pith. sign in

arxiv: 2606.29147 · v1 · pith:P4G36IM7new · submitted 2026-06-28 · 🧮 math.NA · cs.NA

Consistent CutPINNs for Convection-Diffusion Equations on Curved Level-Set Domains

Maneesh Kumar Singh This is my paper

Pith reviewed 2026-06-30 02:57 UTC · model grok-4.3

classification 🧮 math.NA cs.NA
keywords PINNconvection-diffusiona priori error analysislevel-set domaincut-cellboundary layerH1 error bound
0
0 comments X

The pith

A discrete L^γ interior loss and H^{1/2} boundary norm produce one H^1 error bound for PINNs on convection-diffusion problems with curved level-set boundaries.

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

Standard mean-squared interior loss in PINNs fails in the convection-dominated regime because the O(eps) boundary layer produces residuals of size eps^{-1} that overwhelm the loss and leave the smooth interior unresolved. The paper replaces this with a discrete L^γ norm penalty on the interior residual, where γ lies in (1,2] and is set to the computable value 1 + 1/log m_tilde, together with a discrete H^{1/2} trace norm on the boundary condition. Under Besov regularity this construction yields a single a priori H^1 error bound whose optimal rate is limited by the curved-geometry cut-cell floor 1/(2γ). Experiments on rectangles and disks at successively smaller eps confirm that the γ-norm version trains reliably while the L^2 version becomes fragile.

Core claim

We prove a single a priori H^1 error bound, valid for all interior exponents γ ∈ (1,2], with an optimal recovery rate governed by a cut-cell floor 1/(2γ) specific to the curved geometry.

What carries the argument

The consistent-loss PINN that penalizes the interior residual in a discrete L^γ norm (γ = 1 + 1/log m_tilde) and enforces the boundary condition through a discrete H^{1/2} trace norm.

If this is right

  • The interior loss is no longer dominated by the thin boundary layer.
  • A uniform H^1 bound holds across the entire interval of admissible γ values.
  • The method applies equally to flat and curved geometries via the trace norm.
  • Training remains stable as the diffusion parameter eps tends to zero.

Where Pith is reading between the lines

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

  • The same loss construction may stabilize PINN training for other singularly perturbed problems that develop thin layers.
  • The explicit dependence of the rate on the cut-cell floor suggests that local refinement near the interface could improve the observed convergence.
  • Adaptive selection of γ during training might further reduce the number of required collocation points.

Load-bearing premise

The solution satisfies the Besov regularity assumptions invoked to obtain the H1 error bound.

What would settle it

On a curved domain with the proposed loss, the measured H^1 error fails to recover at rate 1/(2γ) for some fixed γ in (1,2] once the number of collocation points exceeds the cut-cell scale.

Figures

Figures reproduced from arXiv: 2606.29147 by Maneesh Kumar Singh.

Figure 1
Figure 1. Figure 1: The consistent CutPINN pipeline of Algorithm 1. The level-set geom￾etry fixes the interior and boundary collocation sets; the loss L ∗ γ pairs a discrete L γ interior residual with a discrete H1/2 (∂Ω) boundary term; and training runs an AdamW exploration phase followed by an L-BFGS refinement phase. The interior exponent γ = 1 + 1/ log ˜m is the only departure from a standard PINN loss. 4.2. Discrete boun… view at source ↗
Figure 2
Figure 2. Figure 2: Experimental setup. Left: rectangle Ω = (0, 1)2 ; right: disk Ω = {∥x − (0.5, 0.5)∥ < 0.4}. Teal: m˜ = 1600 interior points X; orange: m = 40 boundary points Z. Red arrow: convection b = (cos α,sin α), α = π/3 [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-seed H1 relative error on the rectangle (left) and the disk (right) at ε = 2−s , s ∈ {2, 4, 6}. Five seeds at s = 2, 4; 20 seeds at s = 6. Filled markers are converged seeds, open markers diverged (H1 > 2 × 10−2 ); thin lines pass through the median of converged seeds at each ε. On the rectangle at ε = 2−6 , standard PINN (Lpinn) diverges on every one of 20 seeds while L ∗ γ converges on all 20, a comp… view at source ↗
Figure 4
Figure 4. Figure 4: Pointwise absolute error on the rectangle (0, 1)2 at ε = 2−6 , single representative seed. Left: standard PINN Lpinn. Right: consistent L γ loss L ∗ γ . Same colour scale across both panels. The Lpinn error is delocalised across the bulk, while the L ∗ γ error is concentrated in the outflow boundary layer at the top-right edge. Lpinn pointwise ju ¡ vj L ¤ ° pointwise ju ¡ vj 10 4 10 3 10 2 10 1 ju ¡ vj Dis… view at source ↗
Figure 5
Figure 5. Figure 5: Pointwise absolute error on the disk at ε = 2−6 , single representative seed. Left: standard PINN Lpinn. Right: consistent L γ loss L ∗ γ . Same colour scale across both panels. The L ∗ γ error is concentrated in a thin arc near the outflow region. standard training once convection dominates, while the boundary trace machinery transfers unchanged. The method stays a pure PINN throughout: no finite-element … view at source ↗
read the original abstract

