Mix-Topology Design Strategy

• Mix-Topology Design Strategy is a design paradigm that blends heterogeneous topological primitives—such as hierarchical networks, porous microstructures, and multiscale structural assemblies—into one optimization framework.
• It employs integrated methodologies like latent embedding, explicit–implicit parameterizations, and persistent homology to co-optimize design variables and ensure feasibility under performance constraints.
• Applications demonstrate that this approach enhances design flexibility and performance, improving metrics like computational latency, compliance values, and material allocation compared to conventional methods.

Searching arXiv for the cited papers relevant to “Mix-Topology Design Strategy.” Searching arXiv for the cited papers relevant to “Mix-Topology Design Strategy.” Mix-Topology Design Strategy denotes a family of design approaches that combine multiple topological organizations, representations, or structural classes within a single optimization or synthesis workflow. Across the literature, the term appears in materially different technical settings—including mobile edge computing, porous microstructure blending, multiscale topology optimization, structural assemblies, topological circuits, and anonymous communication networks—but a common pattern recurs: a mixed topology is neither purely one topology nor a simple juxtaposition of separate designs. Instead, it integrates heterogeneous topological primitives, such as hierarchical and routed graph structures, explicit and implicit geometric parameterizations, multiple microstructure classes, or discrete material assignments, while coupling that integration to a performance objective and feasibility constraints (Ma et al., 2024, Gao et al., 2024, Pollini et al., 2019, Chan et al., 2021, Piotrowska, 2021).

1. Conceptual scope and defining characteristics

In mobile edge computing, the term refers to a topology that is simultaneously hierarchical, clustered, and routed. The decentralized network topology design for task offloading (DNTD-TO) constructs a three-layer network with a master node, selected cluster-heads, and cluster-members. Its topology is hierarchical because of the layered master→CH→CM organization, clustered within each CH’s local neighborhood, and tree/routed because offloading follows two-hop paths with link-rate-aware formation. The resulting topology is explicitly characterized as neither purely star nor purely tree nor purely mesh, but as a mixed structure that combines hierarchical control, local clustering, and routing-aware link selection and pruning, jointly with task allocation (Ma et al., 2024).

In heterogeneous porous model design, the term refers to the blending of distinct implicit microstructures into a single “mix-topology” blended microstructure inside a designer-specified blending region. The target is not merely geometric interpolation. It is to achieve a smooth transition within the blending region, preserve the non-blending regions exactly, and avoid topological errors such as unintended isolated connected components and isolated voids in the resulting solid (Gao et al., 2024).

In multiscale structural optimization, the concept denotes a hybridization across scales, fidelities, and latent representations. The multifidelity and multiscale topology optimization framework based on phasor-based evolutionary de-homogenization describes “mix-topology” as mixing macro-scale descriptor fields $X=[\mu_1,\mu_2,\theta]$ with micro-scale geometry reconstructed through phasor-based de-homogenization; mixing low-fidelity homogenized models with high-fidelity CAD-based evaluation; and mixing PCA-compressed fields with VAE-based latent crossover and image deformation-based mutation (Xu et al., 9 Oct 2025). In a related but distinct multiscale setting, MR-LVGP-based topology optimization mixes multiple classes of microstructures across a macrostructure so that different regions can adopt the class and parameterization best suited to local stress states, while the discrete class labels are embedded into a continuous latent design space to make inter-class transitions differentiable (Wang et al., 2020).

In structural topology optimization, mixed topology also refers to combining different parameterization paradigms in one design domain. The mixed projection- and density-based topology optimization framework uses an explicit geometric parametrization for selected regions, especially assembly interfaces, and a density-based implicit representation elsewhere. The explicit part controls interface location, shape, local constraints, and local material properties, while the density field retains free-form topology optimization capability in the rest of the domain (Pollini et al., 2019). The phase-field literature uses “mixed” in a variational sense: a three-field Hu–Washizu formulation is combined with a phase-field topology functional so that equilibrium, constitutive, and compatibility relations are enforced directly within the same variational principle (Marino et al., 2021).

Across these uses, a mixed topology is defined less by a single canonical geometry than by a design principle: heterogeneous topological ingredients are co-optimized rather than handled sequentially or independently. This suggests that “Mix-Topology Design Strategy” is best understood as a unifying methodological label rather than a single domain-specific algorithm.

2. Graph, network, and communication interpretations

The most explicit graph-theoretic formulation appears in decentralized MEC offloading. There, the system is modeled as an undirected graph $G=\{V,E\}$ with node set $V=\{0,1,\ldots,N-1\}$, adjacency matrix $A$, communication range $\xi$, and neighbor set $\mathcal{N}_i$. The master node is indexed by $0$. The design variables include binary CH-selection variables $o_i$, binary CM-selection variables $x_{ij}$, and continuous task allocations $y_i\ge 0$. The topology must satisfy disjointness and task conservation constraints while minimizing total completion time $G=\{V,E\}$0, where $G=\{V,E\}$1 (Ma et al., 2024).

Within that formulation, topology design and resource allocation are inseparable. The local cluster formation phase evaluates whether adding a candidate cluster-member improves processing time using a performance indicator $G=\{V,E\}$2, and the master subsequently evaluates candidate cluster-head teams using an analogous criterion based on $G=\{V,E\}$3. The resulting mix-topology is effective because it avoids overloading the master’s links, exploits good mid-tier relays, and prunes nodes that worsen makespan. The paper states that the allocation $G=\{V,E\}$4 returned by Algorithm 3 is optimal for the constructed topology $G=\{V,E\}$5, while also stating that global optimality across all possible topologies is not guaranteed (Ma et al., 2024).

A different graph-theoretic interpretation appears in anonymous communication networks. The mix network simulator study examines network topology choices—cascade, multi-cascade, stratified, and peer-to-peer—and relates them to the anonymity trilemma among anonymity, latency, and bandwidth overhead. In that context, a practical mix-topology strategy is not a single fixed topology but a deployment choice among structured layered graphs, free-route meshes, and cascaded pipelines, each producing different anonymity aggregation and latency behavior. Stratified/layered topology is singled out as the recommended default for scalable anonymity, because routes intersect across layers, anonymity improves with traffic volume, and latency remains tied to exponential per-hop means (Piotrowska, 2021).

The distinction between the two graph settings is important. In MEC, the mixed topology is synthesized as a three-layer tree-like offloading structure from a general undirected graph, and the central performance measure is makespan (Ma et al., 2024). In mixnets, the “mix-topology” design question concerns how different deployment topologies alter anonymity, delay, and cover-traffic burden under a global passive observer, with entropy and sender-receiver unlinkability as anonymity measures (Piotrowska, 2021). The commonality lies in the co-design of topology and operational behavior; the divergence lies in whether the optimization target is computational latency or anonymity under traffic analysis.

3. Geometric and implicit-shape blending formulations

In porous structure design, the central mechanism of mix-topology is implicit-function blending under topology control. A porous solid is represented by a scalar field $G=\{V,E\}$6 and a threshold distribution field $G=\{V,E\}$7, with rod-, pore-, and sheet-type implicit definitions. Given two adjacent regions $G=\{V,E\}$8 and $G=\{V,E\}$9 with distinct microstructures $V=\{0,1,\ldots,N-1\}$0 and $V=\{0,1,\ldots,N-1\}$1, the blended scalar field is

$V=\{0,1,\ldots,N-1\}$2

and the blended solid is $V=\{0,1,\ldots,N-1\}$3 for rod-type structures (Gao et al., 2024).

The blending field $V=\{0,1,\ldots,N-1\}$4 is implemented using trivariate B-spline basis functions, with hard constraints ensuring $V=\{0,1,\ldots,N-1\}$5 outside the blending region on the $V=\{0,1,\ldots,N-1\}$6 side and $V=\{0,1,\ldots,N-1\}$7 outside on the $V=\{0,1,\ldots,N-1\}$8 side. Initialization differs between one-dimensional and fully three-dimensional blending. For simple planar, cylindrical, or spherical interfaces, one-dimensional B-spline ramps are used. For complex blending regions, a constrained least-squares fitting problem is solved using Local-LSPIA, with interior samples assigned values by relative distances to boundary sample sets via KD-tree queries (Gao et al., 2024).

The distinctiveness of the method lies in its topological objective. Persistent homology is computed on a cubical complex derived from a discretization of $V=\{0,1,\ldots,N-1\}$9, and the optimization targets $A$0 and $A$1-type errors in the solid at the evaluation isovalue. The objective

$A$2

moves offending persistence pairs out of the diagram regions associated with unwanted holes and extra components inside $A$3, while keeping non-blending regions unchanged by freezing control coefficients whose B-spline supports intersect the exterior (Gao et al., 2024).

This formulation makes the geometric meaning of mix-topology unusually literal: the strategy blends topologies of porous microstructures while preserving locality and eliminating topological defects. The paper reports that in one-dimensional blending experiments the method achieved $A$4 for several rod-type combinations where baseline blending methods produced many extra connected components or holes, and it maintained exact invariance outside $A$5 in contrast to GRBF-based weights (Gao et al., 2024).

A related geometric blending idea appears in “Remixing Functionally Graded Structures,” but there it operates on classwise signed distance fields rather than on persistent-homology-guided repair. Each class is represented by an SDF family, and a two-step blending procedure forms a weighted cross-dissolve at a representative volume fraction and then performs an activated soft-max union with each class’s lower feasible bound: $A$6

$A$7

The activation $A$8 guarantees that at least one lower-bound shape contributes, which the paper uses to guarantee connectivity and minimum feature size without explicit compatibility constraints between classes (Chan et al., 2021). The geometric and topological motivations therefore overlap with the porous-model case, but the formal machinery is different: persistent-homology optimization in one case, multiclass SDF blending with feasibility bounds in the other.

4. Material, microstructure, and multiscale design strategies

In multiscale structural design, mix-topology is closely tied to hybridization across descriptors, scales, and surrogates. The multifidelity and multiscale topology optimization framework based on phasor-based evolutionary de-homogenization represents each macro element by the descriptor vector $A$9, with $\xi$0 controlling orthogonal bar widths and $\xi$1 the in-plane orientation. These descriptors determine an orthotropic homogenized elasticity tensor $\xi$2, rotated by $\xi$3, while a detailed single-scale lattice is reconstructed by phasor synthesis and thresholding. The topology is therefore “mixed” across macro descriptors and micro geometry rather than represented solely at one scale (Xu et al., 9 Oct 2025).

The same framework also mixes model fidelities. Low-fidelity homogenization-based optimization generates initial designs and physics-informed deformation fields for mutation; high-fidelity evaluations on de-homogenized CAD geometries determine the actual objective values during NSGA-II search (Xu et al., 9 Oct 2025). It further mixes latent representations: PCA compresses descriptor fields, a VAE performs latent crossover, and image deformation-based mutation perturbs reconstructed geometries in a physics-guided way. Numerical results show hypervolume improvements of $\xi$4 and $\xi$5 on double-clamped beam stiffness and buckling tasks, $\xi$6, $\xi$7, and $\xi$8 on three L-bracket stress cases, and $\xi$9 on a multi-loading part (Xu et al., 9 Oct 2025). A plausible implication is that mix-topology here denotes a deliberately heterogeneous optimization stack, not merely a heterogeneous final geometry.

MR-LVGP-based multiscale optimization adopts a different route. Instead of reconstructing geometry from macro descriptors, it learns a latent embedding $\mathcal{N}_i$0 for qualitative microstructure classes $\mathcal{N}_i$1, combines that embedding with quantitative design parameters $\mathcal{N}_i$2, and uses a multi-response Gaussian process to predict homogenized stiffness tensors $\mathcal{N}_i$3. The kernel

$\mathcal{N}_i$4

gives a continuous geometry to the discrete class variable, allowing gradients with respect to class choice (Wang et al., 2020). In the coupled SIMP formulation, each element has density $\mathcal{N}_i$5, microstructure parameter $\mathcal{N}_i$6, and latent class coordinate $\mathcal{N}_i$7, and gradients are propagated through $\mathcal{N}_i$8 and a latent penalization term that encourages convergence to actual library classes (Wang et al., 2020).

The results substantiate the claim that mixing microstructure classes can outperform single-class designs. Reported compliance values improve from $\mathcal{N}_i$9 to $0$0 on a 2D L-beam, from $0$1 to $0$2 on a 2D multi-loading MBB beam, and from $0$3 to $0$4 on a 3D L-beam when multiclass mixing is used instead of single-class topology optimization (Wang et al., 2020).

A third multiscale formulation, “Remixing Functionally Graded Structures,” combines several microstructure families through multiclass SDF blending, feeds the resulting low-dimensional descriptors into a neural-network predictor of $0$5, and couples BESO at the macroscale with MMA at the microscale (Chan et al., 2021). Reported compliance values include $0$6 and $0$7 for 2-class and 3-class truss bases on an MBB beam, and approximately $0$8 and $0$9 for 2-class and 3-class freeform bases on the same problem (Chan et al., 2021). These results suggest that mix-topology may be instantiated either through latent continuous embeddings of discrete classes (Wang et al., 2020) or through direct classwise SDF blending with guaranteed feasibility (Chan et al., 2021).

5. Mixed formulations in structural and material topology optimization

A structurally different meaning of mix-topology appears in optimization formulations that combine representations or materials directly.

The mixed projection- and density-based topology optimization method combines an explicit segmented interface profile $o_i$0 with a classical density field $o_i$1. The geometric coordinates of profile nodes serve as shape variables $o_i$2, and density variables retain global free-form optimization elsewhere. The explicit profile defines a projection field $o_i$3 used to impose local constraints and property projections, including reduced Young’s modulus in the interface strip, local volume constraints, and spatially variable minimum and maximum length scales (Pollini et al., 2019). The optimization problem minimizes compliance $o_i$4 subject to the state equation $o_i$5, volume constraints, interface-strip constraints, and geometric regularization such as a slope constraint $o_i$6. Reported numerical examples include compliance $o_i$7 for an MBB beam with local interface-strip volume control, $o_i$8 when reduced modulus is imposed in the interface strip, $o_i$9 and $x_{ij}$0 for two localized maximum-length-scale cantilever examples, and $x_{ij}$1, $x_{ij}$2, and $x_{ij}$3 for variable minimum and maximum length-scale cases (Pollini et al., 2019).

The unified material interpolation for multi-material topology optimization provides another mixed-topology strategy, this time across materials rather than geometric representations. For each element, a vector $x_{ij}$4 is mapped to effective material weights through

$x_{ij}$5

and the Young’s modulus is interpolated as

$x_{ij}$6

The mapping is explicitly symmetric in the materials and promotes clear one-hot-like per-element assignments as $x_{ij}$7 increases (Yi et al., 2022). The paper reports crisp interfaces and clear 0–1 material selection in cantilever and half-MBB examples, with compliance values including approximately $x_{ij}$8 for a two-material cantilever, $x_{ij}$9 and $y_i\ge 0$0 for three- and five-material cantilevers, and $y_i\ge 0$1, $y_i\ge 0$2, and $y_i\ge 0$3 for two-, three-, and five-material half-MBB beams (Yi et al., 2022). Here, mix-topology denotes mixing material candidates at the element level while using a norm-based interpolation to suppress ambiguous mixtures.

A third mixed structural formulation is variational rather than representational. The mixed Hu–Washizu plus phase-field approach introduces displacement $y_i\ge 0$4, strain $y_i\ge 0$5, stress $y_i\ge 0$6, and phase field $y_i\ge 0$7 into a single functional, together with either a global volume constraint through a Lagrange multiplier $y_i\ge 0$8 or a local volume-penalty term $y_i\ge 0$9 (Marino et al., 2021). The method combines phase-field perimeter regularization, a bounding functional to enforce $G=\{V,E\}$00, and a three-field structural formulation enforcing equilibrium, constitutive, and compatibility equations directly. Numerical investigations show that the volume-minimization formulation yields practically the same compliance and stress metrics as the globally constrained formulation for matched final volume fractions, while often reducing Newton iterations by more than $G=\{V,E\}$01 for $G=\{V,E\}$02 (Marino et al., 2021). In this case, “mixed” refers to the variational structure, but it still fits the broader encyclopedia theme because topology generation is mediated by more than one coupled field representation.

6. Applications, empirical behavior, and limitations

The literature associates mix-topology strategies with several recurring empirical benefits.

