betaFeedback

WorksComputer Science

Multi-Material Topology Optimization for Targeted Structural Dynamic Vibration Attenuation

Abstract

Resonance-induced structural vibrations can compromise the performance, reliability, and service life of engineering structures. Multi-material topology optimization (MMTO) offers a promising approach to vibration attenuation by tailoring the spatial distributions of material constituents to optimize structural stiffness, mass, and natural frequencies. However, existing eigenfrequency-based MMTO methods face several challenges, including the need for flexible interpolation of multiple material properties, non-smooth optimization objectives associated with mode switching and repeated eigenvalues, and numerical difficulties in imposing differentiable constraints to exclude natural frequencies from prescribed operating frequency ranges. To address these challenges, we propose a novel MMTO framework for structural vibration attenuation based on several innovations: (1) a novel multivariate weight-based multi-material interpolation scheme to simultaneously represent material stiffness and mass density, (2) a bound-based optimization formulation to improve the smoothness and stability of gradient-based optimization by addressing mode switching, and (3) a smooth frequency-band formulation to smoothly approximate the discontinuous Heaviside function while enabling differentiable frequency-band constraints. Additionally, we systematically investigate and compare three different topology optimization formulations for dynamic vibration attenuation, including maximizing the fundamental eigenfrequency, maximizing the gap between adjacent eigenfrequencies, and excluding natural frequencies from prescribed frequency bands. Furthermore, we present the detailed derivation of analytical design sensitivities for the three formulations.

Introduction

Dynamic vibrations significantly affect the performance, integrity, and operational lifespan of engineered structures, where unmitigated responses can cause severe structural degradation and failure [1]. Developing effective vibration attenuation strategies has become a central objective in structural design and optimization [2]. Engineering structures, such as beams, plates, and shells, which are widely used in aircraft structures, spacecraft structures, bridges, and machinery, are inevitably subjected to dynamic excitation during operation [3]. The resulting vibration response is strongly influenced by the relationship between the external excitation frequencies and the natural frequencies, and corresponding mode shapes. When an excitation frequency coincides with a natural frequency and couples with corresponding mode shapes, resonance produces large-amplitude oscillations leading to excessive deformation [4]. Resonance-induced vibration relies on the distribution of natural frequencies: when the excitation is close to a single natural frequency, the corresponding mode generally dominates, whereas closely spaced natural frequencies or repeated eigenvalues cause multiple modes to participate simultaneously through modal superposition, resulting in complex dynamic behaviors [5], [6].

Vibration reduction can improve fatigue resistance, lower acoustic emission, enhance reliability, and prolong structure life; yet conventional mitigation strategies generally carry limitations. Passive mitigation approaches, such as structural damping enhancement and vibration isolation techniques, are widely adopted; such approaches, however, need to add additional masses and are generally effective over limited frequency ranges [7]. Active control schemes offer broadband performance, but require sensors, actuators, and energy supplies, thereby increasing complexity and cost [8]. An alternative is to design the structure such that its natural frequencies remain outside the operating frequency range to prevent resonance [9]. This approach is generally pursued by maximizing the fundamental eigenfrequency [10].

Topology optimization (TO) enables the systematic tailoring of dynamic characteristics through the distribution of material constituents and voids within a prescribed domain [11], [12], [13], [14]. Popular methods include the homogenization method [15], density-based formulation [16], [17], evolutionary approach [18], level-set [19], [20], moving morphable components (MMCs) [21], geometric projection method [22], and genetic algorithm [23], [24]. While these methods generally differ in design representation and optimization updates, their objective and constraints can be adapted to account for structural vibration characteristics [25]. The homogenization method was applied to maximize the fundamental natural frequency by optimizing the distribution of reinforcing material [26]. It was subsequently extended from individual-frequency improvement to spectral tailoring through the optimization of multiple eigenfrequencies [27]. Design sensitivity and optimality were established for repeated eigenvalues to address the difficulty of non-differentiability [28], and artificial localized modes in low-density regions in density-based methods [29].

Recent studies have extended eigenfrequency-based topology optimization to targeted vibration suppression [30] and frequency-dependent material properties [4]. Harmonic-response optimization incorporated a modal assurance criterion (MAC)-based mode-tracking strategy, where the excitation frequency is progressively increased toward the prescribed value amid optimization, penalizing structural optimization to avoid resonance [31]. While this approach targets the response at a prescribed excitation frequency, resonance avoidance over an operating frequency band requires the control of natural frequencies whose mode orders may not be known in advance. To address this issue, a frequency-band constraint was proposed to prevent natural frequencies from entering prescribed bands via Heaviside functions [32].

Compared to single-material topology optimization, multimaterial topology optimization (MMTO) expands design space by optimizing structural performance through material constituents and spatial distribution. Early developments included peak-function interpolation to represent materials without increasing the number of design variables [33], rational approximation of material properties (RAMP) to accommodate an arbitrary number of isotropic and anisotropic candidate materials [34], ordered SIMP to represent multiple materials using a single elemental density variable [35], and a univariate mapping-based interpolation scheme to reduce intermediate material densities and prevent undesired layers between material phases by integrating Heaviside projections with multi-variable filtering [36]. Additionally, to enforce discrete material selection per finite element, multiphase level-set methods distinguish non-overlapping material regions [37], and phase-field formulations regularize through energy-based models [38, 39]. MMTO has been investigated for vibration attenuation, where defining proper design objectives relies on the relationship between the structural spectrum and operating frequencies. In a pioneering multi-material beam design [40], two approaches were taken: maximizing the fundamental eigenfrequency to raise the lowest resonance frequency above an operating range, and maximizing the separation between adjacent eigenfrequencies to create an interval between resonance frequencies. Dynamic MMTO has also been extended to multiple materials with frequency constraints using an alternating active phase algorithm (AAPA) interpolation scheme equipped with penalized interpolation for stiffness and linear interpolation for mass to address pseudo modes and mode switching amid optimization [41].

Despite substantial progress, several challenges remain. First, conventional eigenfrequency optimization methods, particularly those based on max-min formulations, are susceptible to non-smoothness and numerical difficulties associated with mode switching and repeated eigenvalues, compromising the stability of gradient-based optimization. Second, although frequency-band constraints have been introduced to prevent natural frequencies from entering prescribed operating ranges, conventional Heaviside-based formulations are discontinuous and unsuitable for gradient-based optimization. Third, existing vibration attenuation strategies primarily focus on maximizing the fundamental eigenfrequency, offering limited flexibility in counteracting against ranges of user-defined operating frequencies, particularly when multiple operating frequency ranges need to be considered. These limitations motivate the development of a comprehensive multi-material topology optimization framework that simultaneously addresses material interpolation, eigenfrequency optimization under mode switching, and smooth enforcement of targeted frequency-band constraints to achieve flexible and effective structural vibration attenuation.

To address these research gaps, we propose a novel multi-material topology optimization for dynamic vibration attenuation with the following major contributions:

  • We formulate a new multivariate weight-based multi-material interpolation scheme to simultaneously represent material stiffness and mass density. Our scheme can accommodate an arbitrary number of candidate materials, enabling flexible spatial distribution of material constituents for structural vibration design.

  • We propose a new bound-based optimization formulation to address mode switching during optimization by introducing an auxiliary scalar design variable to reformulate non-smooth max-min formulations as individual constraints. Our approach improves the smoothness and numerical stability of gradient-based optimization for mode switching and closely spaced eigenfrequencies.

  • We introduce a novel smooth frequency-band constraint via a rational polynomial function to smoothly approximate the discontinuous Heaviside function. It enables differentiable frequency-band constraints and offers flexibility in counteracting user-defined operating frequency ranges with a natural frequency bandgap to prevent resonance-induced excessive vibrations.

  • To demonstrate our method, we systematically study and compare three topology optimization formulations for structural vibration attenuation, including the maximization of the fundamental eigenfrequency, the maximization of the gap between adjacent eigenfrequencies, and the exclusion of natural frequencies from prescribed operating frequency bands. These formulations provide different approaches to tailoring the structural frequency spectrum according to targeted vibration attenuation requirements. Effectiveness and applicability of the proposed method are demonstrated through numerical examples.

We organize the remainder of this paper as follows. In Section 2, we introduce the mathematical formulation of structural vibration. In Section 3, we propose a multi-material interpolation scheme for stiffness and mass matrices. Section 4 presents the topology optimization formulations for eigenfrequency optimization. In Section 5, we present the detailed derivation of analytical design sensitivities for three different topology optimization formulations for dynamic vibration attenuation. We proceed to describe the numerical implementation in Section 6 and demonstrate the proposed approach through numerical examples in Section 7. Finally, we summarize our contributions and highlight future work in Section 8.

Mathematical model of structural vibration

In this section, we model structural vibration by the finite element method (FEM) under the assumptions of linear elasticity, small deformations, and undamped free vibration. Accordingly, we neglect all external forces and damping effects, so that the dynamic response is governed by the mass and stiffness distribution of the structure. This reduces the problem to a generalized eigenvalue problem, from which we extract the natural frequencies and the corresponding mode shapes. In the following sections, we present the governing equations of structural free vibration in § 2.1, followed by the density filtering and Heaviside projection approaches in § 2.2 to prevent checkerboard patterns, reduce mesh dependency, and promote distinct material distributions during optimization.

Governing equation for structural vibration

Undamped free vibration can be described by the second-order equation of motion given in Equation (1):

Mu¨+Ku=0,(1)\mathbf{M}\ddot{\mathbf{u}}+\mathbf{K}\mathbf{u}=\mathbf{0}, \tag*{(1)}

where M\mathbf{M} and K\mathbf{K} denote the global mass and stiffness matrices, respectively, and u\mathbf{u} is the vector of nodal displacements. Since no external excitation or dissipative term appears in Equation (1), the motion is sustained entirely by the exchange between kinetic and strain energy, and the resulting oscillation is harmonic in time. Equation (1) therefore constitutes a second-order differential equation with constant coefficients, whose general solution can be expressed in the separable form u=veiωt\mathbf{u}=\mathbf{v}e^{\mathrm{i}\omega t}, where v\mathbf{v} is an eigenvector describing the spatial distribution of the motion and ω\omega is the associated eigenfrequency. Substituting this solution into Equation (1), followed by elimination of the non-vanishing factor eiωte^{\mathrm{i}\omega t}, we remove the time dependence and reduce the problem to the generalized eigenvalue problem of Equation (2):

(K−ω2M)vj=0;j=1,2,…,nd,(2)(\mathbf{K}-\omega^{2}\mathbf{M})\mathbf{v}_{j}=\mathbf{0};\quad j=1,2,\ldots,n_{d}, \tag*{(2)}

where ndn_d is the degree of freedom in the discretized model. A non-trivial solution exists only when the coefficient matrix becomes non-singular, which yields ndn_d eigenpairs (ωj2,vj)(\omega_j^2,\mathbf{v}_j) corresponding to the natural frequencies and mode shapes of the vibration. The eigenvectors are normalized with respect to the mass matrix according to Equation (3). This orthonormality ensures that the eigenvectors form a linearly independent basis.

viTMvj={0i≠j,1i=j.(3)\mathbf{v}_i^{T}\mathbf{M}\mathbf{v}_j= \begin{cases} 0 & i\ne j,\\ 1 & i=j. \end{cases} \tag*{(3)}

Design variable filtering and projection

In this work, to prevent checkerboard patterns and mesh dependence, density filtering is applied to each design variable within a prescribed filter radius. A smoothed Heaviside projection is then applied to the filtered density design variables to reduce intermediate values and obtain a discrete material type, with the projection steepness gradually increased through a continuation scheme.

Density filter

Element-wise density formulation is prone to checkerboard patterns and mesh dependence. To suppress both, a density filter [42, 43] is applied independently to each of the two design fields, giving the filtered densities (x~e=[x~e(1),x~e(2)])(\widetilde{x}_e=[\widetilde{x}_e^{(1)},\widetilde{x}_e^{(2)}]) in Equation (4):

x~e(f)=∑j∈NeHejxj(f)∑j∈NeHej,f=1,2.(4)\widetilde{x}_e^{(f)}=\frac{\sum_{j\in N_e}H_{ej}x_j^{(f)}}{\sum_{j\in N_e}H_{ej}},\qquad f=1,2. \tag*{(4)}

Here, NeN_e denotes the set of elements whose centroids lie within the filter radius rmin⁡r_{\min} of the element ee, and the linear weighting factor Hej=max⁡(0,rmin⁡−∥re−rj∥)H_{ej}=\max(0,r_{\min}-\lVert\mathbf{r}_e-\mathbf{r}_j\rVert) decays with the distance between the centroids re\mathbf{r}_e and rj\mathbf{r}_j, so that neighboring elements at closer distances contribute more to filtered density.

Heaviside projection

Although filtering suppresses numerical instabilities, it may cause local regions of intermediate density densities. As these grey elements lack clear physical meaning and may complicate manufacturing, a smoothed Heaviside projection [44], given in Equation (5), is applied to each filtered field to obtain the projected physical densities ϕe=[ϕe(1),ϕe(2)]\boldsymbol{\phi}_e=[\phi_e^{(1)},\phi_e^{(2)}].

ϕe(f)=tanh⁡(βfη)+tanh⁡(βf(x~e(f)−η))tanh⁡(βfη)+tanh⁡(βf(1−η)),(5)\phi_e^{(f)}=\frac{\tanh(\beta_f\eta)+\tanh\left(\beta_f(\widetilde{x}_e^{(f)}-\eta)\right)}{\tanh(\beta_f\eta)+\tanh\left(\beta_f(1-\eta)\right)}, \tag*{(5)}

where η\eta is the projection threshold and βf\beta_f controls the sharpness of the projection. For small βf\beta_f, the mapping is nearly linear, whereas for large βf\beta_f, it approaches an ideal step function at η\eta, driving densities towards 0 and 1. Imposing a large βf\beta_f makes the optimization highly non-linear and causes divergence; βf\beta_f is therefore increased gradually during the optimization following the continuation rule of [45], [46], [47], explained in § 6.

