betaFeedback

WorksMathematics

Controlling the energy jump of multistable structures using shape optimization

Abstract

Multistable systems admit distinct stable equilibria for the same loading and boundary conditions. Shape optimization of multistable structures offers a promising route to engineer the mechanical response of advanced materials and metamaterials. In this work, we develop a formulation and algorithm for controlling the energy jump associated with snap-through behaviour in hyperelastic metamaterials by optimizing the shape of the domain. The key challenge is that classical shape optimization theory assumes a single-valued domain-to-solution map, which does not hold in the multistable setting. We address this by extending the theory to settings with non-unique solutions, treating the scaling of the domain deformation as a continuation parameter and applying the implicit function theorem to give sufficient conditions for the continued existence of multiple solution branches as the shape varies. This theoretical foundation underpins a practical algorithm that targets a prescribed energy jump between stable states, enabling the systematic design of structures whose snap-through response can be tuned on demand.

Introduction

Shape optimization lies at the intersection of partial differential equations (PDEs), optimization, physics, and engineering. In most applications, shape optimization aims to choose a domain to optimize a cost functional subject to physical constraints, often expressed as PDEs. This field can be viewed as a branch of optimal control or PDE-constrained optimization, where the control variable is the shape of the domain on which the PDE is defined [1, 10, ?].

In this work, we focus on shape optimization problems constrained by PDEs that admit multiple solutions. Specifically, we apply shape optimization to control the energy jump between pairs of distinct solutions to the same PDE constraint. We build on prior work [8], where we used shape optimization to control the bifurcation point at which multiple solutions emerge. As a motivating application, we consider the shape optimization of neo-Hookean hyperelastic metamaterial structures. However, our approach is easily adaptable to other applications.

Metamaterials are engineered materials with unusual properties not found in natural materials. In particular, we consider elastic metamaterials that exhibit different stable solutions, as shown in Figure 1. These materials have attracted substantial interest for their potential in engineering design for protecting against shocks [11, ?, ?, ?]. Metamaterials also exhibit a very rich bifurcation structure, with many physically achievable stable solutions [25]. This bifurcation structure can be exploited in design. For example, a metamaterial could dissipate impact energy by swapping branches, that is, by transitioning from one stable configuration to another [23, 24, ?]. Controlling this energy dissipation is desirable to achieve optimal protection from impact shocks [23, 24, ?]. Shape optimization has previously been applied to tune stress-strain responses and the (negative) Poisson’s ratio of these materials [26]; in this work we instead aim to control the energy required to snap from one branch to another, rendering the optimization problem qualitatively different.

Initial metamaterial domain and three deformed equilibria

Figure 1. Initial metamaterial domain Ω0\Omega_0 with two holes (left), and deformed domain under three distinct equilibria for the same boundary conditions: u1(Ω0)u_1(\Omega_0) (stable), u2(Ω0)u_2(\Omega_0) (unstable), and u3(Ω0)u_3(\Omega_0) (stable). Dirichlet conditions are imposed on the top and bottom, and natural boundary conditions on all other boundaries.

Most shape optimization methods do not need to address the difficulties introduced by PDE constraints that admit multiple solutions. Instead, it is commonly assumed that the solution to the PDE constraint is unique (possibly by introducing additional constraints). In turn, this induces a unique, single-valued shape-to-solution map, which enables the use of derivative-based optimization algorithms. However, many physical problems exhibit a rich landscape of multiple solutions. For example, PDE constraints stemming from nonconvex energy minimization problems can support multiple solutions in the form of local minima, local maxima, and saddle points. In these instances, correctly accounting for the multiple solutions afforded by the PDE is essential.

The core tool in the bifurcation analysis of problems with multiple solutions is the implicit function theorem (IFT) [9], see also [12, 20, 21]. The IFT gives sufficient but not necessary conditions ensuring that solution branches can be continued locally as a function of a parameter arising in the equations. In classical bifurcation analysis, this parameter is generally a real scalar. By parameterizing the shape deformations in terms of such a parameter, we devise conditions under which the IFT guarantees the local existence of branches of solutions as the domain shape is varied. Building on this insight, we develop a shape optimization algorithm that carefully tracks these solution branches throughout the shape optimization process.

The remainder of this article is organized as follows. In Section 2, we formulate the optimization problem under consideration and highlight that its PDE constraint supports multiple solutions. In Section 3, we propose a tailored shape optimization framework. In Section 4, we detail several numerical experiments. Finally, in Section 5, we draw the main conclusions and discuss open questions.

Problem statement

Let Ω⊂Rd\Omega\subset\mathbb{R}^{d} denote a bounded Lipschitz domain (with d=2d = 2 or d=3d = 3) representing a neo-Hookean hyperelastic material with boundary ∂Ω\partial\Omega. Dirichlet boundary conditions are imposed on ΓD⊂∂Ω\Gamma_D \subset\partial\Omega with positive measure and natural traction-free boundary conditions are imposed on ΓT:=∂Ω∖ΓD\Gamma_T := \partial\Omega\setminus\Gamma_D. The displacement

u∈V(Ω):={u∈H1(Ω,Rd):u∣ΓD=uD}(1)u \in V(\Omega) := \left\{u \in H^{1}(\Omega,\mathbb{R}^{d}) : u|_{\Gamma_D} = u_D\right\} \tag*{(1)}

is a stationary point of the potential energy

E(u):=∫Ωψ(u)−b⋅u dx(2)\mathcal{E}(u) := \int_{\Omega} \psi(u) - b \cdot u \,\mathrm{d}x \tag*{(2)}

where b∈L2(Ω;Rd)b \in L^{2}(\Omega;\mathbb{R}^{d}) is the external force and ψ(u)\psi(u) the internal energy density. Denoting by F:=I+∇uF := I + \nabla u the deformation gradient, the internal energy density ψ(u)\psi(u) takes the form [7]

ψ(u):=μ2(tr⁡(FTF)−d)−μln⁡det⁡(F)+λ2(ln⁡det⁡(F))2(3)\psi(u) := \frac{\mu}{2}\left(\operatorname{tr}(F^{\mathsf{T}}F)-d\right)-\mu\ln\det(F)+\frac{\lambda}{2}\left(\ln\det(F)\right)^{2} \tag*{(3)}

where μ>0\mu> 0 and λ>0\lambda> 0 are the Lamé coefficients. The Lamé coefficients are commonly expressed in terms of the Young’s modulus EE and the Poisson’s ratio ν\nu as μ=E/(2(1+ν))\mu= E/(2(1+\nu)) and λ=Eν/((1+ν)(1−2ν))\lambda= E\nu/((1+\nu)(1-2\nu)).

