Development and Analysis of Chien-Physics-Informed Neural Networks for Singular Perturbation Problems

Gautam Singh; Sofia Haider

REVIEW 4 major objections 5 minor 15 references

Development and Analysis of Chien-Physics-Informed Neural Networks for Singular Perturbation Problems

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims C-PINNs, which split a neural-network solution into a smooth outer part and boundary-layer inner parts joined by exponential weights, solve singularly perturbed convection-diffusion and reaction-diffusion problems more acc

desk verdict Applies an existing C-PINN to coupled and 2D singular perturbation problems but never reports true solution errors, so the accuracy claim is unsupported—desk reject. read the letter →

arxiv 2509.10945 v1 pith:DGGYS7YL submitted 2025-09-13 math.NA cs.NA

Gautam Singh , Sofia Haider This is my paper · ORCID

classification math.NAcs.NA MSC 65L1065L1165L2065L6065L7068T07

keywords physics-informedneuralnetworkssingularperturbationproblemsboundarylayerscompositeasymptoticexpansionconvection-diffusionequationsreaction-diffusioncoupledsystemsC-PINN

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a modified physics-informed neural network, called C-PINN, is more accurate than standard PINNs for singularly perturbed problems. Singular perturbation problems are hard because a tiny parameter multiplies the highest derivative, creating sharp boundary layers that standard networks smear out. C-PINN addresses this by approximating the solution as a smooth outer network plus inner networks that decay exponentially away from the boundary layers. The paper applies this construction to 1D and 2D convection-diffusion, reaction-diffusion, and coupled systems, and reports training losses orders of magnitude lower than standard PINNs. A sympathetic reader would care because this is a mesh-free path to resolving boundary layers without extreme grid refinement.

What carries the argument

The central object is the composite neural network u(x) = u_outer(x;θ1) + u_inner0(x;θ2) exp(-p1(x)/δ1) + u_inner1(x;θ3) exp(-p2(x)/δ1), where p1 and p2 are signed distances to the suspected boundary-layer edges and δ1 is the layer-thickness scale tied to the perturbation parameter. The exponential weights force each inner network to contribute only inside its boundary layer, while the outer network handles the smooth remainder; all subnetworks are trained jointly by minimizing the mean squared error of the PDE residual and boundary conditions.

What would settle it

For the reaction-diffusion example with ε = 1e-5, evaluate both the standard PINN and C-PINN solutions at 2000 uniformly spaced points in [0,1] after the same 9500 training epochs and compute the L2 and L∞ errors against the known analytical solution. If C-PINN's true error is not smaller than the standard PINN's—or if its residual loss is low while its pointwise error exceeds the PINN's—the paper's central accuracy claim collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the composite asymptotic-expansion structure—outer neural network plus layer-specific inner neural networks weighted by exp(-distance/layer-thickness)—allows C-PINN to match the known analytical solutions of several singularly perturbed test problems with far lower residual and boundary-condition MSE than a single standard PINN. The reported losses drop from the order of 1e-2 to order 1e-5 or lower within roughly 10,000 epochs, including tiny perturbation parameters down to 1e-10. The authors claim this makes C-PINN better suited to singularly perturbed problems.

Load-bearing premise

The paper equates training loss (mean squared error of the PDE residual and boundary conditions) with solution accuracy, so if low residual loss can coexist with large true solution error, the reported loss comparisons do not prove C-PINN is more accurate.

Editorial extensions

If this is right

If C-PINN delivers the reported convergence, it gives a mesh-free way to solve convection-dominated and reaction-dominated problems with layers of thickness O(1e-5) or thinner, without mesh refinement.
The same outer-plus-inner construction extends to coupled systems with two different perturbation parameters, as demonstrated by the coupled convection-diffusion and reaction-diffusion examples.
The 2D experiments indicate the architecture transfers to higher dimensions with layer-localized inner networks placed along coordinate directions.
Because layer location and thickness are inputs to the construction, the method presupposes prior knowledge of where boundary layers occur; the paper lists automatic layer detection as future work.
A direct corollary is that residual loss alone, not network width or optimizer choice, is the reported basis for the accuracy claim in these experiments.

