[论文解读] Structure-preserving Randomized Neural Networks for Incompressible Magnetohydrodynamics Equations

Yunlong Li, fei wang+1 more|arXiv (Cornell University)|Mar 1, 2026

Model Reduction and Neural Networks被引用 0

一句话总结

tldr: SP-RaNN automatically enforces divergence-free constraints and reformulates training as a linear least-squares problem, solving incompressible MHD with higher accuracy and faster convergence than traditional NN methods and FEM. It uses a space–time approach and nonlinear iterations to linearize equations.

ABSTRACT

The incompressible magnetohydrodynamic (MHD) equations are fundamental in many scientific and engineering applications. However, their strong nonlinearity and dual divergence-free constraints make them highly challenging for conventional numerical solvers. To overcome these difficulties, we propose a Structure-Preserving Randomized Neural Network (SP-RaNN) that automatically and exactly satisfies the divergence-free conditions. Unlike deep neural network (DNN) approaches that rely on expensive nonlinear and nonconvex optimization, SP-RaNN reformulates the training process into a linear least-squares system, thereby eliminating nonconvex optimization. The method linearizes the governing equations through Picard or Newton iterations, discretizes them at collocation points within the domain and on the boundaries using finite-difference schemes, and solves the resulting linear system via a linear least-squares procedure. By design, SP-RaNN preserves the intrinsic mathematical structure of the equations within a unified space-time framework, ensuring both stability and accuracy. Numerical experiments on the Navier-Stokes, Maxwell, and MHD equations demonstrate that SP-RaNN achieves higher accuracy, faster convergence, and exact enforcement of divergence-free constraints compared with both traditional numerical methods and DNN-based approaches. This structure-preserving framework provides an efficient and reliable tool for solving complex PDE systems while rigorously maintaining their underlying physical laws.

研究动机与目标

• Motivate robust, structure-preserving numerical solvers for incompressible MHD that respect divergence-free constraints for velocity and magnetic fields.
• Introduce SP-RaNN, a randomized neural network that enforces pointwise divergence-free conditions inherently.
• Transform training into linear least-squares problems via space–time formulation and Picard/Newton linearization.
• Demonstrate improved accuracy and convergence of SP-RaNN over traditional NN methods and FEM in several benchmark problems.

提出的方法

• Construct divergence-free neural basis functions using curl representations to guarantee ∇·u = 0 and ∇·B = 0 by design.
• Use randomized neural networks where only the last-layer weights are trained, turning training into a linear least-squares problem.
• Linearize the MHD equations via Picard or Newton iterations and discretize with finite differences at collocation points in a space–time domain.
• Formulate a coupled linear system from collocation of the space–time MHD equations and boundary/initial conditions and solve via QR-based least squares.
• Treat time as an additional dimension (space–time approach) and enforce initial conditions as boundary conditions within the system.
• Ensure pressure uniqueness by incorporating an integral constraint on p via Gauss–Legendre integration.

实验结果

研究问题

• RQ1 Can SP-RaNN enforce pointwise divergence-free constraints for velocity and magnetic fields exactly in incompressible MHD simulations?
• RQ2 Does the SP-RaNN framework yield higher accuracy and faster convergence than traditional NN or FEM methods for Stokes, Navier–Stokes, Maxwell, and MHD problems?
• RQ3 How does the space–time formulation with linearized iterations perform across different Reynolds numbers and problem settings?
• RQ4 What is the impact of using divergence-free basis functions on stability and robustness in high-Reynolds-number MHD simulations?

主要发现

• SP-RaNN consistently achieves higher accuracy with fewer degrees of freedom than RaNN due to exact divergence-free enforcement.
• SP-RaNN maintains robustness across high and low Reynolds numbers, performing well for Re = 1 and Re = 1000 in the reported examples.
• For steady Stokes and unsteady 3D Navier–Stokes and MHD tests, SP-RaNN often yields smaller errors in u and p than comparable RaNN setups and some FEM approaches.
• In Example 5.1, SP-RaNN achieves errors on the order of 7.39E-08 for u and 1.76E-13 to 2.63E-13 for divergence, with comparable or better accuracy at various m (DoFs) than RaNN.
• Table comparisons indicate SP-RaNN requires fewer DoFs to achieve similar or better accuracy than RaNN at Re = 1 and Re = 1000.
• SP-RaNN’s divergence-free construction eliminates the need for post-hoc divergence corrections, contributing to stability and reliability in MHD simulations.