By setting to zero the first order derivative of the potential energy E\mathcal{E}, we obtain the nonlinear variational problem

∫ΩP(u):∇v dx=∫Ωb⋅v dxfor all v∈V0(Ω),(4)\int_{\Omega} P(u) : \nabla v \,\mathrm{d}x = \int_{\Omega} b \cdot v \,\mathrm{d}x \quad\text{for all } v \in V_0(\Omega), \tag*{(4)}

where V0(Ω):={u∈H1(Ω,Rd):u∣ΓD=0}V_0(\Omega) := \left\{u \in H^{1}(\Omega,\mathbb{R}^{d}) : u|_{\Gamma_D} = 0\right\}, and where

P(u):=μ(F−F−T)+λln⁡(det⁡(F))F−T(5)P(u) := \mu(F-F^{-\mathsf{T}})+\lambda\ln(\det(F))F^{-\mathsf{T}} \tag*{(5)}

denotes the first Piola–Kirchhoff stress tensor (see Appendix A). Note that (4) is also the weak formulation of the nonlinear boundary value problem

−∇⋅P(u)=bin Ω,u=uDon ΓD,P(u)⋅n=0on ΓT.(6)\begin{aligned} -\nabla\cdot P(u) &= b && \text{in } \Omega,\\ u &= u_D && \text{on } \Gamma_D,\\ P(u) \cdot n &= 0 && \text{on } \Gamma_T. \tag*{(6)} \end{aligned}

Due to the nonlinearity of the Piola–Kirchhoff stress tensor PP, the variational problem (4) generally admits multiple solutions. For a domain Ω\Omega, let u1(Ω),…,un(Ω)u_1(\Omega), \ldots, u_{n(\Omega)} denote the n(Ω)∈Nn(\Omega) \in\mathbb{N} (distinct, isolated) solutions to (4). To control the energy jump between solutions, we introduce the objective function

J(Ω):=∑1≤i<j≤nCi,j(ΔEi,j(Ω)ΔEi,j(Ω0)−ri,j)2(7)J(\Omega) := \sum_{1 \le i < j \le n} C_{i,j} \left(\frac{\Delta\mathcal{E}_{i,j}(\Omega)}{\Delta\mathcal{E}_{i,j}(\Omega_0)} - r_{i,j}\right)^2 \tag*{(7)}

where Ω0\Omega_0 denotes an initial domain, and where

ΔEi,j(Ω):=E(Ω,ui(Ω))−E(Ω,uj(Ω))(8)\Delta\mathcal{E}_{i,j}(\Omega) := \mathcal{E}(\Omega,u_i(\Omega)) - \mathcal{E}(\Omega,u_j(\Omega)) \tag*{(8)}

denotes the (signed) energy jump between the two solutions ui(Ω)u_i(\Omega) and uj(Ω)u_j(\Omega). This idea is illustrated in Figure 2 where we have used the notation E(Ω,u)\mathcal{E}(\Omega,u) to explicitly represent the dependency on the domain Ω\Omega.

Energy jump before and after optimization

Figure 2. Energy jump before and after optimization.

The coefficient ri,j∈Rr_{i,j} \in\mathbb{R} encodes the desired jump for pair (i,j)(i,j), and the weight Ci,j≥0C_{i,j} \ge0 balances the importance of different solution pairs (setting Ci,j=0C_{i,j} = 0 means that the jump between the solutions ui(Ω)u_i(\Omega) and uj(Ω)u_j(\Omega) is not controlled). Finally, to prevent distinct solutions from collapsing into each other, we consider the penalty term

B(Ω):=p∑1≤i<j≤nCi,j∥ui(Ω)−uj(Ω)∥L2(Ω)2B(\Omega) := p \sum_{1 \le i < j \le n} \frac{C_{i,j}}{\lVert u_i(\Omega) - u_j(\Omega) \rVert_{L^2(\Omega)}^2}

where p>0p > 0 is a penalization parameter.

The associated shape optimization problem can be formulated as follows:

min⁡Ω∈U(Ω0)J(Ω)+B(Ω)(9)\min_{\Omega\in\mathcal{U}(\Omega_0)} J(\Omega) + B(\Omega) \tag*{(9)}

where the dependency of the functionals on Ω\Omega is through the integrals and implicitly through ui(Ω)u_i(\Omega), which are precisely the solutions of (4). Here U(Ω0)\mathcal{U}(\Omega_0) denotes a set of admissible domains constructed from a reference domain Ω0\Omega_0; more details are given in Section 3 below.

Unlike most shape optimization problems, the functions ui(Ω)u_i(\Omega) appearing in the objective (7) are all solutions to the same PDE constraint. Since different domains may admit different numbers of solutions, we tacitly assume that (9) is solved on the initial domain Ω0∈U\Omega_0 \in\mathcal{U}, where all solutions are computed and properly labelled. This initial domain is then updated iteratively, and at each domain update the state solutions are recomputed and properly labelled to ensure consistency. This requires the existence of suitably regular maps Ω↦ui(Ω)\Omega\mapsto u_i(\Omega) for each tracked solution ui(Ω)u_i(\Omega). The existence of such maps is discussed in Section 3.

Iterative shape optimization

For an initial domain Ω0\Omega_0, the set of admissible controls is given by

U(Ω0):={Ω=T(Ω0):T∈F(Θ)}(10)\mathcal{U}(\Omega_0) := \{\Omega= T(\Omega_0) : T \in\mathcal{F}(\Theta)\} \tag*{(10)}

where the set of admissible deformations is given by

F(Θ):={I+θ:θ∈Θ, θ=0 on ΓD, (I+θ)−1 exists, and (I+θ)−1−I∈Θ}(11)\mathcal{F}(\Theta) := \{I+\theta: \theta\in\Theta,\ \theta= 0\text{ on }\Gamma_D,\ (I+\theta)^{-1}\text{ exists, and }(I+\theta)^{-1}-I\in\Theta\} \tag*{(11)}

for the Banach space Θ=W1,∞(Rd,Rd)\Theta= W^{1,\infty}(\mathbb{R}^d,\mathbb{R}^d). This makes U(Ω0)\mathcal{U}(\Omega_0) a complete metric space [10]. Notice in particular that reachable shapes have the same topology as the reference domain.

To solve the shape optimization problem (2.9), we construct a sequence of diffeomorphisms {Tk}k≥1\{T_k\}_{k\geq1} of the form

Tk(x)=x+dTk(x)for x∈Ω0.(12)T_k(x)=x+dT_k(x)\quad\text{for }x\in\Omega_0. \tag*{(12)}

Under suitable selection of the perturbation terms {dTk}k≥1\{dT_k\}_{k\geq1}, the diffeomorphism sequence {Tk}k≥1\{T_k\}_{k\geq1} induces a sequence of domains {Ωk:=Tk(Ω0)}k≥1\{\Omega_k:=T_k(\Omega_0)\}_{k\geq1} such that the sequence {J(Ωk)}k≥1\{J(\Omega_k)\}_{k\geq1} is decreasing. Our focus is on the numerical realization of a shape optimization algorithm, Algorithm 3.1, to tackle (2.9). For this reason, we do not address the theoretical questions of existence and uniqueness of solutions to (2.9) or determining conditions that guarantee that {J(Ωk)}k≥1\{J(\Omega_k)\}_{k\geq1} is a minimizing sequence.

The realization of optimization algorithms based on (3.3) requires a guarantee that the solutions uiu_i to the state constraint (2.4) can be consistently indexed and paired with the corresponding coefficients ri,jr_{i,j} and weights Ci,jC_{i,j} to evaluate (2.7). Under suitable assumptions, the IFT provides this guarantee, as we now explain.

For a fixed vector field A∈W1,∞(Ω0,Rd)A\in W^{1,\infty}(\Omega_0,\mathbb{R}^d) with A=0A=0 on ΓD\Gamma_D, consider the family of transformations {St(x):=x+tA(x)}t≥0\{S_t(x):=x+tA(x)\}_{t\geq0}. It is well known that StS_t is a W1,∞W^{1,\infty}-diffeomorphism if the parameter tt is sufficiently small, say t≤ϵt\leq\epsilon for an ϵ>0\epsilon>0 [10]. In light of this, we introduce the function G:[0,ϵ]×V(Ω0)→V∗(Ω0)G:[0,\epsilon]\times V(\Omega_0)\to V^*(\Omega_0) defined as

⟨G(t,u),v⟩:=∫Ω0(Pt(u):(DSt−⊤∇v)−(b∘St)⋅v)det⁡(DSt) dx,(13)\langle G(t,u),v\rangle:=\int_{\Omega_0}\left(P_t(u):(DS_t^{-\top}\nabla v)-(b\circ S_t)\cdot v\right)\det(DS_t)\,dx, \tag*{(13)}

where V∗(Ω0)V^*(\Omega_0) denotes the dual of V0(Ω0)V_0(\Omega_0), ⟨ , ⟩\langle\ ,\ \rangle denotes the V∗(Ω0)×V0(Ω0)V^*(\Omega_0)\times V_0(\Omega_0) duality pairing,

Pt(u):=μ(Ft−Ft−⊤)+λln⁡(det⁡(Ft))Ft−⊤,(14)P_t(u):=\mu(F_t-F_t^{-\top})+\lambda\ln(\det(F_t))F_t^{-\top}, \tag*{(14)}

and Ft:=I+DSt−⊤∇uF_t:=I+DS_t^{-\top}\nabla u, where DStDS_t denotes the Jacobian of StS_t. Formula (3.4) is obtained by pulling back the weak formulation (2.4) onto Ω0\Omega_0. Therefore, the solutions to

⟨G(t,u),v⟩=0for all v∈V0(Ω0),(15)\langle G(t,u),v\rangle=0\quad\text{for all }v\in V_0(\Omega_0), \tag*{(15)}

are the pullback of the solutions to (2.4), that is, if uu solves (3.6), then the composition u∘St−1u\circ S_t^{-1} solves (2.4), and vice-versa. Moreover, if det⁡(Ft)>0\det(F_t)>0, the function GG is continuous in tt and also continuously differentiable in the variable uu. Therefore, if (0,u)(0,u) is a solution to (3.6) and the Fréchet derivative Gu(0,u)G_u(0,u) is invertible, the IFT [9, 12, 20, 21] guarantees the local existence of a differentiable function t↦u(t)t\mapsto u(t) such that G(t,u(t))=0G(t,u(t))=0. (This also implies that uu is an isolated solution.) By assuming that Bk:=Tk+1∘Tk−1−I∈W1,∞(Ωk)B_k:=T_{k+1}\circ T_k^{-1}-I\in W^{1,\infty}(\Omega_k) is sufficiently small and applying the IFT to Stk:=I+tBkS_t^k:=I+tB_k, we can verify that the solutions ui(Ωk)u_i(\Omega_k) and ui(Ωk+1)u_i(\Omega_{k+1}) correspond to the same index ii by checking that the quantity ∥ui(Ωk)∘S1k−ui(Ωk+1)∥L2(Ωk+1)\lVert u_i(\Omega_k)\circ S_1^k-u_i(\Omega_{k+1})\rVert_{L^2(\Omega_{k+1})} (or, equivalently but more conveniently, ∥ui(Ωk)∘(S1k)−1−ui(Ωk+1)∥L2(Ωk+1)\lVert u_i(\Omega_k)\circ(S_1^k)^{-1}-u_i(\Omega_{k+1})\rVert_{L^2(\Omega_{k+1})}) is also sufficiently small. Note that requiring BkB_k to be sufficiently small is not necessarily detrimental. Indeed, this guarantees that the pushed-forward solution ui(Ωk)∘(S1k)−1u_i(\Omega_k)\circ(S_1^k)^{-1} is a good initial guess for the computation of ui(Ωk+1)u_i(\Omega_{k+1}) using an iterative nonlinear solver like Newton’s method, which may thus converge in just a few iterations.

To ensure that the sequence {J(Ωk)}k≥1\{J(\Omega_k)\}_{k\geq1} is decreasing, we compute the perturbations dTkdT_k using the dogleg trust-region algorithm with BFGS-Hessian updates implemented in the Rapid Optimization Library (ROL) [19]. The underpinning steepest descent directions are based on the shape gradients ∇J\nabla J in V0(Ω0)V_0(\Omega_0), that is, the Riesz representative of the shape derivative of the constrained function JJ [10], with respect to the V(Ω0)V(\Omega_0) inner product [?]. The whole optimization process is automated by the shape optimization library Fireshape [?], which builds on the finite element software Firedrake [16] and leverages the automated adjoint derivation and (shape) differentiation of pyadjoint [14, 27] and of the Unified Form Language (UFL) [2, 17].

Algorithm 3.1 Shape optimization with trust region

Require: Objective functional JJ, initial domain Ω0\Omega_0, initial Hessian approximation H0H_0, initial trust region radius Δ0\Delta_0, maximum radius Δmax⁡\Delta_{\max}, step acceptance threshold η1=0.05\eta_1 = 0.05, radius shrinking threshold η1−=0.05\eta_1^- = 0.05, radius growing threshold η2=0.9\eta_2 = 0.9, shrinking rates γ1−=0.0625\gamma_1^- = 0.0625 (negative ρ\rho), γ1+=0.25\gamma_1^+ = 0.25 (positive ρ\rho), growth rate γ2=1.1\gamma_2 = 1.1, sufficient decrease parameter μ=10−4\mu= 10^{-4}, gradient tolerance ϵg\epsilon_g, step tolerance ϵs\epsilon_s, maximum iterations max_its\mathrm{max\_its}

Initialize k←0k \leftarrow0, Δ←Δ0\Delta\leftarrow\Delta_0, T0←IT_0 \leftarrow I

Load initial solutions (ui(Ω0))(u_i(\Omega_0))

Compute initial objective value J(Ω0)J(\Omega_0) and gradient ∇J(Ω0)\nabla J(\Omega_0)

while k<max_itsk < \mathrm{max\_its} and ∥∇J(Ωk)∥V0(Ω0)>ϵg\lVert\nabla J(\Omega_k) \rVert_{V_0(\Omega_0)} > \epsilon_g do

Construct quadratic trust region model using L-BFGS Hessian approximation HkH_k:

mk(θ)=J(Ωk)+⟨∇J(Ωk),θ⟩V0(Ω0)+12⟨Hkθ,θ⟩V0(Ω0)m_k(\theta) = J(\Omega_k) + \langle\nabla J(\Omega_k), \theta\rangle_{V_0(\Omega_0)} + \frac{1}{2} \langle H_k\theta, \theta\rangle_{V_0(\Omega_0)}

Solve trust region subproblem via Dogleg method:

θk=arg min⁡∥θ∥≤Δ  mk(θ)\theta_k = \underset{\lVert\theta\rVert\le\Delta}{\operatorname{arg\,min}}\; m_k(\theta)

if ∥θk∥V0(Ω0)≤ϵs\lVert\theta_k \rVert_{V_0(\Omega_0)} \le\epsilon_s then

return

end if

Compute candidate shape Ω~=(Tk+θk)(Ω0)\widetilde{\Omega} = (T_k + \theta_k)(\Omega_0)

Compute solutions (ui(Ω~))(u_i(\widetilde{\Omega})) with Newton and initial guesses (ui(Ωk)∘Tk∘(Tk+θk)−1)(u_i(\Omega_k) \circ T_k \circ(T_k + \theta_k)^{-1})

Compute J(Ω~)J(\widetilde{\Omega}) and actual-to-predicted reduction ratio:

ρk=J(Ωk)−J(Ω~)mk(0)−mk(θk)\rho_k = \frac{J(\Omega_k) - J(\widetilde{\Omega})}{m_k(0) - m_k(\theta_k)}

if ρk≥η1\rho_k \ge\eta_1 then

Ωk+1←Ω~\Omega_{k+1} \leftarrow\widetilde{\Omega}, Tk+1←Tk+θkT_{k+1} \leftarrow T_k + \theta_k, (ui(Ωk+1))←(ui(Ω~))(u_i(\Omega_{k+1})) \leftarrow(u_i(\widetilde{\Omega}))

Update gradient ∇J(Ωk+1)\nabla J(\Omega_{k+1}) and L-BFGS Hessian approximation Hk+1H_{k+1}

k←k+1k \leftarrow k + 1

end if

Update trust region radius Δ\Delta according to:

Δ←{γ1−Δif ρk<0γ1+Δif 0≤ρk<η1−Δif η1−≤ρk<η2min⁡(γ2Δ,Δmax⁡)if ρk≥η2 and ∥θk∥=Δ\Delta\leftarrow \begin{cases} \gamma_1^- \Delta& \text{if } \rho_k < 0 \\ \gamma_1^+ \Delta& \text{if } 0 \le\rho_k < \eta_1^- \\ \Delta& \text{if } \eta_1^- \le\rho_k < \eta_2 \\ \min(\gamma_2 \Delta, \Delta_{\max}) & \text{if } \rho_k \ge\eta_2 \text{ and } \lVert\theta_k\rVert= \Delta \end{cases}

end while

Numerical Results

We demonstrate the effectiveness of the proposed methodology on a number of test cases in two and three dimensions. In these numerical experiments, the Young’s modulus is E=106E = 10^6, the Poisson’s ratio is ν=0.3\nu= 0.3, and the internal body force is either b=(0,−103)Tb = (0, -10^3)^{\mathsf{T}} or b=(0,−103,0)Tb = (0, -10^3, 0)^{\mathsf{T}}. The source code for these experiments is available at [15].

The initial computational domain Ω0\Omega_0 is an elastomeric square block with edges of length L=1L = 1 and an embedded N×NN \times N square array of circular holes centered at {(iN+1,jN+1):i,j=1,…,N}\left\{\left(\frac{i}{N+1}, \frac{j}{N+1}\right) : i,j = 1,\ldots,N\right\}. Horizontally (or vertically) adjacent holes have a center-to-center distance a=L/(N+1)a = L/(N+1) and a radius of Ra/2Ra/2 with R∈(0,1)R \in(0,1). The two vertical edges of Ω0\Omega_0 are flanked by a column of semicircles centered at {(x,i/(N+1)):i=1,…,N,x=0,1}\left\{(x,i/(N+1)) : i = 1,\ldots,N, x = 0,1\right\} and of the same radius Ra/2Ra/2. Notice that if R=0R = 0 the circles are reduced to points while adjacent circle pairs touch each other in a single point if R=1R = 1. The parameter RR allows us to choose the complexity of the multistable structure: we observe that the number of solutions increases as RR increases. Hereafter, we consider two cases: (N,R)=(2,0.9)(N,R) = (2,0.9) and (N,R)=(4,0.85)(N,R) = (4,0.85). For three-dimensional simulations, Ω0\Omega_0 is further extruded in the third dimension to a depth equal to h=1h = 1.

The Dirichlet data is set to uD=0u_D = 0 on the bottom of the square, and uD=(0,−0.1)Tu_D = (0,-0.1)^{\mathsf{T}} or uD=(0,−0.1,0)Tu_D = (0,-0.1,0)^{\mathsf{T}} on the top of the square. Natural boundary conditions are imposed on the remaining components of the boundary (side walls and walls of the holes).

To compute the set of initial solutions {ui(Ω0)}\{u_i(\Omega_0)\}, we use the deflated continuation algorithm [13]. The nonlinear equation (4) is solved using PETSc’s Newton algorithm, with the linear systems solved with the sparse direct solver MUMPS [3, 4, 5, 6]. Of the computed solutions, we select three specific solutions: two stable solutions u1(Ω0)u_1(\Omega_0) and u3(Ω0)u_3(\Omega_0) that are Z2\mathbb{Z}_2-symmetric with respect to each other and the saddle point u2(Ω0)u_2(\Omega_0) in between. We check the signs of the eigenvalues of the matrix corresponding to the discrete linearisation of (4). These are depicted in Figure 1 for a two-dimensional example. The approach below extends easily to other choices of solution branches.

We consider two possible families of objective functions:

J1(Ω)=(ΔE1,2(Ω)ΔE1,2(Ω0)−r)2andJ2(Ω)=(ΔE1,2(Ω)ΔE1,2(Ω0)−r)2+(ΔE3,2(Ω)ΔE3,2(Ω0)−r)2.(16)J_1(\Omega) = \left(\frac{\Delta\mathcal{E}_{1,2}(\Omega)}{\Delta\mathcal{E}_{1,2}(\Omega_0)} - r\right)^2 \quad\text{and}\quad J_2(\Omega) = \left(\frac{\Delta\mathcal{E}_{1,2}(\Omega)}{\Delta\mathcal{E}_{1,2}(\Omega_0)} - r\right)^2 + \left(\frac{\Delta\mathcal{E}_{3,2}(\Omega)}{\Delta\mathcal{E}_{3,2}(\Omega_0)} - r\right)^2 . \tag*{(16)}

Here, rr denotes a target ratio. Note that, in this particular case, the symmetry between u1(Ω0)u_1(\Omega_0) and u3(Ω0)u_3(\Omega_0) and the central position of the saddle point u2(Ω0)u_2(\Omega_0) (with respect to the energy functional E\mathcal{E}) imply that ∣ΔE1,2(Ω0)∣=∣ΔE2,3(Ω0)∣\lvert\Delta\mathcal{E}_{1,2}(\Omega_0)\rvert= \lvert\Delta\mathcal{E}_{2,3}(\Omega_0)\rvert.

Figure 3 illustrates, for the case N=2N = 2, the initial domain together with the optimized shapes obtained for the target energy jump ratios r=0.6r = 0.6, 0.40.4 and 0.20.2, using either objective J1J_1 or J2J_2.

Initial domain and optimized shapes for target energy jump ratios

Figure 3. Initial domain Ω0\Omega_0 (leftmost), followed by the optimized shapes for the target energy jump ratios r=0.6r = 0.6, 0.40.4 and 0.20.2 (left to right), with respect to J1J_1 (first row) and J2J_2 (second row).

In both instances, we observe nontrivial domain modifications. When the geometry is optimized by minimizing J2J_2, which accounts for both gaps ΔE1,2\Delta\mathcal{E}_{1,2} and ΔE2,3\Delta\mathcal{E}_{2,3} simultaneously, we observe that the symmetry of the initial domain and its bistable structure are preserved throughout the shape optimization process. In the convergence plots reported in Figure 4, we observe that when J1J_1 is minimized, the energy curves of u1u_1 and u2u_2 converge toward each other, reflecting the targeted reduction of ΔE1,2\Delta\mathcal{E}_{1,2} to 40% of ΔE1,2(Ω0)\Delta\mathcal{E}_{1,2}(\Omega_0) (dashed line). In this simulation, the energy curve of u3u_3, which is not controlled, drifts mildly. In contrast, when minimizing J2J_2, the energy curves of u1u_1 and u3u_3 coincide (consistent with the observed symmetry preservation) and approach the energy curve of u2u_2, reflecting the simultaneous targeted reduction of ΔE1,2\Delta\mathcal{E}_{1,2} and ΔE2,3\Delta\mathcal{E}_{2,3}. Figure 4 also displays convergence plots for N=4N = 4, for which we observe a similar behavior.

Convergence histories for minimizing $J_1$ and $J_2$ with $N=2,4$

Figure 4. Convergence history for r=0.4r = 0.4 and N=2,4N = 2, 4. In the right column, the lines for u1u_1 and u3u_3 lie on top of one another.

We considered the same experiment, changing the target ratio (r=0.2r = 0.2, 0.40.4, 0.60.6). The corresponding convergence histories are shown in Figure 5. For smaller rr values, which demands more aggressive energy-gap reduction, the optimization problems take modestly more iterations. In particular, the optimization for J1J_1 and r=0.2r = 0.2 has more difficulty reaching the optimum and fails to reach the target ratio. Minimizing J2J_2 for the same rr does achieve the target ratio. We also consider the case r=1.6r = 1.6, which corresponds to increasing the energy gap by 60%. In this case, the desired ratio r=1.6r = 1.6 is achieved when minimizing J1J_1, but not when minimizing J2J_2 (not shown). We suspect that this different convergence behavior may be linked to the preservation (or loss) of symmetry, but further research is needed to confirm this observation. Comparing the results for N=2N = 2 and N=4N = 4, we observe that the optimization histories are qualitatively similar, indicating some robustness of the algorithm.

Convergence histories of energy gaps for different target ratios and $N=2,4$

Figure 5. Convergence history of energy gaps for r=0.2,0.4,0.6r = 0.2, 0.4, 0.6 and N=2,4N = 2, 4. Dashed lines indicate target ratios.