Since the objective and constraints are evaluated in terms of the projected densities, the sensitivities must be transferred back to the design variables through the chain rule. The required derivative of the projection is obtained by differentiating Equation (5) and is given in Equation (6).

∂ϕe(f)∂x~e(f)=βf(1−tanh⁡2(βf(x~e(f)−η)))tanh⁡(βfη)+tanh⁡(βf(1−η)).(6)\frac{\partial\phi_e^{(f)}}{\partial\widetilde{x}_e^{(f)}}=\frac{\beta_f\left(1-\tanh^2\left(\beta_f(\widetilde{x}_e^{(f)}-\eta)\right)\right)}{\tanh(\beta_f\eta)+\tanh\left(\beta_f(1-\eta)\right)}. \tag*{(6)}

Multivariate weight-based multi-material interpolation

In this section, we propose a novel formulation, i.e., a multivariate weight-based interpolation scheme to describe the distribution of candidate materials. We discretize the design domain into ne=nelx×nelyn_e=n_{\mathrm{elx}}\times n_{\mathrm{ely}} finite elements, where nelxn_{\mathrm{elx}} and nelyn_{\mathrm{ely}} denote the numbers of elements in the horizontal and vertical directions in a 2D design domain. For nMATn_{\mathrm{MAT}} candidate materials, the scheme requires nMAT−1n_{\mathrm{MAT}}-1 design fields, and consequently, nMAT−1n_{\mathrm{MAT}}-1 design variables are assigned to each finite element. For example, for a three-material structure, it yields two design fields, so that the design variable of the element ee becomes xe=[xe(1),xe(2)]\mathbf{x}_e=[x_e^{(1)},x_e^{(2)}], with xe(1),xe(2)∈[0,1]x_e^{(1)},x_e^{(2)}\in[0,1].

M(xe)={M1,xe(1)<xe,th and xe(2)∈[0,1],M2,xe(1)≥xe,th and xe(2)<xe,th,M3,xe(1)≥xe,th and xe(2)≥xe,th.(7)\mathrm{M}(x_e)= \begin{cases} \mathrm{M1}, & x_e^{(1)}<x_{e,th}\ \text{and}\ x_e^{(2)}\in[0,1],\\ \mathrm{M2}, & x_e^{(1)}\ge x_{e,th}\ \text{and}\ x_e^{(2)}<x_{e,th},\\ \mathrm{M3}, & x_e^{(1)}\ge x_{e,th}\ \text{and}\ x_e^{(2)}\ge x_{e,th}. \end{cases} \tag*{(7)}

During optimization, design variables remain continuous, so the type of the material occupying each element is identified by comparing the design field against a threshold value of xe,th=0.5x_{e,th}=0.5 as expressed in Equation (11). Values below the threshold are interpreted as the preceding candidate material and values above it as the succeeding one. The threshold is applied only to extract and report the material distribution, and does not enter the optimization; therefore, no non-differentiability is induced. For example, in an optimized three-material design, the interpolation scheme assigns the first type of candidate material (M1) per element when xe(1)=0x_e^{(1)}=0, the second material (M2) when xe(1)=1x_e^{(1)}=1 and xe(2)=0x_e^{(2)}=0, and the third material (M3) when xe(1)=1x_e^{(1)}=1 and xe(2)=1x_e^{(2)}=1, as expressed in Equation (11). Figure 1 illustrates this penalization scheme for the three-material example. Specifically, Figure 1(a) shows the interpolation between the first and second materials, while Figure 1(b) shows the interpolation between the second and third materials. The selection of either M2 or M3 requires the corresponding condition in Figure 1(a) to be satisfied first, after which the second design field distinguishes between these two materials. This hierarchical procedure is extended to a system containing user-defined nMATn_{MAT} materials by successively applying the interpolation scheme to the candidate materials.

Figure 1. Illustration of the penalization in the proposed multivariate weight-based interpolation: (a) the first design variable in a three-material design, which selects between the first material and the remaining two materials, and (b) the second design variable, which selects between the second and third materials.

Weights for multi-material interpolation scheme

During optimization, the design variables are firstly filtered using Equation (4) and subsequently projected using Equation (5). The projected fields are then converted into material weights, which define the contribution of each candidate material to the effective element properties. For a mesh of nen_e elements and nMatn_{Mat} candidate materials, the unnormalized weight of the first material follows directly from the first projected field, as shown in Equation (8).

w1un=ϕ1r.(8)w_1^{un}=\phi_1^r. \tag*{(8)}

The unnormalized weights for intermediate materials (2≤m≤nMat−12\le m\le n_{Mat}-1) are designed as in Equation (9):

wmun=(∏j=1m−1(1−ϕjr))ϕmr.(9)w_m^{un}=\left(\prod_{j=1}^{m-1}(1-\phi_j^r)\right)\phi_m^r. \tag*{(9)}

The unnormalized weight for the last material takes the remaining fraction, which is given by Equation (10):

wnMatun=∏j=1nMat−1(1−ϕjr).(10)w_{n_{Mat}}^{un}=\prod_{j=1}^{n_{Mat}-1}(1-\phi_j^r). \tag*{(10)}

The weights are normalized according to Equation (11) to enforce the partition-of-unity condition. We note that this normalization guarantees that the material fractions within an element are mutually exclusive and complementary: an increase in the fraction of one material is necessarily accompanied by a corresponding decrease in the fractions of the remaining materials. Consequently, the interpolated element properties always remain bounded by those of the candidate materials.

wm=wmun∑m=1nMatwmunand∑m=1nMatwm=1.(11)w_m=\frac{w_m^{un}}{\sum_{m=1}^{n_{Mat}}w_m^{un}} \quad\text{and}\quad \sum_{m=1}^{n_{Mat}}w_m=1. \tag*{(11)}

Finite element formulation for structural vibration

We discretize the design domain using four-node bilinear quadrilateral elements, each having two translational degrees of freedom per node. Since the out-of-plane thickness of the domain is small relative to its in-plane dimensions, we assume a plane stress condition, so that the elastic modulus matrix of the mthm^{th} candidate material is described by Equation (12) under the plane stress assumption:

Dm=Em1−νm2[1νm0νm10001−νm2].(12)\mathbf{D}_m=\frac{E_m}{1-\nu_m^2} \begin{bmatrix} 1 & \nu_m & 0\\ \nu_m & 1 & 0\\ 0 & 0 & \frac{1-\nu_m}{2} \end{bmatrix}. \tag*{(12)}

where EmE_m and νm\nu_m denote the Young's modulus and Poisson's ratio of the material mm, respectively. The corresponding element stiffness and mass matrices are obtained from the standard Galerkin formulation in Equation (13):

Ke(m)=∫ΩeBTDmB dΩ;Me(m)=∫ΩeρmNTN dΩ.(13)\mathbf{K}_e^{(m)}=\int_{\Omega_e}\mathbf{B}^{T}\mathbf{D}_m\mathbf{B}\,\mathrm{d}\Omega;\quad \mathbf{M}_e^{(m)}=\int_{\Omega_e}\rho_m\mathbf{N}^{T}\mathbf{N}\,\mathrm{d}\Omega. \tag*{(13)}

in which B\mathbf{B} is the strain-displacement matrix derived from the shape function matrix N\mathbf{N}, ρm\rho_m is the mass density of the material mm, and the integrations are carried out numerically over the element domain Ωe\Omega_e using Gauss quadrature.

We assume all candidate materials share the same Poisson's ratio, allowing the material properties to be decoupled from the element geometry, so that the contribution of each element to the structural response can be expressed through an effective modulus of elasticity and an effective density, as stated in Equation (14):

Eeeff=∑m=1nMatEmwe,m ; ρeeff=∑m=1nMatρmwe,m,(14)E_e^{\mathrm{eff}}=\sum_{m=1}^{n_{\mathrm{Mat}}} E_m w_{e,m}\ ;\ \rho_e^{\mathrm{eff}}=\sum_{m=1}^{n_{\mathrm{Mat}}} \rho_m w_{e,m}, \tag*{(14)}

where we,m\mathbf{w}_{e,m} is the interpolation weight of material mm in element ee, and nMatn_{\mathrm{Mat}} is the number of candidate materials. These weights govern the transition between the constituent materials, thereby determining the amount of each material present in each element.

Consequently, global stiffness and mass matrices are assembled from the elemental contributions scaled by these effective properties in (15):

K=∑e=1neEeeffKe ; M=∑e=1neρeeffMe,(15)\mathbf{K}=\sum_{e=1}^{n_e} E_e^{\mathrm{eff}}\mathbf{K}_e\ ;\ \mathbf{M}=\sum_{e=1}^{n_e} \rho_e^{\mathrm{eff}}\mathbf{M}_e, \tag*{(15)}

where Ke\mathbf{K}_e is the elemental stiffness matrix evaluated for unit modulus of elasticity, Me\mathbf{M}_e is the elemental mass matrix evaluated for unit material density, and nen_e is the total number of elements. Both matrices are computed once at the beginning of the analysis and reused in every iteration, which appreciably reduces the computational effort of the optimization.

With the global matrices assembled and the prescribed boundary conditions imposed, the generalized eigenvalue problem of (2) is solved to obtain the eigenvalues along with associated eigenvectors. The eigenvectors are subsequently mass-normalized by (3), providing a consistent basis required for sensitivity analysis in Section 5.

Frequency band constraint by rational polynomial function

In engineering applications, it may not be sufficient to merely maximize the fundamental frequency to prevent resonance. For this reason, we require no natural structural frequency to reside inside prescribed working frequency ranges by proposing a frequency band constraint in this section.

Let prohibited bands be defined by the lower and upper frequency limits (ωlow(b)\omega_{\mathrm{low}}^{(b)} and ωupp(b)\omega_{\mathrm{upp}}^{(b)}) and indexed by b∈{1,…,nM}b \in\{1,\ldots,n_M\}, where nMn_M is the number of bands. It is convenient to describe each band by its center and half-width by (16):

ωc(b)=ωupp(b)+ωlow(b)2,ωd(b)=ωupp(b)−ωlow(b)2,(16)\omega_c^{(b)}=\frac{\omega_{\mathrm{upp}}^{(b)}+\omega_{\mathrm{low}}^{(b)}}{2},\quad\omega_d^{(b)}=\frac{\omega_{\mathrm{upp}}^{(b)}-\omega_{\mathrm{low}}^{(b)}}{2}, \tag*{(16)}

where this pair of quantities maps the band onto the normalized coordinate (ω−ωc(b))/ωd(b)(\omega-\omega_c^{(b)})/\omega_d^{(b)}, which equals zero at the band center and ±1\pm1 at the band edges. The exclusion requirement can be written compactly as ωj∉[ωlow(b),ωupp(b)]\omega_j \notin[\omega_{\mathrm{low}}^{(b)},\omega_{\mathrm{upp}}^{(b)}] for every eigenfrequency ωj\omega_j by the Heaviside function in (17):

H(ωj)={1,ωj∈[ωlow(b),ωupp(b)],0,ωj∉[ωlow(b),ωupp(b)],(17)H(\omega_j)=\begin{cases} 1, & \omega_j\in[\omega_{\mathrm{low}}^{(b)},\omega_{\mathrm{upp}}^{(b)}],\\ 0, & \omega_j\notin[\omega_{\mathrm{low}}^{(b)},\omega_{\mathrm{upp}}^{(b)}], \end{cases} \tag*{(17)}

which returns unity whenever a frequency violates the exclusion condition and zero otherwise.

However, it cannot be used directly in a gradient-based optimization, as the function is discontinuous at band edges, its derivative vanishes everywhere else, and therefore, it provides little indication about the distance of a given eigenfrequency to the prohibited region. For example, a design whose eigenfrequency sits just inside the band and another design whose eigenfrequency lies at the band center are indistinguishable to the optimizer, and therefore, no descent direction can be constructed.

To overcome this limitation, we propose to replace the Heaviside function with a smooth and differentiable approximation via the rational polynomial function (RPF) in (18):

RPF(ω)=11+(ω−ωcωd)2βFBC,(18)RPF(\omega)=\frac{1}{1+\left(\frac{\omega-\omega_c}{\omega_d}\right)^{2\beta_{FBC}}}, \tag*{(18)}

in which βFBC\beta_{FBC} is a positive integer that controls transition sharpness, as illustrated in Figure 2. At the band center the normalized coordinate vanishes, and the function attains its maximum value of unity; at either band edge the normalized coordinate equals ±1\pm1, so that the denominator becomes exactly two and the function takes the value 1/21/2 irrespective of the choice of βFBC\beta_{FBC}; and as the frequency moves away from the band, the denominator grows without bound, and the function decays monotonically towards zero. We note that the exponent 2βFBC2\beta_{FBC} guarantees that the function is symmetric about ωc\omega_c and remains positive for all arguments. Because the level 1/21/2 is attained precisely at the band limits for every admissible βFBC\beta_{FBC}, it provides a natural and parameter-independent threshold separating admissible from inadmissible frequencies.

Rational polynomial function

Figure 2. Rational polynomial function: Approximation of the Heaviside function in the frequency-band constraint.

