TopoCut: Methods, Models, and Applications

• The paper on structural topology optimization presents a density-based method combining SIMP material distribution with CutFEM, enabling seamless integration of designed and nondesign regions.
• In robotics, TopoCut provides a benchmark and learning framework utilizing spectral rewards and diffusion policies for multi-step cutting of deformable objects.
• In combinatorial topology, the TopoCut framework studies cut complexes that generalize graph duality, offering new insights into shellability and homological properties.

Searching arXiv for papers using the term “TopoCut” and closely related usages. TopoCut is an overloaded term that appears in several technically distinct research programs on arXiv. In computational structural design, it denotes a density-based topology optimization method for linear elasticity that combines SIMP-style material distribution with the Cut Finite Element Method and weak coupling to independently discretized nondesign regions (Burman et al., 2018). In robotic manipulation, it denotes a benchmark and learning framework for goal-conditioned, multi-step cutting of deformable objects, centered on explicit topology discovery, pose-invariant spectral rewards, and discrete diffusion policies (Wang et al., 24 Sep 2025). In topological combinatorics, the closely related cut-complex framework studies simplicial complexes whose facets are complements of disconnected vertex subsets of graphs, thereby generalizing the Alexander-dual graph-complex viewpoint (Bayer et al., 2023). A broader “TopoCut-like” usage also appears in interactive or local-editing formulations of topology optimization, where masked or partial generation enables topology editing rather than full re-solution (Feng et al., 19 Mar 2026). The shared motif is the use of a “cut” operation—geometric, variational, combinatorial, or physical—to expose and control topological structure.

1. TopoCut in structural topology optimization

In structural mechanics, TopoCut is a density-based topology optimization method for linear elasticity that combines two ingredients: standard material distribution optimization on a design region and CutFEM discretization on a fixed background mesh (Burman et al., 2018). The method decomposes the full structure as

$$\Omega = \Omega_0 \cup \Omega_1 \cup \cdots \cup \Omega_n,$$

where $\Omega_0$ is the design domain and $\Omega_i$, $i \ge 1$, are nondesign domain regions whose geometry is already prescribed (Burman et al., 2018). The density field $\chi$ is optimized in $\Omega_0$ while $\chi=1$ is maintained in the nondesign regions (Burman et al., 2018).

The central purpose of this formulation is to allow topology optimization on a domain with an arbitrary boundary without requiring a body-fitted mesh, while simultaneously coupling the optimized region to fixed substructures that may be discretized independently, for example with parametric finite elements or isogeometric analysis (Burman et al., 2018). This places TopoCut within the class of geometry-independent finite-element formulations, but with the added emphasis on interface coupling between optimized and predesigned parts.

The governing model is small-strain linear elasticity with

$$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$

together with Dirichlet, Neumann, and interface continuity conditions (Burman et al., 2018). The optimization objective treated in the paper is compliance minimization under a volume constraint on the design region (Burman et al., 2018). The material coefficients and loads are scaled by density,

$$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$

so material removal simultaneously weakens stiffness and applied loading in low-density zones (Burman et al., 2018).

A distinctive feature is the use of cut finite elements on $\Omega_0$. The boundary of the design domain is permitted to cut arbitrarily through a structured background mesh, with the active mesh defined as

$\Omega_0$0

This avoids repeated mesh generation during optimization and supports complex boundaries on structured grids (Burman et al., 2018). A plausible implication is that the method is particularly attractive in workflows where the optimized region must be embedded into a larger, already engineered assembly.

2. Discretization, coupling, and stability mechanisms

The discrete elasticity problem in TopoCut is written as

$\Omega_0$1

with

$\Omega_0$2

where $\Omega_0$3 is the bulk elasticity term, $\Omega_0$4 is the penalty term, $\Omega_0$5 contains consistency and symmetry terms, and $\Omega_0$6 is a stabilization term (Burman et al., 2018). Interface coupling and Dirichlet enforcement are handled weakly by a weighted Nitsche formulation (Burman et al., 2018).

The interface averages are density-weighted. On $\Omega_0$7,

$\Omega_0$8

The paper emphasizes that this weighting is robust across material contrasts and yields the correct limit behavior when density vanishes (Burman et al., 2018). As $\Omega_0$9, the Nitsche penalty and consistency terms weaken locally, and the formulation approaches traction-free behavior on the corresponding boundary or interface segment (Burman et al., 2018). This is not a heuristic interpretation but an explicit limiting property of the method.

Because cut cells may become arbitrarily small, TopoCut augments the formulation with ghost-penalty stabilization,

$\Omega_i$0

where the sum runs over derivative orders and $\Omega_i$1 denotes faces of elements cut by the boundary (Burman et al., 2018). These terms stabilize the method near cut boundaries, provide the inverse inequalities needed for coercivity, and preserve consistency because they vanish for sufficiently smooth exact solutions (Burman et al., 2018). The paper proves continuity and coercivity of the discrete bilinear form in a suitable energy norm, with

$\Omega_i$2

for sufficiently large penalty parameter $\Omega_i$3 (Burman et al., 2018).

The density field is represented on a refined mesh relative to the physical state discretization. On $\Omega_i$4, the auxiliary variable $\Omega_i$5 is piecewise constant on a refined mesh and the actual density is approximated as

$\Omega_i$6

with $\Omega_i$7 (Burman et al., 2018). This multiresolution design improves geometric detail relative to the state space.

3. Optimization algorithm and practical role in design workflows

The optimization loop in TopoCut follows a steepest-descent or optimal-criteria style procedure: solve elasticity for the current density, evaluate sensitivities, filter the sensitivities, update the auxiliary density, and enforce the volume constraint using a bisection search for the Lagrange multiplier (Burman et al., 2018). The update rule is given explicitly in the paper, with move limit $\Omega_i$8 and multiplicative factor $\Omega_i$9 derived from filtered sensitivities and the volume derivative (Burman et al., 2018). The target volume fraction is ramped gradually during early iterations to improve convergence (Burman et al., 2018).

Sensitivity filtering is adapted to cut cells, nondesign regions, and interfaces, and includes ghost derivatives in nondesign regions so that thin structural layers near interfaces are not incorrectly removed (Burman et al., 2018). The paper notes that the specific filter is not itself the conceptual core of the method, and that alternatives would likely be compatible.

The practical advantages claimed for TopoCut are tightly linked to its discretization strategy: no body-fitted design mesh is needed for the design domain, complex design geometries can be treated on fixed grids, design and nondesign regions may be discretized independently, coupling to parametric or isogeometric components is natural, and stability is maintained even for small cut cells and vanishing densities (Burman et al., 2018). The numerical demonstrations are 2D cantilever-type examples with parametric reinforcements and truss-like nondesign structures, where the optimized design adapts to supplied nondesign geometry and stresses remain continuous across interfaces (Burman et al., 2018).

A common misconception is that the “cut” in TopoCut refers primarily to combinatorial cuts or graph cuts. In this structural setting, it instead refers to the cut finite element treatment of geometry relative to the background mesh (Burman et al., 2018). The coupling and stabilization machinery, rather than a graph-theoretic separation procedure, is the defining mechanism.

4. TopoCut in robotic manipulation of deformable objects

A second, unrelated use of the name appears in robotics, where TopoCut is a benchmark and learning framework for multi-step robotic cutting of deformable objects (Wang et al., 24 Sep 2025). The motivating problem is that multi-step cutting involves topology changes, dense deformable states, and evaluation difficulties: the number of connected components varies over time, the relevant state is particle-wise rather than low-dimensional, and standard geometric metrics such as Chamfer Distance, Earth Mover’s Distance, or Hausdorff Distance are pose-sensitive and can mis-score correct cuts that have undergone rigid motion (Wang et al., 24 Sep 2025).

TopoCut addresses these issues through three components. First, it provides a high-fidelity simulation environment built on FluidLab with an MLS-MPM solver implemented in Taichi, together with a compliant von Mises constitutive model, damage tracking, a 6-DoF knife agent, surface adherence, and progressive softening (Wang et al., 24 Sep 2025). Second, it defines a reward based on Laplace–Beltrami spectral analysis, intended to be pose-invariant (Wang et al., 24 Sep 2025). Third, it introduces an integrated learning pipeline composed of a dynamics-informed perception model and a goal-conditioned discrete diffusion policy called PDDP (Wang et al., 24 Sep 2025).

The environment’s key technical novelty is a damage-driven topology discovery mechanism. Each particle’s deformation gradient $i \ge 1$0 is monitored, and a particle is marked as damaged if

$i \ge 1$1

with $i \ge 1$2, $i \ge 1$3, and $i \ge 1$4 material-dependent thresholds and sensitivity parameter (Wang et al., 24 Sep 2025). For von Mises materials, yielding and softening can also trigger damage (Wang et al., 24 Sep 2025). After damage detection, the knife trajectory is carved into an implicit signed distance field, a surface is extracted with Marching Cubes, smoothed, decomposed into connected components, and particles are assigned cluster identifiers based on proximity to reconstructed surfaces (Wang et al., 24 Sep 2025). This explicit cluster structure is what makes topology-aware reward computation and policy learning possible.

The spectral reward is derived from a $i \ge 1$5-nearest-neighbor graph on a point cloud $i \ge 1$6, with Gaussian weights

$i \ge 1$7

The eigenproblem

$i \ge 1$8

yields eigenvalues and eigenvectors that serve as intrinsic shape descriptors (Wang et al., 24 Sep 2025). The spectral distance between shapes $i \ge 1$9 and $\chi$0 is defined as

$\chi$1

and the reward is obtained by inverse scaling,

$\chi$2

For multi-fragment outputs, fragment-level rewards are summed (Wang et al., 24 Sep 2025). The appendix reportedly shows nearly identical reward-versus-step curves across rotated versions of the same ideal shape, supporting the pose-invariance claim (Wang et al., 24 Sep 2025).

5. Topology-aware policy learning in the robotic TopoCut framework

The perception model in robotic TopoCut learns topological evolution in the form

$\chi$3

where $\chi$4 is the current particle cloud with cluster labels and $\chi$5 is an action represented as a binary cut mask on points (Wang et al., 24 Sep 2025). The model downsamples the particle cloud by FPS, constructs a topology graph and an action graph, encodes both with a shared encoder, and fuses them with a Graph Transformer (Wang et al., 24 Sep 2025). Training uses a geometric loss composed of CD, EMD, and HD together with a topology loss based on Hungarian matching over cluster assignments (Wang et al., 24 Sep 2025).

The downstream policy, PDDP, treats a cutting action as a binary segmentation over particles. It is conditioned on the current topology embedding, a goal embedding, action history, and a noised action mask, and follows a Score-Entropy Discrete Diffusion formulation (Wang et al., 24 Sep 2025). The score network satisfies

$\chi$6

and is trained with a denoising score-entropy objective (Wang et al., 24 Sep 2025). After denoising, the binary mask is converted into a knife pose in $\chi$7 by fitting a cutting plane via SVM (Wang et al., 24 Sep 2025).

The reported experiments use about 1,000 trajectories in the main description, while the appendix mentions approximately 6,000 labeled cuts generated from MPPI and teleoperation combined (Wang et al., 24 Sep 2025). Policies are evaluated on 405 episodes with five sequential cuts each across 9 object geometries, 3 goal types, and randomized pose and scale (Wang et al., 24 Sep 2025). The split includes 5 in-distribution and 4 out-of-distribution geometries (Wang et al., 24 Sep 2025). The main metrics are normalized spectral reward $\chi$8 and cut count $\chi$9, the number of correctly segmented goal-consistent pieces (Wang et al., 24 Sep 2025). Example out-of-distribution results reported for PDDP are: slice, $\Omega_0$0, $\Omega_0$1; stick, $\Omega_0$2, $\Omega_0$3; dice, $\Omega_0$4, $\Omega_0$5 (Wang et al., 24 Sep 2025). The appendix further reports 77.1% seen accuracy and 74.1% unseen accuracy for the final graph-based perception model with sorting, Gaussian noise, and 256 downsampled points (Wang et al., 24 Sep 2025).

Here again, “TopoCut” does not refer to CutFEM or structural compliance optimization. It denotes a topology-aware benchmark for physical cutting. The only commonality with the structural usage is the centrality of topology change.

6. Cut complexes, graph topology, and TopoCut-style combinatorics

A third research line uses “cut” in a topological-combinatorial sense. The paper “Topology of Cut Complexes of Graphs” defines, for a graph $\Omega_0$6 and $\Omega_0$7,

$\Omega_0$8

and then defines the $\Omega_0$9-cut complex as

$\chi=1$0

Its facets are exactly the complements of disconnected $\chi=1$1-vertex subsets (Bayer et al., 2023). The case $\chi=1$2 recovers the Alexander-dual graph-complex setting associated with Fröberg and Eagon–Reiner (Bayer et al., 2023).

The complex is pure when nonvoid, with

$\chi=1$3

and the family satisfies

$\chi=1$4

as $\chi=1$5 increases (Bayer et al., 2023). The paper develops shellability, homotopy type, and homology for numerous graph families using algebraic topology, poset topology, discrete Morse theory, and shellability arguments (Bayer et al., 2023).

Among the principal results are: if $\chi=1$6 is chordal, then $\chi=1$7 is shellable, while for every $\chi=1$8 there exists a chordal graph whose $\chi=1$9-cut complex is not shellable (Bayer et al., 2023). For trees, $$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$0 is shellable for all $$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$1 (Bayer et al., 2023). For cycles with $$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$2,

$$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$3

and $$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$4 is a sphere $$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$5 for $$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$6 (Bayer et al., 2023). For prisms $$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$7,

$$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$8

and this family yields minimal nonshellable examples (Bayer et al., 2023).

Subsequent work studies total cut and cut complexes on grid graphs, especially $$\epsilon(u)=\frac{1}{2}(u\otimes\nabla + \nabla\otimes u), \qquad \sigma(u)=2\mu\,\epsilon(u)+\lambda\,\operatorname{tr}(\epsilon(u))I,$$9 and $$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$0 rectangular grids (Chandrakar et al., 2024). For total cut complexes,

$$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$1

and

$$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$2

are established (Chandrakar et al., 2024). For ordinary cut complexes, $$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$3 is shellable for $$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$4 (Chandrakar et al., 2024). In this literature, the phrase “TopoCut” is best understood as a convenient umbrella for cut-complex topology rather than a named algorithm in the papers themselves. This suggests a semantic family resemblance rather than a single unified method.

7. Related “cut” paradigms and local topology editing

Several adjacent papers clarify how the TopoCut label is used analogically across domains. “SurfCut” is a topology-driven method for extracting a free-boundary surface in a 3D image from a noisy likelihood map and a single seed point (Algarni et al., 2017). Fast Marching is used to compute a weighted distance $$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$5 satisfying

$$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$6

ridge curves of the Euclidean path-length field $$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$7 are extracted on evolving fronts using cubical complexes and Morse theory, a graph cut is used to isolate the boundary curve, and the final surface is recovered as a valley of $$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$8 (Algarni et al., 2017). Boundary extraction and surface extraction each have $$\mu=\chi\widehat{\mu},\qquad \lambda=\chi\widehat{\lambda},\qquad f=\chi\widehat{f},\qquad g_N=\chi\widehat{g}_N,$$9 complexity, and on $\Omega_0$0 synthetic data SurfCut is reported to run in about 2227 seconds while LP and MCNF become infeasible (Algarni et al., 2017). Although not named TopoCut, it is explicitly presented as a topology-preserving cut/extraction method in spirit (Algarni et al., 2017).

In generative structural design, “Sketch2Topo” offers an interactive, sketch-driven topology optimization tool built on the TopoDiff diffusion model (Feng et al., 19 Mar 2026). It supports image-to-image generation, inpainting, and mask-based local optimization, thereby moving topology optimization from a “black-box” batch process to a human-in-the-loop editing workflow (Feng et al., 19 Mar 2026). Users can sketch the material domain and physical constraints with color-coded brushes, optionally apply an inpainting mask to restrict optimization to selected regions, and iteratively regenerate the output (Feng et al., 19 Mar 2026). The paper explicitly frames the mask mechanism as enabling precise selection of specific regions for topology optimization and as a localized editing operation aligned with topology “cutting” (Feng et al., 19 Mar 2026). Reported results compare FEA, image-to-image generation, and two masked variants on minimum compliance and volume fraction under a target volume fraction of $\Omega_0$1 (Feng et al., 19 Mar 2026). For example, FEA yields minimum compliance $\Omega_0$2 and volume fraction $\Omega_0$3, while image-to-image generation gives $\Omega_0$4 and $\Omega_0$5 (Feng et al., 19 Mar 2026). The masked versions trade strict compliance optimality for local control (Feng et al., 19 Mar 2026).

TopoCtrl extends the post-optimization editing theme in a different way by repurposing the latent space of a pretrained topology diffusion model for explicit characteristic-guided editing toward targets such as member thickness, characteristic member length, number of joints, and maximum number of members incident to a joint (Chen et al., 27 Mar 2026). It uses partial noising of an encoded topology latent, a characteristic regressor, and regression-guided reverse diffusion (Chen et al., 27 Mar 2026). This is not TopoCut in name, but it reinforces a broad contemporary meaning of “TopoCut” as selective post hoc topology modification rather than full redesign.

Another distinct “cut” formulation appears in discrete-variable topology optimization using generalized Benders decomposition with a multi-cut master problem and adaptive trust regions (Ye et al., 2024). The paper does not use the TopoCut name explicitly, but its master problem accumulates multiple cuts

$\Omega_0$6

and combines them with trust-region constraints to reduce the number of expensive FEM solves by about one order of magnitude in the reported tests (Ye et al., 2024). This again illustrates that “TopoCut” frequently functions as a descriptive shorthand for cut-based topology methods rather than a single canonical formalism.

Taken together, these uses show that TopoCut is not a unitary concept across arXiv. It names a specific CutFEM-based topology optimization method in structural mechanics (Burman et al., 2018) and a specific topology-aware robotic cutting benchmark in deformable manipulation (Wang et al., 24 Sep 2025), while also serving as a natural descriptor for cut-complex topology (Bayer et al., 2023), topological surface extraction (Algarni et al., 2017), localized topology editing (Feng et al., 19 Mar 2026), and multi-cut optimization decompositions (Ye et al., 2024). The unifying abstraction is the use of a cut operation to expose, preserve, or manipulate topological structure under computational constraints.