To address the convergence failures observed in the previous experiments, we consider a continuation strategy. Instead of directly targeting the ratio r=0.2r = 0.2 (or r=1.6r = 1.6), we solve a sequence of shape optimization problems with decreasing (or increasing) target ratios rr. As evinced in Figure 6 (for N=4N = 4), this strategy successfully achieved the desired target ratios, albeit at the cost of an increased total number of iterations over the continuation steps. The initial and optimized geometries for this experiment are shown in Figure 7. This demonstrates that while direct optimization for large energy gap increases can be challenging due to the competing geometric constraints, a carefully designed continuation strategy can overcome these difficulties and achieve the desired energy dissipation characteristics.

Sequential minimization plots

Figure 6. Left: sequential minimization of J1J_1 with target ratios r=0.6,0.4,0.2r = 0.6, 0.4, 0.2. Right: sequential minimization of J2J_2 with target ratios r=1.1,1.2,1.3,1.6r = 1.1, 1.2, 1.3, 1.6. In both experiments, N=4N = 4. Dashed lines indicate intermediate target ratios.

Domain and its deformed configurations under displacements $u_1$, $u_2$ and $u_3$

Figure 7. Domain Ω\Omega and its deformed configurations under displacements u1u_1, u2u_2 and u3u_3. From top to bottom: initial configuration, after optimization of J2J_2 with target ratio r=1.6r = 1.6, and after optimization of J1J_1 with target ratio r=0.2r = 0.2

Finally, we perform a numerical experiment in three dimensions and minimize J2J_2 first for N=2N = 2 and target ratio r=0.2r = 0.2. The initial and optimized domains (both at rest and displaced) are shown in Figure 8. Consistent with the 2D case, the optimization successfully reduces the energy gap to approximately 20% of its initial value (not shown), confirming the efficacy of the optimization strategy and objective formulation in higher dimensions. The optimized geometry redistributes material to moderate energy dissipation, with adaptations resembling the 2D pattern. In particular, the optimized domain is stretched opposite to the directions of the stable solution branches.

Initial and optimized three-dimensional geometries for the target ratio $r = 0.2$

Figure 8. Initial (top) and optimized (bottom) three-dimensional geometries for the target ratio r=0.2r = 0.2 optimized by minimizing J2J_2. The leftmost geometry (in gray) denotes the optimized geometry at rest. From the left to the right, this geometry is displaced by applying u1u_1, u2u_2, and u3u_3, respectively.

Lastly, we reproduce the same experiment with N=4N = 4; the initial and optimized domains are shown in Figure 9.

Initial and optimized three-dimensional geometries for the target ratio

Figure 9. Initial (top) and optimized (bottom) three-dimensional geometries for the target ratio r=0.2r = 0.2 optimized by minimizing J2J_2 for N=4N = 4. The leftmost geometry (in gray) denotes the optimized geometry at rest. From the left to the right, this geometry is displaced by applying u1u_1, u2u_2, and u3u_3, respectively.

Conclusion

Conclusion. We have presented a shape optimization framework designed to control the energy jump between distinct solutions of a neo-Hookean metamaterial, which has potential applications in impact energy dissipation. The approach is general and can be applied to other shape optimization problems with multi-valued PDE constraints.

The methodology builds on rigorous theoretical foundations and its effectiveness has been validated extensively through numerical experiments in two and three dimensions.

A central assumption of the proposed method is that the number of solution branches remains constant throughout the optimization process. In scenarios where this does not hold, a robust labelling procedure would be required to track branches of interest. Additionally, the current framework cannot be directly applied to self-contact problems (such as those arising from large displacements), in which case the PDE constraint is replaced by a quasi-variational inequality [18, 22].

Declaration. The research presented in this work was conducted by the authors. The use of AI tools was confined to assisting with text editing for spelling and grammar and preparing the code for reproducibility. The authors assume responsibility for all content.

References

References

[?] A. Paganini and F. Wechsung, Fireshape: a shape optimization toolbox for Firedrake, Struct. Multidiscip. Optim., 63 (2021), pp. 2553–2569.

[?] A. Paganini, F. Wechsung, and P. E. Farrell, Higher-order moving mesh methods for PDE-constrained shape optimization, SIAM J. Sci. Comput., 40 (2018), pp. A2356–A2382.

[?] A. Pratelli and G. Leugering, eds., New Trends in Shape Optimization, vol. 166 of International Series of Numerical Mathematics, Springer, 2015.

[?] S. Shan, S. H. Kang, J. R. Raney, P. Wang, L. Fang, F. Candido, J. A. Lewis, and K. Bertoldi, Multistable architected materials for trapping elastic strain energy, Adv. Mater., 27 (2015), pp. 4296–4301.

[?] X. Tan, S. Chen, B. Wang, S. Zhu, L. Wu, and Y. Sun, Design, fabrication, and characterization of multistable mechanical metamaterials for trapping energy, Ext. Mech. Lett., 28 (2019), pp. 8–21.

[?] R. Xu, Y. He, X. Li, M. Lu, and Y. Chen, Snap-fit mechanical metamaterials, Appl. Mater. Today, 30 (2023), p. 101714.

[?] Y. Zhang, M. Tichem, and F. van Keulen, A novel design of multi-stable metastructures for energy dissipation, Mater. Des., 212 (2021), p. 110234.

Appendix A. Derivation of the first Piola–Kirchhoff stress tensor (2.5)

Let Ω0⊂R3\Omega_0 \subset\mathbb{R}^3 be a reference configuration. A deformation φ:Ω0→R3\varphi: \Omega_0 \to\mathbb{R}^3 has deformation gradient F=∇φ∈R3×3F = \nabla\varphi\in\mathbb{R}^{3 \times3}, assumed to satisfy J:=det⁡F>0J := \det F > 0. We write C=FTFC = F^{\mathsf{T}}F for the right Cauchy–Green tensor and I1:=tr⁡(C)=tr⁡(FTF)I_1 := \operatorname{tr}(C) = \operatorname{tr}(F^{\mathsf{T}}F) for its first invariant.

The Frobenius (double-dot) inner product of two matrices A,B∈R3×3A, B \in\mathbb{R}^{3 \times3} is defined by

A:B:=tr⁡(ATB)=∑i,jAijBij.A : B := \operatorname{tr}(A^{\mathsf{T}}B) = \sum_{i,j} A_{ij}B_{ij}.

The first Piola–Kirchhoff (PK1) stress tensor P∈R3×3P \in\mathbb{R}^{3 \times3} is characterised by

ddε∣ε=0W(F+εH)=P:H∀H∈R3×3.\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0} W(F+\varepsilon H) = P : H \qquad\forall H \in\mathbb{R}^{3 \times3}.

i.e. PP is the matrix representative of the derivative of WW at FF.

We recall the neo-Hookean strain energy density (3)

W(F)=μ2(tr⁡(FTF)−3)−μln⁡J+λ2(ln⁡J)2,J=det⁡F,(17)W(F) = \frac{\mu}{2}\left(\operatorname{tr}(F^{\mathsf{T}}F)-3\right) - \mu\ln J + \frac{\lambda}{2}(\ln J)^2,\qquad J = \det F, \tag*{(17)}

where μ>0\mu> 0 is the shear modulus and λ≥0\lambda\ge0 is the first Lamé constant. We split W=W1+W2+W3W = W_1 + W_2 + W_3 according to the three terms and compute each derivative in turn.

Step 1—Linearisation of the trace term.

Expanding (F+εH)T(F+εH)(F+\varepsilon H)^{\mathsf{T}}(F+\varepsilon H) gives

(F+εH)T(F+εH)=FTF+ε(FTH+HTF)+ε2HTH.(F+\varepsilon H)^{\mathsf{T}}(F+\varepsilon H) = F^{\mathsf{T}}F + \varepsilon(F^{\mathsf{T}}H + H^{\mathsf{T}}F) + \varepsilon^2 H^{\mathsf{T}}H.

Taking the trace, the O(ε2)O(\varepsilon^2) term vanishes upon differentiation at ε=0\varepsilon=0, yielding

ddε∣ε=0tr⁡((F+εH)T(F+εH))=tr⁡(FTH)+tr⁡(HTF).\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0}\operatorname{tr}\left((F+\varepsilon H)^{\mathsf{T}}(F+\varepsilon H)\right) = \operatorname{tr}(F^{\mathsf{T}}H) + \operatorname{tr}(H^{\mathsf{T}}F).

Since tr⁡(A)=tr⁡(AT)\operatorname{tr}(A) = \operatorname{tr}(A^{\mathsf{T}}) for any square matrix, tr⁡(HTF)=tr⁡(FTH)\operatorname{tr}(H^{\mathsf{T}}F) = \operatorname{tr}(F^{\mathsf{T}}H), so

ddε∣ε=0W1(F+εH)=μ2⋅2tr⁡(FTH)=μF:H.\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0} W_1(F+\varepsilon H) = \frac{\mu}{2}\cdot2\operatorname{tr}(F^{\mathsf{T}}H) = \mu F : H.

The contribution of W1W_1 to PP is therefore μF\mu F.

Step 2—Linearisation of det⁡\det via Jacobi’s formula.

Factor out FF from the perturbed determinant:

det⁡(F+εH)=det⁡F⋅det⁡(I+εF−1H).\det(F+\varepsilon H) = \det F \cdot\det(I+\varepsilon F^{-1}H).

The first-order Taylor expansion of the determinant near the identity reads det⁡(I+εA)=1+εtr⁡(A)+O(ε2)\det(I+\varepsilon A) = 1+\varepsilon\operatorname{tr}(A)+O(\varepsilon^2), so

ddε∣ε=0det⁡(F+εH)=Jtr⁡(F−1H).\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0}\det(F+\varepsilon H) = J\operatorname{tr}(F^{-1}H).

Since tr⁡(F−1H)=tr⁡((F−T)TH)=F−T:H\operatorname{tr}(F^{-1}H) = \operatorname{tr}((F^{-\mathsf{T}})^{\mathsf{T}}H) = F^{-\mathsf{T}} : H, this is Jacobi’s formula:

ddε∣ε=0det⁡(F+εH)=JF−T:H.(18)\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0}\det(F+\varepsilon H) = J F^{-\mathsf{T}} : H. \tag*{(18)}

Step 3—Linearisation of ln⁡J\ln J and differentiation of W2W_2 and W3W_3.

Applying the chain rule to ln⁡\ln and using (A.2),

ddε∣ε=0ln⁡det⁡(F+εH)=1JJF−T:H=F−T:H.\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0} \ln\det(F+\varepsilon H) = \frac{1}{J} JF^{-\mathsf{T}} : H = F^{-\mathsf{T}} : H.

For W2=−μln⁡JW_2=-\mu\ln J, this directly gives

ddε∣ε=0W2(F+εH)=−μF−T:H,\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0} W_2(F+\varepsilon H) = -\mu F^{-\mathsf{T}} : H,

contributing −μF−T-\mu F^{-\mathsf{T}} to PP.

For W3=λ2(ln⁡J)2W_3=\frac{\lambda}{2}(\ln J)^2, setting g(ε):=ln⁡det⁡(F+εH)g(\varepsilon):=\ln\det(F+\varepsilon H) and applying the chain rule to λ2[g(ε)]2\frac{\lambda}{2}[g(\varepsilon)]^2 at ε=0\varepsilon=0,

ddε∣ε=0W3(F+εH)=λln⁡J⋅g′(0)=λln⁡JF−T:H,\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0} W_3(F+\varepsilon H) = \lambda\ln J\cdot g'(0) = \lambda\ln J F^{-\mathsf{T}} : H,

contributing λln⁡JF−T\lambda\ln J F^{-\mathsf{T}} to PP.

Conclusion.

Summing the three contributions, for every H∈R3×3H\in\mathbb{R}^{3\times3},

ddε∣ε=0W(F+εH)=(μF−μF−T+λln⁡JF−T):H.\left.\frac{\mathrm{d}}{\mathrm{d}\varepsilon}\right|_{\varepsilon=0} W(F+\varepsilon H) = \left(\mu F-\mu F^{-\mathsf{T}}+\lambda\ln J F^{-\mathsf{T}}\right):H.

Since PP is defined by P:HP:H equalling this quantity for all HH, we conclude

P(F)=μ(F−F−T)+λln⁡(det⁡F)F−T.P(F)=\mu(F-F^{-\mathsf{T}})+\lambda\ln(\det F)F^{-\mathsf{T}}.