Based on the rational polynomial function, the frequency band constraint (FBC) for the bthb^{th} prohibited band is imposed in the standard non-positive form as given in (19):

gFBC,b=FBCb−1/2≤0,(19)g_{FBC,b}=FBC_b-1/2\leq0, \tag*{(19)}

so that the satisfaction of the constraint requires FBCb<1/2FBC_b<1/2, which indicates that the eigenfrequencies lie outside [ωlow(b),ωupp(b)][\omega_{\mathrm{low}}^{(b)},\omega_{\mathrm{upp}}^{(b)}], as in Figure 2. Unlike the Heaviside measure, gFBC,bg_{FBC,b} varies continuously with design variables through the eigenfrequencies, and its gradient conveys both the direction and the magnitude of the change required to push a frequency out of the forbidden interval. Applying the constraint to every eigenfrequency would introduce J0×nMJ_0 \times n_M individual constraints and would make the size of the optimization dependent on the number of modes retained. To avoid this, we gathered the contributions of all eigenfrequencies into a single scalar for each band by the p-norm aggregation as in Equation (20):

FBCb=(1J0∑j=1J0FBb,jp)1/p,(20)FBC_b=\left(\frac{1}{J_0}\sum_{j=1}^{J_0}FB_{b,j}^{p}\right)^{1/p}, \tag*{(20)}

where J0J_0 denotes the total number of eigenvalues considered in the analysis and pp is the aggregation parameter. The individual contributions are obtained by evaluating the rational polynomial function at each eigenfrequency ωj\omega_j with respect to the parameters of the bthb^{th} band, leading to an aggregated measure as:

FBb,j=11+(ωj−ωc(b)ωd(b))2βFBC.(21)FB_{b,j}=\frac{1}{1+\left(\frac{\omega_j-\omega_c^{(b)}}{\omega_d^{(b)}}\right)^{2\beta_{FBC}}}. \tag*{(21)}

The aggregated measure in Equation (20) approximates the maximum of the individual terms from below and approaches it as pp increases, so that a single differentiable constraint per band replaces the full set of mode-wise conditions while still being dominated by the most critical eigenfrequency. Since each term FBb,jFB_{b,j} is bounded between zero and unity, the aggregated quantity is bounded, which keeps the constraint well scaled throughout the optimization. The two parameters βFBC\beta_{FBC} and pp together determine how closely the smoothed constraint function reproduces the exact exclusion condition. Larger values of βFBC\beta_{FBC} steepen the transition at the band edges, and thereby reduce the conservatism introduced by the smoothing; larger values of pp make aggregated values closer to true maximum at the cost of increased non-linearity of the constraint (see § 6.2 for detailed parameters).

Topology optimization for vibration attenuation

Using the proposed interpolation scheme, we introduce three topology optimization formulations for vibration attenuation. The first formulation maximizes the fundamental eigenfrequency, the second maximizes the gap between two adjacent eigenfrequencies, and the third creates a frequency gap within a targeted frequency range.

Maximization of fundamental frequency

The first optimization problem considered in this work aims at finding the material distribution that maximizes the fundamental eigenfrequency of the structure. Such a design is desirable in low-frequency dynamic applications, as a high fundamental frequency shifts the structural response away from low-frequency excitation. The problem is a max-min problem over the initial J0J_0 eigenfrequencies given by Equation (22):

max⁡x min⁡ ωj(x)(j=1,…,J0)s.t. vmvm∗−1≤0(m=1,…,nMat)0≤xe(f)≤1,xe=[xe(f=1),xe(f=2),…].(22)\begin{aligned} \max_{\boldsymbol{x}}\ &\min\ \omega_j(\boldsymbol{x}) && (j=1,\ldots,J_0)\\ \mathrm{s.t.}\ &\frac{v_m}{v_m^*}-1\leq0 && (m=1,\ldots,n_{Mat})\\ &0\leq x_e^{(f)}\leq1,\quad x_e=\left[x_e^{(f=1)},x_e^{(f=2)},\ldots\right]. \tag*{(22)} \end{aligned}

where ωj(x)\omega_j(\boldsymbol{x}) is the jthj^{th} natural frequency, vmv_m is the volume fraction of the mthm^{th} material, and vm∗v_m^* is the prescribed upper bound. The volume fraction of each material is evaluated as the element-wise average of the corresponding material weight:

vm=1ne∑e=1newm(e);m=1,…,nMat,v_m=\frac{1}{n_e}\sum_{e=1}^{n_e}w_m(e);\quad m=1,\ldots,n_{Mat},

in which wm(e)w_m(e) denotes the weight of material mm in element ee, obtained from the multi-material interpolation scheme, and nen_e is the total number of finite elements in the design domain. We note that Equation (22) is not suited to gradient-based optimization, as the min⁡\min operator is non-differentiable. To overcome this difficulty, we adopt a bound formulation, in which an auxiliary scalar variable λ\lambda is introduced that acts simultaneously as the objective function and as an additional design variable, as in Equation (23):

max⁡x,λλs.t.λ−ωj2(x)λref≤0(j=1,…,J0)vmvm∗−1≤0(m=1,…,nMat)0≤xe(f)≤1,xe=[xe(f=1),xe(f=2),…].(23)\begin{aligned} \max_{\boldsymbol{x},\lambda}\quad&\lambda\\ \mathrm{s.t.}\quad&\frac{\lambda-\omega_j^2(\boldsymbol{x})}{\lambda_{ref}}\leq0 && (j=1,\ldots,J_0)\\ &\frac{v_m}{v_m^*}-1\leq0 && (m=1,\ldots,n_{Mat})\\ &0\leq x_e^{(f)}\leq1,\quad x_e=\left[x_e^{(f=1)},x_e^{(f=2)},\ldots\right]. \tag*{(23)} \end{aligned}

In this form, the non-smooth operator is replaced by a set of smooth inequality constraints, each of which forces one of the lowest eigenvalues to remain above the bound λ\lambda. Maximizing λ\lambda pushes the lower part of the spectrum upwards, and at convergence, at least one of these constraints is active, so that λ\lambda coincides with the smallest eigenvalue ω12\omega_1^2 where maximizing the lowest eigenvalue is equivalent to maximizing the fundamental frequency. The quantity λref\lambda_{ref} is a scaling factor that normalizes the eigenvalue constraints to the order of unity; it is taken as ω12\omega_1^2 evaluated at the first iteration and kept fixed thereafter.

Another advantage of this formulation is that it prevents mode switching during optimization. Because all lowest modes enter the problem through separate constraints, a change in the ordering of the modes between consecutive iterations does not introduce discontinuity in the objective function. The objective remains smooth, while the reordering is preserved by constraint sets. The optimization therefore remains stable even when repeated or closely spaced eigenvalues occur, where the original max-min statement would fail. Based on the bound variable, the design vector comprises all element-wise design fields of every material together with the scalar λ\lambda:

x=[{xe(1)}e=1ne,{xe(2)}e=1ne,…,λ].\boldsymbol{x}=\left[\left\{x_e^{(1)}\right\}_{e=1}^{n_e},\left\{x_e^{(2)}\right\}_{e=1}^{n_e},\ldots,\lambda\right].

Maximizing adjacent eigenfrequency gap

The second optimization formulation introduces a different design objective to generate a wide gap between two adjacent eigenfrequencies. Such a frequency gap is of practical interest because excitation frequencies falling inside the gap cannot be matched by any structural mode, so the design becomes effectively insensitive to a range of loading frequencies rather than to one.

max⁡x,λ1,λ2λ2−λ1s.t.ωj2λref−λ1≤0(j=1,…,nb)λ2−ωj2λref≤0(j=nb+1,…,J0)vmvm∗−1≤0(m=1,…,nMat)0≤xe(f)≤1,xe=[xe(f=1),xe(f=2),…]0≤λ1≤1,0≤λ2≤1.(24)\begin{aligned} \underset{\boldsymbol{x},\lambda_1,\lambda_2}{\operatorname{max}} \quad& \lambda_2-\lambda_1 \\ \text{s.t.} \quad& \frac{\omega_j^2}{\lambda_{\mathrm{ref}}}-\lambda_1\le0 \qquad(j=1,\ldots,n_b) \\ & \lambda_2-\frac{\omega_j^2}{\lambda_{\mathrm{ref}}}\le0 \qquad(j=n_b+1,\ldots,J_0) \\ & \frac{v_m}{v_m^*}-1\le0 \qquad(m=1,\ldots,n_{\mathrm{Mat}}) \\ & 0\le x_e^{(f)}\le1,\quad x_e=\left[x_e^{(f=1)},x_e^{(f=2)},\ldots\right] \\ & 0\le\lambda_1\le1,\quad0\le\lambda_2\le1. \tag*{(24)} \end{aligned}

Here, nbn_b denotes the index of the mode below which the gap is to be created, so that the modes j=1,…,nbj=1,\ldots,n_b constitute the lower eigenfrequency constraints and the modes j=nb+1,…,J0j=n_b+1,\ldots,J_0 constitute upper eigenfrequency constraints. Following the same reasoning as in the previous formulation, the non-smooth statement of maximizing ωnb+12−ωnb2\omega_{n_b+1}^2-\omega_{n_b}^2 is replaced by a smooth formulation in which two auxiliary scalar variables, i.e., λ1\lambda_1 and λ2\lambda_2, are introduced and treated simultaneously as parts of the objective and as design variables. The variable λ1\lambda_1 acts as an upper bound on the normalized eigenvalues of the lower eigenfrequency constraints, since the first set of constraints forces every ωj2/λref\omega_j^2/\lambda_{\mathrm{ref}} with j≤nbj\le n_b to lie below it, whereas λ2\lambda_2 acts as a lower bound on the normalized eigenvalues of the upper constraint, because the second set of constraints forces every ωj2/λref\omega_j^2/\lambda_{\mathrm{ref}} with j>nbj>n_b to lie above it. Maximizing the difference λ2−λ1\lambda_2-\lambda_1 simultaneously pushes the lower constraint downwards and the upper constraint upwards.

As in the maximization of fundamental frequency, the eigenvalues are normalized by λref\lambda_{\mathrm{ref}}, which keeps all eigenvalue constraints of the order of unity and permits the side constraints 0≤λ1≤10\le\lambda_1\le1 and 0≤λ2≤10\le\lambda_2\le1 to be imposed consistently. The remaining constraints include the volume fraction constraints, which limit the amount of each of the nMatn_{\mathrm{Mat}} materials, and the box constraints on the element-wise design variables. Consequently, the vector of design variables contains all element-wise design variables per material together with the two bound variables:

x=[{xe(1)}e=1ne,{xe(2)}e=1ne,…,λ1,λ2].\boldsymbol{x}=\left[\left\{x_e^{(1)}\right\}_{e=1}^{n_e},\left\{x_e^{(2)}\right\}_{e=1}^{n_e},\ldots,\lambda_1,\lambda_2\right].

Maximizing targeted frequency gap

The two aforementioned formulations do not allow the frequency gap to be positioned in a targeted frequency range. In many engineering applications, however, the excitation spectrum is known in advance. Our formulation in this section introduces frequency maximization with an explicit band constraint located at a prescribed interval as in Equation (25):

max⁡x,λλs.t.λ−ωj2λref≤0(j=1,…,J0)vmvm∗−1≤0(m=1,…,nMat)FBCb−12≤0(b=1,…,nM)0≤xe(f)≤1,xe=[xe(f=1),xe(f=2),…].(25)\begin{aligned} \underset{\boldsymbol{x},\lambda}{\operatorname{max}} \quad& \lambda\\ \text{s.t.} \quad& \lambda-\frac{\omega_j^2}{\lambda_{\mathrm{ref}}}\le0 \qquad(j=1,\ldots,J_0) \\ & \frac{v_m}{v_m^*}-1\le0 \qquad(m=1,\ldots,n_{\mathrm{Mat}}) \\ & FBC_b-\frac{1}{2}\le0 \qquad(b=1,\ldots,n_M) \\ & 0\le x_e^{(f)}\le1,\quad x_e=\left[x_e^{(f=1)},x_e^{(f=2)},\ldots\right]. \tag*{(25)} \end{aligned}

As in the maximization of fundamental frequency, the bound variable λ\lambda serves simultaneously as the objective function and as an additional design variable. The constraints force every normalized eigenvalue ωj2/λref\omega_j^2/\lambda_{\mathrm{ref}} of the J0J_0 modes to remain above it. Maximizing λ\lambda therefore lifts the lower part of the spectrum, while the normalization by λref\lambda_{\mathrm{ref}} keeps the eigenvalue constraints of the order of unity.

In the third constraint, FBCbFBC_b denotes the frequency-band constraint associated with the bthb^{\mathrm{th}} band target. This function is constructed such that it takes a value larger than one half whenever the corresponding eigenfrequency falls inside the target frequency range, and a value smaller than one half when it lies outside. Enforcing FBCb−1/2≤0FBC_b-1/2\le0 for all monitored modes expels the eigenfrequencies from the prescribed interval, and creates a gap located at the targeted spectrum. The design vector for this formulation collects the element-wise design fields of individual materials together with the bound variable.

Design sensitivity derivations

To solve our optimization with a gradient-based optimizer, the sensitivities of the objective and constraint functions need to be provided at each optimization iteration [48, 49, 50]. These sensitivities are derived analytically in the following.

Design sensitivity of material candidates

Sensitivities are obtained by the chain rule, so the gradients of the interpolation weights with respect to design variables need to be derived. We note that the index ff denotes a design field, 1≤f≤nMat−11\le f\le n_{\mathrm{Mat}}-1, and mm denotes a material, 2≤m≤nMat2\le m\le n_{\mathrm{Mat}}, while rr is the penalization exponent of the SIMP interpolation. Because each un-normalized weight wmw_m is a product of design variables raised to the power rr, its derivative with respect to a given variable is non-zero when the variable appears in the product, which leads to three distinct cases.

