Physics-Informed Patchwork
- 1Scientific Center for Information Technologies and Artificial Intelligence, Sirius University of Science and Technology, 1 Olimpijskij Avenue, 354340 Sirius Federal Territory, Russia
- 2Institute of Applied Mathematics, Far Eastern Branch of the Russian Academy of Sciences, 7 Radio Street, 690041 Vladivostok, Russia
Abstract
We propose Physics-Informed Patchwork (PIP), a domain-decomposition framework intended for a broad class of stationary boundary-value problems through the reuse of the same pretrained boundary-conditioned local neural solver across subdomains. In this study, we assess its feasibility for the Laplace equation with Dirichlet boundary conditions. The general idea is to train the neural model to approximate solutions of a stationary partial differential equation (PDE) on the unit square, with boundary profiles provided as inputs. A global solution is then assembled using the model with frozen weights by optimizing the internal interface profiles between adjacent subdomains so as to reduce mismatches in solution values and normal derivatives across their interfaces. We demonstrate the method using a fully connected neural network as the local solver. In local-solver tests, the mean relative error is 2.064% on the in-distribution (ID) dataset and 6.853–6.954% on two out-of-distribution (OOD) datasets. For the tested global boundary-value problems, computational scaling is investigated up to 104 subdomains. The measured computational time of global-domain assembly exhibits near-linear scaling with the number of subdomains. For the tested nonrectangular global domains assembled from square subdomains, the mean relative error with respect to finite-difference reference solutions is 0.44–1.14%, while for problems with analytical harmonic solutions it is 0.4–1.3%. For the stationary Laplace equation considered here, these results demonstrate the feasibility of assembling global solutions on large and geometrically nontrivial domains by repeatedly reusing a single local neural model without retraining its weights for each new global problem.
Related
- Lazy training of quantum physics informed neural networks
- AKAPINN: Adaptive Kolmogorov-Arnold Physics-Informed Neural Networks for approximating solutions to quasilinear partial differential equations
- Comparison of a Parametric Physics-Informed Neural Network and a Tensorial Reduced-Order Model for the Shallow-Water Dam-Break Problem
- Comparison of neural and spline representations for physics-informed learning
- Physics-Informed Basis Functions for Nonlinear Response Curve Decomposition: A Parsimonious Alternative to Splines
BibTeX
@article{klepcova2026physicsinformed,
title = {{Physics-Informed Patchwork}},
author = {Lubov A. Klepcova and Yuriy A. Shevchenko and Anastasia P. Koroleva},
journal = {Mathematics},
year = {2026},
month = nov,
doi = {10.3390/math14214008},
url = {https://www.mdpi.com/2227-7390/14/21/4008},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Physics-Informed Patchwork
AU - Lubov A. Klepcova
AU - Yuriy A. Shevchenko
AU - Anastasia P. Koroleva
PY - 2026
DA - 2026/11/03
JO - Mathematics
DO - 10.3390/math14214008
UR - https://www.mdpi.com/2227-7390/14/21/4008
AB - We propose Physics-Informed Patchwork (PIP), a domain-decomposition framework intended for a broad class of stationary boundary-value problems through the reuse of the same pretrained boundary-conditioned local neural solver across subdomains. In this study, we assess its feasibility for the Laplace equation with Dirichlet boundary conditions. The general idea is to train the neural model to approximate solutions of a stationary partial differential equation (PDE) on the unit square, with boundary profiles provided as inputs. A global solution is then assembled using the model with frozen weights by optimizing the internal interface profiles between adjacent subdomains so as to reduce mismatches in solution values and normal derivatives across their interfaces. We demonstrate the method using a fully connected neural network as the local solver. In local-solver tests, the mean relative error is 2.064% on the in-distribution (ID) dataset and 6.853–6.954% on two out-of-distribution (OOD) datasets. For the tested global boundary-value problems, computational scaling is investigated up to 104 subdomains. The measured computational time of global-domain assembly exhibits near-linear scaling with the number of subdomains. For the tested nonrectangular global domains assembled from square subdomains, the mean relative error with respect to finite-difference reference solutions is 0.44–1.14%, while for problems with analytical harmonic solutions it is 0.4–1.3%. For the stationary Laplace equation considered here, these results demonstrate the feasibility of assembling global solutions on large and geometrically nontrivial domains by repeatedly reusing a single local neural model without retraining its weights for each new global problem.
SN - 2227-7390
ER -