Mesh Field Theory: Port–Hamiltonian Formulation of Mesh-Based Physics
Satoshi Noguchi, Yoshinobu Kawahara · 2026
Read Mesh Field Theory: Port–Hamiltonian Formulation of Mesh-Based Physics by Satoshi Noguchi, Yoshinobu Kawahara (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-00394.txt.
Introduction
Satoshi Noguchi and Yoshinobu Kawahara present Mesh Field Theory, a framework for simulating physics on meshes that keeps the physics honest by construction. Their starting move is a clean split: the topological structure of a mesh, which cells touch which, is separated from its metric structure, the sizes and angles. Minimal principles, such as locality, symmetry under relabeling, and energy balance, drive the theory.
The authors prove a reduction theorem. Under those principles, mesh dynamics must factor into a port-Hamiltonian form. The conservative coupling is fixed uniquely by topology. Only the material response and dissipation remain free. Their neural version learns only those free parts. In tests it showed near-zero energy drift, correct wave behavior, and momentum conservation. It also extrapolated strongly to unseen cases with high data efficiency.
Catalog record
| CREATOR | Satoshi Noguchi, Yoshinobu Kawahara |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 116 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.00394
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Mesh Field Theory: Port–Hamiltonian Formulation of Mesh-Based Physics Satoshi Noguchi, Yoshinobu Kawahara Abstract We present Mesh Field Theory (MeshFT) and its neural realization, MeshFT-Net: a structure-preserving framework for mesh-based continuum physics that cleanly separates the physics’ topological structure from its metric structure. Imposing minimal physical principles (locality, permutation equivariance, orientation covariance, and energy balance/dissipation inequality), we prove a reduction theorem for mesh-based physics. Under these conditions, the physical dynamics admit a local factorization into a port–Hamiltonian form: the conservative interconnection is fixed uniquely by mesh topology, whereas metric effects enter only through constitutive relations and dissipation. This reduction clarifies what must be fixed and what should be learned, directly informing MeshFT-Net’s design. Across evaluations on analytic and realistic datasets, physics-consistency tests, and out-of-distribution validation, MeshFT-Net achieves near-zero energy drift and strong physical fidelity—correct dispersion and momentum conservation—along with robust extrapolation and high data efficiency. By eliminating non-physical degrees of freedom and learning only metric-dependent structure, MeshFT provides a principled inductive bias for stable, faithful, and data-efficient physical simulation. The code is available at https://github.com/noguchisatoshi/MeshFT-Net . 1 Introduction Machine learning is increasingly used to accelerate continuum simulations, from data-driven surrogates and weak-form training to operator learning that transfers across meshes and parameters ( 28 ; 24 ; 14 ; 27 ; 19 ; 35 ; 39 ; 43 ; 4 ; 2 ) . Another line of work represents a discretized domain as a graph built from a mesh and learns time evolution by message passing, yielding strong results in analysis of fluid and deformable solids while remaining flexible in resolution and topology ( 34 ; 38 ) . However, a key structural point is often left implicit. In exterior calculus on a manifold ( 16 ) , the exterior derivative d is topological (metric-independent) while geometry and material properties appear only through metric-dependent operators such as the Hodge star \star . In many existing learned mesh simulators, these roles are conflated, letting non-physical effects contaminate predicted fields and produce spurious modes and instabilities. Discrete exterior calculus (DEC) ( 23 ; 12 ; 13 ) argues that mesh topology provides the algebraic backbone for differential operators describing mesh-based physics. Also, metric-dependent part is clearly divided from such topological structures. Classical structure-preserving numerical schemes (e.g., finite-difference time-domain (FDTD)) exploits the same geometric structure and, as a result, yield stable and conservative updates ( 45 ; 41 ; 32 ; 3 ) . These observations suggest a clean division of roles in mesh-based physics simulators: hard-code topological structures based on the algebraic backbone given by fixed mesh topology, and learn only metric-dependent structures. We develop this idea as Mesh Field Theory (MeshFT) and its neural realization MeshFT-Net. First, we formalize four physical requirements, locality, permutation equivariance, orientation covariance, and an energy balance/passivity inequality. Then, under these conditions, we prove that physical dynamics on a mesh follow a local port–Hamiltonian reduction in which the conservative interconnection is fixed by mesh topology, while metric effects enter only through constitutive and dissipative operators. Importantly, this local port–Hamiltonian structure