Sensitivity of the first material candidate

The unnormalized weight depends on the first design variable, and its gradient is given by Equation (26):

∂w1un∂ϕ1=rϕ1r−1(26)\frac{\partial w_1^{\mathrm{un}}}{\partial\phi_1}=r\phi_1^{r-1} \tag*{(26)}

Sensitivity of intermediate material candidate

The gradient for intermediate materials 2≤m≤nMat−12\le m\le n_{\mathrm{Mat}}-1, depends on two cases. The first case occurs when the design field index is lower than the material index, f<mf<m, the gradient follows Equation (27):

∂wmun∂ϕf=(∏j=1j≠fm−1(1−ϕjr))(−rϕfr−1)ϕmr(27)\frac{\partial\mathbf{w}^{un}_{m}}{\partial\phi_f}= \left(\prod_{\substack{j=1\\j\ne f}}^{m-1}(1-\phi_j^r)\right)(-r\phi_f^{r-1})\phi_m^r \tag*{(27)}

and the second case is that when the design variable index coincides with the material index, i.e., f=mf=m:

∂wmun∂ϕf=(∏j=1m−1(1−ϕjr))rϕmr−1.(28)\frac{\partial\mathbf{w}^{un}_{m}}{\partial\phi_f}= \left(\prod_{j=1}^{m-1}(1-\phi_j^r)\right)r\phi_m^{r-1}. \tag*{(28)}

Sensitivity of the last material candidate

For the last material, i.e., m=nMatm=n_{Mat}, the gradient of the unnormalized weight in tensor form is given by Equation (29):

∂wmun∂ϕf=(∏j=1j≠fm−1(1−ϕjr))(−rϕfr−1).(29)\frac{\partial\mathbf{w}^{un}_{m}}{\partial\phi_f}= \left(\prod_{\substack{j=1\\j\ne f}}^{m-1}(1-\phi_j^r)\right)(-r\phi_f^{r-1}). \tag*{(29)}

Since the weights used in the interpolation are normalized, the normalized gradient of weights is given by Equation (30):

∂wm∂ϕf=(∂wmun∂ϕf)∑m=1nMatwmun−wmun(∑m=1nMat∂wmun∂ϕf)(∑m=1nMatwmun)2.(30)\frac{\partial\mathbf{w}_m}{\partial\phi_f}= \frac{\left(\frac{\partial\mathbf{w}^{un}_m}{\partial\phi_f}\right)\sum_{m=1}^{nMat}\mathbf{w}^{un}_m-\mathbf{w}^{un}_m\left(\sum_{m=1}^{nMat}\frac{\partial\mathbf{w}^{un}_m}{\partial\phi_f}\right)} {\left(\sum_{m=1}^{nMat}\mathbf{w}^{un}_m\right)^2}. \tag*{(30)}

Sensitivity for fundamental frequency maximization

As the bound formulation introduces an auxiliary variable (λ\lambda) alongside the density design variables (x\mathbf{x}), the sensitivities of the objective and constraints must be derived with respect to both sets of variables.

Sensitivity of objective function

Our optimization is posed as a maximization, which is solved using the method of moving asymptotes (MMA). A negative sign is introduced to recast the objective in an equivalent minimization compatible with the MMA framework:

f0=−λ.(31)f_0=-\lambda. \tag*{(31)}

The sensitivity of the objective with respect to the design variables is given by:

∂f0∂xe=0;∂f0∂λ=−1.(32)\frac{\partial f_0}{\partial x_e}=0;\qquad\frac{\partial f_0}{\partial\lambda}=-1. \tag*{(32)}

Combined sensitivity with respect to design variables x\mathbf{x} and λ\lambda is given by:

∂f0∂xmma=[∂f0∂x, ∂f0∂λ].(33)\frac{\partial f_0}{\partial\mathbf{x}_{mma}}= \left[\frac{\partial f_0}{\partial\mathbf{x}},\ \frac{\partial f_0}{\partial\lambda}\right]. \tag*{(33)}

Sensitivity of eigenfrequency constraints

Our design variables are modified following the sequence of x→x~→ϕ→w→K,M\mathbf{x}\to\widetilde{\mathbf{x}}\to\boldsymbol{\phi}\to\mathbf{w}\to\mathbf{K},\mathbf{M}. We apply the chain rule for sensitivity calculation. After calculating sensitivity with respect to the projected design variables, density filtering is applied using the chain rule. For eigen mode indices j=1,…,J0j=1,\ldots,J_0; the eigenfrequency constraint is given by Equation (34):

gj=λ−ωj2λref≤0.(34)g_j=\frac{\lambda-\omega_j^2}{\lambda_{\mathrm{ref}}}\le0. \tag*{(34)}

Its sensitivity is given by:

∂gj∂xe=−1λref∂ωj2∂xe.(35)\frac{\partial g_j}{\partial x_e}=-\frac{1}{\lambda_{\mathrm{ref}}}\frac{\partial\omega_j^2}{\partial x_e}. \tag*{(35)}

As the sensitivity follows the chain rule, it is simplified as:

∂ωj2∂xe(f)=∂ωj2∂ϕe(f)∂ϕe(f)∂x~e(f)∂x~e(f)∂xe(f).(36)\frac{\partial\omega_j^2}{\partial x_e^{(f)}}= \frac{\partial\omega_j^2}{\partial\phi_e^{(f)}} \frac{\partial\phi_e^{(f)}}{\partial\widetilde{x}_e^{(f)}} \frac{\partial\widetilde{x}_e^{(f)}}{\partial x_e^{(f)}}. \tag*{(36)}

where the first term ∂ωj2/∂ϕe(f)\partial\omega_j^2/\partial\phi_e^{(f)} is further simplified:

∂ωj2∂ϕe(f)=dKe(j)∂Eeeff∂ϕe(f)+dMe(j)∂ρeeff∂ϕe(f),(37)\frac{\partial\omega_j^2}{\partial\phi_e^{(f)}}= \mathbf{d}K_e^{(j)}\frac{\partial E_e^{eff}}{\partial\phi_e^{(f)}}+ \mathbf{d}M_e^{(j)}\frac{\partial\rho_e^{eff}}{\partial\phi_e^{(f)}}, \tag*{(37)}

where we have dKe(j)=(ue(j))TKeue(j)\mathbf{d}K_e^{(j)}=(\mathbf{u}_e^{(j)})^{T}K_e\mathbf{u}_e^{(j)} and dMe(j)=−ωj2(ue(j))TMeue(j)\mathbf{d}M_e^{(j)}=-\omega_j^2(\mathbf{u}_e^{(j)})^{T}M_e\mathbf{u}_e^{(j)}.

Sensitivities of effective material properties are as:

∂Eeeff∂ϕe(f)=∑m=1nMatEm∂we,m∂ϕe(f);\frac{\partial E_e^{eff}}{\partial\phi_e^{(f)}}= \sum_{m=1}^{nMat}E_m\frac{\partial w_{e,m}}{\partial\phi_e^{(f)}};
∂ρeeff∂ϕe(f)=∑m=1nMatρm∂we,m∂ϕe(f).(38)\frac{\partial\rho_e^{eff}}{\partial\phi_e^{(f)}}= \sum_{m=1}^{nMat}\rho_m\frac{\partial w_{e,m}}{\partial\phi_e^{(f)}}. \tag*{(38)}

Projection derivative term ∂ϕe(f)/∂x~e(f)\partial\phi_e^{(f)}/\partial\widetilde{x}_e^{(f)} of Equation (36) is obtained from Equation (6). The term ∂x~e(f)/∂xe(f)\partial\widetilde{x}_e^{(f)}/\partial x_e^{(f)} can be calculated through density filtering as:

∂g~j∂x=H(∂gj∂x~1Hs).(39)\frac{\partial\widetilde{g}_j}{\partial\mathbf{x}}= \mathbf{H}\left(\frac{\partial g_j}{\partial\widetilde{\mathbf{x}}}\frac{1}{H_s}\right). \tag*{(39)}

Sensitivity with respect to the auxiliary variable λ\lambda is:

∂gj∂λ=1λref.(40)\frac{\partial g_j}{\partial\lambda}=\frac{1}{\lambda_{\mathrm{ref}}}. \tag*{(40)}

Final sensitivity is computed by combining Equation (39) and Equation (40) as in Equation (41). When computing sensitivities, filtering and Heaviside projection are applied only to the density design variables, whereas neither filtering nor projection is applied to the auxiliary variables. Sensitivity for the repeated-eigenvalue case is provided in § 5.8.

∂gj∂xmma=[∂gj∂x, ∂gj∂λ].(41)\frac{\partial g_j}{\partial\mathbf{x}_{mma}}= \left[\frac{\partial g_j}{\partial\mathbf{x}},\ \frac{\partial g_j}{\partial\lambda}\right]. \tag*{(41)}

Sensitivity of volume constraints

Sensitivity to volume constraints for each material is calculated from the material's volume fraction. For material m=1,…,nMatm=1,\ldots,n_{Mat}, the volume constraint is given by Equation (42):

gvol,m=vmvmtarget−1≤0.(42)g_{\mathrm{vol},m}=\frac{v_m}{v_m^{\mathrm{target}}}-1\leq0. \tag*{(42)}

The element-wise sensitivity of the volume constraint is given by (43):

∂gvol,m∂xe=1vmtarget∂vm∂xe,(43)\frac{\partial g_{\mathrm{vol},m}}{\partial x_e}=\frac{1}{v_m^{\mathrm{target}}}\frac{\partial v_m}{\partial x_e}, \tag*{(43)}

which is simplified as (44):

∂vm∂xe=1ne∂we,m∂xe.(44)\frac{\partial v_m}{\partial x_e}=\frac{1}{n_e}\frac{\partial w_{e,m}}{\partial x_e}. \tag*{(44)}

Applying chain rules for projection and density filtering to sensitivity gives (45):

∂we,m∂xe(f)=∑f=1nfields∂we,m∂ϕe(f)∂ϕe(f)∂x~e(f)∂x~e(f)∂xe(f).(45)\frac{\partial w_{e,m}}{\partial x_e^{(f)}}=\sum_{f=1}^{n_{\mathrm{fields}}}\frac{\partial w_{e,m}}{\partial\phi_e^{(f)}}\frac{\partial\phi_e^{(f)}}{\partial\widetilde{x}_e^{(f)}}\frac{\partial\widetilde{x}_e^{(f)}}{\partial x_e^{(f)}}. \tag*{(45)}

Sensitivity of volume constraint with respect to auxiliary variable λ\lambda is given in (46):

∂gvol,m∂λ=0.(46)\frac{\partial g_{\mathrm{vol},m}}{\partial\lambda}=0. \tag*{(46)}

Combining sensitivity with respect to both design variables yields (47):

∂gvol,m∂xmma=[∂gvol,m∂x,∂gvol,m∂λ].(47)\frac{\partial g_{\mathrm{vol},m}}{\partial\mathbf{x}_{mma}}=\left[\frac{\partial g_{\mathrm{vol},m}}{\partial\mathbf{x}},\frac{\partial g_{\mathrm{vol},m}}{\partial\lambda}\right]. \tag*{(47)}

Sensitivity for adjacent frequency gap maximization

For this formulation, sensitivities of the objective and constraints must be derived with respect to the density design variable (x\mathbf{x}) and two auxiliary variables (λ1\lambda_1 and λ2\lambda_2).

Sensitivity of the objective function

The objective indicates the separation between the two auxiliary variables and is independent of the density design variables, as given by (48):

f0=−(λ2−λ1).(48)f_0=-(\lambda_2-\lambda_1). \tag*{(48)}

The sensitivity of the objective function with respect to the respective design variables is given by (49):

∂f0∂xe=0;∂f0∂λ1=1;∂f0∂λ2=−1,(49)\frac{\partial f_0}{\partial x_e}=0;\quad\frac{\partial f_0}{\partial\lambda_1}=1;\quad\frac{\partial f_0}{\partial\lambda_2}=-1, \tag*{(49)}

where the vanishing derivative with respect to xex_e confirms that the material distribution is driven purely by the eigenfrequency bound constraints, whereas the two non-zero sensitivities push λ2\lambda_2 upward and λ1\lambda_1 downward, widening the gap. Collecting these terms in the order required by the optimizer yields the combined sensitivity:

∂f0∂xmma=[∂f0∂x,∂f0∂λ1,∂f0∂λ2].(50)\frac{\partial f_0}{\partial\mathbf{x}_{mma}}=\left[\frac{\partial f_0}{\partial\mathbf{x}},\frac{\partial f_0}{\partial\lambda_1},\frac{\partial f_0}{\partial\lambda_2}\right]. \tag*{(50)}

Sensitivity of lower eigenfrequency constraints

For eigen mode indices j=1,…,nbj=1,\ldots,n_b, the lower frequency constraint is given by (51):

gj(1)=ωj2λref−λ1≤0,(51)g_j^{(1)}=\frac{\omega_j^2}{\lambda_{\mathrm{ref}}}-\lambda_1\leq0, \tag*{(51)}

where the normalization by the reference eigenvalue λref\lambda_{\mathrm{ref}} keeps all constraint values of comparable magnitude. Differentiating (51) with respect to an element density gives:

∂gj(1)∂xe=1λref∂ωj2∂xe(52)\frac{\partial g_j^{(1)}}{\partial x_e}=\frac{1}{\lambda_{\mathrm{ref}}}\frac{\partial\omega_j^2}{\partial x_e} \tag*{(52)}

in which the eigenvalue sensitivity ∂ωj2/∂xe\partial\omega_j^2/\partial x_e is evaluated exactly as in (36). Sensitivity with respect to auxiliary variables is given by:

∂gj(1)∂λ1=−1;∂gj(1)∂λ2=0.(53)\frac{\partial g_j^{(1)}}{\partial\lambda_1}=-1;\quad\frac{\partial g_j^{(1)}}{\partial\lambda_2}=0. \tag*{(53)}

Final sensitivity is given by (54):

∂gj(1)∂xmma=[∂gj(1)∂x,∂gj(1)∂λ1,∂gj(1)∂λ2].(54)\frac{\partial g_j^{(1)}}{\partial\mathbf{x}_{mma}}=\left[\frac{\partial g_j^{(1)}}{\partial\mathbf{x}},\frac{\partial g_j^{(1)}}{\partial\lambda_1},\frac{\partial g_j^{(1)}}{\partial\lambda_2}\right]. \tag*{(54)}

Sensitivity of upper eigenfrequency constraints

For upper eigenmodes u=nb+1,…,J0u=n_b+1,\ldots,J_0, constraint is given by:

gu(2)=λ2−ωu2λref≤0.(55)g_u^{(2)}=\lambda_2-\frac{\omega_u^2}{\lambda_{\mathrm{ref}}}\leq0. \tag*{(55)}

Sensitivity of this upper bound is given by:

∂gu(2)∂xe=−1λref∂ωu2∂xe,(56)\frac{\partial g_u^{(2)}}{\partial x_e}=-\frac{1}{\lambda_{\mathrm{ref}}}\frac{\partial\omega_u^2}{\partial x_e}, \tag*{(56)}

where ∂ωu2/∂xe\partial\omega_u^2/\partial x_e is the same as (36). Sensitivity with respect to auxiliary variables is given by (57):

