Coordinate Transformations for Neural Solvers on Irregular Domains

This direction is split into other directions: geometric domain transformation. Its last version stays readable and citable here.

Version 1 of 1126 papers149 conceptsfrozen 2026-10-04

Deep learning has shown immense promise in accelerating the solution of partial differential equations (PDEs). However, most high-performance architectures—like Convolutional Neural Networks (CNNs) and Fourier Neural Operators (FNOs)—are "grid-bound." They rely on the regularity of Cartesian grids to perform efficient convolutions or Fast Fourier Transforms.

Real-world physics rarely happens in a box. Whether it is the airflow over a curved aerofoil, groundwater moving through a jagged basin, or the electrical signals in a twisting human heart, the "physical space" is almost always irregular. To bridge this gap, researchers have converged on a powerful strategy: Coordinate Transformation. By mapping complex physical domains into standardized "computational spaces," we can treat irregular geometries as regular grids without losing the underlying physics.

Mapping to Computational Space

The core intuition is to find a differentiable, invertible mapping (a diffeomorphism) between a complex physical domain $\Omega$ and a simple reference domain $\Xi$, such as a unit square or cube.

In this framework, every point $(x, y)$ in the messy physical world corresponds to a point $(\xi, \eta)$ in a neat, rectangular grid. As Angran Li and colleagues demonstrate in their work on reaction-diffusion systems, this allows a standard CNN to process non-rectangular geometries by simply operating on the "parametric unit square." The network sees a regular grid, but that grid represents a warped version of the real world.

This approach is not merely a pre-processing trick; it is a fundamental shift in how the neural network perceives space. Jian-Xun Wang and the team behind PhyGeoNet use elliptic coordinate transformations to ensure that boundary surfaces align perfectly with the grid axes. This "body-fitted" approach eliminates the aliasing errors and "staircase" artifacts that occur when trying to represent a curve using square pixels (voxelization).

Physics in a Warped World

Mapping the domain is only half the battle. If you stretch a rubber sheet, the distances and angles on that sheet change. To solve a PDE on a transformed grid, the neural network must account for this stretching.

This is achieved by incorporating the Jacobian matrix and transformation metrics into the model. Liwei Chen and Nils Thuerey showed that providing a network with non-dimensional metrics—like the Jacobian and unit normals—allows it to achieve unprecedented accuracy in turbulent flow predictions. These metrics tell the network exactly how much the "computational" derivative needs to be scaled to represent the "physical" derivative.

In Physics-Informed Neural Networks (PINNs), this is handled by reformulating the differential operators. Instead of calculating $\nabla u$ in physical space, the loss function calculates it in the reference space using the chain rule and the transformation metrics. This ensures that even though the computation happens on a square, the physical laws—like the conservation of mass or momentum—are strictly preserved, as seen in the FloodCast framework for modeling winding river systems.

Learning the Warp

While many methods use pre-defined mathematical mappings (like elliptic or algebraic transformations), a new frontier involves learned deformations.

In Geo-FNO, Daniel Zhengyu Huang and Anima Anandkumar propose a model that learns to "warp" the irregular physical domain into a uniform latent space (like a unit torus). This allows the Fourier Neural Operator to use the Fast Fourier Transform (FFT) on geometries where it would otherwise be impossible. By learning the optimal mapping, the model can reduce error rates by 50% compared to standard interpolation methods.

Similarly, the DIMON framework by Minglang Yin and others uses diffeomorphic mappings to transport varying, patient-specific heart geometries onto a single "template" domain. This turns a problem with moving or changing boundaries into a fixed-boundary problem, allowing a single neural operator to generalize across a wide family of shapes.

Handling Discontinuities and Moving Boundaries

Coordinate transformations also offer a solution to "sharp" features like shock waves. Standard networks often struggle with the jagged edges of a shock. Yameng Zhu and colleagues introduced R-adaptive DeepONet, which uses a moving-mesh strategy to transform a discontinuous physical solution into a smooth computational one. By "stretching" the grid to follow the shock, the network can resolve the discontinuity with much higher precision.

This concept of "domain flattening" is also a staple in classical fluid mechanics for free boundary problems. As Yoshihiro Shibata and David Altizio demonstrate, mapping a time-varying fluid surface to a static reference domain allows researchers to apply standard analytical tools to systems that are constantly changing shape, such as oscillating Faraday waves.

Summary

The shift from raw physical coordinates to transformed computational spaces represents a major step toward "geometry-aware" AI. By treating the domain's shape as a coordinate transformation rather than just a boundary condition, these models can:

  1. Leverage efficient architectures (CNNs, FNOs) on irregular shapes.
  2. Maintain physical rigor by incorporating transformation metrics into differential operators.
  3. Generalize across geometries by mapping diverse shapes to a common reference template.

Papers in this version

Showing 60 of 126 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_coordinate_transformations_for_neural_so_50e23570,
  title = {Coordinate Transformations for Neural Solvers on Irregular Domains},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/coordinate-transformations-for-neural-solvers-on-irregular-domains/rdr_50e2357026414ab7b7b8c44ea1b0e83a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_coordinate_transformations_for_neural_so_50e23570_v1,
  title = {Coordinate Transformations for Neural Solvers on Irregular Domains},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/coordinate-transformations-for-neural-solvers-on-irregular-domains/rdr_50e2357026414ab7b7b8c44ea1b0e83a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}