[Paper Review] Guiding continuous operator learning through Physics-based boundary constraints

Nadim Saad, Gaurav Gupta+2 more|arXiv (Cornell University)|Dec 14, 2022

Model Reduction and Neural Networks4 citations

TL;DR

This paper proposes Boundary-enforcing Operator Network (BOON), a method that embeds physics-based boundary conditions (Dirichlet, Neumann, periodic) directly into the kernel architecture of neural operators to guarantee their satisfaction. By applying structural corrections with linear space complexity and no increase in computational cost, BOON achieves up to 20× relative $L^2$ error reduction—e.g., 0.000084 error on Burgers’ equation—while preserving resolution-invariance and improving accuracy across diverse PDEs.

ABSTRACT

Boundary conditions (BCs) are important groups of physics-enforced constraints that are necessary for solutions of Partial Differential Equations (PDEs) to satisfy at specific spatial locations. These constraints carry important physical meaning, and guarantee the existence and the uniqueness of the PDE solution. Current neural-network based approaches that aim to solve PDEs rely only on training data to help the model learn BCs implicitly. There is no guarantee of BC satisfaction by these models during evaluation. In this work, we propose Boundary enforcing Operator Network (BOON) that enables the BC satisfaction of neural operators by making structural changes to the operator kernel. We provide our refinement procedure, and demonstrate the satisfaction of physics-based BCs, e.g. Dirichlet, Neumann, and periodic by the solutions obtained by BOON. Numerical experiments based on multiple PDEs with a wide variety of applications indicate that the proposed approach ensures satisfaction of BCs, and leads to more accurate solutions over the entire domain. The proposed correction method exhibits a (2X-20X) improvement over a given operator model in relative $L^2$ error (0.000084 relative $L^2$ error for Burgers' equation).

Motivation & Objective

• To address the lack of guaranteed boundary condition (BC) satisfaction in neural operator models for PDEs, which leads to non-physical and unstable solutions.
• To develop a method that enforces BCs explicitly through architectural modifications, rather than relying on implicit learning from data.
• To maintain the resolution-invariant and efficient nature of neural operators while ensuring physical consistency.
• To demonstrate state-of-the-art performance across diverse PDEs with complex BCs, including Dirichlet, Neumann, and periodic conditions.

Proposed method

• Introduce a systematic kernel architecture modification that embeds boundary condition constraints directly into the neural operator's integral kernel.
• Design three numerically efficient algorithms to apply boundary corrections with linear space complexity and no increase in computational cost.
• Use a correction procedure that modifies the kernel weights based on the physical boundary constraints, ensuring the solution satisfies BCs by construction.
• Prove bounded error estimates showing that the correction induces only controlled changes in the solution, preserving model stability.
• Apply the correction to multiple neural operator variants, including FNO and MWT, demonstrating generalization across architectures.
• Preserve the resolution-invariant property of neural operators by ensuring the correction does not depend on the input resolution.

Experimental results

Research questions

• RQ1 Can physics-based boundary conditions be embedded directly into the kernel of a neural operator to guarantee their satisfaction without relying on data-driven learning?
• RQ2 How can boundary condition enforcement be achieved with minimal computational overhead while preserving the resolution-invariance of neural operators?
• RQ3 What is the impact of explicit BC enforcement on solution accuracy and generalization across different PDEs and boundary types?
• RQ4 Can the proposed correction method be generalized to multiple neural operator architectures, such as FNO and MWT?
• RQ5 How does the performance of the corrected model (BOON) compare to baseline neural operators in terms of error reduction and boundary consistency?

Key findings

• BOON achieves a relative $L^2$ error of 0.000084 on Burgers’ equation—representing a 20× improvement over the baseline FNO model.
• For the 1D heat equation with Neumann BCs, BOON reduces relative $L^2$ error to 0.00811 at resolution $N=500$, compared to 0.01730 for FNO.
• BOON maintains high accuracy when generalizing to higher resolutions: when trained at $N=1024$, it achieves only 1.25% relative $L^2$ error at $N=4096$, demonstrating strong resolution invariance.
• The method ensures exact satisfaction of Dirichlet, Neumann, and periodic boundary conditions, as verified by boundary error metrics consistently reported as zero.
• BOON improves performance across multiple PDEs, including the 2D Navier-Stokes lid cavity problem, with consistent error reduction and physical consistency.
• The correction procedure is generalizable, as demonstrated by successful application to both FNO and MWT architectures, with significant error reduction in all tested benchmarks.