∂gu(2)∂λ1=0;∂gu(2)∂λ2=1.(57)\frac{\partial g_u^{(2)}}{\partial\lambda_1}=0;\quad\frac{\partial g_u^{(2)}}{\partial\lambda_2}=1. \tag*{(57)}

Final sensitivity is given by:

∂gu(2)∂xmma=[∂gu(2)∂x,∂gu(2)∂λ1,∂gu(2)∂λ2](58)\frac{\partial g_u^{(2)}}{\partial\mathbf{x}_{mma}}=\left[\frac{\partial g_u^{(2)}}{\partial\mathbf{x}},\frac{\partial g_u^{(2)}}{\partial\lambda_1},\frac{\partial g_u^{(2)}}{\partial\lambda_2}\right] \tag*{(58)}

Sensitivity of volume constraints

Sensitivity of the volume constraint for each material with respect to the density design variables, ∂gvol,m/∂x\partial g_{\mathrm{vol},m}/\partial\mathbf{x} is the same as (43). Sensitivity of volume constraints with respect to auxiliary variables λ1\lambda_1 and λ2\lambda_2 is given by (59):

∂gvol,m∂λ1=0;∂gvol,m∂λ2=0.(59)\frac{\partial g_{\mathrm{vol},m}}{\partial\lambda_1}=0;\quad\frac{\partial g_{\mathrm{vol},m}}{\partial\lambda_2}=0. \tag*{(59)}

Combining sensitivity with respect to all design variables yields:

∂gvol,m∂xmma=[∂gvol,m∂x,∂gvol,m∂λ1,∂gvol,m∂λ2].(60)\frac{\partial g_{\mathrm{vol},m}}{\partial\mathbf{x}_{mma}}=\left[\frac{\partial g_{\mathrm{vol},m}}{\partial\mathbf{x}},\frac{\partial g_{\mathrm{vol},m}}{\partial\lambda_1},\frac{\partial g_{\mathrm{vol},m}}{\partial\lambda_2}\right]. \tag*{(60)}

Sensitivity for targeted frequency gap

For this formulation, sensitivities of the objective function and constraints must be derived with respect to the density design variable (x\mathbf{x}) and one auxiliary variable (λ\lambda).

Sensitivity of the objective function

The objective function for this formulation is defined as:

f0=−λ.(61)f_0=-\lambda. \tag*{(61)}

The sensitivity of the objective function with respect to the respective design variables is given by:

∂f0∂xe=0;∂f0∂λ=−1.(62)\frac{\partial f_0}{\partial x_e}=0;\qquad\frac{\partial f_0}{\partial\lambda}=-1. \tag*{(62)}

Combined sensitivity with respect to both design variables x\mathbf{x} and λ\lambda is given by Equation (63):

∂f0∂xmma=[∂f0∂x,∂f0∂λ].(63)\frac{\partial f_0}{\partial\mathbf{x}_{mma}}=\left[\frac{\partial f_0}{\partial\mathbf{x}},\frac{\partial f_0}{\partial\lambda}\right]. \tag*{(63)}

Sensitivity of eigenfrequency constraints

For eigen mode indices j=1,…,J0j=1,\ldots,J_0, the eigenfrequency constraint is given by Equation (64):

gj=λ−ωj2λref≤0.(64)g_j=\lambda-\frac{\omega_j^2}{\lambda_{ref}}\leq0. \tag*{(64)}

The sensitivity of this constraint is given by Equation (65):

∂gj(1)∂xe=−1λref∂ωj2∂xe.(65)\frac{\partial g_j^{(1)}}{\partial x_e}=-\frac{1}{\lambda_{ref}}\frac{\partial\omega_j^2}{\partial x_e}. \tag*{(65)}

where ∂ωj2/∂xe\partial\omega_j^2/\partial x_e is the same as Equation (36). Sensitivity with respect to auxiliary variables is given by:

∂gj(1)∂λ=1.(66)\frac{\partial g_j^{(1)}}{\partial\lambda}=1. \tag*{(66)}

Final sensitivity is given by:

∂gj(1)∂xmma=[∂gj(1)∂x,∂gj(1)∂λ].(67)\frac{\partial g_j^{(1)}}{\partial\mathbf{x}_{mma}}=\left[\frac{\partial g_j^{(1)}}{\partial\mathbf{x}},\frac{\partial g_j^{(1)}}{\partial\lambda}\right]. \tag*{(67)}

The sensitivity of the volume constraint given in Equation (47) is the same for both the fundamental frequency maximization and the prescribed frequency gap formulations, as they use the same design variables.

Sensitivity for frequency band constraint

The frequency band constraint is imposed, so that no eigenfrequency of the structure is permitted inside a prescribed band. For the bthb^{th} band, the constraint is expressed through a rational polynomial function, and is written as:

gj(3)=FBCb−12≤0;b=1,…,nM.(68)g_j^{(3)}=FBC_b-\frac{1}{2}\leq0;\qquad b=1,\ldots,n_M. \tag*{(68)}

Sensitivity of the frequency band constraint is given by:

∂gj(3)∂x=∂FBCb∂x.(69)\frac{\partial g_j^{(3)}}{\partial\mathbf{x}}=\frac{\partial FBC_b}{\partial\mathbf{x}}. \tag*{(69)}

where it is further simplified:

∂FBCb∂x=FBCb1−pJ0∑j=1J0FBb,jp−1∂FBb,j∂ωj∂ωj∂x;∂FBb,j∂ωj=−2∗βFBb,j2ωd(ωj−ωcωd)2β−1.(70)\frac{\partial FBC_b}{\partial\mathbf{x}}=\frac{FBC_b^{1-p}}{J_0}\sum_{j=1}^{J_0}F B_{b,j}^{p-1}\frac{\partial F B_{b,j}}{\partial\omega_j}\frac{\partial\omega_j}{\partial\mathbf{x}};\qquad\frac{\partial F B_{b,j}}{\partial\omega_j}=-2*\beta\frac{F B_{b,j}^{2}}{\omega_d}\left(\frac{\omega_j-\omega_c}{\omega_d}\right)^{2\beta-1}. \tag*{(70)}

where the terms FBCbFBC_b and FBb,jF B_{b,j} are obtained from Equations (20) and (21), respectively. The term ∂ωj/∂x\partial\omega_j/\partial\mathbf{x} in Equation (70) is expanded as:

∂ωj∂xe(f)=12ωj∂ωj2∂xe(f).(71)\frac{\partial\omega_j}{\partial x_e^{(f)}}=\frac{1}{2\omega_j}\frac{\partial\omega_j^2}{\partial x_e^{(f)}}. \tag*{(71)}

where ∂ωj2/∂xe\partial\omega_j^2/\partial x_e is the same as Equation (36). The chain rule for projection and sensitivity is similar to the maximization of fundamental frequency. Sensitivity with respect to auxiliary variables is given by:

∂gj(3)∂λ=0,(72)\frac{\partial g_j^{(3)}}{\partial\lambda}=0, \tag*{(72)}

The final sensitivity is given by:

∂gj(3)∂xmma=[∂gj(3)∂x,∂gj(3)∂λ].(73)\frac{\partial g_j^{(3)}}{\partial\mathbf{x}_{mma}}=\left[\frac{\partial g_j^{(3)}}{\partial\mathbf{x}},\frac{\partial g_j^{(3)}}{\partial\lambda}\right]. \tag*{(73)}

Sensitivity for repeated eigenvalues

Mode tracking is generally needed when the optimizer drives multiple eigenfrequencies close. In such scenarios, the standard sensitivity derived for distinct eigenvalues loses validity, as the sensitivity of an eigenvalue would have multiple distinct values (due to the different eigenvectors), as seen in Equation (37). In fact, it is only directionally differentiable, and its derivatives can be extracted from the subspace spanned by associated eigenvectors, as discussed in prior studies [28, 32].

Let a repeated eigenvalue of multiplicity mcm_c be denoted by ωj12=ωj22=⋯=ωjmc2=ωˉ2\omega_{j_1}^2=\omega_{j_2}^2=\cdots=\omega_{j_{m_c}}^2=\bar{\omega}^2, where ωˉ2\bar{\omega}^2 is the value of the coincident eigenvalues. Since exact coincidence is generally not attained numerically, two eigenvalues are treated as repeated in the present work whenever they satisfy the closeness criterion ∣(ωj+12−ωj2)/ωj2∣<10−4\left|(\omega_{j+1}^2-\omega_j^2)/\omega_j^2\right|<10^{-4}. This measure is preferred over an absolute one so that the same threshold remains meaningful across the entire spectrum, irrespective of the magnitude of the eigenvalues being compared.

Once a cluster of mcm_c coincident eigenvalues has been identified, the directional derivatives are obtained by projecting elemental stiffness and mass operators onto corresponding eigenvector subspaces, where the elemental stiffness and mass perturbation matrices are assembled as:

(DK(e))ab=(ua(e))TKeub(e);\left(\mathbf{D}_{K}^{(e)}\right)_{ab}=\left(\mathbf{u}_a^{(e)}\right)^{\mathsf{T}}\mathbf{K}_e\mathbf{u}_b^{(e)};
(DM(e))ab=(ua(e))TMeub(e),for a,b=1,…,mc,(74)\left(\mathbf{D}_{M}^{(e)}\right)_{ab}=\left(\mathbf{u}_a^{(e)}\right)^{\mathsf{T}}\mathbf{M}_e\mathbf{u}_b^{(e)},\qquad\text{for }a,b=1,\ldots,m_c, \tag*{(74)}

in which ua(e)\mathbf{u}_a^{(e)} and ub(e)\mathbf{u}_b^{(e)} are the restrictions of the atha^{th} and bthb^{th} eigenvectors of the repeated set to the degrees of freedom of element ee, and Ke\mathbf{K}_e and Me\mathbf{M}_e are the corresponding elemental stiffness and mass matrices. Both DK(e)\mathbf{D}_{K}^{(e)} and DM(e)\mathbf{D}_{M}^{(e)} are symmetric matrices of order mcm_c, and they describe how a perturbation of element ee redistributes energy within the degenerate subspace. Combining the two elemental perturbation matrices gives the perturbation matrix associated with the repeated eigenvalue:

D(e)=DK(e)−ωˉ2DM(e).(75)\mathbf{D}^{(e)}=\mathbf{D}_{K}^{(e)}-\bar{\omega}^{2}\mathbf{D}_{M}^{(e)}. \tag*{(75)}

We note that we can compute the eigenvectors pk\mathbf{p}_k (k=1,…,mck = 1, \ldots, m_c) from the reduced D(e)D^{(e)} matrix that spans the subspace of repeated eigenvalues, and pk\mathbf{p}_k define the directions within the degenerate subspace along which repeated eigenvalues split under an infinitesimal design perturbation. The vectors pk\mathbf{p}_k restore a well-defined, unique set of directional derivatives.

The sensitivity of each eigenvalue of the cluster with respect to the projected design variable follows the chain rule through the effective material properties:

∂ωk2∂ϕe(f)=∂ωk2∂Eeeff∂Eeeff∂ϕe(f)+∂ωk2∂ρeeff∂ρeeff∂ϕe(f).(76)\frac{\partial\omega_k^2}{\partial\phi_e^{(f)}} = \frac{\partial\omega_k^2}{\partial E_e^{\mathrm{eff}}} \frac{\partial E_e^{\mathrm{eff}}}{\partial\phi_e^{(f)}} + \frac{\partial\omega_k^2}{\partial\rho_e^{\mathrm{eff}}} \frac{\partial\rho_e^{\mathrm{eff}}}{\partial\phi_e^{(f)}}. \tag*{(76)}

where the stiffness and mass contributions are given by:

∂ωk2∂Eeeff=pkTDK(e)pk;(77)\frac{\partial\omega_k^2}{\partial E_e^{\mathrm{eff}}} = \mathbf{p}_k^{T}D_K^{(e)}\mathbf{p}_k; \tag*{(77)}
∂ωk2∂ρeeff=−ωˉ2pkTDM(e)pk,\frac{\partial\omega_k^2}{\partial\rho_e^{\mathrm{eff}}} = -\bar{\omega}^{2}\mathbf{p}_k^{T}D_M^{(e)}\mathbf{p}_k,

Finally, the sensitivity is propagated backwards through the projection and filtering operations to compute physical design variables as:

∂ωk2∂xe(f)=H(∂ωk2∂ϕe(f)∂ϕe(f)∂x~e(f)1Hs).(78)\frac{\partial\omega_k^2}{\partial x_e^{(f)}} = H\left( \frac{\partial\omega_k^2}{\partial\phi_e^{(f)}} \frac{\partial\phi_e^{(f)}}{\partial\tilde{x}_e^{(f)}} \frac{1}{H_s} \right). \tag*{(78)}

For mc=1m_c = 1 the perturbation matrices reduce to scalars, pk\mathbf{p}_k becomes unity, and Equations (76)–(77) collapse exactly to the classical sensitivity expression for a simple eigenvalue. Treating repeated eigenvalues in this manner suppresses oscillatory behavior and mode-switching that otherwise occur when eigenfrequencies cross during optimization iterations and thereby improves convergence stability.

Numerical implementations

Our proposed multi-material topology optimization is implemented in three optimization formulations and validated by multiple numerical experiments, including maximization of the fundamental frequency, maximization of the gap between two adjacent eigenfrequencies, and the creation of frequency gaps within a targeted frequency range. The MATLAB implementation was developed with reference to the 99-line [43] and 88-line [51] topology optimization codes and univariate mapping-based code [36], with necessary modifications to accommodate the proposed multi-material interpolation scheme and eigenfrequency-based optimization formulations. All the numerical studies were implemented in MATLAB R2025b and executed on a workstation equipped with an Intel® Core™ i7-11700K CPU @ 2.50 GHz and 16 GB of RAM.