Reading between the lines

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

Editorial inference: the reported comparisons use training loss, not true solution error on a hold-out grid; a direct pointwise error test against the given analytical solutions would strengthen or qualify the accuracy claim.
Editorial inference: in the 2D example an auxiliary inner network was added near y=1 where no layer theoretically exists; this suggests C-PINN also acts as a flexible residual-reduction tool, so its benefit may not be limited to known layer locations.
Editorial inference: the exponential-decay weighting implicitly encodes a length scale, so C-PINN may be interpretable as a physics-informed network with built-in coordinate stretching; testing it on problems with interior layers or turning points would reveal how sensitive the method is to layer-location input.
Editorial inference: no runtime or cost comparison is given, so the claim of computational efficiency is not established; residual-loss speedups do not guarantee wall-clock or memory advantages.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Dear colleague,

This paper is a straightforward extension of Wang et al.'s C-PINN to coupled 1D and 2D singular perturbation problems. That extension is legitimate and the test problems (convection-diffusion, reaction-diffusion, coupled systems, 2D) are well chosen. The authors are also candid that C-PINN requires knowing boundary-layer locations and thicknesses in advance. But the central accuracy claim—that C-PINNs provide more accurate solutions than conventional methods—is not established.

The papers' quantitative evidence is training loss only. Tables 1, 3, 4, and 6 compare residual MSE between C-PINN and PINN, and Section 3 explicitly equates accuracy with the L2 loss computed as MSE. For singularly perturbed problems, a low residual loss does not imply small solution error. The paper never reports actual L2 or L-infinity errors against the known exact solutions, even though those exact solutions are available for every test case. That is a straightforward omission.

The stress-test note is correct, and it is worse than a generic 'loss is not error' point. With 600 uniform collocation points on [0,1] and epsilons from 1e-5 down to 1e-10, the exponential factors in the C-PINN ansatz are effectively zero at every interior collocation point (safe_exp clamps at exp(-20)). So the inner subnetworks are invisible to the residual loss; the loss is minimized by the outer network alone. The reported loss comparisons do not even show that the boundary-layer subnetworks are being trained, let alone that C-PINN is more accurate.

There is also a post-hoc element in 2D Example 1, where an auxiliary inner network at y=1 is added after the model struggled. The authors disclose this, but it means the experiment is not a clean test of the method. No code or data is provided, so the experiments cannot be reproduced.

On the positive side, the paper follows the prior C-PINN work closely and cites it properly. The idea of applying the method to coupled systems and 2D problems is worth a technical note, but not in this form. A revision that reports true solution errors, includes error bars, uses epsilons where the layer terms are actually visible to the collocation points, and releases code could be useful.

Who is this for? Someone working on PINNs for singular perturbation problems might skim it as a catalog of test cases, but no one should rely on the accuracy claims. I would desk-reject this version.

Referee Report

4 major / 5 minor

Summary. This manuscript applies Chien-PINNs (C-PINNs), a composite neural-network ansatz combining an outer network with exponentially weighted inner subnetworks, to singularly perturbed convection-diffusion, reaction-diffusion, and coupled systems in one and two dimensions. For each test problem, the source term is chosen from a known analytical solution, and the paper reports training-loss curves comparing C-PINN with standard PINN (or standalone C-PINN loss). The central claim is that C-PINNs 'provide a more accurate solution' than conventional methods. The paper includes algorithms and solution plots, but no true solution-error metrics are reported.

