Checkerboards as Spurious Four-Body Interactions in Topology Optimization

Topology optimization asks a computer where to put material inside a block so that the resulting part is as stiff as possible for a given weight. The standard approach divides the block into a grid of square cells, lets each cell hold a "density" between empty and solid, and uses gradients to move density around. It has a well-known failure. The optimizer fills regions with a checkerboard of solid and empty cells touching only at their corners. The physics simulation wrongly rates these corner-only joints as stiff, so the optimizer exploits them. The usual fix is to blur the design with a spatial filter. Blurring removes checkerboards, but it also imposes a minimum feature size, produces gray (half-solid) boundaries, and needs continuation schedules that are tuned by hand. That is a blunt tool for what is really a localized error in the simulator.

We propose to treat that error the way geneticists treat epistasis, meaning the non-additive interaction between mutations. We will measure the simulator's error as a function of the on/off pattern of the few cells around each grid node. We will expand this function in the Walsh–Hadamard basis, an orthogonal decomposition of any function of binary variables into additive, pairwise, three-way and higher-order interaction terms. We will then subtract the learned spurious terms from the objective. Our hypothesis is concrete and checkable: the numerical "reward" for checkerboards is a sparse, local, four-body interaction concentrated in one or two symmetry-related coefficients per node. If it is, a correction with a handful of parameters, calibrated once, should remove checkerboards without imposing any length scale. That would separate two problems that filters currently conflate: numerical artifacts and well-posedness.

Core Claim

Existing fixes treat checkerboarding either as noise to smooth or as a geometric shape to forbid. Spatial filters and Heaviside projection smooth the design. The normalized field product method of Normalized field product approach builds a length scale into fixed neighborhood products and gets grayness below 1%. Coordinate neural networks act as implicit low-pass filters, as the infinite-width analysis in DNN-based Topology Optimisation: Spatial Invariance and Neural Tangent Kernel shows. The resonator work in Topology optimization in micromechanical resonator design hand-designs an "H-constraint" against one-node hinges. Closest in spirit is Robust regularization of topology optimization problems with a posteriori error estimators, which penalizes a computable upper bound on the finite-element error. That work motivates ours, because it shows that penalizing the error alone removes checkerboards. It also constrains our design. Its bound is generic rather than calibrated, so it needs weights above $10^8$ that swamp the true objective. It also uses greedy hard-kill element removal rather than continuous densities, and only heat conduction is shown.

Our claim is different. The discretization error is itself a structured pseudo-Boolean function, and its Walsh spectrum can be estimated and cancelled with a correction whose natural weight is one. Write $\rho_{n,i}\in[0,1]$, $i=1,\dots,4$, for the densities of the four elements around interior node $n$, and let $s_{n,i}=2\rho_{n,i}-1$. The corrected compliance is

$$\tilde C(\rho) = C_h(\rho) - \sum_{n} \bar\varepsilon_n(\rho)\sum_{S\subseteq\{1,2,3,4\}} \beta_S \prod_{i\in S} s_{n,i},$$

where $C_h$ is the coarse finite-element compliance (a measure of how much the structure deflects), $\bar\varepsilon_n$ is the strain energy stored near node $n$, and $\beta_S$ are interaction coefficients tied together by the square's eight symmetries. On binary designs the inner sum is exactly a Walsh expansion. On gray designs it is the multilinear extension of that expansion, so it is differentiable and fits directly inside an ordinary SIMP loop. SIMP, short for Solid Isotropic Material with Penalization, sets stiffness to density to a power $p\ge 3$ so that gray cells are uneconomical. The prediction is that the four-body coefficient $\beta_{\{1,2,3,4\}}$, together with the diagonal pair terms, carries most of the fitted error energy. That is the "fourth-order" structure that Beyond Tokens found hardest for learned models to represent in proteins, here arising from the simulator rather than from biology.

Research Program

The method side comes from epistasis modeling. Epistatic Net showed that L1 penalties on Walsh coefficients recover sparse higher-order interactions from limited data. The Walsh-surrogate work in Surrogate-assisted multi-objective combinatorial optimization fits Lasso surrogates up to a chosen interaction order. That paper also supplies a warning we take seriously: high-order surrogates degraded optimization because of approximation error. Both operate on abstract strings. We apply the same estimator to a spatial, translation-invariant node patch, so the number of unknowns is tiny: sixteen terms in 2D, fewer after symmetry.

The first step is calibration. On small square domains in Firedrake, we will sample random and adversarially structured binary layouts under several load cases. For each, we compute compliance on the design mesh with bilinear elements and on an eight-times-refined reference mesh. We then fit $\beta_S$ by Lasso, regressing the energy discrepancy on the energy-weighted node-patch Walsh features. Using strain-energy weighting is our modeling choice. Spurious hinge stiffness only matters where the hinge carries load, and the fit will tell us whether that is enough or whether the correction depends on the direction of stress.

The second step is spectral analysis. We will report what fraction of the error variance each interaction order explains. We will also test whether enlarging the patch to a 3×3 neighborhood of elements adds anything, which checks locality. Finally, we will repeat the fit for heat conduction and for 3D trilinear elements, where each node patch has 256 Walsh terms before symmetry reduction.

The third step is optimization. We minimize $\tilde C$ with SIMP and the Method of Moving Asymptotes, a standard gradient optimizer for constrained design problems, using automatic adjoints. We run either no filter or a filter radius below one element, so that any absence of checkerboards cannot be credited to blurring.

Mesh dependence is a separate question. Correcting the simulator error does not stop the optimizer from adding ever-finer microstructure as the mesh is refined, because the continuous problem still lacks a length scale. We will measure mesh dependence rather than claim it is solved. The intended product is a design pipeline in which feature size is a deliberate manufacturing choice, not an accidental side effect of fixing a numerical bug.

Baselines

The naive reference point is unfiltered SIMP with penalization exponent 3, optimized with the Method of Moving Asymptotes on the same meshes. It should produce checkerboards, and it lets us confirm that our evaluation metrics detect them.

The standard engineering baseline is SIMP with a density filter and Heaviside projection under β-continuation. We will tune filter radius and continuation schedule with the same number of full optimization runs as our method gets for its calibration. It counts as beaten if we match its fine-mesh compliance while allowing thinner members and lower grayness, or if we achieve better compliance at equal minimum feature size.

The strongest existing parameter-free method is the normalized field product approach of Normalized field product approach, run on the cantilever, MBB beam (a standard simply supported beam benchmark) and displacement inverter. It reports grayness below 1%. No compliance values for these benchmarks are reported on Lacuna, so we will reproduce it at matched mesh, volume fraction and iteration budget. Beating it means equal or lower grayness and lower compliance evaluated on the reference mesh. We will also show that our method produces checkerboard-free designs at length scales the field product approach cannot represent.

For error-based regularization we reimplement the a posteriori estimator penalty of Robust regularization on its own heat conduction Problems A and B. These problems use 4,700 to 5,600 unstructured triangles at volume fractions 0.4 and 0.5. The paper reports that results stabilize for weights above $10^8$ and that Problem B reaches functional values half those of earlier hard-kill studies. We run their hard-kill scheme and also a SIMP variant of their penalty, and count it as beaten if we reach a lower reference-mesh functional without a weight sweep.

We also include the H-constraint of Topology optimization in micromechanical resonator design as the hand-crafted hinge penalty. Finally, we include analysis-mesh refinement, which runs the simulation on a mesh twice as fine as the design grid. It is given the same wall-clock budget.

The key ablation truncates the correction to interaction orders of two or lower, removing the four-body term. A second ablation replaces the fitted correction with a perimeter penalty of matched average magnitude. If either matches the full correction, then the higher-order structure is not doing the work, and we will say so.

Why not just…?

Why not just use a density filter with Heaviside projection? It works for removing checkerboards, and we expect it to remain competitive on compliance at moderate length scales. Its cost is that it ties artifact removal to a forced minimum feature size and a continuation schedule. Normalized field product approach documents that aggressive continuation traps the optimizer in poor gray minima. The decisive experiment is a sweep of filter radius from two elements down to zero. If our correction stays checkerboard-free and gains compliance as the radius shrinks, it adds something filters do not. If filtered designs dominate at every radius manufacturing would accept, the contribution narrows to a diagnostic one.

Why not just penalize the discretization error with an a posteriori estimator? This is the most direct competitor, because it also targets error rather than shape. Its known weakness is calibration. A generic upper bound needs weights so large that it competes with the physics (Robust regularization). Our correction estimates the error's actual sign and structure, so its weight should be near one. Running both on that paper's own heat problems settles which approach distorts the objective less. We will measure distortion as the gap between the coarse objective and the reference-mesh objective at convergence.

Why not just refine the analysis mesh or use higher-order elements? This is a real threat, because it attacks the root cause directly. It costs roughly four times the degrees of freedom in 2D and eight times in 3D, and the simulation solve is already the bottleneck, as DNN-based Topology Optimisation and NTopo both note. Our correction adds only local terms per node. The comparison is therefore made at equal wall-clock time. If refinement wins at equal time in 3D, the correction has no practical role.

Why not just let a coordinate neural network do the filtering? Spectral bias does suppress checkerboards, but it does so by being a low-pass filter. That reintroduces an implicit length scale and requires engineered coordinate embeddings (DNN-based Topology Optimisation). NTopo needed about 37 hours of training and an auxiliary loss to escape local minima (NTopo). We treat this as a secondary comparison on the 2D benchmarks, not as the main competitor.

Proposed Experiments

All designs are scored on the eight-times-refined reference mesh, never on the mesh they were optimized on. That is where checkerboards are exposed as weak. We report four metrics. The first is reference compliance. The second is the coarse-to-reference compliance gap. The third is a checkerboard index, defined as the energy of each design in the node-patch four-body Walsh coefficient. The fourth is the grayness measure, and we also report wall-clock time.

There are three decisive checks. The first is spectral concentration: if orders three and four do not explain most of the fitted error, or if the fit fails to transfer from the calibration domains to the MBB beam and cantilever, the core hypothesis is wrong. The second is the filter-radius sweep in 2D elasticity against filtered SIMP and the normalized field product approach. The third is the 3D trilinear transfer at equal time against analysis-mesh refinement, which determines whether the idea matters beyond 2D.