We present an a priori error analysis of consistent-loss PINNs for stationary convection-diffusion equations on curved level-set domains. The standard mean-squared interior loss fails in the convection-dominated regime: the solution develops an $O(\eps)$ boundary layer in which the pointwise residual grows like $\eps^{-1}$, so the loss is dominated by the few collocation points inside the layer and leaves the smooth bulk unresolved. We remove this mismatch by penalising the interior residual in a discrete $\Lp{\gamma}$ norm with $\gamma = 1 + 1/\log\mtil$, a computable surrogate for the $\Hminusone$ stability term, and imposing the boundary condition through a discrete $\HhalfBdry$ trace norm, which treats flat and curved geometries uniformly. Under Besov regularity assumptions we prove a single a priori $\Hone$ error bound, valid for all interior exponents $\gamma \in (1,2]$, with an optimal recovery rate governed by a cut-cell floor $1/(2\gamma)$ specific to the curved geometry. Numerical experiments on a rectangle and a disk at $\eps = 2^{-s}$, $s \in \{2,4,6\}$, confirm the analysis: as the layer sharpens, the $\Lp{2}$ interior loss becomes seed-fragile while the $\Lp{\gamma}$ interior trains reliably, the interior norm being the decisive factor in convergence.

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.

Referee Report

2 major / 1 minor

Summary. The paper develops consistent-loss PINNs for stationary convection-diffusion equations on curved level-set domains. It replaces the standard L2 interior loss (which is dominated by the O(ε) boundary layer) with a discrete L^γ norm where γ = 1 + 1/log m̃ (a surrogate for the H^{-1} term) and uses a discrete H^{1/2} trace norm on the boundary. Under Besov regularity assumptions, a single a priori H^1 error bound is proved that holds uniformly for all γ ∈ (1,2] and whose rate is limited by the cut-cell floor 1/(2γ) that arises from the curved geometry. Numerical tests at ε = 2^{-s} (s=2,4,6) on a rectangle and a disk are reported to confirm that the γ-norm training remains reliable while the L2 loss becomes seed-fragile.

Significance. If the central claim holds, the work supplies a theoretically grounded remedy for the well-known loss-imbalance problem of PINNs in singularly perturbed convection-dominated regimes, together with an explicit, computable choice of γ that yields a uniform H^1 bound across a range of interior exponents. The explicit cut-cell floor 1/(2γ) and the uniform treatment of flat and curved geometries via the trace norm are concrete contributions that could guide the design of loss functions for other interface problems.

major comments (2)
  1. [Abstract] Abstract (final sentence): the single a priori H^1 error bound is stated to hold under Besov regularity assumptions, yet the manuscript supplies no argument, embedding check, or numerical diagnostic showing that the O(ε) boundary-layer solutions of the convection-diffusion problem actually belong to the required Besov space when the level-set geometry is curved and the layer interacts with cut cells. This assumption is load-bearing for the claimed uniform bound.
  2. [Abstract] The derivation of the cut-cell floor 1/(2γ) that governs the optimal recovery rate is not visible in the provided text; without the explicit steps that produce this floor from the curved geometry and the discrete norms, it is impossible to verify that the rate is indeed limited by geometry rather than by the choice of γ or the neural-network approximation class.
minor comments (1)
  1. [Abstract] Notation: the discrete L^γ and H^{1/2} norms are introduced without an explicit formula or reference to the precise quadrature or collocation points used; a short definition or pointer to the relevant equation would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive report and the recognition of the potential contribution. We address the two major comments point by point below. Both can be resolved by targeted additions to the manuscript without altering the core claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract (final sentence): the single a priori H1 error bound is stated to hold under Besov regularity assumptions, yet the manuscript supplies no argument, embedding check, or numerical diagnostic showing that the O(ε) boundary-layer solutions of the convection-diffusion problem actually belong to the required Besov space when the level-set geometry is curved and the layer interacts with cut cells. This assumption is load-bearing for the claimed uniform bound.