Significance. The composite-asymptotic ansatz is a plausible idea for problems with known boundary-layer structure, and the paper extends it to coupled and 2D systems, which is a useful test bed. Credit is due for giving explicit ansatz formulas and algorithmic pseudocode. However, the central accuracy claim is not established: all quantitative evidence is residual/boundary loss, and for the reported parameter ranges the layer terms are numerically invisible at interior collocation points. If corrected with layer-adapted collocation and true error metrics, the results could be a useful benchmark; in the current form, the claims outrun the evidence.

major comments (4)

[Section 3, Tables 1-6, Eq. (6)] The paper measures accuracy by the residual/boundary loss (6), not by comparing the computed solution with the known exact solution. Section 3 states "The model's accuracy is evaluated using the L2 loss, calculated as the MSE", conflating the training objective with solution error. For singular perturbation problems, small residual loss does not imply small solution error, especially when layer terms are invisible to the residual. Tables 1, 3, and 4 therefore do not support the claim of "more accurate solution"; Tables 2 and 5 have no baseline. Please report L2/L∞ errors against the exact solutions.
[§3.1-§3.3, Eqs. (7)-(8), §3.2 safe_exp] With 600 uniformly distributed collocation points on [0,1], the nearest interior point is roughly 8e-4 from a boundary. For ε=1e-5 to 1e-10, exp(-x/ε) is below machine precision at all interior points, and the safe_exp clamp (-20) makes these terms constants (~2e-9). Thus the inner subnetworks contribute almost nothing to the residual loss and receive no gradient; the reported loss minimization is effectively performed by the outer network alone. No conclusion about the composite C-PINN structure can be drawn from these runs. Use layer-adapted sampling or larger ε, and verify that the inner-network parameters actually change during training.
[§3.3 Example 1, §2.4] The "auxiliary inner network around y=1" is an ad hoc modification added after the model "struggled" near y=1, with no theoretical justification and only a vague claim that "global relative error" was reduced. More generally, §2.4 requires knowing boundary-layer locations and thicknesses in advance, and the exponentials in Eqs. (7)-(8) are chosen to match the exact solutions. The experiments therefore demonstrate fitting known layer shapes rather than discovering layers. This limitation should be stated prominently, and at least one example with unknown or not-pre-specified layer structure should be included if generality is claimed.
[§3.2.1, Eqs. (12)-(13)] The coupled-system ansatz is written as a single scalar u(x) = outer_nn1 + inner_nn10·safe_exp(-x/ε) + outer_nn2 + inner_nn20·safe_exp(-x/µ), which is ambiguous for a two-component solution u=(u1,u2). Algorithm 2 instead defines separate uθ1 and uθ2, each with left and right layer terms, whereas the analytical solution has layers only at x=0. The actual network output structure is therefore not reproducible as written. Please specify exactly which ansatz is used for u1 and u2.

minor comments (5)

[Section 2.2, Eq. (4)] p1, p2, and δ1 are used before being defined; notation for layer thickness varies among δ1, δ(ε), and ∂(ε) in Section 2.4.
[§3.1.1] The soft-residual weight w(x)=e^{-λ|R(x)|} is described, but λ=0.8 is mentioned only once and no experiment isolating its effect is shown; the loss-weight parameters λ_Dε, λ_B, λ_I in Eq. (6) are never specified.
[Table 1] At epoch 4500 the C-PINN loss jumps to 0.001955 from ~2.6e-5 at neighboring epochs, contradicting the text that C-PINN "consistently achieves a lower loss across all epochs".
[Section 3] The term "L2 loss" is used for the residual MSE; this should be relabeled to avoid confusion with the L2 solution error.
[Throughout] The name "PI-PINN" is used for the standard PINN baseline; the acronym is nonstandard and should be defined or replaced. Code and full hyperparameters are not provided, hampering reproducibility.

Circularity Check

2 steps flagged · score 6.0 of 10

C-PINN's claimed superiority is partly built in: the ansatz exponentials are taken from the known exact solutions and one 2D architecture is patched after failure; the accuracy metric is residual loss only.

self definitional [Section 3.1.1, Eq. (7); Section 2.4 Algorithm step 1]
"The source term f(x) is chosen such that the analytical solution is given by: y(x) = (1 − e^{(x−1)/ε}) sin(x). ... we use the following composite neural network to represent a uniform solution: y(x; θ1, θ2) = outer nn(x, θ1) + inner nn(x, θ2)·exp(−(1−x)/ε) (Eq. 7)."

