pith. sign in

arxiv: 2606.18032 · v1 · pith:YWCOVXDFnew · submitted 2026-06-16 · 🧮 math.NA · cs.LG· cs.NA· physics.comp-ph

INI-VPINN: A Variational Physics-Informed Neural Network with Implicit Neumann and Interface Handling for Multi-Material Domains with Geometric Singularities

Shayan Dodge (1) , Alessandro Formisano (2) , Sami Barmada (1) ((1) DESTeC , University of Pisa , Pisa , Italy , (2) Department of Engineering , University of Campania Luigi Vanvitelli
show 2 more authors
Aversa Italy)
This is my paper

Pith reviewed 2026-06-26 23:53 UTC · model grok-4.3

classification 🧮 math.NA cs.LGcs.NAphysics.comp-ph
keywords variational physics-informed neural networkmulti-material domainsNeumann boundary conditionsinterface conditionsgeometric singularitiesPoisson equationLaplace equationweak formulation
0
0 comments X

The pith

INI-VPINN incorporates Neumann boundary and interface conditions directly into the variational formulation for multi-material domains.

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

The paper introduces INI-VPINN as a variational physics-informed neural network that solves Poisson and Laplace problems in domains with multiple materials and sharp geometric features. By relying on compact support weighting functions together with integration by parts, the approach folds Neumann boundary values and material-interface continuity into the weak-form loss itself. This removes the usual need for separate penalty terms or separate networks per subdomain. The resulting single-network model produces higher accuracy and smoother convergence on the tested problems than other PINN variants. The framework is presented as a general route to mixed-boundary multi-material problems.

Core claim

INI-VPINN naturally incorporates Neumann boundary and interface conditions into the variational formulation. It removes the need for additional loss terms or multiple subdomain networks. This framework employs compact support weighting functions and integration by parts to implicitly impose flux and continuity constraints, ensuring physical consistency across material boundaries without explicit enforcement.

What carries the argument

Compact support weighting functions combined with integration by parts, which transfer Neumann flux and interface continuity terms into the variational loss.

If this is right

  • A single network suffices for mixed Neumann-Dirichlet multi-material problems without subdomain splitting.
  • Higher pointwise accuracy is obtained on Poisson and Laplace equations with sharp interfaces compared with other PINN formulations.
  • Convergence becomes smoother and requires fewer iterations because interface penalties are absent.
  • The same weak-form construction extends immediately to any linear elliptic operator whose weak form admits integration by parts.

Where Pith is reading between the lines

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

  • The method could be applied to linear elasticity or Stokes flow by replacing the scalar test functions with appropriate vector or tensor versions.
  • Because no interface-specific loss weights are tuned, the approach may reduce the hyperparameter search space in practical deployments.
  • The compact-support construction might allow automatic differentiation libraries to evaluate the loss on unstructured meshes without additional coding for jump conditions.

Load-bearing premise

The weighting functions possess compact support so that integration by parts moves all boundary and interface contributions into the loss without leaving residual explicit terms that must still be enforced separately.

What would settle it

A manufactured multi-material Poisson problem with an exact analytic solution where the network output violates continuity of the normal derivative across the interface by more than the observed discretization error.

Figures

Figures reproduced from arXiv: 2606.18032 by (2) Department of Engineering, Alessandro Formisano (2), Aversa, Italy, Italy), Pisa, Sami Barmada (1) ((1) DESTeC, Shayan Dodge (1), University of Campania Luigi Vanvitelli, University of Pisa.

Figure 1
Figure 1. Figure 1: Test function families: (a) odd cosines, (b) odd sines, and (c) Jacobi difference modes, shown for ranks [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Arrangement of test functions over a 1D domain: Odd cosine modes combined with Jacobi functions for [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Localized weighting function v (k) 1,1 (x, y) on a partitioned domain with mixed boundary conditions. These functions vanishes at the Dirichlet boundaries (ΓD) and are normalised to one at the Neumann boundaries (ΓN ). Element labels indicate the chosen local basis type (J = Jacobi, C = odd cosine, S = odd sine). (a) m = 1, n = 1, (b [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Definition of the local coordinate system at the interface between two media. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Material coefficient κ(x) represented by a sharp Heaviside step function (a,c) and a smooth tanh approxima￾tion with smoothing parameter δ = 0.01 (b,d). Panels (a,b) show the 1D case, while (c,d) illustrate the corresponding 2D extensions. Here, [ · ]ΓI denotes the variation across the interface. The first condition enforces continuity of the function u (often referred as potential function), while the sec… view at source ↗
Figure 6
Figure 6. Figure 6: Schematic of the INI-VPINN framework. The fully connected neural network, sketched in [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Schematic representation of the geometry of the test cases and distribution of the test functions within local [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of predicted u(x, y) for the L-shaped domain with homogeneous material: (a) FEM (Ground Truth), (b) INI-VPINN, (d) VPINN (Ex), with explicitly enforced Neumann boundary condition, (f) PINN. Streamlines in each case illustrate the direction of the field (−∇u(x, y)). (c-e-g) Corresponding absolute error distributions |uPred − uGT|. 14 [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Training loss convergence of PINN, VPINN (Ex), and INI-VPINN models for the homogeneous L-shaped [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison of predicted u(x, y) in the T-shaped domain with homogeneous material using (a) FEM (Ground Truth), (b) INI-VPINN, and (c) VPINN (Ex). The colormap represents the scalar potential, while the streamlines illustrate the flux direction (−∇u). Error distributions relative to FEM are shown in (c–e). 16 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Training loss convergence comparison between VPINN (Ex) and INI-VPINN for the homogeneous T-shaped [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Comparison of predicted u(x, y) in the non-homogeneous L-shaped domain with two materials, where κ = 1 for x < 0 and κ = 8 for x > 0. The solutions obtained with (a) FEM (Ground Truth), (b) INI-VPINN, and (d) cPINN (conservative physics-informed neural network) are shown. The colormap represents the scalar potential, while the streamlines illustrate the flux direction (−∇u). Error distributions relative t… view at source ↗
Figure 13
Figure 13. Figure 13: Comparison of predicted u(x, y) in the non-homogeneous nested square domain, with κ = 8 inside the square inclusion and κ = 1 outside. The solutions obtained with (a) FEM (Ground Truth), (b) INI-VPINN, and (d) cPINN (conservative physics-informed neural network) are shown. The colormap represents the scalar potential, while the streamlines illustrate the field direction (−∇u). Error distributions relative… view at source ↗
Figure 14
Figure 14. Figure 14: Convergence histories of (a) cPINN – first network ( [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of the predicted potential u(x, y) in the non-homogeneous square domain with circular inclusion, where κ = 8 inside the inclusion and κ = 1 outside. The solutions obtained using (a) FEM (ground truth), (b) INI￾VPINN, and (d) CPINN (conservative physics-informed neural network) are presented. The colormap represents the scalar potential, while streamlines indicate the field direction (−∇u). Pane… view at source ↗
Figure 16
Figure 16. Figure 16: Comparison between the FEM reference and the proposed INI-VPINN solution for the Poisson problem [PITH_FULL_IMAGE:figures/full_fig_p022_16.png] view at source ↗
read the original abstract

We propose a new weak-form Physics-Informed Neural Network approach (named INI-VPINN). INI-VPINN naturally incorporates Neumann boundary and interface conditions into the variational formulation. It removes the need for additional loss terms or multiple subdomain networks. This framework employs compact support weighting functions and integration by parts to implicitly impose flux and continuity constraints. In this way, it implicitly ensures physical consistency across material boundaries. The proposed method is tested on Poisson and Laplace problems with sharp interfaces and complex geometries. Results show that, compared with several other Physics Informed Neural Networks-based formulations, the INI-VPINN consistently achieves higher accuracy, smoother and faster convergence. The proposed framework provides a general approach for solving multimaterial problems with complex geometries and mixed Neumann-Dirichlet boundary conditions using neural networks. The implementation is publicly available in a GitHub repository.

