Boundary-Informed Method of Lines for Physics Informed Neural Networks

Maximilian Cederholm, Siyao Wang, Haochun Wang, Rui‐Hua Xu, Yuefan DengarXiv 2025 · DOI · arXiv

Abstract

We propose a hybrid solver that fuses the dimensionality-reduction strengths of the Method of Lines (MOL) with the flexibility of Physics-Informed Neural Networks (PINNs). Instead of approximating spatial derivatives with fixed finite-difference stencils - whose truncation errors force extremely fine meshes - our method trains a neural network to represent the initial spatial profile and then employs automatic differentiation to obtain spectrally accurate gradients at arbitrary nodes. These high-fidelity derivatives define the right-hand side of the MOL-generated ordinary-differential system, and time integration is replaced with a secondary temporal PINN while spatial accuracy is retained without mesh refinement. The resulting "boundary-informed MOL-PINN" matches or surpasses conventional MOL in accuracy using an order of magnitude fewer collocation points, thereby shrinking memory footprints, lessening dependence on large data sets, and increasing complexity robustness. Because it relies only on automatic differentiation and standard optimizers, the framework extends naturally to linear and nonlinear PDEs in any spatial dimension.

Introduction

Solving partial differential equations (PDEs) is a cornerstone of scientific computing, yet both traditional numerical methods and modern deep learning approaches face significant hurdles. Standard Physics-Informed Neural Networks (PINNs) often struggle with high-dimensional loss landscapes in complex systems like turbulence, requiring vast amounts of collocation points for convergence. Conversely, the classical Method of Lines (MOL) typically relies on finite-difference (FD) stencils, which introduce truncation errors that necessitate memory-intensive fine meshes.

The "Boundary-Informed MOL-PINN" addresses these challenges by fusing MOL's dimensionality reduction with the mesh-free flexibility of PINNs. By replacing error-prone FD stencils with automatic differentiation on a neural surrogate, the method achieves high spatial accuracy and "complexity robustness" while utilizing significantly fewer data points than standard PINNs.

Methodology

The proposed framework operates in two distinct stages to decouple spatial and temporal learning:

  1. Stage I: Spatial Surrogate Learning: A feed-forward neural network (FFNN) is trained to represent the initial spatial profile of the system at $t=0$. As a smooth, continuous function, this surrogate allows for the use of Automatic Differentiation (AD) to calculate spatial derivatives. This provides spectrally accurate gradients at arbitrary points, bypassing the limitations of traditional grid-based stencils.
  2. Stage II: Temporal Evolution: The PDE is transformed into a system of ordinary differential equations (ODEs). Instead of using classical integrators like Runge-Kutta, a secondary, compact temporal PINN learns the evolution of the system over time along specific trajectories or "lines."
Comparison of Normal PINN and the proposed two-stage MOL-PINN architecture.

Experimental Evaluation

The framework was evaluated on several classic benchmarks, including the non-linear Burgers’ Equation, the 3D Taylor-Green Vortex for decaying turbulence, and the 4D spatio-temporal ABC Flow.

Key Results

  • Accuracy and Robustness: In the 3D Taylor-Green Navier-Stokes test, MOL-PINN achieved a Mean Squared Error (MSE) of $2.23 \times 10^{-2}$, significantly outperforming the standard PINN's $5.87 \times 10^{-1}$. Notably, the method exhibits "complexity robustness," where accuracy degrades much more slowly than standard PINNs as the PDE dimensionality increases.
  • Data Efficiency: The hybrid model matches or surpasses conventional MOL accuracy while using approximately 10x fewer residual evaluations, leading to a reduced memory footprint.
Error distribution for the Burgers' equation benchmark using MOL-PINN, showing high fidelity to the exact solution.

Discussion and Future Work

The MOL-PINN achieves spectral accuracy without a grid, allowing for arbitrary placement of trajectories. However, a primary trade-off is temporal efficiency; the computational overhead per training epoch increases with the number of trajectories. Additionally, the decomposition into lines can introduce a "directional grain bias," where accuracy is lower for regions far from the anchored trajectories.

Future research will focus on multi-directional MOL extensions to mitigate grain bias and the implementation of adaptive refinement pipelines to concentrate computational resources on high-error regions.

Summary updated

Abstract

We propose a hybrid solver that fuses the dimensionality-reduction strengths of the Method of Lines (MOL) with the flexibility of Physics-Informed Neural Networks (PINNs). Instead of approximating spatial derivatives with fixed finite-difference stencils - whose truncation errors force extremely fine meshes - our method trains a neural network to represent the initial spatial profile and then employs automatic differentiation to obtain spectrally accurate gradients at arbitrary nodes. These high-fidelity derivatives define the right-hand side of the MOL-generated ordinary-differential system, and time integration is replaced with a secondary temporal PINN while spatial accuracy is retained without mesh refinement. The resulting "boundary-informed MOL-PINN" matches or surpasses conventional MOL in accuracy using an order of magnitude fewer collocation points, thereby shrinking memory footprints, lessening dependence on large data sets, and increasing complexity robustness. Because it relies only on automatic differentiation and standard optimizers, the framework extends naturally to linear and nonlinear PDEs in any spatial dimension.
BibTeX
@misc{cederholm2025boundary,
  author = {Maximilian Cederholm and Siyao Wang and Haochun Wang and Rui‐Hua Xu and Yuefan Deng},
  title = {{Boundary-Informed Method of Lines for Physics Informed Neural Networks}},
  year = {2025},
  eprint = {2510.15852},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.15852},
  howpublished = {arXiv},
}

Loading related content...

Figure