Elucidating the Conformal Structure of the Brinkman Penalisation Method for Geometry-Adapted, Structure-Preserving Operator Learning of Hamiltonian PDEs

Teo Deveney, Baige Xu, Takaharu YaguchiarXiv 2026 · arXiv

TL;DR

This paper reveals that the Brinkman penalisation method—traditionally used as a computational trick to handle complex geometries on simple grids—uncovers an exact geometric structure called a multi-conformal symplectic structure. Leveraging this insight, the authors design structure-preserving numerical integrators and a novel neural operator architecture (CSNO) that can learn physics-consistent solutions across varying domain geometries without suffering from unphysical energy drift.

What Problem Does This Problem Address?

Scientific Machine Learning (SciML) has made massive strides in accelerating physical simulations using neural operators. However, two persistent obstacles limit their deployment in real-world engineering:

  1. Complex and Changing Geometries: Standard numerical methods rely on body-fitted meshes, which are expensive to regenerate whenever a shape changes. Neural networks trained on a fixed geometry typically fail to generalize and require complete retraining for new shapes.
  2. Data Quality and Physical Fidelity: Standard neural operators (like the Fourier Neural Operator, or FNO) lack physical constraints. When trained on data, they often introduce unphysical energy drift, leading to catastrophic errors in long-term predictions.

The Brinkman penalisation method solves the geometry problem by embedding complex shapes into a simple Cartesian grid, treating solid obstacles as highly dissipative media using a geometry indicator function $\chi(x)$. However, until now, it was unknown whether this penalisation destroyed the underlying geometric and Hamiltonian structures of the PDEs, making it difficult to build physics-conditional neural surrogates.

Methods

The authors bridge this gap by analyzing the geometric structure of penalised Hamiltonian PDEs and translating their findings into algorithms:

  • Theoretical Discovery: They prove that if a multi-symplectic PDE penalised by partial friction satisfies an algebraic compatibility condition ($K = PK + KP$), it retains a multi-conformal symplectic structure. This yields an exact local conservation law where energy/symplectic density is conserved perfectly in the fluid and decays exponentially inside the solid.
  • Strang Splitting Integrator: They propose a second-order structure-preserving numerical scheme. It splits the PDE into two parts: an exact, pointwise linear ODE for the dissipative (solid) step, and any standard multi-symplectic integrator for the conservative (fluid) step. This avoids the time-step constraints typically imposed by stiff penalisation parameters.
  • Conformal Symplectic Neural Operator (CSNO): They lift finite-dimensional "SympNets" to infinite-dimensional function spaces. The CSNO architecture hardcodes the exact parameter-free dissipation semigroup to sandwich a learnable, multi-symplectic evolution operator (built using kick and drift shears). Because the geometry indicator function $\chi(x)$ is treated as an explicit network input, a single trained model can generalize to entirely new shapes.
Input Geometry (χ) ──> [ Exact Dissipation (D) ] ──> [ Learnable Multi-Symplectic (Sθ) ] ──> [ Exact Dissipation (D) ] ──> Output

Datasets

The authors generated high-quality, physics-consistent training and evaluation datasets using their proposed splitting integrators:

  • 1D Wave Reflection: A Gaussian pulse interacting with a 1D solid block.
  • 2D Acoustic Wave Scattering: A Gaussian pulse colliding with a circular obstacle.
  • 2D Airfoil Wave Propagation: Wave trajectories past randomly sampled symmetric NACA airfoils with varying thicknesses, chords, and angles of attack.
  • 3D Maxwell Electromagnetic Scattering: 3D trajectories of electromagnetic waves scattering past randomly sampled conducting spheres.

Key Insights and Ideas

  • Brinkman is More Than a Trick: Brinkman penalisation isn't just a convenient computational hack; it is an mathematically rigorous geometric framework that naturally models localized dissipation.
  • The Price of Partial Damping: When damping only a subset of physical variables (e.g., damping velocity but not position in waves, or the electric field but not the magnetic field in Maxwell's equations), you must apply an explicit quadratic correction to the Hamiltonian density to preserve the underlying geometry.
  • Hard Architectural Constraints Win: By embedding the exact dissipation mechanism directly into the neural network's architecture, the CSNO is physically guaranteed to obey the conformal conservation law, completely preventing the unphysical energy explosions common in unconstrained models.

Key Results

  • Mathematical Validation: The splitting scheme demonstrated strict second-order time accuracy and remained stable uniformly as the penalisation parameter $\eta \to 0$.
  • Eliminating Energy Explosion: In 2D acoustic scattering, the proposed splitting scheme tracked the exact energy budget perfectly, whereas a standard explicit Euler baseline exhibited catastrophic energy growth of about 100 orders of magnitude by $T=6$.
  • Superior Operator Learning:
  • In the 2D airfoil wave propagation task, the CSNO closely tracked the ground-truth energy over the entire rollout, while the FNO baseline suffered from severe unphysical energy growth and artificial oscillations.
  • For 3D Maxwell scattering, the CSNO achieved a relative $L^2$ error of $0.073 \pm 0.017$, which is over an order of magnitude smaller than the FNO baseline ($0.988 \pm 0.015$). Furthermore, the FNO baseline spuriously lost over 56% of the system's electromagnetic energy, whereas the CSNO tracked the true energy budget to within 0.5%.
Model 3D Maxwell Relative $L^2$ Error Energy Deviation at Final Time
FNO Baseline $0.988 \pm 0.015$ ~56% (Spurious Loss)
CSNO (Proposed) $0.073 \pm 0.017$ < 0.5%

Limitations and Future Work

  • Focus on Linear Systems: The detailed proofs and examples in this work focus primarily on linear equations (wave and Maxwell's). Extending this framework to nonlinear Hamiltonian PDEs (like the nonlinear Schrödinger equation) is an open theoretical challenge.
  • Static Geometries: The current formulation assumes time-independent geometries ($\chi(x)$). Adapting the framework to handle moving or deforming obstacles ($\chi(t, x)$) will be necessary for applications like aeroelasticity or flapping wings.
  • Approximation Theory: While the architecture performs exceptionally well, a rigorous universal approximation theory for infinite-dimensional conformal symplectic neural operators remains to be fully developed.

Summary updated

Abstract

The Brinkman penalisation method embeds boundary-value problems on complex domains into a simple computational box by modeling the solid region as a strongly dissipative medium, avoiding body-fitted mesh generation. We show that multi-symplectic Hamiltonian PDEs regularised by Brinkman-type penalisation retain a multi-conformal symplectic structure under a compatibility condition linking the symplectic matrix and the penalisation projection. This yields an exact local conservation law, under which the multi-symplectic two-form is conserved in the fluid region and decays exponentially inside the solid. The linear wave equation with Brinkman friction and Maxwell's equations with artificial Ohmic conductivity satisfy this condition, with explicit modified Hamiltonian densities. Building on this, we propose (i) structure-preserving numerical integrators via Strang splitting that satisfy a discrete conformal conservation law, and (ii) conformal symplectic neural operators that interleave exact dissipative flows with learnable multi-symplectic evolution operators, allowing geometry-dependent operator learning. Numerical experiments on wave and electromagnetic scattering demonstrate that our methods reproduce correct local energy budgets and avoid unphysical energy drift, providing a principled framework for physics-consistent scientific machine learning on complex domains.
BibTeX
@misc{deveney2026elucidating,
  author = {Teo Deveney and Baige Xu and Takaharu Yaguchi},
  title = {{Elucidating the Conformal Structure of the Brinkman Penalisation Method for Geometry-Adapted, Structure-Preserving Operator Learning of Hamiltonian PDEs}},
  year = {2026},
  eprint = {2609.29847},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.29847},
}
Figure