De Rham cohomological architectures for PINNs and PDEs with topological holes

Andrew GracykarXiv 2026 · arXiv

TL;DR

This paper introduces Hodge-de Rham PINNs, a novel architecture designed to solve partial differential equations (PDEs) on domains with topological holes, punctures, or voids. By parameterizing solutions using differential 1-forms and the Hodge-Morrey-Friedrichs decomposition, these networks can capture complex physical phenomena—like fluid vortices and quantum interference—that cause standard "vanilla" PINNs to fail catastrophically.

What Problem Does This Paper Address?

Physics-Informed Neural Networks (PINNs) have become a staple for solving PDEs, but they have a major blind spot: topology. Standard PINNs typically learn a single-valued scalar function $u_\theta$. Mathematically, the gradient of this function corresponds to an "exact" differential form.

The problem is that on domains with holes (like a pipe with a cylinder inside it), physics often requires "non-exact" components. For example, if you have a fluid vortex or a magnetic field wrapping around a wire, the circulation around that hole is non-zero. However, the circulation of any "exact" form around a hole is always zero.

As a result, vanilla PINNs often encounter "topological pathologies." They struggle with:

  • Branch cuts: Discontinuities that appear when trying to map a circular coordinate (like an angle) to a flat scalar value.
  • Path dependence: Integration results that change depending on the route taken around a hole.
  • Failure to converge: The network simply cannot find a valid global scalar potential that satisfies the boundary conditions and the physics simultaneously.
Standard PINNs are mathematically forced to have zero circulation, leading to failure in "holey" domains.

Methods

The authors propose a "decomposed ansatz" based on the Hodge-Morrey-Friedrichs (HMF) decomposition. Instead of solving for a scalar potential directly, they parameterize a differential 1-form $\alpha_\theta$ (which you can think of as a vector field) using three distinct components:

  1. The Exact Part ($d\phi_\theta$): This represents the standard, irrotational gradient field that vanilla PINNs use.
  2. The Coexact Part ($\delta\beta_\theta$): This captures rotational or "solenoidal" flows (vortices and eddies).
  3. The Harmonic Part ($\sum c_j h_j$): This is the "secret sauce." It uses pre-defined "harmonic representatives"—mathematical functions that naturally "wrap" around the holes in the domain. The network learns the coefficients $c_j$, which represent the physical strength of the topological obstruction (e.g., the total circulation).

To handle potential numerical "blow-ups" near singular points (like the center of a puncture), the authors also introduce exhaustion functions. These act as weights in the loss function, effectively shielding the network from training on regions where the physics might diverge.

Datasets

The researchers evaluated their method on several challenging topological benchmarks:

  • Unit Radial Vortex: A classic fluid dynamics problem in an annular (ring-shaped) domain.
  • Arctangent Jump: A test case involving a jump discontinuity across a branch cut.
  • Aharonov-Bohm Effect: A famous quantum mechanics problem where a magnetic field affects particles in a region they never actually enter—a purely topological effect.
  • Kähler Manifolds: Complex-valued domains used to test the method's extensibility to higher-order geometries.

Key Insights and Ideas

  • Topology as an Architectural Constraint: You can't fix topological failures just by adding more layers or training longer. The failure is baked into the calculus of standard PINNs. You must change the architecture to allow for non-zero circulation.
  • Learning the "Flow" instead of the "Potential": By solving for the 1-form (the flow field) and then integrating to find the scalar potential, the model bypasses the need to learn discontinuous functions directly.
  • Harmonic Representatives: Using known topological "basis functions" ($h_j$) allows the network to "know" about the holes in the domain without needing to see "inside" them.

Key Results

The Hodge-de Rham PINN showed dramatic improvements over baseline models:

  • Annulus Success: In the radial vortex test, the proposed method achieved high accuracy, while the vanilla PINN failed completely because it was mathematically incapable of representing non-zero circulation (see Figure 1).
  • Aharonov-Bohm Effect: The model successfully captured the complex interference patterns of the Aharonov-Bohm effect with one and two poles, a result the authors note is largely non-existent in previous PINN literature.
  • Theoretical Proof: The paper includes a theorem proving that the error of a standard PINN is bounded from below by the magnitude of the topological obstruction. In other words, if the hole matters, a standard PINN is guaranteed to fail.
Figure 1: We plot success of a (1) Hodge-de Rham PINN and failure of a (2) vanilla PINN on a domain {(x, y)|1 <

Limitations and Future Work

  • High Dimensions: The method currently faces a "combinatorial bottleneck" in 4D or higher dimensions because the number of components in the coexact term grows rapidly.
  • Prior Knowledge: The user must provide the harmonic representatives ($h_j$) for the domain. While these are well-known for simple punctures, they might be difficult to derive for highly irregular or "messy" real-world geometries.
  • Integration Complexity: Extracting the final scalar solution from the learned 1-form often requires numerical integration, which can introduce its own set of challenges if the model isn't perfectly trained.

Summary updated

Abstract

We attempt to alleviate failure modes of classical physics-informed neural networks (PINNs) in scenarios of topological pathologies corresponding to punctures, holes, poles, and voids by introducing de Rham cohomological augmentations, or Hodge-de Rham PINNs. Vanilla PINNs often encounter pitfalls in hole-based settings due to discontinuous interfaces, nonzero closed path integration, path dependence, correspondence to exact forms via the exterior derivative, etc., therefore we absolve traditional failures via parameterizing a 1-form by the Hodge-Morrey-Friedrichs (HMF) decomposition for manifolds with boundary, drawing a connection to de Rham cohomology such as through the HMF isomorphism theorem. 1-forms can act as representations of scalar-valued PDE solutions through integral operator mappings. We propose performing a PINN-based loss on the parameterized 1-form through an associated PDE, which subsequently solves the potential-based PINN problem and moreover remedies the failures of vanilla PINNs via the Sobolev-regular operator under a Fréchet differentiation interconnection. The operator relationship often has a nontrivial closed-form and often requires integration or a case-by-case analysis, which we address thoroughly in our experiments. When an integrability condition is not met on our losses, we can equip exhaustion functions to train the networks. We demonstrate our methods on the Aharonov-Bohm effect, which, to our understanding, is largely nonexistent in PINN literature. We develop theoretical guarantees of our methods and demonstrate our pathological safeguards on a broad selection of topologically-adverse problems, including scenarios of inexact forms, and those of nontrivial harmonics and coexact terms.
BibTeX
@misc{gracyk2026de,
  author = {Andrew Gracyk},
  title = {{De Rham cohomological architectures for PINNs and PDEs with topological holes}},
  year = {2026},
  eprint = {2611.05276},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05276},
}
Figure