Structure-preserving Randomized Neural Networks for Incompressible Magnetohydrodynamics Equations

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.

💡 Research Summary

The paper introduces a novel computational framework called Structure‑Preserving Randomized Neural Network (SP‑RaNN) for solving the incompressible magnetohydrodynamics (MHD) equations, which couple the Navier‑Stokes and Maxwell systems and impose two divergence‑free constraints (∇·u = 0 for velocity and ∇·B = 0 for magnetic field). Traditional solvers, including finite‑element methods (FEM) and recent physics‑informed neural networks (PINNs), either enforce these constraints only weakly or require costly non‑convex optimization, leading to reduced accuracy and stability.

Key innovations of SP‑RaNN are:

1. Divergence‑free basis construction – By exploiting the vector calculus identities ∇·(∇×Ψ)=0 (3‑D) and ∇·(curl ψ)=0 (2‑D), the authors build basis functions that are analytically divergence‑free. Randomly initialized hidden‑layer weights define scalar potentials ψ_i (or vector potentials Ψ_i). Applying curl to these potentials yields vector fields φ_i that automatically satisfy the pointwise divergence‑free condition.

2. Randomized neural network architecture – All weights and biases in the hidden layers are fixed after random initialization, leaving only the output‑layer weights as trainable parameters. Consequently, the solution is expressed as a linear combination of the divergence‑free basis functions, and learning reduces to solving a linear least‑squares problem for the output coefficients. This eliminates the non‑convex optimization that plagues deep neural network training.

3. Linearization of the nonlinear MHD system – The authors employ Picard or Newton iteration to linearize the governing equations at each outer iteration. The linearized PDEs are then discretized in a unified space‑time collocation framework, treating time and space variables on equal footing and avoiding traditional time‑stepping error accumulation. Finite‑difference stencils (second‑ or fourth‑order) are used to approximate spatial derivatives at collocation points, while Dirichlet, Neumann, or mixed boundary conditions are incorporated directly into the linear system.

4. Efficient solution of the resulting linear system – Because the only unknowns are the output‑layer coefficients, the assembled system is solved by standard linear least‑squares solvers (QR decomposition, SVD), guaranteeing a global optimum and dramatically reducing computational cost.

The methodology is validated on three benchmark problems: (i) steady and unsteady Navier‑Stokes flow, (ii) Maxwell’s equations in a homogeneous medium, and (iii) the full incompressible MHD system in two and three dimensions. Numerical results demonstrate that SP‑RaNN achieves:

• Higher accuracy – L2 errors are 1–2 orders of magnitude lower than those of PINNs and comparable or better than high‑order FEM on the same mesh.
• Exact enforcement of divergence constraints – The computed velocity and magnetic fields satisfy ∇·u = 0 and ∇·B = 0 up to machine precision, eliminating spurious pressure modes or magnetic monopole artifacts.
• Faster convergence – Newton‑based SP‑RaNN converges in 30–50 % fewer outer iterations than Picard, and the overall wall‑clock time is roughly half that of deep‑learning‑based approaches because only a linear solve is required.
• Robustness at high Reynolds and magnetic Reynolds numbers – The structure‑preserving nature of the basis functions yields stable solutions even in regimes where traditional FEM may suffer from numerical instability.

The authors conclude that SP‑RaNN provides a mathematically consistent, physically faithful, and computationally efficient alternative for solving complex, multi‑physics PDE systems with intrinsic divergence‑free structures. Future work is outlined to extend the approach to adaptive collocation point selection, more intricate geometries, and higher‑dimensional problems, further broadening the applicability of structure‑preserving randomized neural networks in scientific computing.