The boundary-layer factor in the C-PINN ansatz, exp(−(1−x)/ε), is exactly the layer term e^{(x−1)/ε} in the analytical solution (up to sign). The algorithm instructs the user to 'identify p(x) and ∂(ϵ)' before construction, so layer location and thickness are inputs, not predictions. The network only has to fit coefficients on exponentials that are already correct for the known solution; the low loss and superiority over PINN are substantially guaranteed by construction rather than discovered from the PDE.
fitted input called prediction [Section 3.3, 2D Example 1]
"Our model struggled near y=1, despite no theoretical boundary layer there. To address this, we added an auxiliary inner network around y=1, not to represent a physical layer, but to enhance flexibility and improve accuracy. Through empirical observations during training, this adjustment..."

The architecture is modified after seeing the model's failure on the target problem; the 'auxiliary inner network around y=1' is a post-hoc structural fit to this specific test case. The paper then reports the resulting low loss (Table 5) as evidence of C-PINN accuracy. The claimed prediction of good performance is therefore partly built from the test outcome itself, not from an independent C-PINN construction.

full rationale

Two steps make the central comparative claim partly circular. In every 1D/coupled example, the C-PINN ansatz inserts exponential terms exp(-x/ε), exp(-(1-x)/ε), or exp(-x/µ) that are exactly the boundary-layer terms appearing in the manufactured analytical solutions (e.g., §3.1.1: exact y = (1 - e^{(x-1)/ε}) sin x vs Eq. (7): outer + inner·exp(-(1-x)/ε)). Since Section 2.4 makes 'identify p(x) and δ(ε)' an explicit input, the layer structure is not a prediction of the method; it is an assumption matched to the answer. Additionally, in 2D Example 1 an auxiliary inner network is added only after the model 'struggled near y=1', i.e., the architecture is fitted to the test instance, and that fitted version is then used to demonstrate accuracy. I do not count the residual-loss-as-accuracy issue (Section 3: 'The model's accuracy is evaluated using the L2 loss, calculated as the MSE') as a circular step; that is a validity/correctness concern about whether training loss equals solution error. No load-bearing self-citation is present: C-PINN is attributed to Wang et al. [13] and Opschoor et al. [8], and the authors' own prior work [11,12] is used only to supply test problems. Nevertheless, because the ansatz is tailored to the known solution form and one test is post-hoc patched, the abstract's claim that 'C-PINNs provide a more accurate solution' is substantially built into the construction, warranting a 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim rests on several unproven assumptions: the composite ansatz is valid for all problems, training loss reflects solution accuracy, layer positions are known, and post-hoc architectural additions do not bias results. The free parameters are chosen by hand or from the known exact solutions, so the method is tuned to the test cases.

free parameters (4)

Soft residual weighting parameter lambda = 0.8
Appears in Section 3.1.1: w(x) = e^{-lambda|R(x)|}. No sensitivity analysis; chosen to stabilize training.
Boundary condition weight lambda_BC = 3.0
Used in Section 3.1.1 specifically near x=1 for the convection-diffusion problem. Chosen ad hoc.
Boundary layer thickness delta(epsilon) = epsilon for all examples
Set by assuming layers have thickness O(epsilon), matching the known manufactured solutions. No method to determine thickness for unknown problems.
Auxiliary inner network at y=1 (2D Example 1) = Present
Added in Section 3.3 Example 1 after observing large errors near y=1; a post-hoc architectural choice with no physical basis.

assumptions (5)

ad hoc to paper Composite asymptotic expansion (equation 4) is a valid ansatz for all considered problems
The representation as outer plus exponentially weighted inner solutions is asserted without derivation or error analysis; it is borrowed from reference [13] and assumed to hold for coupled and 2D cases.
domain assumption Low training loss implies accurate solution
The paper evaluates accuracy solely via the MSE loss (Section 3). This assumes the loss landscape is such that a small residual implies a small solution error, which is not generally true for singularly perturbed problems.
domain assumption Boundary layer location and thickness are known a priori
Each test problem is designed so that the layer positions and widths are known from the constructed exact solution. The method does not provide a mechanism to discover them.
ad hoc to paper Soft residual weighting improves solution accuracy rather than merely reducing loss
The weighting w(x) = e^{-lambda|R(x)|} is introduced in Section 3.1.1 without any theoretical justification that it improves the actual solution, only that it stabilizes training.
ad hoc to paper Auxiliary inner network at y=1 does not bias the solution
In 2D Example 1, the extra network is added to fix observed errors; the paper provides no argument that it preserves the true solution's structure.