Material properties

We define the material properties in Table 1 for the sets of three, four, five, and seven candidate materials, respectively. Young’s modulus and mass density are uniformly spaced over E∈[1,10]E \in[1, 10] MPa and ρ∈[0.1,1]\rho\in[0.1, 1] kgm−3^{-3}, respectively, and Poisson’s ratio is fixed at ν=0.3\nu= 0.3. In all four material sets, we assume that the stiffest material has the highest density, and the most compliant material has the lowest density. We note that these two extreme materials have the same stiffness-to-density ratio, i.e., E/ρ=10E/\rho= 10 MPa m3^3kg−1^{-1} in every example. Because materials of the extreme properties are identical across experiments, differences in optimized designs can be attributed to the number and distribution of the materials of intermediate properties which have a lower stiffness-to-density ratio than the extreme materials. We note that the properties listed in the table serve as numerical test cases and do not represent real engineering materials.

Number of materialsMaterial propertiesMaterial1 (M1)Material2 (M2)Material3 (M3)Material4 (M4)Material5 (M5)Material6 (M6)Material7 (M7)
Three materialsModulus of Elasticity (MPa)105.0501----
Density (Kg/m³)10.5500.100----
Four materialsModulus of Elasticity (MPa)106.7003.4001---
Density (Kg/m³)10.7000.4000.100---
Five materialsModulus of Elasticity (MPa)107.5255.0502.5751--
Density (Kg/m³)10.7750.550.3250.100--
Seven materialsModulus of Elasticity (MPa)108.3506.7005.0503.4001.7501
Density (Kg/m³)10.8500.7000.5500.4000.2500.100

Table 1. Material properties of candidate materials (Poisson's ratio ν=0.3\nu= 0.3 for all materials)

Experiment parameter setup

For implementation, the following parameters are adopted to obtain well-defined optimized topologies.

  • Penalization power: For all numerical examples, the penalization powers for both the stiffness and mass matrices are initially set to 1.0 and are subsequently updated every 25 iterations by a factor of 1.2 until reaching a maximum value of 5.0.

If mod⁡(loop⁡,25)=0\operatorname{mod}(\operatorname{loop}, 25) = 0, update r←min⁡(1.2∗p,5)r \leftarrow\min(1.2 \ast p, 5).

  • Initial design variable: For all examples, the initial value of design variables per element is set to 0.5. For example, the initial value for the three-material example is set as xe=[0.5,0.5]x_e = [0.5, 0.5] and for four-material as xe=[0.5,0.5,0.5]x_e = [0.5, 0.5, 0.5].

  • Scalar variable λ\lambda: For the pinned beam example, the scalar variable λ\lambda in the maximization of fundamental frequency (maximization of fundamental frequency) is initialized to 5000 with bounds defined as [0,106][0, 10^6]. In the maximization of the gap between adjacent eigenfrequencies, the scalar variables λ1\lambda_1 and λ2\lambda_2 are initialized with 0.4 and 0.6, respectively, with bounds [0,1][0, 1]. In frequency gap generation between the targeted range, the scalar λ\lambda is initialized to 0.5 with bounds of [0,1][0, 1]. The scalar variable used for the square plate and C-clamp examples is described in Sections 7.2 and § 8, respectively.

  • Filter radius: For the pinned beam example, the filter radius in the maximization of fundamental frequency is set to 8.0 for the five-material case and 7.0 for the seven-material design case. In maximizing the gap between adjacent eigenfrequencies, the filter radius is set to 7.0 for both the three-material and four-material cases. For creating a gap between the targeted frequency range, the filter radius is chosen as 7.0 for both single and double FBC. To obtain optimized topologies, different filter radii are adopted. The filter radius used for the square plate and C-clamp examples is described in Sections 7.2 and § 8, respectively.

  • Projection parameter: For the pinned beam example, the projection parameter (βf\beta_f) is initially set to 1.0 for all three formulations. In the maximization of fundamental frequency, for the five-material design case, the projection parameter is increased by a factor of 3 to 32, then by a factor of 1.5 every 25 iterations until reaching 300, i.e.,

If βf≤32, update βf←min⁡(3βf,32),\text{If }\beta_f \le32,\ \text{update }\beta_f \leftarrow\min(3\beta_f, 32),
else min⁡(1.5βf,300).\text{else }\min(1.5\beta_f, 300).

For the seven-material design case, the projection parameter is increased by a factor of 1.5 every 25 iterations, reaching a maximum of 300. For maximization of the gap between adjacent eigenfrequencies, the projection parameter is increased by a factor of 1.5 to 32, then by a factor of 3 every 25 iterations until reaching a maximum value of 300. To create a gap between the targeted range, it is increased by 1.5 times every 25 iterations until reaching a maximum value of 300. The projection parameter update schemes adopted for the square plate and C-clamp examples are described in Sections 7.2 and § 8, respectively.

  • FBC steepness parameter: For all three numerical examples, to apply the frequency band constraint, the projection parameter (βFBC\beta_{FBC}) of the Heaviside function is increased by a factor of 1 every 25 iterations until it reaches a maximum of 32, i.e., if mod⁡(loop⁡,25)=0\operatorname{mod}(\operatorname{loop}, 25) = 0, update βFBC←min⁡(βFBC+1,32)\beta_{FBC} \leftarrow\min(\beta_{FBC} + 1, 32).

  • Projection threshold: For all numerical examples, the threshold value of the projection function is taken as 0.5.

  • P-norm parameter: For all numerical examples, the parameter in the p-norm approximation for the maximum functions is set to 40.

  • Number of eigenfrequency modes: For all examples, the number of eigenmodes considered in the analysis is 6.

  • Volume fractions: For maximization of both fundamental frequency and gap between adjacent eigenfrequencies, the volume fraction of the stiffest material is set to 0.3, while the remaining volume is distributed equally among the other materials. To create a gap between the targeted frequency range, the volume fractions of all materials are identical.

  • Stopping criteria: For all examples, the optimization is terminated once the objective function shows no further change between successive iterations and the continuation parameter has reached its maximum value.

Optimization algorithm

We adopt MMA [52] in our topology optimization. The complete procedure is summarized in Algorithm 1.

Algorithm 1: Multi-material topology optimization for dynamic vibration attenuation

  1. Inputs: number of candidate materials nmn_m; material data {Em,νm,ρm}\{E_m, \nu_m, \rho_m\}, m=1,…,nmm = 1, \ldots, n_m; mesh data; number of eigenmodes to compute nbn_b; maximum number of iterations Nmax⁡N_{\max}.

  2. Output: optimized design variables x\mathbf{x}.

  3. Initialization: x←x0\mathbf{x} \leftarrow\mathbf{x}_0; loop ←0\leftarrow0;

  4. rmin⁡←rmin⁡,0r_{\min} \leftarrow r_{\min,0}, β←β0\beta\leftarrow\beta_0, p←p0p \leftarrow p_0 ▷ filter radius, projection and P-norm parameters

  5. initialize MMA parameters (xloop−2,xloop−1,L,U)(\mathbf{x}^{loop-2}, \mathbf{x}^{loop-1}, \mathbf{L}, \mathbf{U})

  6. while loop <Nmax⁡< N_{\max} do

  7. x←{xe(f=1),xe(f=2),…,xe(f=nm−1)}\mathbf{x} \leftarrow\{x_e^{(f=1)}, x_e^{(f=2)}, \ldots, x_e^{(f=n_m-1)}\} ▷ element-wise design fields

  8. x~←Filter⁡(x,rmin⁡)\widetilde{\mathbf{x}} \leftarrow\operatorname{Filter}(\mathbf{x}, r_{\min})

  9. ϕ←Projection⁡(x~,β,η)\boldsymbol{\phi} \leftarrow\operatorname{Projection}(\widetilde{\mathbf{x}}, \beta, \eta)

  10. w←weight⁡(ϕ)\mathbf{w} \leftarrow\operatorname{weight}(\boldsymbol{\phi})

  11. K(x),M(x)←assemble⁡(w)\mathbf{K}(\mathbf{x}), \mathbf{M}(\mathbf{x}) \leftarrow\operatorname{assemble}(\mathbf{w}) ▷ finite element analysis

  12. solve [K(x)−ωn2M(x)]un=0[\mathbf{K}(\mathbf{x}) - \omega_n^2 \mathbf{M}(\mathbf{x})]\mathbf{u}_n = 0, n=1,…,nbn = 1, \ldots, n_b

  13. evaluate objective and sensitivity: f0(x),∂f0/∂xf_0(\mathbf{x}), \partial f_0/\partial\mathbf{x}

  14. evaluate constraints and sensitivities: f(x),∂f/∂x\mathbf{f}(\mathbf{x}), \partial\mathbf{f}/\partial\mathbf{x}

  15. xmma←MMA⁡(x,f0(x),f(x),∂f0/∂x,∂f/∂x)\mathbf{x}_{mma} \leftarrow\operatorname{MMA}(\mathbf{x}, f_0(\mathbf{x}), \mathbf{f}(\mathbf{x}), \partial f_0/\partial\mathbf{x}, \partial\mathbf{f}/\partial\mathbf{x})

  16. x←xmma\mathbf{x} \leftarrow\mathbf{x}_{mma} ▷ update design variables

  17. (β,p,rmin⁡)←update⁡(β,p,rmin⁡)(\beta, p, r_{\min}) \leftarrow\operatorname{update}(\beta, p, r_{\min}) ▷ continuation scheme

  18. If loop =Nmax⁡= N_{\max} & f0loop−1≈f0loopf_0^{loop-1} \approx f_0^{loop} then

  19. converged ←\leftarrow true ▷ check convergence

  20. end

  21. loop ←\leftarrow loop + 1

  22. end while

  23. return x\mathbf{x}

Numerical examples

To validate the proposed method, three numerical examples are presented, including a pinned vibrating beam, a square plate, and a C-clamp. For all examples, the candidate materials and their corresponding properties, including Young's modulus, density, and Poisson's ratio, are listed in Table 1.

Vibrating beam

In this study, a two-dimensional beam with dimensions 8 m×1 m8\ \mathrm{m} \times1\ \mathrm{m} is considered (see Figure 3), where the midpoint of both ends of the beam is fixed. The design domain is discretized into 400×50400 \times50 finite elements.

Figure 3. Boundary condition of the pinned vibrating beam

Maximum fundamental frequency

In this example, the objective is to maximize the fundamental eigenfrequency. The evolution of the fundamental frequency over the optimization by five material candidates is shown in Figure 4(c). The maximum fundamental eigenfrequency attained is 167 rad/s. The corresponding material distribution is presented in Figure 4(d), where the color bar denotes the materials ordered from the stiffest and densest (M1) to the least stiff and dense (M5).

Comparison of fundamental eigenfrequency maximization with five candidate materials, showing convergence histories and optimized multi-material designs

Figure 4. Comparison of fundamental eigenfrequency maximization with five candidate materials. (a) Convergence history of the fundamental eigenfrequency of the ordered SIMP interpolation, (b) optimized multi-material design, (c) convergence history of the fundamental eigenfrequency of the proposed multivariate weight-based interpolation, and (d) optimized multi-material design.

To assess the performance of the proposed framework, results are compared with those obtained using the ordered SIMP scheme [35]. The ordered SIMP interpolation was originally developed for compliance minimization problems; however, for a fair comparison, it is extended here to the eigenfrequency maximization problem. The convergence history and the optimized topologies are shown in Figure 4. Both methods converge rapidly and yield similar layered material distributions (Figure 4(b) and Figure 4(d)), while our method achieves a slightly higher fundamental eigenfrequency of 167 rad/s, compared with 166 rad/s for the ordered SIMP method as seen in Figure 4(a) and Figure 4(c).

The beam is further optimized using seven candidate materials. As shown in Figure 5(a), the convergence is similar to that of the five-material case, with a maximum fundamental eigenfrequency of 166 rad/s. This indicates that increasing the number of candidate materials does not necessarily result in a higher fundamental eigenfrequency. The material distribution is presented in Figure 5(b).

Fundamental eigenfrequency maximization with seven candidate materials, showing convergence and a multi-material design

Figure 5. Fundamental eigenfrequency maximization with seven candidate materials: (a) convergence of the fundamental eigenfrequency and (b) multi-material design.

We note that the maximization of the fundamental eigenfrequency is governed by the Rayleigh quotient, i.e., ω2=vTKvvTMv\omega^2 = \frac{\mathbf{v}^{T}\mathbf{K}\mathbf{v}}{\mathbf{v}^{T}\mathbf{M}\mathbf{v}}, in which the numerator represents the modal strain energy and the denominator represents the modal kinetic energy. The eigenvalue sensitivity in Equation (79) with respect to the design variable indicates that the increase in stiffness raises the eigenvalue if the resulting gain in strain energy exceeds ω2\omega^2 times the accompanying gain in kinetic energy.

∂ω2∂x=vT(∂K∂x−ω2∂M∂x)v.(79)\frac{\partial\omega^2}{\partial x}=\mathbf{v}^{T}\left(\frac{\partial K}{\partial x}-\omega^2\frac{\partial M}{\partial x}\right)\mathbf{v}. \tag*{(79)}

We find that the optimization tends to assign the stiffest and densest material to regions with high modal strain energy. Conversely, the least stiff and dense material is retained in the interior region, where reducing mass is more beneficial than increasing stiffness.

Gap maximization between adjacent eigenfrequencies

In this formulation, the objective is to maximize the frequency gap between the third and fourth eigenfrequencies. Figure 6(a) and Figure 7(a) present the evolution of the two bounding eigenfrequencies over the optimization history for the three- and four-material cases, respectively. In both cases, the gap widens, i.e., the third eigenfrequency moves downward while the fourth is driven upward. The optimized gaps obtained are 847 rad/s for the three-material design and 800 rad/s for the four-material design, as shown in Figure 6(b) and Figure 7(b). The topologies differ from those obtained in the preceding formulation. In particular, the stiffest and densest material is no longer concentrated in the regions adjacent to the support regions. The difference stems from the objective, which is governed not by a single eigenmode, but by the relative positions of two adjacent modes. Material layout is dictated by the difference between the modal energy distributions of the third and fourth modes.

Figure 6. Gap maximization of adjacent eigenfrequencies with three candidate materials: (a) convergence history of the eigenfrequency gap, and (b) multi-material design.