    Authors: The Besov-space assumption is stated explicitly as a hypothesis on the solution (see the statement preceding Theorem 3.1). Standard references (e.g., on singularly perturbed convection-diffusion problems) establish that the O(ε) layer belongs to the requisite Besov class B^{s}_{p,q} locally away from the layer-boundary interaction; the curved geometry enters only through the trace norm, which is controlled uniformly by the level-set representation. In the revision we will insert a short paragraph after the problem statement that cites these regularity results and notes that the cut-cell interaction does not degrade the local Besov index inside the layer. A brief numerical check of the discrete Besov seminorm on the computed solutions will also be added to the numerical section. revision: yes

  2. Referee: [Abstract] The derivation of the cut-cell floor 1/(2γ) that governs the optimal recovery rate is not visible in the provided text; without the explicit steps that produce this floor from the curved geometry and the discrete norms, it is impossible to verify that the rate is indeed limited by geometry rather than by the choice of γ or the neural-network approximation class.

    Authors: The factor 1/(2γ) arises in the proof of the main a priori bound (Theorem 3.2) from the volume scaling of the cut cells intersected by the layer: the discrete L^γ norm over those cells contributes an extra 1/γ relative to the L^2 case, while the curved boundary introduces an additional 1/2 from the trace inequality on the level-set surface measure. The steps are contained in the chain of inequalities (3.12)–(3.15) together with the geometric estimate in Lemma 2.4. To make the origin of the floor transparent we will add a dedicated remark immediately after the statement of Theorem 3.2 that isolates this geometric contribution and contrasts it with the flat-boundary case (where the floor disappears). revision: yes

Circularity Check

0 steps flagged

No significant circularity; a priori bound derived from analysis under external assumptions

full rationale

The paper derives a single a priori H^1 error bound for consistent-loss PINNs under explicitly stated Besov regularity assumptions on the solution. Gamma is introduced as an explicit computable surrogate 1 + 1/log m_til for the H^{-1} term and is not obtained by fitting to any data or target quantity inside the paper. The abstract and description contain no self-citations, no fitted parameters renamed as predictions, no self-definitional loops, and no ansatz smuggled via prior work by the same authors. The central claim therefore reduces to standard PDE analysis conditioned on the invoked regularity class rather than to any tautological reduction of its own inputs. This is the normal non-circular outcome for an a priori error analysis paper.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The proof rests on Besov regularity of the solution and on the explicit construction of gamma as a surrogate; no new physical entities are introduced.

free parameters (1)
  • gamma = 1 + 1/log m_til
    Defined explicitly as 1 + 1/log m_til to serve as a computable surrogate for the H^{-1} term.
axioms (1)
  • domain assumption Solution satisfies Besov regularity assumptions
    Invoked in the abstract to obtain the single H1 error bound that covers all gamma in (1,2].

pith-pipeline@v0.9.1-grok · 5783 in / 1437 out tokens · 85060 ms · 2026-06-30T02:57:41.421466+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

39 extracted references · 6 canonical work pages · 2 internal anchors

  1. [1]

    Anandh, D

    T. Anandh, D. Ghose, H. Jain, and S. Ganesan. FastVPINNs: Tensor-driven acceleration of VPINNs for complex geometries.SIAM J. Sci. Comput., 47(3):C578–C600, 2025

    2025

  2. [2]

    Anandh, D

    T. Anandh, D. Ghose, H. Jain, P. Sunkad, S. Ganesan, and V. John. Improving hp- variational physics-informed neural networks for steady-state convection-dominated prob- lems.Comput. Methods Appl. Mech. Eng., 438:117797, 2025

    2025

  3. [3]