invented entities (1)

Auxiliary inner network at y=1 (inner-nn-y1)
purpose: Added to improve accuracy near y=1 in 2D Example 1, where the model struggled despite no theoretical boundary layer.
Introduced post hoc after observing training failure, with no physical or theoretical basis. It is a numerical device, not a physical entity.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Development and Analysis of Chien-Physics-Informed Neural Networks for Singular Perturbation Problems." pith.science (2026). https://pith.science/paper/DGGYS7YL

@misc{pith2026250910945,
  author       = {Pith},
  title        = {Pith review of: Development and Analysis of Chien-Physics-Informed Neural Networks for Singular Perturbation Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DGGYS7YL}},
  note         = {Machine review of arXiv:2509.10945}
}

read the original abstract

In this article, we employ Chien-Physics Informed Neural Networks (C-PINNs) to obtain solutions for singularly perturbed convection-diffusion equations, reaction-diffusion equations, and their coupled forms in both one and two-dimensional settings. While PINNs have emerged as a powerful tool for solving various types of differential equations, their application to singular perturbation problems (SPPs) presents significant challenges. These challenges arise because a small perturbation parameter multiplies the highest-order derivatives, leading to sharp gradient changes near the boundary layer. To overcome these difficulties, we apply C-PINNs, a modified version of the standard PINNs framework, which is specifically designed to address singular perturbation problems. Our study shows that C-PINNs provide a more accurate solution for SPPs, demonstrating better performance than conventional methods.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 4 linked inside Pith

[1]

Al-Aradi, A

A. Al-Aradi, A. Correia, D. Naiff, and G. Jardim, Solving Nonlinear and High- Dimensional PDEs via Deep Learning,arXiv preprint, arXiv:1811.08782, https: //arxiv.org/abs/1811.08782, 2018

arXiv 2018
[2]

F. Cao, F. Gao, X. Guo, and D. Yuan, Physics-informed neural networks with parameter asymptotic strategy for learning singularly perturbed convection- dominated problem,Comput. Math. Appl., 150: 229-242, 2023

2023
[3]

Cen, Parameter-Uniform Finite Difference Scheme for a System of Coupled Sin- gularly Perturbed Convection-Diffusion Equations,Int

Z. Cen, Parameter-Uniform Finite Difference Scheme for a System of Coupled Sin- gularly Perturbed Convection-Diffusion Equations,Int. J. Comput. Math., 82(2): 177-192, 2005

2005
[4]

Goodfellow, Y

I. Goodfellow, Y. Bengio, and A. Courville,Deep Learning, MIT Press, 2016

2016
[5]

T. G. Grossmann, U. J. Komorowska, J. Latz, and C. B. Sch¨ onlieb, Can Physics- Informed Neural Networks beat the Finite Element Method?,arXiv preprint, arXiv:2302.04107, https://arxiv.org/abs/2302.04107, 2023

arXiv 2023
[6]

Kumar, B

S. Kumar, B. V. R. Kumar, and S. V. S. S. N. V. G. K Murthy, A deep learning- based numerical approach for the natural convection inside a porous media,Int. J. Adv. Eng. Sci. Appl. Math., 16: 233–243, 2024

2024
[7]

Kumar, B

S. Kumar, B. V. R. Kumar, and S. V. S. S. N. V. G. K Murthy, ANN-based deep collocation method for natural convection in porous media,Neural Comput. Appl., 36: 6067–6083, 2024. 20

2024
[8]

J. A. A. Opschoor, Ch. Schwab, and C. Xenophontos, Neural Networks for Singu- lar Perturbations,arXiv preprint, arXiv:2401.06656, https://arxiv.org/abs/2401. 06656, 2024

arXiv 2024

Show all 15 references

[9]

