Eliminating Numerical Artifacts in Structural Topology Optimization

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 222 papers27 conceptsfrozen 2026-10-04

In the quest to design the lightest and strongest structures—from aircraft wings to micro-resonators—engineers rely on topology optimization. The goal is simple: given a design space and a set of loads, where should we place material to maximize performance? However, when we translate this continuous physical problem into a discrete computer simulation, the optimizer often "cheats." Instead of finding a smooth, manufacturable shape, it discovers numerical loopholes that look like high-frequency noise or checkerboard patterns.

Recent research has shifted from simply filtering these errors to fundamentally rethinking how we represent and solve the optimization problem to ensure that the results are physically sound and independent of the underlying grid.

The Checkerboard Problem and Mesh Dependency

The most persistent ghost in the machine of topology optimization is the checkerboard artifact. As noted by Wei He and Sanjay Govindjee, numerical material layouts frequently produce alternating solid and void pixels. This happens because the finite element method (FEM) tends to overestimate the stiffness of these patterns. To a numerical solver, a checkerboard looks artificially efficient, even though it is physically impossible to manufacture.

Closely related is the issue of mesh dependency. In a well-posed physical problem, refining your grid should give you a more accurate version of the same shape. However, as Igor Ostanin and Denis Zorin observe, traditional methods often produce entirely different topologies as the resolution increases. This lack of consistency suggests that the optimizer is chasing discretization errors rather than a true physical optimum.

Neural Reparameterization as a Natural Filter

One of the most promising shifts involves moving away from "pixel-based" optimization. In traditional setups, every cell in a grid is an independent variable. Stephan Hoyer and colleagues at Google Research introduced neural reparameterization to solve this. By using a convolutional neural network (CNN) as a "generator" for the structure, the design inherits the internal logic and smoothness of the network. Because CNNs are inherently biased toward low-frequency, spatially correlated patterns, they naturally suppress checkerboarding without the need for manual filters.

Similarly, Ajendra Singh and Souvik Chakraborty proposed DPNN-TO, a dual physics-informed neural network. By linking a displacement network with a density network, they minimize energy functionals directly. This approach bypasses the discrete grid entirely during the optimization phase, producing high-resolution designs that remain stable across different scales.

Penalizing Numerical Error Directly

If the optimizer exploits regions where the simulation is inaccurate, why not penalize those inaccuracies? Vladislav Pimanov and George Ovchinnikov developed a regularization technique using a posteriori error estimators.

The intuition is elegant: if a proposed material distribution causes a high degree of numerical error in the FEM solver, the optimizer is "taxed." By incorporating the upper bound of the true physical functional into the objective, the system learns to avoid configurations (like checkerboards) that the underlying math cannot accurately represent. This forces the design toward "high-regularity" shapes that are both robust and mesh-independent.

Shifting the Mathematical Ground

Other researchers are tackling the problem by changing the discretization scheme itself. Igor Ostanin and I. V. Tsybulin advocate for the Boundary Element Method (BEM) combined with the Kernel-Independent Fast Multipole Method (KIFMM). By focusing only on the surface of a structure rather than its entire volume, they eliminate many of the volumetric instabilities that plague FEM-based approaches.

Furthermore, Miguel A. Salazar de Troya and Daniel A. Tortorelli argue that the root of mesh dependency lies in treating design variables as simple vectors. They propose reformulating optimization algorithms within the $L^2$ function space. This ensures that the "weight" of a design variable scales correctly with the size of the mesh element, preventing Adaptive Mesh Refinement (AMR) from skewing the final topology.

Toward Binary and Manufacturable Designs

The ultimate goal of these advancements is to produce "black-and-white" designs that require no post-processing. Prabhat Kumar and Anupam Saxena introduced the normalized field product (nFP) method, which achieves close-to-binary solutions with an embedded minimum length scale. This ensures that the resulting structures are not only numerically stable but also physically manufacturable, avoiding the "semi-solid" gray regions that often plague density-based models like SIMP.

By moving from discrete pixel-pushing to continuous, physics-aware representations, the field is transforming topology optimization from a "black box" prone to numerical artifacts into a robust tool for high-fidelity engineering.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 13 papers
  2. Version 1 · 2026-10-04 · 22 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_checkerboard_patterns_d5c961ad,
  title = {checkerboard patterns},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/checkerboard-patterns/rdr_d5c961ad03104df5bb68b245f13eb6ea},
  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_eliminating_numerical_artifacts_in_struc_d5c961ad_v1,
  title = {Eliminating Numerical Artifacts in Structural Topology Optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/eliminating-numerical-artifacts-in-structural-topology-optimization/rdr_d5c961ad03104df5bb68b245f13eb6ea/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}