Computational models excel at solving physics equations on simple, uniform geometries like rectangles, circles, and spheres. However, the real world is messy, filled with sharp corners, curved boundaries, and complex multi-material interfaces. When we try to force these irregular shapes onto standard computational grids, we run into a fundamental problem: voxelization or naive discretization introduces aliasing errors, breaks boundary integrity, and destroys the efficiency of optimized solvers.
To bypass this limitation, a powerful strategy has emerged across classical mechanics, numerical analysis, and deep learning: warping the space itself. Instead of modifying the solver to fit a complex geometry, researchers use coordinate transformations to map irregular physical domains into standardized, canonical reference spaces (such as a unit square, a unit disk, or a uniform rectangular grid).
By shifting the geometric complexity from the shape of the boundary into the coefficients of the differential equations, this approach allows highly optimized, grid-based architectures to operate with perfect boundary fidelity.
The Core Intuition: Shifting Complexity from Boundary to Operator
To understand why this works, consider the challenge of enforcing a boundary condition on a curved surface using a standard Cartesian grid. Because the grid lines do not align with the surface, you must approximate the boundary using a jagged "staircase" pattern or employ complex interpolation schemes. This introduces significant errors right where the physics is often most critical—at the interface.
A geometric domain transformation solves this by finding a bijective mathematical mapping between the physical space and a simplified computational space.
$$\mathbf{x} = (x, y) \in \text{Physical Domain} \quad \longleftrightarrow \quad \boldsymbol{\xi} = (\xi, \eta) \in \text{Canonical Domain}$$
In this canonical space, the irregular boundary becomes a perfectly straight line or a smooth circle. The price we pay for this simplification is that our spatial derivatives must be modified using the Jacobian matrix of the transformation (via the chain rule). However, solving a slightly more complex partial differential equation (PDE) on a perfectly uniform grid is computationally superior to solving a simple PDE on a messy, unstructured mesh. It preserves boundary integrity, eliminates aliasing, and unlocks the use of tensor products and uniform grid operations.
Classical Foundations: Analytical Simplification via Conformal Mapping
The mathematical bedrock of this idea lies in complex analysis and conformal mapping, which preserves angles locally. In early structural engineering, Hideo Saito demonstrated how conformal transformations could render intractable elasticity problems solvable. In On the Stress Distribution in a Regular Polygon Plate under Forces on their Sides, Hideo Saito utilized the Schwarz-Christoffel transformation to map the interior of regular polygons onto a unit circle. This transformed difficult boundary-value problems featuring sharp corners into manageable analytical calculations. Similarly, in On Bending and Torsion of Perforated Thin Plates, Hideo Saito mapped complex, rounded hypotrochoidal cutouts to a unit circle to derive closed-form expressions for internal stress concentrations.
This ability to smooth out geometric discontinuities extends naturally to thermal and inverse problems. For instance, Nan Yang et al. in Influence of Circular through Hole in Pt–Rh Bushing on Temperature Propagation at High Temperature mapped complex through-hole geometries to a unit circle to analytically solve non-Fourier heat conduction equations. In inverse problems, where boundaries are unknown or damaged, A. L. Bukhgeĭm et al. demonstrated in Conditional Stability in an Inverse Problem of Determining a Non-smooth Boundary that conformal mappings could regularize harmonic functions, enabling the stable determination of non-smooth Lipschitz boundaries from electrostatic measurements.
Numerical Discretization and Mesh Standardization
In numerical methods like the Finite Element Method (FEM), Boundary Element Method (BEM), and spectral methods, coordinate transformations are essential for standardizing integration and isolating geometric singularities.
Body-Fitted and Hybrid Grids
To solve large-scale continuum mechanics equations, researchers map physical domains to body-fitted coordinate systems. For example, Mathias Funk et al. in Mechanisms of fast flow in Jakobshavns Isbræ, West Greenland: Part II. Modeling of englacial temperatures mapped highly irregular glacier geometries onto a rectangular computational grid. This mapping simplified the application of boundary conditions at the moving surface and irregular bed while maintaining the mathematical integrity of the underlying heat-transfer equations.
For element-based solvers, local transformations allow complex shapes to be evaluated using standard 1D quadrature rules. Tim Warburton et al. developed a framework in A discontinuous Galerkin spectral/hp method on hybrid grids that recursively maps hybrid element shapes into simpler reference domains. To achieve high-speed interpolation on simplicial shapes (like triangles and tetrahedra), Edward Laughton et al. showed in Fast Barycentric-Based Evaluation Over Spectral/hp Elements that collapsing specific coordinates maps these simplices into standard rectangular domains, unlocking fast, structured tensor-product evaluation algorithms.
Handling Singularities and Interfaces
When boundaries exhibit extreme proximity or multi-material jumps, local coordinate transformations isolate the geometric complexity:
- Near-Field Singularities: In acoustics, Shen Qu et al. in Adaptive integration technique for nearly singular integrals in near-field acoustics boundary element analysis applied Jacobian transformations to map subdivided boundary patches into a normalized coordinate space, ensuring standard Gaussian quadrature remains accurate near boundaries.
- Material Interfaces: For problems with internal discontinuities, Bao Wang et al. in Second order method for solving 3D elasticity equations with complex interfaces applied local transformations at the intersections of grid lines and material interfaces. Aligning the local frame with the interface normal simplifies the enforcement of complex 3D jump conditions into a manageable set of 1D equations.
- Global Flow Simplification: Transformations can also simplify the physics of the domain itself. Fabien Casenave et al. demonstrated in Coupled BEM-FEM for the convected Helmholtz equation with non-uniform flow in a bounded domain that a global Prandtl–Glauert transformation can map an entire simulation domain to simplify exterior uniform flow into a standard wave equation, allowing standard boundary element kernels to be used in the far-field.
Unlocking Grid-Based Deep Learning for Irregular Shapes
Modern deep learning architectures, such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), are highly optimized for uniform, rectangular tensor grids. Standard convolutional kernels cannot natively process irregular boundaries without voxelization, which wastes computation and introduces geometric errors. Mapping coordinates provides a bridge, allowing standard neural architectures to generalize to arbitrary shapes.
Geometry-Adaptive Neural Operators
Instead of designing complex graph-based networks, researchers use coordinate transformations to standardize the inputs to grid-based networks. Angran Li et al. demonstrated in Reaction diffusion system prediction based on convolutional neural network that mapping curved physical domains into parametric unit squares allows standard grid-based CNNs to process non-rectangular geometries without losing essential spatial relationships.
To eliminate discretization errors entirely in deep learning solvers, Jian-Xun Wang et al. introduced PhyGeoNet in PhyGeoNet: Physics-informed geometry-adaptive convolutional neural networks for solving parameterized steady-state PDEs on irregular domain. By using elliptic coordinate transformations, the authors mapped irregular physical domains to rectangular reference grids. This allowed standard convolutional filters to compute exact spatial derivatives via the chain rule, ensuring strict adherence to physical laws along complex boundaries.
Physical Domain (Irregular) Canonical Domain (Uniform Grid)
.-------. .-----------.
/ \ Coordinate Mapping | | Standard CNN / ViT
| Fixed | ====================> | Uniform | ===================> Accurate Predictions
\ Boundary/ (Jacobian/Chain) | Grid | (No Aliasing)
'-------' '-----------'
This strategy is highly effective across fluid dynamics and medical imaging:
- Aerodynamics: In Prediction of transonic flow over supercritical airfoils using geometric-encoding and deep-learning strategies, Zhiwen Deng et al. transformed physical coordinates into uniform curvilinear coordinates, standardizing the input for a Vision Transformer to accurately predict transonic shock waves.
- Sensor Alignment: Thien Nguyen et al. showed in A212 SENSITIVITY OF PIV/INTERFACE GRADIOMETRY TO ESTIMATED WALL POSITION that conformal mapping can translate curvilinear boundary geometries into a standardized rectangular coordinate system, enabling uniform grid processing near wavy interfaces.
- Tomography: For medical diagnostics, Juan Pablo Agnelli et al. in Stroke classification using virtual hybrid edge detection from in silico electrical impedance tomography data used conformal mapping to translate sensor data from irregular head geometries into a standardized unit disk, enabling the application of highly efficient analytical solutions.
- Non-Uniform Point Clouds: To accelerate mesh-free operators, Songming Liu et al. developed the Non-Uniform Neural Operator (NUNO) in NUNO: A General Framework for Learning Parametric PDEs with Non-Uniform Data. NUNO uses K-D tree domain decomposition to map irregular point clouds into localized uniform grids, allowing fast mesh-based operators (like the Fast Fourier Transform) to run on irregular geometries.
Geometry-Aware Sampling in PINNs
In Physics-Informed Neural Networks (PINNs), coordinate transformations improve the learning of boundary conditions. Daehee Cho et al. developed CoCo-PINNs in Conformal mapping Coordinates Physics-Informed Neural Networks (CoCo-PINNs): learning neural networks for designing neutral inclusions. By embedding conformal mapping directly into the network architecture, the authors transformed complex exterior geometries into a canonical circular domain. This provides a natural coordinate system that automatically aligns collocation points with the curvature of the boundary, allowing the network to capture sharp physical jump conditions that geometry-blind Cartesian sampling misses.
Regularizing Free Boundaries and Generative Constraints
Beyond static domains, coordinate transformations can map evolving or highly constrained spaces into fixed, simple geometries.
Free Boundary Regularization
In free boundary problems where the physical domain changes dynamically, tracking the boundary explicitly is computationally expensive. Andrew M. Stuart demonstrated in Singular Limits in Free Boundary Problems that mapping a shrinking, unknown reaction zone onto a fixed geometric reference (such as a unit ball) regularizes the mathematical formulation. This transformation maps the moving boundary to a static coordinate line, enabling the application of classical linear analysis and bifurcation theory to complex systems like combustion and plasma equilibrium.
Enforcing Hard Constraints in Generative Models
In generative modeling, ensuring that a model generates samples strictly within a complex, bounded geometric domain (like a specific convex set) is a difficult optimization challenge.
Xinpeng Li et al. solved this in Gauge Flow Matching: Efficient Constrained Generative Modeling over General Convex Set and Beyond by introducing a bijective "gauge mapping." This function transforms an arbitrary, complex constrained geometric domain into a standardized unit ball.
Generative Modeling Gauge Mapping
(Inside Simple Unit Ball) ===============================> Strictly Feasible Samples
[ Guaranteed Bounds ] x = g^-1(ξ) (Map Back to Space) [ Inside Complex Bounds ]
By training the generative model entirely within this simplified, spherical space, the model can operate without complex boundary penalties. Once a sample is generated inside the unit ball, it is mapped back to the target physical domain via the inverse gauge transformation. This mathematical framework guarantees 100% sample feasibility while bypassing the computational overhead of traditional projection methods.
Key Foundational Work
On the Stress Distribution in a Regular Polygon Plate under Forces on their Sides (1952) — contributed 2 concept elements to this direction.
On Bending and Torsion of Perforated Thin Plates (1951) — contributed 1 concept elements to this direction.