Exact Geometry for High Precision Numerical Analysis

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

Version 1 of 1109 papers123 conceptsfrozen 2026-10-04

In most of computer graphics and engineering, we represent the world as a collection of flat tiles. Whether they are triangles in a mesh or voxels in a grid, these piecewise linear approximations are the "lingua franca" of 3D data. However, as research moves toward high-precision simulation and deep learning, this "approximation tax" is becoming a fundamental bottleneck.

A growing body of research suggests that the errors introduced by discretizing smooth surfaces into flat segments are not just aesthetic—they fundamentally limit the accuracy of physical simulations, the stability of inverse problems, and the reliability of geometric deep learning.

The Approximation Tax

When we represent a curved surface with a triangle mesh, we introduce a geometric discretization error. For simple visualization, this error is negligible. But for numerical analysis, the consequences are severe.

As Yi‐King Choi, Bernard Mourrain, and colleagues point out in Continuous Collision Detection for Composite Quadric Models, linear approximations can lead to false positives or missed contacts in high-fidelity simulations. If the geometry is "wrong" at the micro-scale, the physics calculated on top of it will be wrong as well.

This issue is particularly acute when calculating derivatives, such as surface normals or curvature. Grady B. Wright and Robert M. Kirby demonstrate that simulating chemical reactions on surfaces requires highly accurate normal vectors; traditional piecewise approximations fail to provide the precision needed for these coupled reaction-diffusion systems A Radial Basis Function (RBF)-Finite Difference Method....

The Convergence Ceiling

The most striking insight from this research direction is that geometric errors can "cap" the performance of even the most sophisticated numerical solvers.

In planetary-scale fluid simulations, D. Rhodri Davies and C. R. Wilson show that representing curved boundaries with flat-sided elements restricts velocity convergence to second-order accuracy. No matter how much you refine the solver's internal math, the linear approximation of the boundary becomes a bottleneck that prevents higher-order precision Analytical solutions for mantle flow in cylindrical and spherical shells.

Similarly, Alec Jacobson and Oded Stein highlight that solving partial differential equations (PDEs) on surfaces involves a dual source of error: the functional approximation (the math) and the geometric discretization (the shape). If the mesh distorts the manifold, the numerical accuracy suffers in ways that traditional 2D analysis cannot easily predict A mixed finite element method with piecewise linear elements for the biharmonic equation on surfaces.

Bridging the CAD to CAE Bottleneck

In industrial workflows, there is a painful "translation" process. Designers create smooth parts in Computer-Aided Design (CAD) using mathematical curves, but engineers must convert these into meshes for Computer-Aided Engineering (CAE) simulations.

Lei Zhang 0164, Thomas J. R. Hughes, and Wing Kam Liu argue for a movement called "Engineering Software 2.0," which uses Isogeometric Analysis (IGA). The intuition is simple: use the same mathematical basis (like B-splines or NURBS) for both the design and the simulation. This eliminates the need for meshing and preserves the exact curvature of the original design Multi-Patch Isogeometric Convolution Hierarchical Deep-learning Neural Network.

Fehmi Cirak has applied this philosophy to structural optimization, showing that maintaining an exact mathematical representation is critical for high-precision manufacturing tasks like CNC machining Filletting Quador Lattice Structures.

Beyond the Mesh: Implicit and Algebraic Solutions

If meshes are the problem, what are the alternatives? Researchers are exploring several paths to represent geometry more faithfully:

  1. Global Polynomials: Sachin K. Thekke Veettil, Michael Hecht, and colleagues propose representing shapes as "Global Polynomial Level Sets" (GPLS). By turning shapes into pure algebraic equations, they can calculate high-order geometric properties with machine precision, far outperforming mesh-based methods Global Polynomial Level Sets for Numerical Differential Geometry of Smooth Closed Surfaces.
  2. Neural Code Generation: Instead of generating "fuzzy" 3D pixels, Maximilian Mews and Alan Akbik fine-tune LLMs to write Constructive Solid Geometry (CSG) code. This produces precise, mathematically defined parts rather than approximate meshes Don't Mesh with Me....
  3. Orthogonal Distance Fields: To solve the interpolation errors inherent in standard Signed Distance Fields (SDFs), Yujie Lu and Long Wan introduced Unsigned Orthogonal Distance Fields (UODFs), which allow for high-precision reconstruction of complex, non-watertight structures Unsigned Orthogonal Distance Fields....

Why It Matters for Deep Learning

The shift toward exact geometry is also transforming how we train AI. Marc Alexa and the creators of the ABC-Dataset provided over a million CAD models with exact analytic ground truth for normals and curvature. Their work revealed a sobering reality: when exact connectivity and geometric information are available, many current deep learning architectures struggle to outperform classical algorithms ABC: A Big CAD Model Dataset for Geometric Deep Learning.

By prioritizing geometric fidelity over simple discretization, this research direction is building a foundation for a new generation of engineering and AI tools that are as precise as the smooth, continuous world they aim to model.

Papers in this version

Showing 60 of 109 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_exact_geometry_for_high_precision_numeri_626126d2,
  title = {Exact Geometry for High Precision Numerical Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/exact-geometry-for-high-precision-numerical-analysis/rdr_626126d2694f472c89bbc3cb8f938378},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_exact_geometry_for_high_precision_numeri_626126d2_v1,
  title = {Exact Geometry for High Precision Numerical Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/exact-geometry-for-high-precision-numerical-analysis/rdr_626126d2694f472c89bbc3cb8f938378/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}