Singh and R

A. Singh and R. K. Sinha, SS-DNN: A hybrid strang splitting deep neural net- work approach for solving the Allen–Cahn equation,Eng. Anal. Bound. Elem., 169:105944, 15, 2024

2024
[10]

Singh and R

A. Singh and R. K. Sinha, Multi output FOSLS Deep Neural Network for Solving Allen–Cahn Equation,Math. Comput. Simul., 15: 1132–1146, 2023

2023
[11]

Singh and S

G. Singh and S. Natesan, Super convergence of discontinuous Galerkin method with interior penalties for singularly perturbed two-point boundary-value prob- lems,Calcolo: 55(4):54, 30, 2018

2018
[12]

Singh and S

G. Singh and S. Natesan, A uniformly convergent numerical scheme for a coupled system of singularly perturbed reaction-diffusion equations,Numerical Functional Analysis and Optimization, 41:1-18, 03, 2020

2020
[13]

L. Wang, L. Zhang, and G. He , Chien-physics-informed neural networks for solving singularly perturbed boundary-layer problems,Appl. Math. Mech., 45(9): 1467–1480, 2024

2024
[14]

S. Wang, P. Zhao, Q. Ma, and T. Song, General-Kindered Physics-Informed Neu- ral Network to the Solutions of Singularly Perturbed Differential Equations,arXiv preprint, arXiv:2408.14734, https://arxiv.org/abs/2408.14734, 2018

2018 arXiv
[15]

Yadav and S

S. Yadav and S. Ganesan, How Deep Learning Performs with Singularly Per- turbed Problems,IEEE Second International Conference on Artificial Intelligence and Knowledge Engineering (AIKE): 293-297, 10.1109/AIKE.2019.00058, 2019. 21

2019

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Al-Aradi, A

A. Al-Aradi, A. Correia, D. Naiff, and G. Jardim, Solving Nonlinear and High- Dimensional PDEs via Deep Learning,arXiv preprint, arXiv:1811.08782, https: //arxiv.org/abs/1811.08782, 2018

arXiv 2018

[2] [2]

F. Cao, F. Gao, X. Guo, and D. Yuan, Physics-informed neural networks with parameter asymptotic strategy for learning singularly perturbed convection- dominated problem,Comput. Math. Appl., 150: 229-242, 2023

2023

[3] [3]

Cen, Parameter-Uniform Finite Difference Scheme for a System of Coupled Sin- gularly Perturbed Convection-Diffusion Equations,Int

Z. Cen, Parameter-Uniform Finite Difference Scheme for a System of Coupled Sin- gularly Perturbed Convection-Diffusion Equations,Int. J. Comput. Math., 82(2): 177-192, 2005

2005

[4] [4]

Goodfellow, Y

I. Goodfellow, Y. Bengio, and A. Courville,Deep Learning, MIT Press, 2016

2016

[5] [5]

T. G. Grossmann, U. J. Komorowska, J. Latz, and C. B. Sch¨ onlieb, Can Physics- Informed Neural Networks beat the Finite Element Method?,arXiv preprint, arXiv:2302.04107, https://arxiv.org/abs/2302.04107, 2023

arXiv 2023

[6] [6]

Kumar, B

S. Kumar, B. V. R. Kumar, and S. V. S. S. N. V. G. K Murthy, A deep learning- based numerical approach for the natural convection inside a porous media,Int. J. Adv. Eng. Sci. Appl. Math., 16: 233–243, 2024

2024

[7] [7]

Kumar, B

S. Kumar, B. V. R. Kumar, and S. V. S. S. N. V. G. K Murthy, ANN-based deep collocation method for natural convection in porous media,Neural Comput. Appl., 36: 6067–6083, 2024. 20

2024

[8] [8]

J. A. A. Opschoor, Ch. Schwab, and C. Xenophontos, Neural Networks for Singu- lar Perturbations,arXiv preprint, arXiv:2401.06656, https://arxiv.org/abs/2401. 06656, 2024

arXiv 2024