References

  1. [1]G. Allaire, Conception Optimale de Structures, vol. 58 of Mathématiques & Applications, Springer-Verlag, Berlin, 2007.
  2. [2]M. S. Alnæs, A. Logg, K. B. Ølgaard, M. E. Rognes, and G. N. Wells, Unified form language: A domain-specific language for weak formulations of partial differential equations, ACM Trans. Math. Softw., 40 (2014).
  3. [3]P. R. Amestoy, I. S. Duff, J. Koster, and J.-Y. L’Excellent, A fully asynchronous multifrontal solver using distributed dynamic scheduling, SIAM J. Matrix Anal. Appl., 23 (2001), pp. 15–41.
  4. [4]S. Balay, A. Abhyankar, M. Adams, J. Brown, P. Brune, K. Buschelman, L. Dalcin, V. Eijkhout, W. D. Gropp, D. Kaushik, M. G. Knepley, L. C. McInnes, K. Rupp, B. F. Smith, S. Zampini, and H. Zhang, PETSc users manual, Tech. Rep. ANL-95/11 - Revision 3.6, Argonne National Laboratory, 2015.
  5. [5]——, PETSc Web page. http://www.mcs.anl.gov/petsc, 2015.
  6. [6]S. Balay, W. D. Gropp, L. C. McInnes, and B. F. Smith, Efficient management of parallelism in object oriented numerical software libraries, in Modern Software Tools in Scientific Computing, E. Arge, A. M. Bruaset, and H. P. Langtangen, eds., Birkhäuser Press, 1997, pp. 163–202.
  7. [7]J. Bonet and R. D. Wood, Nonlinear Continuum Mechanics for Finite Element Analysis, Cambridge University Press, Cambridge, second ed., 2008.
  8. [8]N. Boullé, P. E. Farrell, and A. Paganini, Control of bifurcation structures using shape optimization, SIAM J. Sci. Comput., 44 (2022), pp. A57–A76.DOI
  9. [9]P. G. Ciarlet, Linear and Nonlinear Functional Analysis with Applications, vol. 130 of Other Titles in Applied Mathematics, Society for Industrial & Applied Mathematics, Philadelphia, USA, 1st ed., 2013.
  10. [10]M. C. Delfour and J.-P. Zolésio, Shapes and Geometries, vol. 22 of Advances in Design and Control, Society for Industrial & Applied Mathematics, Philadelphia, USA, second ed., 2011.
  11. [11]J. C. Dixon, The Shock Absorber Handbook, John Wiley & Sons, 2007.
  12. [12]C. H. Edwards, Jr., Advanced Calculus of Several Variables, Academic Press, New York-London, 1973.
  13. [13]P. E. Farrell, Computing multiple solutions of systems of nonlinear equations with deflation, in Proceedings of the International Congress of Mathematicians (2026), Vol. 7, Society for Industrial and Applied Mathematics and International Mathematical Union, Philadelphia, USA, 2026.
  14. [14]P. E. Farrell, D. A. Ham, S. W. Funke, and M. E. Rognes, Automated derivation of the adjoint of high-level transient finite element programs, SIAM J. Sci. Comput., 35 (2013), pp. C369–C393.
  15. [15]A. Hadj Slimane, P. E. Farrell, A. Paganini, and À. Ferrer, Control of the energy jump of multistable structures using shape optimization, GitHub repository, https://github.com/ArselaneHS/Control-of-the-energy-jump-of-multistable-structures-using-shape-optimization, 2026.
  16. [16]D. A. Ham, P. H. J. Kelly, L. Mitchell, C. J. Cotter, R. C. Kirby, K. Sagiyama, N. Bouziani, S. Vorderwuelbecke, T. J. Gregory, J. Betteridge, D. R. Shapero, R. W. Nixon-Hill, C. J. Ward, P. E. Farrell, P. D. Brubeck, I. Marsden, T. H. Gibson, M. Homolya, T. Sun, A. T. T. McRae, F. Luporini, A. Gregory, M. Lange, S. W. Funke, F. Rathgeber, G.-T. Bercea, and G. R. Markall, Firedrake User Manual, Imperial College London and University of Oxford and Baylor University and University of Washington, 2023.DOI
  17. [17]D. A. Ham, L. Mitchell, A. Paganini, and F. Wechsung, Automated shape differentiation in the unified form language, Struct. Multidiscip. Optim., 60 (2019), pp. 1813–1820.DOI
  18. [18]M. Hintermüller and T. Surowiec, First-order optimality conditions for elliptic mathematical programs with equilibrium constraints via variational analysis, SIAM J. Optim., 21 (2011), pp. 1561–1593.
  19. [19]D. Kouri and D. Ridzel, Rapid optimization library, 2014. https://www.osti.gov/biblio/code-3316.
  20. [20]S. G. Krantz and H. R. Parks, The Implicit Function Theorem, Modern Birkhäuser Classics, Birkhäuser/Springer, New York, 2013. Reprint of the 2003 edition.
  21. [21]S. Lang, Fundamentals of Differential Geometry, vol. 191 of Graduate Texts in Mathematics, Springer-Verlag, New York, 1999.
  22. [22]Z.-Q. Luo, J.-S. Pang, and D. Ralph, Mathematical Programs with Equilibrium Constraints, Cambridge University Press, Cambridge, 1996.
  23. [23]H. Ma, K. Wang, H. Zhao, Y. Hong, Y. Zhou, J. Xue, Q. Li, G. Wang, and B. Yan, Energy dissipation in multistable auxetic mechanical metamaterials, Compos. Struct., 304 (2023), p. 116410.DOI
  24. [24]H. Ma, K. Wang, H. Zhao, W. Shi, J. Xue, Y. Zhou, Q. Li, G. Wang, and B. Yan, Energy dissipation and shock isolation using novel metamaterials, Int. J. Mech. Sci., 228 (2022), p. 107464.
  25. [25]E. Medina, P. E. Farrell, K. Bertoldi, and C. H. Rycroft, Navigating the landscape of nonlinear mechanical metamaterials for advanced programmability, Phys. Rev. B, 101 (2020), p. 064101.
  26. [26]E. Medina, C. H. Rycroft, and K. Bertoldi, Nonlinear shape optimization of flexible mechanical metamaterials, Ext. Mech. Lett., 61 (2023), p. 12015.
  27. [27]S. K. Mitusch, S. W. Funke, and J. S. Dokken, dolfin-adjoint 2018.1: automated adjoints for FEniCS and Firedrake, J. Open Source Softw., 4 (2019), p. 1292.

Paper details

vlnc:W_25jdcxbk
Version
v1 · latest
Published
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.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents