Cellular Topology Optimization on Differentiable Voronoi Diagrams

💡 Research Summary

This paper addresses a fundamental bottleneck in the computational design of complex cellular structures: the lack of a differentiable geometric representation that can be directly incorporated into large‑scale topology optimization pipelines. Traditional Voronoi diagrams provide a discrete partition of space into polygonal cells, but their hard “if‑else” definition prevents gradient‑based sensitivity analysis. The authors propose a novel differentiable Voronoi formulation that converts a set of site points into a continuous density field using a softmax‑based partition. By defining a soft assignment function Sₘ(x) = softmax(−d₁(x,x₁), …, −d_Nc(x,x_Nc)) and a density ρ(x)=1−∑ₘSₘ(x)^β, the method yields smooth cell interiors and finite‑thickness boundaries whose sharpness is controlled by the scalar β.

To capture the anisotropic and spatially adaptive nature of many natural cellular patterns, the distance metric dₘ is generalized to a Mahalanobis distance dₘ(x,xₘ)= (x−xₘ)ᵀ Aₘ (x−xₘ) with Aₘ = Dₘ Dₘᵀ, where Dₘ is a learnable symmetric matrix. Each cell therefore carries its own metric tensor, allowing arbitrary aspect ratios, orientations, and curvature of cell faces. This heterogeneous anisotropic Voronoi tessellation can represent structures such as insect wing veins or bone trabeculae, where cell geometry varies dramatically across the domain.

A further innovation is the introduction of a “free‑boundary” capability. By adding a virtual site point with a constant εₛ and extending the index set to include this point, the formulation can generate cells that lie on the outer boundary of the design domain, effectively allowing the cellular lattice to have open or irregular perimeters rather than being forced to fill the entire space.

Computationally, the naïve evaluation of all pairwise distances scales as O(N²). The authors mitigate this by employing a k‑nearest neighbor search, reducing the complexity to O(k log N) per iteration, which makes the approach feasible for thousands of cells. The continuous density field is then embedded into the classic SIMP (Solid Isotropic Material with Penalization) topology optimization framework. Design variables consist of site point coordinates, metric tensor entries, and the free‑boundary constant. Automatic differentiation provides exact gradients of compliance, volume, and stress constraints, enabling the use of standard gradient‑based optimizers such as MMA or ADAM.

The methodology is validated on two biologically inspired case studies: a femur bone interior and an Odonata (dragonfly) wing. In both cases, the authors optimize structures comprising 1,000–5,000 anisotropic cells. Compared with conventional grid‑based topology optimization, the resulting designs achieve 15–30 % higher stiffness‑to‑mass ratios and exhibit realistic spatial variation in cell size, shape, and orientation. The experiments also demonstrate that increasing β sharpens cell walls, while the learned metric tensors naturally adapt to load paths, confirming the effectiveness of the anisotropic formulation.

Key contributions of the work are: (1) a mathematically rigorous differentiable Voronoi representation based on softmax; (2) extension to heterogeneous anisotropic cells via Mahalanobis metrics; (3) support for free boundaries through a virtual site point; (4) a hybrid grid‑particle discretization that enables efficient large‑scale optimization; and (5) empirical validation on realistic biological structures. Limitations include sensitivity to the choice of β and initial metric tensors, current focus on linear elasticity (excluding multi‑physics effects), and the need for post‑processing to obtain manufacturable geometries. Future directions suggested are multi‑physics integration, real‑time design interfaces, and meta‑learning strategies for automatic hyper‑parameter tuning.