Convergence history and optimized multi-material design for gap maximization with four candidate materials

Figure 7. Gap maximization of adjacent eigenfrequencies with four candidate materials: (a) convergence history of the eigenfrequency gap, and (b) optimized multi-material design.

We show the first six mode shapes of the four-material design in Figure 8, where colorbars represent mass-normalized displacement magnitudes. We note that modes ω1\omega_1, ω2\omega_2, ω4\omega_4, ω5\omega_5 and ω6\omega_6 are transverse bending modes, and mode ω3\omega_3 is a longitudinal (axial) mode, in which the beam stays essentially straight and moves along its axis. Unlike the global bending modes ω1\omega_1, ω2\omega_2 and ω4\omega_4, modes ω5\omega_5 and ω6\omega_6 are localized near the supports, where the end segments deflect significantly.

Six vibrational eigenmodes of the four-material pinned beam

Figure 8. Vibrational eigenmodes: First six eigenmodes of the four-material pinned beam obtained from eigenfrequency-gap maximization in § 7.2.

Creating a gap between targeted frequency range

Unlike the preceding formulations, in which the target eigenfrequencies are identified by their modal index, this formulation imposes a frequency-band constraint that prevents any natural frequency of the structure from lying within a user-specified operating range, thereby creating a gap over that range. The proposed rational polynomial function is employed to approximate the Heaviside function to avoid non-differentiability in the FBC formulation for gradient-based optimization. Two cases are considered. The first involves a single operating range of [500, 1000] rad/s, and the second considers two different operating ranges of [300, 500] rad/s and [700, 1000] rad/s. In both cases, the structure must be designed such that none of its natural frequencies lies within the prescribed intervals.

Design for a single frequency range

We start by designing with three candidate materials, such that all natural frequencies are outside the prescribed frequency interval of [500, 1000] rad/s. The resulting eigenfrequencies are reported in Table 2, the convergence history in Figure 9(a), and the corresponding material distribution in Figure 9(b), respectively. The attained gap slightly exceeds the prescribed interval, which is attributable to the approximation error of the pp-norm aggregation.

Eigenmode index (ii)123456
Single FBC, ωi\omega_i (rad/s)1333351000100110861386
Double FBC, ωi\omega_i (rad/s)1332586056297001011

Table 2. First six eigenfrequencies (ωi\omega_i) of the three-material beam by excluding natural frequencies from a single targeted frequency range of [500, 1000] rad/s and from two targeted frequency ranges of [300, 500] and [700, 1000] rad/s

Convergence of the frequency gap

(a)

Optimized multi-material design

(b)

Figure 9. Targeted frequency gap for a three-material beam: (a) convergence of the frequency gap and (b) optimized multi-material design.

::::

(c) Two targeted frequency gaps: (a) convergence of the frequency gap, and (b) optimized multi-material design.

::::

The convergence history shows minor variation in the eigenfrequencies during the initial iterations, followed by a sharp drop in eigenfrequencies at the 150th iteration, caused by the update of the continuation parameters; thereafter, the spectrum stabilizes rapidly and satisfies the frequency-band constraints for both intervals. It is noted that the transition occurs earlier than in a single frequency band, which is consistent with the additional constraint acting to drive the design away from the initial configuration at an earlier stage of the optimization.

Square plate design {#sec:7.2}

A two-dimensional square plate with dimensions of 1 m × 1 m is shown in Figure 11, with all four corner points fixed. The design domain is discretized using 200×200 finite elements. The continuation for the projection parameter βf\beta_f is adjusted to obtain optimized topologies with minimal intermediate-density regions. The initial value of βf\beta_f is set to 2 for the maximization of fundamental frequency and 1 for maximizing the gap between adjacent eigenfrequencies and creating a gap between the targeted eigenfrequency range. In the maximization of fundamental frequency, βf\beta_f is increased by a factor of 3 until it reaches 32, and thereafter by a factor of 1.5 every 25 iterations up to a maximum value of 300. In maximizing gap between adjacent eigenfrequencies, βf\beta_f is increased by a factor of 1.5 every 25 iterations until it reaches 300. In creating a gap between the targeted frequency range, βf\beta_f is increased by a factor of 1.5 until it reaches 32, and then by a factor of 3 every 25 iterations up to 300. In addition, the filter radius is set to 8 for all three formulations. The scalar variable λ\lambda in the maximization of fundamental frequency is initialized to 5000, with bounds [0,108][0, 10^{8}]. In maximizing the gap between adjacent eigenfrequencies, scalar variables λ1\lambda_1 and λ2\lambda_2 are initialized with 0.4 and 0.6, respectively, with bounds [0,1][0, 1]. To create a gap between the targeted range, the variable λ\lambda is initialized to 0.5 with bounds of [0,1][0, 1].

(f) Square plate: boundary condition and design domain.

Maximum fundamental frequency {#sec:7.2.1}

In this numerical example, the fundamental eigenfrequency of a two-dimensional square plate is maximized, and the optimized material distribution for the five-material case is presented in Figure 12(b). As in the pinned beam example, the stiffest and densest material concentrates near the supports. The convergence history is shown in Figure 12(a).

Figure 12. Fundamental eigenfrequency maximization with five candidate materials: (a) convergence history of the fundamental eigenfrequency and (b) optimized multi-material design.

Design for two frequency ranges

In this case, the structure is designed to create frequency gaps over two disjoint operating ranges, [300,500][300, 500] rad/s and [700,1000][700, 1000] rad/s. The frequency-band constraint is accordingly extended to act on both intervals, so that no eigenfrequency is admitted into either gap. The resulting eigenfrequencies for the three-material design are reported in Table 2, the convergence in Figure 10(a), and material distributions in Figure 10(b), respectively.

Maximum gap between adjacent eigenfrequencies

In this numerical example, gap maximization between adjacent eigenfrequencies is applied to the square plate to maximize the gap between the third and fourth eigenfrequencies, and a maximized frequency gap of 8,499 rad/s is obtained for the five-material case. The convergence history in Figure 13(a) shows that the gap is opened through a reduction of the lower eigenfrequencies. The first three eigenfrequencies decrease gradually up to approximately the 150th iteration, undergo a pronounced drop thereafter, and subsequently stabilize at values close to the lower bound of the spectrum, whereas the fourth and higher eigenfrequencies vary only slightly throughout the optimization. The material distribution is presented in Figure 13(b), in which the candidate materials are placed about the center of the domain.

Convergence history of the eigenfrequency gap and optimized multi-material design

Figure 13. Maximizing the gap between adjacent eigenfrequencies with five candidate materials: (a) convergence history of the eigenfrequency gap, and (b) optimized multi-material design.

Creating the gap between targeted frequency range

In this study, a single operating frequency range of [2000,10000][2000, 10000] rad/s is considered for the square plate. The resulting eigenfrequencies for the five-material design are reported in Table 3, and the convergence history is shown in Figure 14(a), from which a clear frequency gap is observed. The corresponding optimized material distribution is presented in Figure 14(b), where materials form concentric layers ordered from the stiffest material at the center to the least dense one at the boundary.

Mode (ii)123456
ωi\omega_i (rad/s)18041804.518059992999410799

Table 3. First six eigenfrequencies (ωi\omega_i) of five-material square plate design by excluding natural frequencies from a single targeted frequency range of [2000,10000][2000, 10000] rad/s

Targeted frequency gap with five candidate materials

Figure 14. Targeted frequency gap with five candidate materials: (a) convergence history of the frequency gap and (b) optimized multi-material design.

C-clamp design

We design a C-clamp of non-rectangular geometry in this section, as shown in Figure 15, in which the right arms are fixed. The design domain is discretized using finite elements of size 5 mm×5 mm5\ \mathrm{mm} \times5\ \mathrm{mm}. The three optimization formulations are considered. As in the previous examples, a continuation strategy is applied to the projection parameter βf\beta_f to obtain optimized topologies with minimal intermediate-density regions. The initial value of βf\beta_f is set to 1 for all three formulations. In maximization of both fundamental frequency and gap of adjacent eigenfrequencies, βf\beta_f is increased by a factor of 3 until it reaches 32, and thereafter by a factor of 1.5 every 25 iterations up to a maximum value of 300.

Boundary condition and geometry of C-clamp

Figure 15. Boundary condition and geometry of C-clamp.

In creating a gap between the targeted frequency range, βf\beta_f is increased by a factor of 1.5 until it reaches 32, and then by a factor of 3 every 25 iterations up to 300. In addition, the filter radius is set to 8 for all three formulations. The variable λ\lambda in the maximization of fundamental frequency is initialized to 5000, with bounds [0,108][0, 10^8]. In maximizing the gap between adjacent eigenfrequencies, the scalar variables λ1\lambda_1 and λ2\lambda_2 are initialized with 0.4 and 0.6, respectively, with bounds [0,1][0, 1]. To create a gap between the targeted frequency range, the variable λ\lambda is initialized to 0.5 with bounds of [0,1][0, 1].

Maximum fundamental frequency

The fundamental eigenfrequency is maximized for a 2D C-clamp, and the optimized material distribution for the five-material case is presented in Figure 16(b). The convergence history is shown in Figure 16(a).

Convergence history of the fundamental eigenfrequency and optimized multi-material design

Figure 16. Fundamental eigenfrequency maximization with five candidate materials: (a) convergence history of the fundamental eigenfrequency and (b) optimized multi-material design.

Maximum gap between adjacent eigenfrequencies

In this numerical example, gap maximization between adjacent eigenfrequencies is applied to the C-clamp in order to maximize the gap between the third and fourth eigenfrequencies, and a maximized frequency gap of 16285 rad/s is obtained for the five-material case. The convergence history in Figure 17(a) indicates that the gap is opened predominantly through a reduction of the third eigenfrequency rather than an elevation of the fourth. The optimized material distribution is presented in Figure 17(b). The layout is symmetric about the horizontal mid-plane. Unlike the maximization of the first eigenfrequency, the objective is to increase the fourth eigenfrequency while decreasing the third, widening the frequency gap between them.

Figure 17. Maximizing the gap between adjacent eigenfrequencies with five candidate materials: (a) convergence history of the eigenfrequency gap and (b) optimized multi-material design.

Creating gaps between targeted frequency ranges

A single frequency gap of [10000,25000][10000, 25000] rad/s is considered. The resulting eigenfrequencies for the five-material configuration are reported in Table 4, and the convergence history is shown in Figure 18(a), from which a clear frequency gap spanning the interval [10000,25000][10000, 25000] rad/s is observed. The optimized material distribution is presented in Figure 18(b).

Mode (ii)123456
ωi\omega_i (rad/s)41241033224472247252599230637

Table 4. First six eigenfrequencies (ωi\omega_i) of five-material C-clamp obtained by excluding natural frequencies from a single targeted frequency range of [10000,25000][10000, 25000] rad/s

Convergence history of the frequency gap and optimized multi-material design

Figure 18. Targeted frequency gap with five candidate materials: (a) convergence history of the frequency gap and (b) optimized multi-material design.

Conclusions

This study presented a novel MMTO framework with multivariate weight-based interpolation for structural vibration attenuation. We addressed several methodological challenges in MMTO for structural vibration attenuation, including the representation of multiple material properties, the non-smoothness of eigenfrequency optimization in the presence of mode switching and repeated eigenvalues, and the difficulty of enforcing differentiable constraints to exclude natural frequencies from prescribed operating frequency bands. We investigated three different optimization formulations to achieve structural vibration attenuation. First, fundamental eigenfrequency maximization was formulated using an auxiliary scalar design variable and individual eigenvalue constraints, replacing the non-smooth max-min objective, which improves optimization stability in the presence of mode switching. Second, a bound-based formulation was introduced to maximize the separation between adjacent eigenfrequencies by simultaneously controlling the eigenvalues below and above the targeted gap. Third, a targeted frequency-band constraint was developed using a rational polynomial function to smoothly approximate the discontinuous Heaviside function. Combined with p-norm aggregation, this formulation enables differentiable constraints for excluding natural frequencies from prescribed operating frequency ranges, including multiple prohibited intervals. The proposed framework was evaluated through multiple numerical experiments. The results demonstrate the capability of the framework to generate distinct multi-material distributions according to the specified frequency optimization objectives.

Future work will extend the proposed framework toward programmable vibration attenuation, where structural dynamic characteristics can be adjusted in response to external stimuli. It can be extended to the design of multi-material metamaterials and periodic structures for vibration and wave attenuation.

Acknowledgements

The authors would also like to thank the constructive suggestions from anonymous reviewers. The authors would like to thank discussions with Dr. Behzad Rankouhi from the University of Minnesota, Dr. Cheng Huang from the University of Kansas, and Dr. Nathan Hertlein from Air Force Research Laboratory. The authors would like to thank Prof. Krister Svanberg for the MATLAB codes for the MMA optimizer. Shiguang Deng would like to thank the support of the Faculty Startup Grant, the New Faculty Research Development Grant, and the General Research Grant from the University of Kansas.

Replication of results

The results presented should be reproducible through the implementation of the methodologies described in this publication.

Not applicable.

Data availability

The data supporting the findings of this study will be made available by the corresponding author upon reasonable request.

Conflict of interest

No competing financial or non-financial interests.