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

0 major / 2 minor

Summary. The manuscript proposes INI-VPINN, a variational physics-informed neural network that implicitly incorporates Neumann boundary conditions and interface continuity constraints for multi-material domains with geometric singularities. It achieves this via compact-support weighting functions and integration by parts in the weak form, eliminating explicit loss terms for fluxes or the need for multiple subdomain networks. The approach is demonstrated on Poisson and Laplace problems with sharp interfaces, claiming higher accuracy and smoother/faster convergence relative to other PINN formulations, with public code availability.

Significance. If the variational construction and numerical results hold, the method offers a streamlined, parameter-light framework for interface problems that aligns with standard weak-form treatments of flux continuity. This could reduce architectural complexity in PINN applications to multi-physics and multi-material settings while maintaining physical consistency, with the open implementation supporting reproducibility.

minor comments (2)
  1. [Abstract] Abstract: the claim of 'consistently achieves higher accuracy, smoother and faster convergence' is stated without any quantitative metrics, error tables, or convergence plots; these should be added or referenced to a specific results section/figure for substantiation.
  2. [Methods] The description of 'compact support weighting functions' would benefit from an explicit definition or example in the methods section to clarify how support is chosen relative to interface geometry.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the INI-VPINN manuscript and the recommendation for minor revision. No specific major comments were provided in the report.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central contribution is a reformulation of the variational loss for PINNs that uses compact-support test functions and integration by parts to embed Neumann and interface conditions directly. This is a standard weak-form construction and does not reduce any claimed prediction or uniqueness result to a fitted parameter or to a self-citation chain. No equations are shown to be equivalent by definition to their inputs, and the abstract and supplied context contain no load-bearing self-citations or ansatz smuggling. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the method relies on standard variational principles and neural network training.

pith-pipeline@v0.9.1-grok · 5740 in / 1069 out tokens · 23025 ms · 2026-06-26T23:53:00.117387+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

18 extracted references · 13 canonical work pages

  1. [1]

    Shayan Dodge, Sami Barmada, and Alessandro Formisano

    doi:10.1109/TMAG.2025.3626152. Shayan Dodge, Sami Barmada, and Alessandro Formisano. A stacked adaptive residual pinn (star-pinn) approach to 2d time-domain magnetic diffusion in nonlinear materials. IEEE Access, 13:141380 – 141394,

    work page doi:10.1109/tmag.2025.3626152 2025

  2. [2]

    doi:10.1109/ACCESS.2025.3597869

    ISSN 2169-3536. doi:10.1109/ACCESS.2025.3597869. S Barmada, S Dodge, M Tucci, A Formisano, P Di Barba, and ME Mognaschi. A novel hybrid boundary element–physics informed neural network method for numerical solutions in electromagnetics. IEEE Access, 12:171444–171457,

    work page doi:10.1109/access.2025.3597869 2025

  3. [3]

    Y Chen and L Dal Negro

    doi:10.1109/ACCESS.2024.3500039. Y Chen and L Dal Negro. Physics-informed neural networks for imaging and parameter retrieval of photonic nanostructures from near-field data. APL Photonics, 7(1):010802,

    work page doi:10.1109/access.2024.3500039 2024

  4. [4]

    Physics informed neural network for option pricing

    A Dhiman and Y Hu. Physics informed neural network for option pricing. arXiv preprint arXiv:2312.06711,

    arXiv

  5. [5]

    Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations

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

    2002

  6. [6]

    and Yu, X

    doi:10.1016/j.jcp.2021.110768. LD McClenny and UM Braga-Neto. Self-adaptive physics-informed neural networks. Journal of Computational Physics, 474:111722,

    work page doi:10.1016/j.jcp.2021.110768 2021

  7. [7]

    Ehsan Kharazmi, Zhongqiang Zhang, and George E

    doi:10.1016/j.cma.2019.112790. Ehsan Kharazmi, Zhongqiang Zhang, and George E. Karniadakis. Variational physics-informed neural networks for solving partial differential equations. arXiv preprint arXiv:1912.00873,

    work page doi:10.1016/j.cma.2019.112790 2019

  8. [8]

    Reza Khodayi-Mehr and Michael M

    URL https://arxiv.org/abs/ 1912.00873. Reza Khodayi-Mehr and Michael M. Zavlanos. Varnet: Variational neural networks for the solution of partial differential equations. In Proceedings of the 2nd Conference on Learning for Dynamics and Control, volume 120 ofProceedings of Machine Learning Research, pages 298–307,

    arXiv 1912

  9. [9]

    Yaohua Zang, Gang Bao, Xiaojing Ye, and Haomin Zhou

    URLhttps://arxiv.org/abs/1912.07443. Yaohua Zang, Gang Bao, Xiaojing Ye, and Haomin Zhou. Weak adversarial networks for high-dimensional partial differential equations. Journal of Computational Physics, 411:109409,

    arXiv 1912

  10. [10]

    24 INI-VPINNDODGE ET AL

    doi:10.1016/j.jcp.2020.109409. 24 INI-VPINNDODGE ET AL. Rui Xu, Dongxiao Zhang, Miao Rong, and Nanzhe Wang. Weak form theory-guided neural network (tgnn-wf) for deep learning of subsurface single- and two-phase flow. Journal of Computational Physics, 436:110318,

    work page doi:10.1016/j.jcp.2020.109409 2020

  11. [11]

    Chuang Liu and HengAn Wu

    doi:10.1016/j.jcp.2021.110318. Chuang Liu and HengAn Wu. cv-pinn: Efficient learning of variational physics-informed neural network with domain decomposition. Extreme Mechanics Letters, 63:102051,

    work page doi:10.1016/j.jcp.2021.110318 2021

  12. [12]

    Stefano Berrone, Claudio Canuto, Moreno Pintore, and N

    doi:10.1016/j.eml.2023.102051. Stefano Berrone, Claudio Canuto, Moreno Pintore, and N. Sukumar. Enforcing dirichlet boundary conditions in physics-informed neural networks and variational physics-informed neural networks. Heliyon, 9(8):e18820,

    work page doi:10.1016/j.eml.2023.102051 2023

  13. [13]

    Stefano Berrone and Moreno Pintore

    doi:10.1016/j.heliyon.2023.e18820. Stefano Berrone and Moreno Pintore. Meshfree variational-physics-informed neural networks (mf-vpinn): An adaptive training strategy. Algorithms, 17(9):415,

    work page doi:10.1016/j.heliyon.2023.e18820 2023

  14. [14]

    Sergio Rojas, Paweł Maczuga, Judit Muñoz-Matute, David Pardo, and Maciej Paszy ´nski

    doi:10.3390/a17090415. Sergio Rojas, Paweł Maczuga, Judit Muñoz-Matute, David Pardo, and Maciej Paszy ´nski. Robust variational physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 425:116904,

    work page doi:10.3390/a17090415

  15. [15]

    doi:10.1016/j.cma.2024.116904. Vivek G. Patel and Nikunj V . Rachchh. Meshless method–review on recent developments.Materials Today: Proceedings, 26:1598–1603,

    work page doi:10.1016/j.cma.2024.116904 2024

  16. [16]

    Ben-Yu Guo, Jie Shen, and Li-Lian Wang

    doi:10.1016/j.matpr.2019.12.333. Ben-Yu Guo, Jie Shen, and Li-Lian Wang. Optimal spectral-galerkin methods using generalized jacobi polynomials. Journal of Scientific Computing, 27(1):305–322,

    work page doi:10.1016/j.matpr.2019.12.333 2019

  17. [17]

    arXiv:1412.6980. Eid H. Doha, Ali H. Bhrawy, Mohamed A. Abdelkawy, and Robert A. Van Gorder. Jacobi–gauss–lobatto collocation method for the numerical solution of 1+1 nonlinear schrödinger equations. Journal of Computational Physics, 261: 244–255,

    Pith/arXiv arXiv

  18. [18]

    doi:10.1016/j.jcp.2014.01.003. 25

    work page doi:10.1016/j.jcp.2014.01.003 2014