In MEC offloading, DNTD-TO outperforms Unequal, Leach-C, LBAS, and a two-hop-pruned Dijkstra baseline in experiments with $G=\{V,E\}$03 and $G=\{V,E\}$04 over ten random topologies each, and its advantage becomes more pronounced at larger $G=\{V,E\}$05. Increasing communication range $G=\{V,E\}$06 from $G=\{V,E\}$07 m to $G=\{V,E\}$08 m improves performance, and DNTD-TO remains best across $G=\{V,E\}$09 values. The gains are attributed to balanced computation via optimal allocations, equal-bandwidth OFDM sharing within clusters, and pruning of long or slow routes (Ma et al., 2024). At the same time, the method assumes static topology, arbitrarily decomposable tasks, negligible downlink result size, no interference, and no energy model; performance may degrade under heavy interference, high mobility, or task-dependency constraints (Ma et al., 2024).

In porous blending, the persistent-homology-guided method is validated on one-dimensional, free-form, and disconnected blending regions, and is explicitly reported to preserve non-blending geometry while eliminating isolated components and holes in the blending region. Reported runtimes range from approximately $G=\{V,E\}$10–$G=\{V,E\}$11 s for $G=\{V,E\}$12 grids to around $G=\{V,E\}$13 s for $G=\{V,E\}$14 grids, with PH recomputation dominating cost (Gao et al., 2024). The method’s limitations include runtime scaling, sensitivity to grid resolution and B-spline settings, and the absence of integrated geometric or physical objectives such as curvature or pore-size constraints (Gao et al., 2024).

In multiscale optimization, the benefits are typically framed in terms of Pareto quality and manufacturability. The phasor-based evolutionary de-homogenization framework claims improved Pareto fronts, robust performance under damage, and modest generative-model cost relative to HF finite-element analysis, but also states that HF evaluations dominate runtime, exact property matching is not guaranteed, and the demonstrated setting is primarily 2D and linear elastic (Xu et al., 9 Oct 2025). MR-LVGP achieves multiclass microstructure selection with runtimes in minutes rather than hours relative to FE$G=\{V,E\}$15-style concurrent optimization, but the approach depends on data quality, scale separation, and uncertainty management in the surrogate (Wang et al., 2020). The multiclass shape blending framework achieves designs approaching or surpassing published FE$G=\{V,E\}$16 baselines at a fraction of the cost, yet still spends substantial time in bisection to match exact volume fraction and remains grounded in linear homogenized elasticity (Chan et al., 2021).

In structural assemblies, explicit–implicit mixed parameterization allows local control of joints and interfaces but introduces additional parameters, tighter move limits, and a more intricate continuation strategy (Pollini et al., 2019). In multi-material interpolation, norm-based mapping removes material-order bias and promotes discrete assignments, but stability depends on continuation in $G=\{V,E\}$17, $G=\{V,E\}$18, and Heaviside parameters (Yi et al., 2022). In phase-field Hu–Washizu formulations, the monolithic scheme is more robust than staggered NAND, but parameter setting remains central and extension to more constrained constitutive behavior is identified as future work (Marino et al., 2021).

A recurring misconception is that “mixed topology” merely means mixing shapes or materials. The surveyed papers indicate a broader and more technical meaning. It may refer to mixed graph organizations (Ma et al., 2024), mixed topology classes with continuous latent interpolation (Wang et al., 2020), mixed geometric bases with guaranteed feasibility (Chan et al., 2021), mixed explicit–implicit parameterizations (Pollini et al., 2019), mixed field formulations (Marino et al., 2021), or mixed material interpolation schemes (Yi et al., 2022). Another misconception is that mixing necessarily sacrifices rigor. Several of these methods pair the mixed representation with explicit optimization principles: KKT-based optimality within fixed clusters in MEC (Ma et al., 2024), persistent-homology objectives for topology repair (Gao et al., 2024), adjoint or chain-rule sensitivities in multiscale optimization (Wang et al., 2020), and monolithic variational stationarity in phase-field topology optimization (Marino et al., 2021).

Taken together, the literature supports a concise but technically precise characterization: Mix-Topology Design Strategy is a design paradigm in which heterogeneous topological primitives, classes, or representations are blended within a single optimization framework so that the final design space is richer than any one constituent topology, while feasibility and performance are maintained through explicit constraints, surrogates, or variational structure (Ma et al., 2024, Gao et al., 2024, Pollini et al., 2019, Xu et al., 9 Oct 2025, Wang et al., 2020, Chan et al., 2021).