References

  1. [1]S. S. Rao, Mechanical vibrations, 5th ed. Upper Saddle River, N.J: Prentice Hall, 2011.
  2. [2]S. Zargham, T. A. Ward, R. Ramli, and I. A. Badruddin, “Topology optimization: a review for structural designs under vibration problems,” Struct Multidisc Optim, vol. 53, no. 6, pp. 1157–1177, Jun. 2016, doi: 10.1007/s00158-015-1370-5.DOI
  3. [3]A. Zhu, X. Gong, J. Zhou, X. Zhang, and D. Zhang, “Efficient Vibration Measurement and Modal Shape Visualization Based on Dynamic Deviations of Structural Edge Profiles,” Sensors, vol. 24, no. 13, p. 4413, Jul. 2024, doi: 10.3390/s24134413.DOI
  4. [4]Q. Li, Q. Wu, S. Dou, J. Wang, S. Liu, and W. Chen, “Nonlinear eigenvalue topology optimization for structures with frequency-dependent material properties,” Mechanical Systems and Signal Processing, vol. 170, p. 108835, May 2022, doi: 10.1016/j.ymssp.2022.108835.DOI
  5. [5]R. Palej and A. Krowiak, “MODAL ANALYSIS OF MULTI-DEGREE-OF-FREEDOM SYSTEMS WITH REPEATED FREQUENCIES – ANALYTICAL APPROACH”.
  6. [6]B. B. Chaudhary, B. Chhantyal, and M. C. Luintel, “Transient response analysis of simply supported Pelton turbine during starting and shutdown,” Jnl Innov. Engin. Educ., vol. 7, no. 1, pp. 124–130, Dec. 2024, doi: 10.3126/jiee.v7i1.72340.DOI
  7. [7]C. Beards, Structural Vibration: Analysis and Damping. London : New York: Butterworth-Heinemann, 1996.
  8. [8]A. Molter, J. S. O. Fonseca, and L. D. S. Fernandez, “Simultaneous topology optimization of structure and piezoelectric actuators distribution,” Applied Mathematical Modelling, vol. 40, no. 9–10, pp. 5576–5588, May 2016, doi: 10.1016/j.apm.2016.01.023.DOI
  9. [9]Y. Maeda, S. Nishiwaki, K. Izui, M. Yoshimura, K. Matsui, and K. Terada, “Structural topology optimization of vibrating structures with specified eigenfrequencies and eigenmode shapes,” Numerical Meth Engineering, vol. 67, no. 5, pp. 597–628, Jul. 2006, doi: 10.1002/nme.1626.DOI
  10. [10]Z. Meng, G. Yang, Q. Wu, S. Xiao, and Q. Li, “Reliability-based topology optimization for fundamental frequency maximization with frequency band constraints,” Mechanical Systems and Signal Processing, vol. 195, p. 110295, Jul. 2023, doi: 10.1016/j.ymssp.2023.110295.DOI
  11. [11]S. Deng and K. Suresh, “Topology optimization under thermo-elastic buckling,” Struct Multidisc Optim, vol. 55, no. 5, pp. 1759–1772, May 2017, doi: 10.1007/s00158-016-1611-2.DOI
  12. [12]S. Deng and K. Suresh, “Stress constrained thermo-elastic topology optimization with varying temperature fields via augmented topological sensitivity based level-set,” Struct Multidisc Optim, vol. 56, no. 6, pp. 1413–1427, Dec. 2017, doi: 10.1007/s00158-017-1732-2.DOI
  13. [13]S. Deng and K. Suresh, “Multi-constrained topology optimization via the topological sensitivity,” Struct Multidisc Optim, vol. 51, no. 5, pp. 987–1001, May 2015, doi: 10.1007/s00158-014-1188-6.DOI
  14. [14]S. Deng and K. Suresh, “Multi-constrained 3D topology optimization via augmented topological level-set,” Computers & Structures, vol. 170, pp. 1–12, Jul. 2016, doi: 10.1016/j.compstruc.2016.02.009.DOI
  15. [15]M. P. Bendsøe and N. Kikuchi, “Generating optimal topologies in structural design using a homogenization method,” Computer Methods in Applied Mechanics and Engineering, vol. 71, no. 2, pp. 197–224, Nov. 1988, doi: 10.1016/0045-7825(88)90086-2.DOI
  16. [16]M. P. Bendsøe, “Optimal shape design as a material distribution problem,” Structural optimization, vol. 1, no. 4, pp. 193–202, Dec. 1989, doi: 10.1007/BF01650949.DOI
  17. [17]M. Zhou and G. I. N. Rozvany, “The CO algorithm, Part II: Topological, geometrical and generalized shape optimization,” Computer Methods in Applied Mechanics and Engineering, vol. 89, no. 1–3, pp. 309–336, Aug. 1991, doi: 10.1016/0045-7825(91)90046-9.DOI
  18. [18]Y. M. Xie and G. P. Steven, “A simple evolutionary procedure for structural optimization,” Computers & Structures, vol. 49, no. 5, pp. 885–896, Dec. 1993, doi: 10.1016/0045-7949(93)90035-C.
  19. [19]G. Allaire, F. Jouve, and A.-M. Toader, “Structural optimization using sensitivity analysis and a level-set method,” Journal of Computational Physics, vol. 194, no. 1, pp. 363–393, Feb. 2004, doi: 10.1016/j.jcp.2003.09.032.DOI
  20. [20]M. Y. Wang, X. Wang, and D. Guo, “A level set method for structural topology optimization,” Computer Methods in Applied Mechanics and Engineering, vol. 192, no. 1–2, pp. 227–246, Jan. 2003, doi: 10.1016/S0045-7825(02)00559-5.DOI
  21. [21]W. Zhang, J. Yuan, J. Zhang, and X. Guo, “A new topology optimization approach based on Moving Morphable Components (MMC) and the ersatz material model,” Struct Multidisc Optim, vol. 53, no. 6, pp. 1243–1260, Jun. 2016, doi: 10.1007/s00158-015-1372-3.DOI
  22. [22]H. Smith and J. Norato, “Topology optimization of structures made of fiber-reinforced plates,” Struct Multidisc Optim, vol. 65, no. 2, p. 58, Feb. 2022, doi: 10.1007/s00158-021-03164-y.DOI
  23. [23]S. Y. Wang and K. Tai, “Structural topology design optimization using Genetic Algorithms with a bit-array representation,” Computer Methods in Applied Mechanics and Engineering, vol. 194, no. 36–38, pp. 3749–3770, Sep. 2005, doi: 10.1016/j.cma.2004.09.003.DOI
  24. [24]I. Pehnec, D. Sedlar, I. Marinic-Kragic, and D. Vučina, “A New Approach to Topology Optimization with Genetic Algorithm and Parameterization Level Set Function,” Computation, vol. 13, no. 7, p. 153, Jun. 2025, doi: 10.3390/computation13070153.DOI
  25. [25]J. Du and N. Olhoff, “Topological design of freely vibrating continuum structures for maximum values of simple and multiple eigenfrequencies and frequency gaps,” Struct Multidisc Optim, vol. 34, no. 2, pp. 91–110, Jun. 2007, doi: 10.1007/s00158-007-0101-y.DOI
  26. [26]A. R. Díaz and N. Kikuchi, “Solutions to shape and topology eigenvalue optimization problems using a homogenization method,” International Journal for Numerical Methods in Engineering, vol. 35, no. 7, pp. 1487–1502, Oct. 1992, doi: 10.1002/nme.1620350707.
  27. [27]Z.-D. Ma, N. Kikuchi, and H.-C. Cheng, “Topological design for vibrating structures,” Computer Methods in Applied Mechanics and Engineering, vol. 121, no. 1–4, pp. 259–280, Mar. 1995, doi: 10.1016/0045-7825(94)00714-X.DOI
  28. [28]A. P. Seyranian, E. Lund, and N. Olhoff, “Multiple eigenvalues in structural optimization problems,” Structural Optimization, vol. 8, no. 4, pp. 207–227, Dec. 1994, doi: 10.1007/BF01742705.DOI
  29. [29]N. L. Pedersen, “Maximization of eigenvalues using topology optimization,” Struct Multidisc Optim, vol. 20, no. 1, pp. 2–11, Aug. 2000, doi: 10.1007/s001580050130.DOI
  30. [30]L. Xue, G. Wen, H. Wang, and J. Liu, “Eigenvectors-guided topology optimization to control the mode shape and suppress the vibration of the multi-material plate,” Computer Methods in Applied Mechanics and Engineering, vol. 391, p. 114560, Mar. 2022, doi: 10.1016/j.cma.2021.114560.
  31. [31]T. Liu, J.-H. Zhu, F. He, H. Zhao, Q. Liu, and C. Yang, “A MAC based excitation frequency increasing method for structural topology optimization under harmonic excitations,” Int. J. Simul. Multisci. Des. Optim., vol. 8, p. A4, 2017, doi: 10.1051/smdo/2016012.DOI
  32. [32]Q. Li, Q. Wu, J. Liu, J. He, and S. Liu, “Topology optimization of vibrating structures with frequency band constraints,” Struct Multidisc Optim, vol. 63, no. 3, pp. 1203–1218, Mar. 2021, doi: 10.1007/s00158-020-02753-7.DOI
  33. [33]L. Yin and G. K. Ananthasuresh, “Topology optimization of compliant mechanisms with multiple materials using a peak function material interpolation scheme,” Structural and Multidisciplinary Optimization, vol. 23, no. 1, pp. 49–62, Dec. 2001, doi: 10.1007/s00158-001-0165-z.DOI
  34. [34]C. F. Hvejsel and E. Lund, “Material interpolation schemes for unified topology and multi-material optimization,” Struct Multidisc Optim, vol. 43, no. 6, pp. 811–825, Jun. 2011, doi: 10.1007/s00158-011-0625-z.DOI
  35. [35]W. Zuo and K. Saitou, “Multi-material topology optimization using ordered SIMP interpolation,” Struct Multidisc Optim, vol. 55, no. 2, pp. 477–491, Feb. 2017, doi: 10.1007/s00158-016-1513-3.DOI
  36. [36]W. Ding, “An easy-to-use univariate mapping-based method for multi-material topology optimization with implementation in MATLAB,” Struct Multidisc Optim, vol. 68, no. 3, p. 48, Mar. 2025, doi: 10.1007/s00158-025-03983-3.DOI
  37. [37]M. Y. Wang and X. Wang, “‘Color’ level sets: a multi-phase method for structural topology optimization with multiple materials,” Computer Methods in Applied Mechanics and Engineering, vol. 193, no. 6–8, pp. 469–496, Feb. 2004, doi: 10.1016/j.cma.2003.10.008.DOI
  38. [38]R. Tavakoli, “Multimaterial topology optimization by volume constrained Allen–Cahn system and regularized projected steepest descent method,” Computer Methods in Applied Mechanics and Engineering, vol. 276, pp. 534–565, Jul. 2014, doi: 10.1016/j.cma.2014.04.005.DOI
  39. [39]S. Zhou and M. Y. Wang, “Multimaterial structural topology optimization with a generalized Cahn–Hilliard model of multiphase transition,” Struct Multidisc Optim, vol. 33, no. 2, pp. 89–111, Dec. 2006, doi: 10.1007/s00158-006-0035-9.DOI
  40. [40]J. P. Blasques, “Multi-material topology optimization of laminated composite beams with eigenfrequency constraints,” Composite Structures, vol. 111, pp. 45–55, May 2014, doi: 10.1016/j.compstruct.2013.12.021.DOI
  41. [41]M.-N. Nguyen and D. Lee, “Improving the performance of a multi-material topology optimization model involving stress and dynamic constraints,” Composite Structures, vol. 324, p. 117532, Nov. 2023, doi: 10.1016/j.compstruct.2023.117532.DOI
  42. [42]M. Zhou, Y. K. Shyy, and H. L. Thomas, “Checkerboard and minimum member size control in topology optimization,” Struct Multidisc Optim, vol. 21, no. 2, pp. 152–158, Apr. 2001, doi: 10.1007/s001580050179.DOI
  43. [43]O. Sigmund, “A 99 line topology optimization code written in Matlab,” Struct Multidisc Optim, vol. 21, no. 2, pp. 120–127, Apr. 2001, doi: 10.1007/s001580050176.DOI
  44. [44]F. Wang, B. S. Lazarov, and O. Sigmund, “On projection methods, convergence and robust formulations in topology optimization,” Struct Multidisc Optim, vol. 43, no. 6, pp. 767–784, Jun. 2011, doi: 10.1007/s00158-010-0602-y.DOI
  45. [45]J. K. Guest, A. Asadpour, and S.-H. Ha, “Eliminating beta-continuation from Heaviside projection and density filter algorithms,” Struct Multidisc Optim, vol. 44, no. 4, pp. 443–453, Oct. 2011, doi: 10.1007/s00158-011-0676-1.DOI
  46. [46]T. Borrvall and J. Petersson, “Topology optimization using regularized intermediate density control,” Computer Methods in Applied Mechanics and Engineering, vol. 190, no. 37–38, pp. 4911–4928, Jun. 2001, doi: 10.1016/S0045-7825(00)00356-X.DOI
  47. [47]O. Sigmund and J. Petersson, “Numerical instabilities in topology optimization: A survey on procedures dealing with checkerboards, mesh-dependencies and local minima,” Structural Optimization, vol. 16, no. 1, pp. 68–75, Aug. 1998, doi: 10.1007/BF01214002.DOI
  48. [48]S. Deng, “Augmented Topological Level-Set for Large-Scale Thermo-Elastic Topology Optimization”.
  49. [49]S. Deng, K. Suresh, and J. Joo, “Stress-Constrained Thermo-Elastic Topology Optimization: A Topological Sensitivity Approach,” in Volume IA: 34th Computers and Information in Engineering Conference, Buffalo, New York, USA: American Society of Mechanical Engineers, Aug. 2014, p. V01AT02A015. doi: 10.1115/DETC2014-34385.DOI
  50. [50]S. Deng and K. Suresh, “Topology Optimization Under Linear Thermo-Elastic Buckling,” in Volume 1A: 36th Computers and Information in Engineering Conference, Charlotte, North Carolina, USA: American Society of Mechanical Engineers, Aug. 2016, p. V01AT02A029. doi: 10.1115/DETC2016-59408.DOI
  51. [51]E. Andreassen, A. Clausen, M. Schevenels, B. S. Lazarov, and O. Sigmund, “Efficient topology optimization in MATLAB using 88 lines of code,” Struct Multidisc Optim, vol. 43, no. 1, pp. 1–16, Jan. 2011, doi: 10.1007/s00158-010-0594-7.DOI
  52. [52]K. Svanberg, “The method of moving asymptotes—a new method for structural optimization,” Numerical Meth Engineering, vol. 24, no. 2, pp. 359–373, Feb. 1987, doi: 10.1002/nme.1620240207.DOI

Paper details

vlnc:W_qebkbgew
Version
v2 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents