PaperlayerView exact revision on arXiv

Source-linked AI summary

Sensitivity Hot Spot Penalization: A Robust Topology Optimization Framework against First-Order Worst-Case Perturbations

Junpeng Wang, Niels Aage, Ole Sigmund

arXiv:2608.30499v1cs.CE

TL;DR

Deterministic topology optimization does not explicitly control localized fragility from manufacturing and geometric uncertainty, while conventional robust formulations can require multiple costly realizations. This paper derives SHoSP as a first-order worst-case approximation under a norm-bounded material-mass perturbation budget and interprets its penalty weight accordingly. The analysis connects sensitivity hot spots with stress-energy concentrations and localized mechanism interactions, while extensions demonstrate applicability and scalability across several settings.

  • Problem

    The robustness interpretation of SHoSP was unspecified, including the meaning of its penalty weight, aggregation exponent, normalization, and connection to stress concentrations and hinge localization.

  • Method

    The paper combines a first-order Taylor expansion with a norm-bounded material-mass perturbation model to derive SHoSP as a worst-case robust approximation.

  • Results

    The analysis identifies κ as a dimensionless perturbation budget and relates sensitivity hot spots to density-weighted stress-energy concentrations and localized state–adjoint interaction.

  • Takeaways & Limitations

    SHoSP provides a low-cost robustness formulation whose interpretation covers compliance minimization, compliant mechanisms, porous infill, multiple load cases, and large-scale 3D examples.

  • Takeaways & Limitations

    The original SHoSP formulation lacked a specified uncertainty set and a direct uncertainty-based interpretation of its penalty parameter.

Abstract

from arXiv · show

Deterministic topology optimization can efficiently generate high-performance structural designs, but it does not explicitly control localized fragility induced by manufacturing variations and geometric uncertainties. Such fragility often appears as stress concentrations or hinge-like deformation mechanisms. Conventional robust topology optimization can suppress these features, but typically requires multiple design realizations and substantially increased computational cost. Recent work by Sigmund et al. (2026) introduced sensitivity hot spot penalization (SHoSP), which augments the nominal objective by a smooth maximum of its sensitivities and suppresses localized fragile features at low additional cost while promoting more even stress distributions. This paper establishes a general connection between SHoSP and a first-order worst-case robust approximation under a material-mass perturbation budget using a Taylor expansion and H{"o}lder's duality. The original penalty weight is identified as a dimensionless norm-bounded perturbation budget relative to the area (2D) or volume (3D) of the design domain. This interpretation explains the suppression of sensitivity hot spots, hinge localization, and stress concentrations by limiting the maximum first-order degradation under budgeted perturbations. It is investigated for compliance minimization and compliant mechanism design. Robustness is assessed using a relaxed worst-case density perturbation and by comparing deterministic and SHoSP designs under equivalent perturbation budgets. Stress-related effects observed on structured meshes are cross-validated by stress-oriented topology optimization and body-fitted finite element analyses. Extensions to porous infill optimization, multiple load cases, and large-scale 3D examples further demonstrate the applicability and superior scalability of the SHoSP framework.

1. Introduction

The introduction positions SHoSP as a low-cost alternative to explicit uncertainty realizations and establishes its first-order worst-case interpretation under a material-mass perturbation budget. It connects sensitivity hot spots with fragile mechanical features and outlines theoretical and numerical validation.

  • Existing robust formulations: Scenario-based erosion–intermediate–dilation formulations improve manufacturing tolerance, control minimum feature sizes, suppress localized hinges, and promote nearly discrete layouts.They optimize against the worst response among prescribed geometric realizations.
  • Motivation: Repeated uncertainty realizations increase computational cost, especially for large-scale 3D compliant mechanisms where erosion can weaken or disconnect slender hinges and transmission paths.The cost also depends on the numerical properties of each realization, not only their number.
  • Sensitivity-based robustness: Sensitivity-based approaches use Taylor expansions and derivative information to approximate uncertainty effects without explicitly resolving many uncertain realizations.Related first-order formulations can require only one additional adjoint system per optimization iteration.
  • SHoSP: SHoSP augments the nominal objective with a smooth maximum of area- or volume-normalized density sensitivities, requiring only two additional adjoint right-hand sides.Prior results associated the method with reduced stress concentrations and suppressed hinge-like mechanisms.
  • Contribution: This paper identifies SHoSP as the first-order worst-case approximation for a norm-bounded perturbation of element-wise material mass, interpreting κ as a dimensionless area- or volume-relative budget.The aggregation exponent controls how spatially concentrated the admissible perturbation can be.
  • Validation and scope: The mechanical analysis relates sensitivity hot spots to density-weighted stress-energy concentrations in compliance minimization and localized state–adjoint interaction in compliant mechanisms.The framework is assessed with associated density perturbations and extended to porous infill, multiple load cases, and large-scale 3D examples.

2. Density-based topology optimization

The paper formulates density-based topology optimization with filtered and projected design variables, SIMP material interpolation, volume constraints, and adjoint sensitivities for compliance and compliant mechanisms.

  • Design variables are transformed into physical densities through sequential filtering and projection operations.
  • SIMP interpolates finite-element material properties using a penalization factor typically set to q=3.
  • The Helmholtz-type PDE filter regularizes the optimization problem and suppresses mesh-dependent patterns.
  • A global volume-fraction constraint uses element areas or volumes and a prescribed volume fraction.
  • Compliance minimization optimizes structural stiffness, while compliant mechanism design optimizes a signed output displacement.
  • Objective sensitivities with respect to physical densities are computed by adjoint analysis and back-propagated through filtering and projection.

3. SHoSP as a first-order worst-case approximation

SHoSP augments topology optimization with a smooth sensitivity maximum that suppresses localized fragile features. The paper shows that its penalty weight is a normalized material-mass budget for first-order worst-case perturbations.

  • Deterministic designs exhibit sensitivity peaks associated with stress concentrations and hinge-like deformation, while SHoSP redistributes these effects over larger regions.
  • SHoSP augments the nominal objective with a sensitivity hot-spot measure scaled by penalty parameter κ.
  • The area- or volume-normalized sensitivity removes element-size dependence, and the p-norm approaches a maximum operator as p→∞.
  • The dual exponent r controls whether perturbations concentrate on a few elements or distribute spatially across the design domain.
  • A first-order Taylor expansion expresses objective degradation under physical-density perturbations as a material-mass sensitivity term.
  • Hölder duality identifies the SHoSP aggregation as the maximum first-order objective deterioration over a norm-bounded material-mass uncertainty set.
  • κ is the dimensionless material-mass perturbation budget relative to the design-domain area or volume, with adversarial distribution determining the worst first-order degradation.
  • The formal Hölder worst-case perturbation may violate point-wise density bounds, so bounded approximations are needed for finite-perturbation robustness assessment.

4. Mechanical interpretation of sensitivity hot spots

Sensitivity hot spots have mechanical interpretations: in compliance minimization they reflect density-weighted local stress energy, while in compliant mechanisms they indicate localized force-transmission mechanisms.

  • Penalizing sensitivity hot spots often alleviates stress concentrations without including a stress measure in the augmented objective.
  • In compliant mechanisms, localized hinge-like deformation patterns are replaced by more distributed compliant regions.
  • Compliance sensitivity hot spots are closely related to local stress-energy concentrations through density-weighted sensitivity expressions.
  • In compliant mechanism design, hot spots are governed by the interaction between state and adjoint stress fields, indicating localized force-transmission mechanisms.
  • Sensitivity hot spots and von Mises stress concentrations are related but not identical because of density weighting and different scalarizations of stress.
  • The stress-energy measure includes deviatoric and volumetric contributions, whereas von Mises stress uses only the deviatoric component; therefore no general one-to-one correspondence is implied.

Appendix C. The density-dependent factor 𝑐𝜌further modifies this relation in intermediate-density

SHoSP has distinct mechanical interpretations in compliance minimization and compliant mechanism design. It suppresses localized structural responsibility, producing more distributed responses and improved regularity and robustness.

  • In compliant mechanisms, hot spots arise from local interaction between state and adjoint stress fields, not from a single stress field.The strongest interaction can occur between distinct state- and adjoint-stress peaks around a hinge.
  • Penalizing compliant-mechanism hot spots discourages localized transmission and redistributes deformation and force transfer over a larger compliant region.The resulting distributed hinge pattern weakens the associated state-stress concentration without directly minimizing state stress.
  • In compliance minimization, sensitivity hot spots are closely related to density-weighted local stress-energy concentrations.
  • In both problem classes, hot spots identify localized features on which the global response depends disproportionately.Suppressing these features improves structural regularity and robustness, with manifestations including stress reduction, hinge redistribution, or more distributed load transfer.

5. Implementation and scalability

The implementation extends SHoSP to scalable solvers, multiple load cases, and constrained design spaces. Its perturbation-budget meaning remains, but practical effects depend on objective scale, activation, weighting, and redistribution freedom.

  • Implementation and computational cost: SHoSP adds auxiliary adjoint systems, requiring two additional solves for compliant mechanisms and one for compliance minimization using shared stiffness matrices.Fixed-mesh factorization reuse and common matrix-specific preconditioners support the added right-hand sides.
  • Implementation and computational cost: Auxiliary systems are well-suited to warm starts, with iteration counts generally no larger and often slightly smaller than corresponding state and nominal systems.
  • Extended formulations: Multiple loading conditions use a min–max formulation that bounds SHoSP-augmented compliance across independently evaluated load-specific hot-spot measures.The governing load case, or multiple active cases, is selected directly by optimization.
  • Extended formulations: Local volume constraints restrict material redistribution, making the effect of a given κ more problem-dependent while preserving the first-order interpretation.This constrained setting tests SHoSP under strongly limited feasible redistribution, including porous infill design.
  • Practical choice of penalization strength: The same κ has the same budget interpretation but can produce different nominal-performance trade-offs depending on objective scale, Ψ, and redistribution freedom.Activation that is too weak or late may merely regularize an established topology, whereas excessive or early activation may interfere with intended response formation.

6. Results

Across 2D and 3D benchmarks, SHoSP reduces sensitivity localization and localized structural mechanisms, with robustness gains verified under adverse perturbations and body-fitted reanalysis. The method also produces distinct mechanisms from direct stress control.

  • Canonical 2D benchmarks: SHoSP weakens localized sensitivity hot spots in both benchmarks, reducing L-beam stress concentration and compliant-mechanism hinge localization.
  • Verification and robustness assessment: Body-fitted reanalysis keeps objective discrepancies below 3% for the L-beam and below 1% for the inverter while preserving SHoSP trends.Body-fitted peak von Mises stresses are consistently lower than structured-mesh values.
  • Verification and robustness assessment: For every tested SHoSP design with κ_test ≤ κ, measured objective deterioration remains below the corresponding first-order reference value κΨ∕|Φ|.Deterministic designs may exceed these references, especially at larger κ and κ_test.
  • Comparison with stress-constrained topology optimization: SHoSP and stress-constrained optimization can attain similar stress levels while producing different sensitivity fields and mechanisms.SHoSP replaces localized hinges with more distributed compliant regions, showing that stress control and sensitivity-localization suppression are distinct.
  • Three-dimensional examples: In 3D, increasing κ reduces sensitivity peaks and redistributes material, with the L-beam peak decreasing from 28.9 to 2.72 at κ = 0.5%.For that L-beam, maximum von Mises stress decreases from 5.11 to 2.46 while compliance increases from Φ = 0.078 to Φ = 0.085.
  • Multiple loading conditions: For the five-load bridge, the largest sensitivity decreases from approximately 1.38 × 10^4 to 2.07 × 10^3 at κ = 0.2%.The min–max formulation redistributes material among principal load-carrying members and differentiates nominal compliances across load cases.

6.4. Large-scale practical demonstrations

Large-scale demonstrations show that SHoSP remains effective for compliant grippers, multi-load brackets, and porous infill despite substantial problem size and, in porous designs, restricted material redistribution.

  • Large-scale practical demonstrations: In the large-scale gripper, SHoSP replaces localized hinge-like connections with thicker, spatially extended compliant regions.The hot-spot measure decreases from Ψ = 1.32 × 10^5 to Ψ = 2.51 × 10^4, and maximum von Mises stress from 95.7 to 50.4, while Φ changes from −37.18 to −28.98.
  • Large-scale practical demonstrations: In the GE bracket, SHoSP reduces hot-spot measures across the individual load cases while retaining the same governing cases and incurring a moderate stiffness trade-off.The four load cases are evaluated with corresponding stress fields, compliances, and sensitivity hot-spot measures.
  • Large-scale practical demonstrations: In porous molar infill, the hot-spot measure decreases from Ψ = 6.21 × 10^5 to Ψ = 1.56 × 10^5 and maximum von Mises stress from 280 to 184.Compliance increases from Φ = 42.96 to Φ = 44.49 while final volume fraction remains essentially unchanged.
  • Large-scale practical demonstrations: The porous-infill example demonstrates SHoSP effectiveness even when local volume constraints strongly restrict material redistribution.Its larger relative computational overhead is associated with the heterogeneous stiffness distribution of porous designs.

7. Conclusion and future work

The paper interprets SHoSP as a first-order worst-case robust topology optimization method and reports reduced sensitivity hot spots with moderate nominal-objective changes across applications and scales.

  • Conclusion and future work: SHoSP’s augmented objective corresponds to a worst-case first-order approximation under a budgeted material-mass uncertainty set.The derivation uses a first-order Taylor expansion and Hölder’s duality.
  • Conclusion and future work: The penalty parameter κ represents a dimensionless norm-bounded material-mass perturbation budget relative to the design-domain area or volume.The large-p aggregation corresponds to a near-ℓ1 material-mass uncertainty mode.
  • Conclusion and future work: SHoSP designs exhibit smaller objective deterioration under density-bounded adverse material-mass perturbations.This perturbation-based assessment supports the worst-case robustness interpretation.
  • Conclusion and future work: Sensitivity hot spots relate to density-weighted local stress-energy concentrations in compliance minimization and localized force-transmission mechanisms in compliant mechanisms.These interpretations explain observed suppression of stress concentrations and hinge-like mechanisms.
  • Conclusion and future work: The formulation extends to multiple loading conditions and porous infill optimization, while large-scale examples remain computationally feasible.Additional adjoint systems share the original stiffness matrix and support warm-started iterative solves in the examples considered.
  • Conclusion and future work: Future work includes alternative perturbation modes, application-specific uncertainty models, higher-order extensions for non-small perturbations, and adaptive κ strategies.Other aggregation exponents may represent different admissible perturbation geometries, while finite-amplitude extensions may be relevant when perturbations are no longer small.

A. Robin boundary parameter in the PDE filter

The PDE filter uses Robin boundary compensation to address missing neighborhood information near free boundaries. The selected parameter ls = 2l0 is presented as a practical compromise between compensation and excessive rounding.

  • A. Robin boundary parameter in the PDE filter: The Robin treatment compensates for missing neighborhood information outside non-box-shaped design domains.Its compensation strength is controlled by ls.
  • A. Robin boundary parameter in the PDE filter: Figure 19 compares deterministic compliance minimization results for 2D and 3D L-beam benchmarks across ls = 0, l0, 2l0, and 3l0.All other parameters are kept unchanged in the comparison.
  • A. Robin boundary parameter in the PDE filter: Without Robin compensation, material tends to accumulate near free boundaries because neighborhood information is missing.The compensation changes filtering behavior near free design-domain boundaries.
  • A. Robin boundary parameter in the PDE filter: The choice ls = 2l0 balances boundary compensation against excessive rounding near design-domain creases.The paper uses this value throughout its computations.

B. Derivation of dΨ

The appendix derives the sensitivity of the hot-spot measure Ψ with respect to physical density variables using total differentiation and auxiliary adjoint systems.

  • B. Derivation of dΨ: The resulting derivation is equivalent to the original SHoSP adjoint-variable procedure but is presented in a compact direct-differentiation form.The formulation starts from the sensitivity-density aggregation adopted in this work.
  • B. Derivation of dΨ: The derivative of Ψ with respect to physical density includes explicit dependence and implicit dependence through the state and nominal adjoint fields U and λ.The derivation differentiates the state and nominal adjoint equations with respect to ρe.
  • B. Derivation of dΨ: The derivation uses the sensitivity-density aggregation, Kronecker-delta relations, and stiffness-matrix symmetry to obtain a compact derivative expression.The compact form is written after identifying auxiliary adjoints.
  • B. Derivation of dΨ: For compliance minimization, λ = −U collapses the two auxiliary adjoint systems into a single system.Compliance minimization is treated as a self-adjoint special case of the compliant mechanism formulation under the paper’s sign convention.

C. Relation between strain-energy density and von Mises stress

The appendix relates compliance sensitivity hot spots to local strain-energy density while distinguishing this quantity from von Mises stress. Their spatial localization may be similar in particular stress states, but no general one-to-one correspondence is expected.

  • C. Relation between strain-energy density and von Mises stress: The appendix expresses the compliance sensitivity in terms of principal stresses and gives the von Mises stress relation.The principal stresses are σ1, σ2, and σ3.
  • C. Relation between strain-energy density and von Mises stress: Strain-energy density contains deviatoric and volumetric contributions, whereas von Mises stress depends only on the deviatoric stress state.The decomposition uses the shear and bulk moduli of the solid material.
  • C. Relation between strain-energy density and von Mises stress: Similar spatial localization of compliance sensitivity and von Mises stress may occur when the deviatoric contribution dominates, but no general one-to-one correspondence should be expected.The plane-stress reduction sets σ3 = 0, while plane strain retains generally nonzero out-of-plane stress.
  • C. Relation between strain-energy density and von Mises stress: Figure 20 reports the evolution of the relative SHoSP correction rκ = κΨ/|Φ| for benchmark examples with nonzero κ.The histories provide numerical reference values for penalization levels and show transient increases after continuation updates.

D. Evolution of the relative SHoSP correction

The relative SHoSP correction is generally moderate but can become temporarily dominant after continuation updates. In multi-load optimization, the correction remains active across all distinct loading conditions, indicating suppression across the full load set.

  • Relative correction: Moderate SHoSP settings generally produce a non-negligible but non-dominant correction relative to the nominal objective.The relative correction r_κ is used to assess the penalty strength relative to the nominal objective.
  • Relative correction: Sharp transient increases in r_κ can occur after continuation updates, especially in compliant mechanism and porous infill optimization.These increases make the correction useful to monitor during optimization.
  • Multi-load convergence: Figure 21 compares nominal load-case responses Φ_i with SHoSP-augmented responses Φ_i + κΨ_i for three distinct loads at κ = 0.2%.Symmetry reduces the displayed cases to F1–F3.
  • Multi-load convergence: The SHoSP contribution remains active throughout optimization for every distinct loading condition.This supports suppression of sensitivity localization across the full set of load cases rather than only in the governing response.
Loading 2608.30499v1…