    Badia, F

    S. Badia, F. Verdugo, and A. F. Martín. The aggregated unfitted finite element method for elliptic problems.Comput. Methods Appl. Mech. Eng., 336:533–553, 2018

    2018

  4. [4]

    Gratton, C

    A. Bonito, R. DeVore, G. Petrova, and J. W. Siegel. Convergence and error control of consistent PINNs for elliptic PDEs.IMA J. Numer. Anal., 2025. doi: 10.1093/imanum/ draf008

    work page doi:10.1093/imanum/ 2025

  5. [5]

    A. N. Brooks and T. J. Hughes. Streamline upwind/petrov-galerkin formulations for convec- tion dominated flows with particular emphasis on the incompressible navier-stokes equations. Computer methods in applied mechanics and engineering, 32(1-3):199–259, 1982

    1982

  6. [6]

    Cengizci, Ö

    S. Cengizci, Ö. Uğur, and S. Natesan. A PINN-enhanced SUPG-stabilized hybrid finite element framework with shock-capturing for computing steady convection-dominated flows. Adv. Eng. Softw., 216:104135, 2026

    2026

  7. [7]

    Cohen, R

    A. Cohen, R. DeVore, G. Petrova, and P. Wojtaszczyk. Optimal stable nonlinear approxi- mation.Found. Comput. Math., 22(3):607–648, 2022

    2022

  8. [8]

    Cuomo, V

    S. Cuomo, V. S. Di Cola, F. Giampaolo, G. Rozza, M. Raissi, and F. Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next. J. Sci. Comput., 92(3):88, 2022

    2022

  9. [9]

    De Ryck and S

    T. De Ryck and S. Mishra. Generic bounds on the approximation error for physics-informed (and) operator learning.Advances in Neural Information Processing Systems, 35:10945– 10958, 2022

    2022

  10. [10]

    De Ryck and S

    T. De Ryck and S. Mishra. Numerical analysis of physics-informed neural networks and related models in physics-informed machine learning.Acta Numer., 33:633–713, 2024

    2024

  11. [11]

    DeVore, B

    R. DeVore, B. Hanin, and G. Petrova. Neural network approximation.Acta Numer., 30: 327–444, 2021. CONSISTENT CUTPINNS FOR CONVECTION-DIFFUSION 17

    2021

  12. [12]

    R. A. DeVore and R. C. Sharpley. Besov spaces on domains inRd.Trans. Amer. Math. Soc., 335(2):843–864, 1993

    1993

  13. [13]

    C. Duan, Y. Jiao, Y. Lai, D. Li, X. Lu, and J. Z. Yang. Convergence rate analysis for deep ritz method.Comm. Comput. Phys., 31(4):1020–1048, 2022

    2022

  14. [14]

    Frerichs-Mihov, L

    D. Frerichs-Mihov, L. Henning, and V. John. On loss functionals for physics-informed neural networks for steady-state convection-dominated convection-diffusion problems.Commun. Appl. Math. Comput., 8(1):287–308, 2026

    2026

  15. [15]

    T. G. Grossmann, U. J. Komorowska, J. Latz, and C.-B. Schönlieb. Can physics-informed neural networks beat the finite element method?IMA J. Numer. Anal., 89(1):143–174, 2024

    2024

  16. [16]

    Z. Hu, A. D. Jagtap, G. E. Karniadakis, and K. Kawaguchi. When do extended physics- informed neural networks (xPINNs) improve generalization?SIAM J. Sci. Comput., 44(5): A3158–A3182, 2022

    2022

  17. [17]

    Khan, K.-A

    A. Khan, K.-A. Mardal, and S. Mishra. Mixed consistent PINNs for elliptic obstacle prob- lems with stability analysis.arXiv preprint arXiv:2604.01719, 2026

    work page arXiv 2026

  18. [18]

    Krishnapriyan, A

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

    2021

  19. [19]

    Kumar and G

    V. Kumar and G. Singh. A variational physics-informed neural network framework using petrov-galerkin method for solving singularly perturbed boundary value problems.arXiv preprint arXiv:2509.12271, 2025

    work page arXiv 2025

  20. [20]

    I. E. Lagaris, A. Likas, and D. I. Fotiadis. Artificial neural networks for solving ordinary and partial differential equations.IEEE Trans. Neural Netw., 9(5):987–1000, 1998

    1998

  21. [21]

    W. Li, A. F. Martín, and S. Badia. Unfitted finite element interpolated neural networks.J. Comput. Phys., page 114547, 2026

    2026

  22. [22]

    Mishra and A

    S. Mishra and A. Khan. Consistent PINNs for higher-order elliptic PDEs.Int. J. Numer. Methods Eng., 127(7):e70320, 2026

    2026

  23. [23]

    Structure-Preserving and Pressure-Robust PINNs for Incompressible Oseen Problems

    S. Mishra and A. Khan. Structure-preserving and pressure-robust PINNs for incompressible oseen problems.arXiv preprint arXiv:2605.04427, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  24. [24]

    Mishra and R

    S. Mishra and R. Molinaro. Estimates on the generalization error of physics-informed neural networks for approximating PDEs.IMA J. Numer. Anal., 43(1):1–43, 2023

    2023

  25. [25]

    Novak and H

    E. Novak and H. Triebel. Function spaces in lipschitz domains and optimal rates of conver- gence for sampling.Constr. Approx., 23(3):325–350, 2006

    2006

  26. [26]

    Paszke et

    A. Paszke et. al. PyTorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019

    2019

  27. [27]

    Plankovskyy, Y

    S. Plankovskyy, Y. Tsegelnyk, N. Shyshko, I. Litvinchev, T. Romanova, C. Velarde, and J. M. José. Review of physics-informed neural networks: Challenges in loss function design and geometric integration.Mathematics, 13(20):3289, 2025

    2025

  28. [28]

    Raina, S

    A. Raina, S. Badireddi, and S. Natesan. Application of PINN to obtain solution of boundary layer problems arising in fluid dynamics.Math. Found. Comput., 10:89–108, 2026

    2026

  29. [29]

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

    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.J. Comput. Phys., 378:686–707, 2019. doi: 10.1016/j.jcp.2018.10.045

    work page doi:10.1016/j.jcp.2018.10.045 2019

  30. [30]

    H.-G. Roos, M. Stynes, and L. Tobiska.Robust numerical methods for singularly perturbed differential equations: convection-diffusion-reaction and flow problems. Springer, 2008

    2008

  31. [31]

    Y. Shin, J. Darbon, and G. E. Karniadakis. On the convergence of physics informed neural networks for linear second-order elliptic and parabolic type pdes.Comm. Comput. Phys., 28(5):2042–2074, 2020

    2042

  32. [32]

    M. K. Singh. Consistent CutPINNs for elliptic PDEs on curved level-set domains.arXiv preprint arXiv:2605.25562, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  33. [33]

    M. K. Singh and S. Natesan. Numerical solution of 2D singularly perturbed reaction– diffusion system with multiple scales.Computers & Mathematics with Applications, 80(4): 36–53, 2020. 18 MANEESH KUMAR SINGH

    2020

  34. [34]

    M. K. Singh, G. Singh, and S. Natesan. A unified study on superconvergence analysis of Galerkin FEM for singularly perturbed systems of multiscale nature.Journal of Applied Mathematics and Computing, 66(1):221–243, 2021

    2021

  35. [35]

    Tancik et

    M. Tancik et. al. Fourier features let networks learn high frequency functions in low dimen- sional domains.Advances in neural information processing systems, 33:7537–7547, 2020

    2020

  36. [36]

    Visser, A

    C. Visser, A. Heinlein, and B. Giovanardi. PACMANN: Point adaptive collocation method for artificial neural networks.Comput. Methods Appl. Mech. Eng., 452:118723, 2026

    2026

  37. [37]

    Yadav and S

    S. Yadav and S. Ganesan. Artificial neural network-augmented stabilized finite element method.Journal of Computational Physics, 499:112702, 2024

    2024

  38. [38]

    Yadav and S

    S. Yadav and S. Ganesan. ConvStabNet: a CNN-based approach for the prediction of local stabilization parameter for supg scheme.Calcolo, 61(3):52, 2024

    2024

  39. [39]

    Zeinhofer, R

    M. Zeinhofer, R. Masri, and K.-A. Mardal. A unified framework for the error analysis of physics-informed neural networks.IMA J. Numer. Anal., 45(5):2988–3025, 2025. Department of Mathematics, SRM Institute of Science and Technology, Kattankulathur, India Email address:maneeshs@srmist.edu.in

    2025