is deduced from the given principles rather than assumed as a prescribed global template. Guided by the theorem, MeshFT-Net can implement the principled inductive bias by fixing the topology-wired conservative coupling and learning only metric and dissipation factors, without assuming a known partial differential equation (PDE) form or a predefined global structure template. Across evaluation on analytic datasets and acoustic-scattering data from The Well ( 33 ; 30 ) , together with physics-consistency tests and out-of-distribution (OOD) validation, MeshFT-Net achieves near-zero energy drift while preserving correct dispersion and momentum behavior, and it improves generalization and data efficiency relative to baselines. These results support that the principle-driven approach, which induces an explicit topology–metric separation, provides an inductive bias for stable and faithful mesh-based simulation. Our contributions can be summarized as follows: Reduction theorem. We formalize four physical requirements and prove a principle-driven local reduction theorem. Without assuming a known PDE form or committing to a prescribed global dynamics template, the dynamics reduce locally to a port–Hamiltonian form in which the conservative interconnection has topology-fixed signed-incidence sparsity, while only metric-dependent constitutive and dissipative effects are learnable. Neural architecture. Based on the theorem, we design MeshFT-Net to fix the topology-wired incidence structure and learn only metric-dependent constitutive/dissipative operators. Empirical results. Across evaluation on acoustic-scattering data from The Well , physics-consistency tests, and OOD validation, MeshFT-Net shows near-zero energy drift with strong physical fidelity, robust extrapolation performance, and higher data efficiency. 2 Related Work By comparison with related work, we position MeshFT as a principle-derived hypothesis class for mesh-based physics. Starting from physical principles, MeshFT deduces a Jacobian-level local reduction in which conservative coupling is constrained to a signed-incidence wiring determined by mesh topology, while learning is confined to local metric and dissipation operators. This differs from approaches that begin by postulating a global Hamiltonian or port–Hamiltonian template and then learning the remaining components within that form. Starting from principles avoids committing to a single global template, which improves robustness to model mismatch. MeshGraphNets (MGN). MGN learn mesh-based dynamics via permutation-equivariant message passing on mesh graphs ( 34 ; 2 ) . From our viewpoint, MeshGraphNets can be seen as relaxations of MeshFT: by not enforcing the orientation and energy constraints, they can represent a broader class of interactions, which offers flexibility but does not guarantee a part of physical requirements and can admit spurious modes (Fig. 1 ). Structure-Preserving Learning of Dynamics. Hamiltonian and Lagrangian neural networks learn an energy and induce a structured field, e.g., (\partial_{p}H_{\theta},-\partial_{q}H_{\theta}) in canonical coordinates ( 17 ; 9 ; 10 ; 15 ) . In addition, some approaches adopt a port–Hamiltonian and learn ingredients such as a Hamiltonian and dissipation operators from data ( 11 ; 8 ; 31 ) . Thermodynamically structured models such as metriplectic or GENERIC typically fit operators within a prescribed global template ( 46 ; 21 ; 25 ; 22 ; 20 ) . Symplectic neural ODE methods combine learned dynamics with symplectic integration ( 47 ) . These approaches are effective when an appropriate template is available. MeshFT instead deduces a local port–Hamiltonian structure from fundamental physical principles, which constrains conservative coupling by topology and leaves only local metric and dissipation factors to learn. Neural Constitutive Laws. Constitutive-learning approaches often assume a known continuum PDE and learn unknown parameter fields ( 29 ) . MeshFT does not assume an explicit continuum PDE or PDE-residual supervision. It fixes the topology-wired coupling and learns metric-dependent operators, which can be interpreted as learning constitutive relations in some cases. DEC and Data-Driven Exterior Calculus. DEC formalizes structure-preserving discretizations in which cochains represent fields on k -cells and the coboundary operator D_{k} satisfies D_{k+1}D_{k}=0 ( 23 ; 12 ; 1 ) . Metric-dependent effects enter through discrete Hodge operators, defining inner products, energies, and constitutive relations. Data-driven exterior calculus similarly fixes incidence operators and learns metric-dependent operators such as Hodge stars ( 42 ) , typically within a discretized operator-learning setting. MeshFT is aligned with this topology–metric split, but starts from physical principles and uses the resulting Jacobian-level reduction to constrain admissible dynamics. Exterior-Calculus Related Networks. Several network architectures incorporate exterior-calculus or bracket-inspired structure, for example GRAND ( 5 ) and bracket-based GNNs (