, but no theoretical analysis or empirical verification is supplied to show that the added constraints produce a net benefit rather than offsetting trade-offs.

minor comments (1)

[Method] Notation for the variance-controlled sampling strategy in deeper layers should be defined explicitly with a formula or pseudocode to allow reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful review and constructive comments on our manuscript. The points raised correctly identify that the central claims require stronger supporting evidence. We will revise the manuscript to address these concerns by adding the requested numerical experiments, error tables, condition-number measurements, and theoretical analysis.

read point-by-point responses

Referee: [Abstract] Abstract: the assertion that accuracy and stability improve for highly oscillatory solutions rests on the design of the symmetry constraint and variance-controlled sampling, yet the manuscript contains no numerical experiments, error tables, condition-number measurements, or approximation bounds comparing GTransNet to TransNet on any test problem; this is load-bearing for the central claim.

Authors: We agree that the current manuscript does not contain the comparative numerical experiments, error tables, condition-number measurements, or approximation bounds needed to substantiate the claims for highly oscillatory solutions. In the revised version we will add a dedicated numerical experiments section that includes test problems with highly oscillatory solutions, direct comparisons of GTransNet against TransNet, error tables, condition-number results, and any available approximation bounds. revision: yes
Referee: [Method] The description of the first-layer symmetry constraint and subsequent-layer variance sampling (detailed after the abstract) is presented as simultaneously raising accuracy and avoiding saturation/conditioning degradation, but no theoretical analysis or empirical verification is supplied to show that the added constraints produce a net benefit rather than offsetting trade-offs.

Authors: We acknowledge that the manuscript currently lacks both theoretical analysis and empirical verification demonstrating a net benefit from the symmetry constraint and variance-controlled sampling. We will expand the methods section with a theoretical discussion of how these modifications improve accuracy while avoiding saturation and conditioning degradation, and we will support this with the empirical results from the new numerical experiments section. revision: yes

Circularity Check

0 steps flagged

No circularity: GTransNet is an explicit architectural construction from TransNet rules

full rationale

The paper defines GTransNet directly via construction rules: first hidden layer keeps TransNet's uniform partition hyperplane sampling but adds a symmetry constraint on biases; subsequent layers drop biases and use variance-controlled weight sampling. These choices are presented as design decisions to address oscillatory solutions, not as predictions or derivations that reduce to fitted parameters or prior results by construction. No equations equate a claimed improvement to an input quantity, no self-citation is invoked as a uniqueness theorem or load-bearing ansatz, and the feature-generation mechanism is preserved by explicit rule rather than tautology. The derivation chain consists of independent architectural specifications.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The proposal rests on the domain assumption that the original TransNet parameter-sampling principle remains beneficial when extended to deeper networks and that the new symmetry and variance controls will not degrade conditioning. No explicit free parameters or invented entities are named in the abstract.

axioms (1)

domain assumption The uniformly distributed partition hyperplanes principle from TransNet remains a sound basis for feature construction even after adding layers and symmetry constraints.
Invoked when the abstract states that the first hidden layer retains TransNet's parameter sampling strategy.

pith-pipeline@v0.9.0 · 5510 in / 1368 out tokens · 69148 ms · 2026-05-13T18:01:13.327781+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

first hidden layer ... symmetry constraint on the neuron biases ... subsequent hidden layers omit bias terms and employ a variance-controlled sampling strategy
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Theorem 3 (Controlled Variance Propagation ... Var[ψ(l)_i] ≤ δ^{l-1} σ_0²

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · 1 internal anchor

[1]

Bridging traditional and machine learning- based algorithms for solving PDEs: The random feature method.Journal of Machine Learning, 1:268– 298, 06 2022

Jingrun Chen, Xurong Chi, Weinan E, and Zhouwang Yang. Bridging traditional and machine learning- based algorithms for solving PDEs: The random feature method.Journal of Machine Learning, 1:268– 298, 06 2022

work page 2022
[2]

A micro-macro decomposition-based asymptotic-preserving random feature method for multiscale radiative transfer equations.Journal of Computational Physics, 537:114103, 2025

Jingrun Chen, Zheng Ma, and Keke Wu. A micro-macro decomposition-based asymptotic-preserving random feature method for multiscale radiative transfer equations.Journal of Computational Physics, 537:114103, 2025

work page 2025
[3]

The random feature method for solving interface problems.Computer Methods in Applied Mechanics and Engineering, 420:116719, 2024

Xurong Chi, Jingrun Chen, and Zhouwang Yang. The random feature method for solving interface problems.Computer Methods in Applied Mechanics and Engineering, 420:116719, 2024

work page 2024
[4]

Suchuan Dong and Zongwei Li. Local extreme learning machines and domain decomposition for solv- ing linear and nonlinear partial differential equations.Computer Methods in Applied Mechanics and Engineering, 387:Paper No. 114129, 63, 2021

work page 2021
[5]

A modified batch intrinsic plasticity method for pre-training the random coefficients of extreme learning machines.Journal of Computational Physics, 445:110585, 2021

Suchuan Dong and Zongwei Li. A modified batch intrinsic plasticity method for pre-training the random coefficients of extreme learning machines.Journal of Computational Physics, 445:110585, 2021

work page 2021
[6]

The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems.Communications in Mathematics and Statistics, 6(1):1–12, 2018

Weinan E and Bing Yu. The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems.Communications in Mathematics and Statistics, 6(1):1–12, 2018

work page 2018
[7]

Least squares with equality constraints extreme learning machines for the resolution of PDEs.Journal of Computational Physics, 547:114553, 2026

Davide Elia De Falco, Enrico Schiassi, and Francesco Calabr` o. Least squares with equality constraints extreme learning machines for the resolution of PDEs.Journal of Computational Physics, 547:114553, 2026

work page 2026
[8]

Xavier Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. Journal of Machine Learning Research - Proceedings Track, 9:249–256, 01 2010

work page 2010
[9]

A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics.Computer Methods in Applied Mechanics and Engineering, 379:Paper No

Ehsan Haghighat, Maziar Raissi, Adrian Moure, Hector Gomez, and Ruben Juanes. A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics.Computer Methods in Applied Mechanics and Engineering, 379:Paper No. 113741, 22, 2021

work page 2021
[10]

Deep residual learning for image recognition

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016

work page 2016
[11]

Huang, Q

G. Huang, Q. Zhu, and C. Siew. Extreme learning machine: Theory and applications.Neurocomputing, 70(1):489–501, 2006

work page 2006
[12]

Weinberger

Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q. Weinberger. Densely connected con- volutional networks. In2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261–2269, 2017

work page 2017
[13]

Jagtap, Ehsan Kharazmi, and George Em Karniadakis

Ameya D. Jagtap, Ehsan Kharazmi, and George Em Karniadakis. Conservative physics-informed neural networks on discrete domains for conservation laws: applications to forward and inverse problems. Computer Methods in Applied Mechanics and Engineering, 365:113028, 27, 2020

work page 2020
[14]

The principles of diffusion models, 2025

Chieh-Hsin Lai, Yang Song, Dongjun Kim, Yuki Mitsufuji, and Stefano Ermon. The principles of diffusion models, 2025. 22

work page 2025
[15]

Subspace decomposition based dnn algorithm for elliptic type multi-scale PDEs.Journal of Computational Physics, 488:112242, 2023

Xi-An Li, Zhi-Qin John Xu, and Lei Zhang. Subspace decomposition based dnn algorithm for elliptic type multi-scale PDEs.Journal of Computational Physics, 488:112242, 2023

work page 2023
[16]

Solving a class of multi-scale elliptic PDEs by fourier-based mixed physics informed neural networks.Journal of Computational Physics, 508:113012, 2024

Xi’an Li, Jinran Wu, Xin Tai, Jianhua Xu, and You-Gan Wang. Solving a class of multi-scale elliptic PDEs by fourier-based mixed physics informed neural networks.Journal of Computational Physics, 508:113012, 2024

work page 2024
[17]

Fourier Neural Operator for Parametric Partial Differential Equations

Zong-Yi Li, Nikola B. Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, An- drew M. Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.ArXiv, abs/2010.08895, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2010
[18]

Liang, Q

Y. Liang, Q. Zhang, and S. Zeng. A piecewise extreme learning machine for interface problems.Math- ematics and Computers in Simulation, 227:303–321, 2025

work page 2025
[19]

Deep Nitsche method: deep Ritz method with essential boundary conditions.Communications in Computational Physics, 29(5):1365–1384, 2021

Yulei Liao and Pingbing Ming. Deep Nitsche method: deep Ritz method with essential boundary conditions.Communications in Computational Physics, 29(5):1365–1384, 2021

work page 2021
[20]

Multi-scale deep neural network (MscaleDNN) for solving Poisson- Boltzmann equation in complex domains.Journal of Computational Physics, 28:1970–2001, 2020

Ziqi Liu, Wei Cai, and Zhiqin Xu. Multi-scale deep neural network (MscaleDNN) for solving Poisson- Boltzmann equation in complex domains.Journal of Computational Physics, 28:1970–2001, 2020

work page 1970
[21]

Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3):218–229, 2021

Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators.Nature Machine Intelligence, 3(3):218–229, 2021

work page 2021
[22]

A multiple transferable neural network method with domain decomposition for elliptic interface problems.Journal of Computational Physics, 530:113902, 2025

Tianzheng Lu, Lili Ju, and Liyong Zhu. A multiple transferable neural network method with domain decomposition for elliptic interface problems.Journal of Computational Physics, 530:113902, 2025

work page 2025
[23]

The old and the new: Can physics-informed deep-learning replace traditional linear solvers?Frontiers in Big Data, 4, 2021

Stefano Markidis. The old and the new: Can physics-informed deep-learning replace traditional linear solvers?Frontiers in Big Data, 4, 2021

work page 2021
[24]

McClenny and Ulisses M

Levi D. McClenny and Ulisses M. Braga-Neto. Self-adaptive physics-informed neural networks.Journal of Computational Physics, 474:Paper No. 111722, 23, 2023

work page 2023
[25]

fPINNs: fractional physics-informed neural networks

Guofei Pang, Lu Lu, and George Em Karniadakis. fPINNs: fractional physics-informed neural networks. SIAM Journal on Scientific Computing, 41(4):A2603–A2626, 2019

work page 2019
[26]

Hamprecht, Yoshua Bengio, and Aaron C

Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Dr¨ axler, Min Lin, Fred A. Hamprecht, Yoshua Bengio, and Aaron C. Courville. On the spectral bias of neural networks. InInternational Conference on Machine Learning, 2018

work page 2018
[27]

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

work page 2019
[28]

Matched asymptotic expansions-based transferable neural networks for singular perturbation problems, 2025

Zhequan Shen, Lili Ju, and Liyong Zhu. Matched asymptotic expansions-based transferable neural networks for singular perturbation problems, 2025

work page 2025
[29]

Pfnn: A penalty-free neural network method for solving a class of second- order boundary-value problems on complex geometries.Journal of Computational Physics, 428:110085, 2021

Hailong Sheng and Chao Yang. Pfnn: A penalty-free neural network method for solving a class of second- order boundary-value problems on complex geometries.Journal of Computational Physics, 428:110085, 2021

work page 2021
[30]

Dgm: a deep learning algorithm for solving partial differential equations.Journal of Computational Physics, 375:1339–1364, 2018

Justin Sirignano and Konstantinos Spiliopoulos. Dgm: a deep learning algorithm for solving partial differential equations.Journal of Computational Physics, 375:1339–1364, 2018

work page 2018
[31]

Local randomized neural networks with discontinuous galerkin methods for partial differential equations.Journal of Computational and Applied Mathematics, 445:115830, 2024

Jingbo Sun, Suchuan Dong, and Fei Wang. Local randomized neural networks with discontinuous galerkin methods for partial differential equations.Journal of Computational and Applied Mathematics, 445:115830, 2024

work page 2024
[32]

Two-level random feature methods for elliptic partial differential equa- tions over complex domains.Computer Methods in Applied Mechanics and Engineering, 441:Paper No

Yifei Sun and Jingrun Chen. Two-level random feature methods for elliptic partial differential equa- tions over complex domains.Computer Methods in Applied Mechanics and Engineering, 441:Paper No. 117961, 15, 2025. 23

work page 2025
[33]

Timemixer++: A general time series pattern machine for universal predictive analysis

Shiyu Wang, Jiawei LI, Xiaoming Shi, Zhou Ye, Baichuan Mo, Wenze Lin, Ju Shengtong, Zhixuan Chu, and Ming Jin. Timemixer++: A general time series pattern machine for universal predictive analysis. InThe Thirteenth International Conference on Learning Representations, 2025

work page 2025
[34]

When and why PINNs fail to train: a neural tangent kernel perspective.Journal of Computational Physics, 449:Paper No

Sifan Wang, Xinling Yu, and Paris Perdikaris. When and why PINNs fail to train: a neural tangent kernel perspective.Journal of Computational Physics, 449:Paper No. 110768, 28, 2022

work page 2022
[35]

An extreme learning machine-based method for computational PDEs in higher dimensions.Journal of Computational Physics, 418:Paper No

Yiran Wang and Suchuan Dong. An extreme learning machine-based method for computational PDEs in higher dimensions.Journal of Computational Physics, 418:Paper No. 116578, 31, 2024

work page 2024
[36]

A practical pinn framework for multi-scale problems with multi-magnitude loss terms.Journal of Computational Physics, 510:113112, 2024

Yong Wang, Yanzhong Yao, Jiawei Guo, and Zhiming Gao. A practical pinn framework for multi-scale problems with multi-magnitude loss terms.Journal of Computational Physics, 510:113112, 2024

work page 2024
[37]

Solving multiscale elliptic problems by sparse radial basis function neural networks.Journal of Computational Physics, 492:Paper No

Zhiwen Wang, Minxin Chen, and Jingrun Chen. Solving multiscale elliptic problems by sparse radial basis function neural networks.Journal of Computational Physics, 492:Paper No. 112452, 16, 2023

work page 2023
[38]

The random feature method for elliptic eigenvalue problems.Advances in Applied Mathematics and Mechanics, 17(4):1133–1170, 2025

Kexiang Xiong, Zhen Zhang, Jingrun Chen, and Yixin Luo. The random feature method for elliptic eigenvalue problems.Advances in Applied Mathematics and Mechanics, 17(4):1133–1170, 2025

work page 2025
[39]

Weak transnet: A petrov-galerkin based neural network method for solving elliptic PDEs, 2025

Zhihang Xu, Min Wang, and Zhu Wang. Weak transnet: A petrov-galerkin based neural network method for solving elliptic PDEs, 2025

work page 2025
[40]

Frequency principle: Fourier analysis sheds light on deep neural networks.Journal of Computational Physics, 28:1746–1767, 2020

Zhiqin Xu, Yaoyu Zhang, Tao Luo, Yanyang Xiao, and Zheng Ma. Frequency principle: Fourier analysis sheds light on deep neural networks.Journal of Computational Physics, 28:1746–1767, 2020

work page 2020
[41]

B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data.Journal of Computational Physics, 425:Paper No

Liu Yang, Xuhui Meng, and George Em Karniadakis. B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data.Journal of Computational Physics, 425:Paper No. 109913, 23, 2021

work page 2021
[42]

Weak adversarial networks for high- dimensional partial differential equations.Journal of Computational Physics, 411:109409, 14, 2020

Yaohua Zang, Gang Bao, Xiaojing Ye, and Haomin Zhou. Weak adversarial networks for high- dimensional partial differential equations.Journal of Computational Physics, 411:109409, 14, 2020

work page 2020
[43]

Huajian Zhang, Chao Li, Rui Xia, Xinhai Chen, Tiaojie Xiao, Xiao-Wei Guo, and Jie Liu. Fe-pirbn: Feature-enhanced physics-informed radial basis neural networks for solving high-frequency electromag- netic scattering problems.Journal of Computational Physics, 527:113798, 2025

work page 2025
[44]

Zhang, F

Z. Zhang, F. Bao, L. Ju, and G. Zhang. Transferable neural networks for partial differential equations. Journal of Scientific Computing, 99(1):2, 2024. 24

work page 2024