betaFeedback

WorksComputer Science

Component-wise hyperreduction for nonlinear solid mechanics problems

Abstract

Hyperreduced nonlinear solid-mechanics components can be generated offline and reused as transferable building blocks across different assemblies, boundary conditions, meshes, material parameters, and constitutive models. We use proper orthogonal decomposition (POD) and energy conserving sampling and weighting (ECSW) on the component level and connect the substructures by mortar mesh tying. The POD modes and the ECSW weights and elements are computed offline from simulations of single components and the finite rigid body motions are treated by 12 additional rigid body modes per substructure. The numerical examples demonstrate errors below 1 % for large quasi-static assemblies while evaluating less than 10 % of the elements. The same component bases and ECSW elements are successfully reused for finite-strain viscoelastic dynamics, although they were trained only on elastic Neo-Hookean component simulations. These results indicate that component-wise hyperreduction can provide reusable reduced building blocks for modular nonlinear solid-mechanics simulations.

Introduction

In nonlinear solid mechanics, parametric projection based model order reduction (MOR) methods are widely used. Typical applications are computational homogenization, e.g. [15, 31, 45, 43, 36, 42], nonlinear vibrations, e.g. [33, 39, 35] and also history-dependent material behavior involving viscoelasticity, plasticity and damage, e.g. [34, 16, 22].

For nonlinear problems, a common method is the proper orthogonal decomposition (POD) [5]. In POD, full-order solutions for samples of input parameters are first computed in an offline stage. These so-called snapshots are then used to compute a projection matrix, using which the FOM can be projected onto a low-dimensional subspace. In the online stage, the reduced order model (ROM) can be solved for new input parameters. Because of the reduced dimension, the solution of the equation system is much faster than the FOM, but the assembly of the equation system still depends on the original size of the problem. To accelerate the assembly, different hyperreduction methods were developed in the past two decades. The main idea of hyperreduction methods is to approximate the nonlinear terms in the reduced system of equations by evaluating only a subset of the elements or integration points. There are two main families of hyperreduction method, the approximate-then-project and the project-then-approximate hyperreduction methods. In approximate-then-project methods, the nonlinear term is first approximated using sparse reconstruction and then projected onto the reduced space. The most popular method of this family is the discrete empirical interpolation method (DEIM) developed by Chaturantabut and Sorensen [ [6] and its unassembled version (UDEIM) developed by Tiso and Rixen [ [38]. The full nonlinear term is approximated by a linear combination of modes that are computed from the snapshots of the nonlinear term. The coefficients are obtained by evaluating the nonlinear term in a subset of the elements or integration points. In solid mechanics, DEIM has been used for example in Radermacher and Reese [ [30], Bonomi et al. [ [3], Ghavamian et al. [ [16], Kehls et al. [ [21]. In the project-then-approximate methods, the projected nonlinear terms of the reduced set of elements or integration points are weighted to approximate the full nonlinear term. The first method that can be classed as part of this family was developed in An et al. [ [1]. In Farhat et al. [ [12], also the projected element quantities are weighted to approximate the full projected nonlinear term. The difference to An et al. [ [1] is that the weights are computed by a greedy solution of a sparse non-negative least-squares problem. The empirical cubature method (ECM) developed in Hernández et al. [ [17] and its continuous version Hernández et al. [ [18] evaluates and weights the nonlinear terms in a subset of integration points. In both methods, the weights are computed by a greedy solution of a sparse non-negative least-squares problem. The difference is, that in ECSW the least-squares problem is formulated using element contributions to the projected internal forces, whereas ECM approximates the full quadrature by selecting and weighting individual integration point contributions. Yano and Patera [ [44] developed a method where the weights are computed by solving a linear programming problem. For nonlinear computational homogenization problems, recently, two new methods were developed. Wulfinghoff and Hauck [ [43], Wulfinghoff [ [42] developed a method where statistical representatives of the strain mode distribution are used instead of a subset of the integration points. This approach proved to be more efficient in computational homogenization problems. Recently, Faust and Scheunemann [ [13] developed a method, which uses a cluster-wise linearisation of the constitutive response around one cluster-wise reference point, at which stresses are evaluated.

All of these hyperreduction methods are tied to the full system used to generate the snapshots. Applying hyperreduction at the component level in substructured systems makes the approach spatially adaptive: component ROMs can be reused and assembled in different system configurations without recomputation. Because each substructure has fewer degrees of freedom than the full model, the offline cost is reduced. The design of the offline sampling strategy is challenging because the reduced models must handle varying boundary and interface coupling conditions.

To the authors’ knowledge, the oldest works on component-wise MOR for nonlinear solid mechanics problems are Barbić and Zhao [ [2], Kim and James [ [23] in the field of computer graphics. Barbić and Zhao [ [2] coupled the substructures using a penalty method. They solve the Newton-Euler equations for each substructure, where the rigid body motions are treated as independent degrees of freedom (DOF). To reduce the computational cost, they apply the hyperreduction method developed by An et al. [ [1] to each substructure. Kim and James [ [23] treat the rotation using the gradient and the Hessian of the rotation matrix. They couple the substructures for open-loop systems using Featherstone’s algorithm [14]. Both methods from computer graphics do not fulfill the tied contact conditions exactly. Zhou et al. [ [46] use a penalty method for coupling and POD modes of the substructures. In structural dynamics, there exist some methods that extend the well-established Craig-Bampton method for geometric nonlinearities [24, 25, 4]. Kuether et al. [ [24, 25], which extends the modal equations by using a quadratic and a cubic term. The nonlinear coefficients are computed by evaluating the nonlinear forces for the modal displacements. Bui et al. [ [4] uses static modal derivatives to capture the nonlinearities and a Taylor series expansion of the nonlinear forces.

Related contributions for nonlinear thermal problems were made by Ebrahimi and Yano [ [10, 11], who propose a component-wise hyperreduction method. They treated interface and interior degrees of freedom (DOFs) separately and developed a component-wise empirical quadrature method based on the work by Yano and Patera [ [44]. In the offline stage, they generated component snapshots from randomly sampled systems and tested the method on different systems assembled from the same components.

The present work extends the method developed by Ritzert et al. [ [32], where the substructures are reduced by a POD method and the coupling is done by a mortar method that fulfills the tied contact conditions exactly. The DOFs of the substructures are split into slave-side interface DOFs, master-side interface DOFs and interior DOFs. Only the interior and master side interface DOFs are reduced, the slave side interface DOFs are projected onto the reduced space of the master side interface DOFs. The mortar coupling has the advantage that non-matching interfaces can be handled (see e.g. [41, 29, 26, 28]). The novelty of this paper is that we extend the method by a component-wise ECSW hyperreduction method. The reduced sets of elements and the corresponding weights are computed in the offline stage for each substructure type individually and do not require any information about the coupling conditions. Additionally, we show that finite rotations can be represented by a linear combination of 9 modes, which can be derived from the general description of the rigid body rotation of a 3D body. The advantage of this approach is the simplicity compared to the treatment of rotations by Kim and James [ [23] and Barbić and Zhao [ [2].

Research gap Existing hyperreduction methods for nonlinear solid mechanics are mostly constructed for one fixed, fully assembled system. This limits their reuse in modular simulations, where the same component may appear in different assemblies, boundary conditions, or material models. A key open question is therefore whether hyperreduced component models can be trained independently and still be assembled reliably into new nonlinear systems.

Aim of this study This study develops a component-wise hyperreduction framework for nonlinear solid mechanics. Component level POD bases and ECSW element weights are computed offline for individual substructures and then reused in different assemblies coupled by mortar mesh tying. The aim is to show that such hyperreduced components can serve as transferable building blocks while retaining accuracy and computational efficiency under changing assemblies, boundary conditions, and material behavior.

Component-wise model order reduction

The component-wise model order reduction method (CWMOR) (see Ritzert et al. [ [32]) is based on a mortar tied-contact full-order model where the slave-side contact DOFs and the Lagrange multipliers are condensed out. The remaining DOFs are projected into their respective substructure reduced spaces to obtain a reduced system of equations. In this chapter, we briefly explain the full-order model and the CWMOR method. We use a different notation than in Ritzert et al. [ [32] to have a more compact notation for the further derivations.

Full-order model

The full-order model (FOM) is governed by a tied-contact problem discretized using the finite element method.

Weak formulation

The weak form of the problem is

∑i=1nsδginti(u,u¨,δu)+∑j=1ncδgc,uj(λ,δuc1,δuc2)=0∑j=1ncδgc,λj(uc1,uc2,δλ)=0(1)\sum_{i=1}^{n_s}\delta g_{\mathrm{int}}^{i}(\mathbf{u},\ddot{\mathbf{u}},\delta\mathbf{u})+\sum_{j=1}^{n_c}\delta g_{c,u}^{j}(\boldsymbol{\lambda},\delta\mathbf{u}_{c}^{1},\delta\mathbf{u}_{c}^{2})=0 \\ \sum_{j=1}^{n_c}\delta g_{c,\lambda}^{j}(\mathbf{u}_{c}^{1},\mathbf{u}_{c}^{2},\delta\boldsymbol{\lambda})=0 \tag*{(1)}

where δginti\delta g_{\mathrm{int}}^{i} is the virtual work done by the internal forces of substructure ii and δgc,uj\delta g_{c,u}^{j} and δgc,λj\delta g_{c,\lambda}^{j} are the virtual work contributions of the tied-contact conditions of interface jj. The virtual work done by the internal forces of each substructure ii with its domain Ω0i\Omega_{0}^{i}, the inertia forces, and the external forces is

δginti=∫Ω0i(S:δE+ρ0u¨⋅δu−ρ0b⋅δu) dV−∫Γtδu⋅t dA(2)\delta g_{\mathrm{int}}^{i}=\int_{\Omega_{0}^{i}}\left(\mathbf{S}:\delta\mathbf{E}+\rho_{0}\ddot{\mathbf{u}}\cdot\delta\mathbf{u}-\rho_{0}\mathbf{b}\cdot\delta\mathbf{u}\right)\,\mathrm{d}V-\int_{\Gamma_{t}}\delta\mathbf{u}\cdot\mathbf{t}\,\mathrm{d}A \tag*{(2)}

where S\mathbf{S} is the second Piola-Kirchhoff stress tensor, δE\delta\mathbf{E} is the variation of the Green-Lagrange strain tensor, ρ0\rho_{0} is the mass density in the reference configuration, b\mathbf{b} is the body force per unit mass and t\mathbf{t} is the traction vector on the Neumann boundary Γt\Gamma_{t}. The virtual work contributions of the tied contact conditions at interface jj are defined as

δgc,uj=∫Γcj(δuc1−δuc2)⋅λ dA(3)\delta g_{c,u}^{j}=\int_{\Gamma_{c}^{j}}\left(\delta\mathbf{u}_{c}^{1}-\delta\mathbf{u}_{c}^{2}\right)\cdot\boldsymbol{\lambda}\,\mathrm{d}A \tag*{(3)}
δgc,λj=∫Γc1,2δλ⋅(uc1−uc2) dA(4)\delta g^{j}_{c,\lambda}=\int_{\Gamma^{1,2}_{c}}\delta\lambda\cdot\left(u^{1}_{c}-u^{2}_{c}\right)\,\mathrm{d}A \tag*{(4)}

Here, Γc1,2\Gamma^{1,2}_{c} is the contact interface between substructure 1 and 2, uc1u^{1}_{c} and uc2u^{2}_{c} are the displacements on the slave and master side of the interface, respectively, and λ\lambda are the Lagrange multipliers that enforce the tied contact condition.

Discretization

We use the finite element method to discretize the problem in space and the Newmark-β\beta method for the time discretization, producing a nonlinear residual vector GiG^{i} for each substructure. The residual vector is the sum of the internal forces, the inertial forces, and the external forces

Gi=MρiU¨i+Ri(Ui)−Fexti(5)G^{i}=M^{i}_{\rho}\ddot{U}^{i}+R^{i}\left(U^{i}\right)-F^{i}_{\mathrm{ext}} \tag*{(5)}

where MρiM^{i}_{\rho} is the mass matrix, RiR^{i} is the vector of internal forces, and FextiF^{i}_{\mathrm{ext}} is the vector of external dead forces. The current nodal velocity and acceleration vectors U˙i\dot{U}^{i} and U¨i\ddot{U}^{i} are computed by the Newmark-β\beta method

U˙i=γβΔt(Ui−Uni)−γ−ββU˙ni−Δtγ−2β2βU¨ni\dot{U}^{i}=\frac{\gamma}{\beta\Delta t}\left(U^{i}-U^{i}_{n}\right)-\frac{\gamma-\beta}{\beta}\dot{U}^{i}_{n}-\Delta t\frac{\gamma-2\beta}{2\beta}\ddot{U}^{i}_{n}
U¨i=1βΔt2(Ui−Uni)−1βΔtU˙ni−1−2β2βU¨ni(6)\ddot{U}^{i}=\frac{1}{\beta\Delta t^{2}}\left(U^{i}-U^{i}_{n}\right)-\frac{1}{\beta\Delta t}\dot{U}^{i}_{n}-\frac{1-2\beta}{2\beta}\ddot{U}^{i}_{n} \tag*{(6)}

where β\beta and γ\gamma are the Newmark parameters, that we chose as β=0.25\beta=0.25 and γ=0.5\gamma=0.5 in this article.

Because of the nonlinear material behavior and the finite strain kinematics, the residual vector GiG^{i} is a nonlinear function of the displacements UiU^{i}. The system is solved by the Newton-Raphson method, where the displacement vector is updated iteratively by Ui←Uni+ΔUiU^{i}\leftarrow U^{i}_{n}+\Delta U^{i} until convergence is reached. In each iteration, the following linear system has to be solved for the displacement increment ΔUi\Delta U^{i} and the Lagrange multipliers Λj\Lambda^{j}

∑i=1ns(Gi(t,Uni)+Ki(Uni)ΔUi)+∑j=1nc(DjTΛnj−MjTΛnj)=0(7)\sum_{i=1}^{n_s}\left(G^{i}\left(t,U^{i}_{n}\right)+K^{i}\left(U^{i}_{n}\right)\Delta U^{i}\right)+\sum_{j=1}^{n_c}\left(D^{T}_{j}\Lambda^{j}_{n}-M^{T}_{j}\Lambda^{j}_{n}\right)=0 \tag*{(7)}
∑j=1nc(DjΔUc,j1−MjΔUc,j2)=0(8)\sum_{j=1}^{n_c}\left(D_{j}\Delta U^{1}_{c,j}-M_{j}\Delta U^{2}_{c,j}\right)=0 \tag*{(8)}

The matrices DjD^{j} and MjM^{j} stem from the finite element discretization of the tied contact conditions using the mortar method (see e.g. Wohlmuth [ [41], Puso [ [29], Laursen et al. [ [26], Popp and Wriggers [ [28]). The mortar method allows for non-matching meshes on the contact interface. The tangential stiffness matrix KiK^{i} is the Jacobian of the residual vector GiG^{i} with respect to the displacements UiU^{i}. For the dynamic case, the Newmark-beta time discretization results in the following expression for the tangential stiffness matrix

Ki=∂Gi∂Ui=1βΔt2Mρi+KRi(Ui).(9)K^{i}=\frac{\partial G^{i}}{\partial U^{i}}=\frac{1}{\beta\Delta t^{2}}M^{i}_{\rho}+K^{i}_{R}\left(U^{i}\right). \tag*{(9)}

Here, KRiK^{i}_{R} is the partial derivative of the internal forces RiR^{i} with respect to the displacements UiU^{i}.

In the following, we consider the case of two substructures with one interface, to have a simple notation for the further derivations. The master substructure will be denoted by a superscript MM and the slave substructure by a superscript SS. The extension to more substructures and interfaces is straightforward. The linear system that has to be solved in each iteration of the Newton-Raphson method, is then given by

[KIIMKICM000KCIMKCCM00−MT00KIISKICS000KCISKCCSDT0−M0D0][ΔUIMΔUCMΔUISΔUCSΔΛ]=−[GIMGCMGISGCS0](10)\begin{bmatrix} K^{M}_{II} & K^{M}_{IC} & 0 & 0 & 0 \\ K^{M}_{CI} & K^{M}_{CC} & 0 & 0 & -M^{T} \\ 0 & 0 & K^{S}_{II} & K^{S}_{IC} & 0 \\ 0 & 0 & K^{S}_{CI} & K^{S}_{CC} & D^{T} \\ 0 & -M & 0 & D & 0 \end{bmatrix} \begin{bmatrix} \Delta U^{M}_{I} \\ \Delta U^{M}_{C} \\ \Delta U^{S}_{I} \\ \Delta U^{S}_{C} \\ \Delta\Lambda \end{bmatrix} = - \begin{bmatrix} G^{M}_{I} \\ G^{M}_{C} \\ G^{S}_{I} \\ G^{S}_{C} \\ 0 \tag*{(10)} \end{bmatrix}

The DOFs of each substructure are separated into internal DOFs (denoted by the subscript II) and contact DOFs (denoted by the subscript CC). The stiffness matrix and the residual vector are separated into blocks corresponding to the internal and contact DOFs. The mortar matrices DD and MM are only defined for the contact interface DOFs.

The system of equations 10 can be condensed by expressing the slave side contact DOFs by the master side contact DOFs

ΔUCS=D−1MΔUCM=PΔUCM(11)\Delta U_C^S = D^{-1}M\Delta U_C^M = P\Delta U_C^M \tag*{(11)}

Additionally, the Lagrange multiplier increments ΔΛ\Delta\Lambda can be expressed as

ΔΛ=−D−T(GCS+KCISΔUIS+KCCSΔUCS)(12)\Delta\Lambda= -D^{-T}\left(G_C^S + K_{CI}^S\Delta U_I^S + K_{CC}^S\Delta U_C^S\right) \tag*{(12)}

To evaluate Eq. (11) and Eq. (12), we need the inverse of the mortar matrix DD. By choosing dual shape functions for the Lagrange multipliers the mortar matrix DD becomes diagonal and is therefore easy to invert, as described e.g. in the works by Wohlmuth [ [41], Popp and Wiggers [ [28], Ritzert et al. [ [32]. With the above equations, the system of equations Eq. (10) can be condensed to

[KIIMKICM0KCIMKCCM+PTKCCSPPTKCIS0KICSPKIIS][ΔUIMΔUCMΔUIS]=−[GIMGCM+PTGCSGIS](13)\begin{bmatrix} K_{II}^M & K_{IC}^M & 0 \\ K_{CI}^M & K_{CC}^M + P^T K_{CC}^S P & P^T K_{CI}^S \\ 0 & K_{IC}^S P & K_{II}^S \end{bmatrix} \begin{bmatrix} \Delta U_I^M \\ \Delta U_C^M \\ \Delta U_I^S \end{bmatrix} = - \begin{bmatrix} G_I^M \\ G_C^M + P^T G_C^S \\ G_I^S \tag*{(13)} \end{bmatrix}

where the discrete interface coupling operator P=D−1MP = D^{-1}M is used.

For the application of MOR methods it is advantageous to rewrite the condensed system of Eq. (13) using a transformation matrix TT. The transformation matrix TT maps the condensed displacement vector ΔUcond\Delta U_{\mathrm{cond}} to the original displacement vector ΔU\Delta U by

[UIMUCMUISUCS]=[I000I000I0P0]⏟T[UIMUCMUIS](14)\begin{bmatrix} U_I^M \\ U_C^M \\ U_I^S \\ U_C^S \end{bmatrix} = \underbrace{\begin{bmatrix} I & 0 & 0 \\ 0 & I & 0 \\ 0 & 0 & I \\ 0 & P & 0 \end{bmatrix}}_{T} \begin{bmatrix} U_I^M \\ U_C^M \\ U_I^S \tag*{(14)} \end{bmatrix}

The transpose of the transformation matrix TTT^T maps the original residual vector GG to the condensed residual vector GcondG_{\mathrm{cond}} by

[GIMGCM+PTGCSGIS]=[I0000I0PT00I0]⏟TT[GIMGCMGISGCS](15)\begin{bmatrix} G_I^M \\ G_C^M + P^T G_C^S \\ G_I^S \end{bmatrix} = \underbrace{\begin{bmatrix} I & 0 & 0 & 0 \\ 0 & I & 0 & P^T \\ 0 & 0 & I & 0 \end{bmatrix}}_{T^T} \begin{bmatrix} G_I^M \\ G_C^M \\ G_I^S \\ G_C^S \tag*{(15)} \end{bmatrix}

The residual vector GG contains the residuals of both substructures SS and MM. The condensed stiffness matrix KcondK_{\mathrm{cond}} can be obtained by

Kcond=[I0000I0PT00I0][KIIMKICM00KCIMKCCM0000KIISKICS00KCISKCCS][I000I000I0P0](16)K_{\mathrm{cond}} = \begin{bmatrix} I & 0 & 0 & 0 \\ 0 & I & 0 & P^T \\ 0 & 0 & I & 0 \end{bmatrix} \begin{bmatrix} K_{II}^M & K_{IC}^M & 0 & 0 \\ K_{CI}^M & K_{CC}^M & 0 & 0 \\ 0 & 0 & K_{II}^S & K_{IC}^S \\ 0 & 0 & K_{CI}^S & K_{CC}^S \end{bmatrix} \begin{bmatrix} I & 0 & 0 \\ 0 & I & 0 \\ 0 & 0 & I \\ 0 & P & 0 \end{bmatrix} \tag*{(16)}

The transformation matrix TT will later be used to derive the reduced system of equations.

Component-wise model order reduction

In this article, we derive a component-wise hyperreduction method for nonlinear solid mechanics problems. This hyperreduction method is an extension of the component-wise model order reduction method that was introduced in our previous workRitzert et al. [ [32]. The idea of the component-wise model order reduction method is to compute a projection matrix for each substructure and to use these bases to derive a reduced system of equations for the condensed system (13).

In this work, the projection matrices are computed for each substructure separately by using the POD method on a matrix composed of displacement snapshots, see e.g. [5]. For each substructure ii, we compute ll snapshots of the displacements UiU^i in an offline stage by solving the substructure problems for different loading and boundary conditions. These snapshots are collected in a snapshot matrix

SSnapi=[U1i,U2i,…,Uli].S_{\mathrm{Snap}}^i = [U_1^i,U_2^i,\ldots,U_l^i].

To compute the projection matrix Ψi\Psi^i for substructure ii, the snapshot matrix is decomposed by the singular value decomposition (SVD).

SSnapi=ViXiWi(17)S_{\mathrm{Snap}}^i = V^i X^i W^i \tag*{(17)}

The first mim_i left singular vectors are used as the projection matrix Ψi\Psi^i for substructure ii

Ψi=[V1i,V2i,…,Vmii].(18)\Psi^i = [V_1^i,V_2^i,\ldots,V_{m_i}^i]. \tag*{(18)}

With these substructure projection matrices Ψi\Psi^i, we can derive a reduced system of equations for the full-order condensed system Eq. (13).

Each projection matrix Ψi\Psi^i is split into internal and master contact DOF parts ΨIi\Psi_I^i and ΨCi\Psi_C^i. All parts of the projection matrix that correspond to the slave-side contact DOFs are set to zero, since these DOFs are expressed by the master side contact DOFs. To apply Dirichlet boundary conditions, the corresponding rows of the projection matrix are also set to zero. Considering the case of two substructures the condensed displacement vector is approximated as

[UIMUCMUCSUIS]≈[ΨIM0ΨCM0000ΨIS][aMaS](19)\begin{bmatrix} U_I^M\\ U_C^M\\ U_C^S\\ U_I^S \end{bmatrix} \approx \begin{bmatrix} \Psi_I^M & 0\\ \Psi_C^M & 0\\ 0 & 0\\ 0 & \Psi_I^S \end{bmatrix} \begin{bmatrix} a^M\\ a^S \tag*{(19)} \end{bmatrix}

Here, aMa^M and aSa^S are the reduced displacement vectors of the master and slave substructure, respectively. Using the transformation matrix TT from Eq. (14) the original displacement vector can be approximated as

\begin{bmatrix} U_I^M\\ U_M^M\\ U_C^S\\ U_I^S \end{bmatrix} \approx \begin{bmatrix} I & 0 & 0\\ 0 & I & 0\\ 0 & 0 & I\\ 0 & P & 0 \end{bmatrix} \begin{bmatrix} \Psi_I^M & 0\\ \Psi_C^M & 0\\ 0 & \Psi_I^S \end{bmatrix} \begin{bmatrix} a^M\\ a^S \end{bmatrix} = \begin{bmatrix} \Psi_I^M & 0\\ \Psi_C^M & 0\\ 0 & \Psi_I^S\\ P\Psi_C^M & 0 \end{bmatrix} \begin{bmatrix} a^M\\ a^S \end{bmatrix} = \Phi a \tag*{(20)}\end{bmatrix}

The global projection matrix Φ\Phi maps the reduced displacement vector aa to the original displacement vector UU. For systems with many substructures, it becomes a sparse matrix, since the projection matrices of each substructure only have nonzero entries corresponding to the DOFs of this substructure and the neighboring master substructures.

We can compute the reduced stiffness matrix K^cond=ΦTKΦ\hat{K}_{\mathrm{cond}} = \Phi^T K\Phi and the reduced residual vector G^cond=ΦTG\hat{G}_{\mathrm{cond}} = \Phi^T G by a Galerkin projection of the global stiffness matrix and residual with the global projection matrix Φ\Phi. In matrix notation, the reduced condensed stiffness matrix is given by

K^cond=[ΨIMTΨCMT0ΨCMTPT00ΨIST0][KIIMKICM00KCIMKCCM0000KIISKICS00KCISKCCS][ΨIM0ΨCM000PΨCM0](21)\hat{K}_{\mathrm{cond}} = \begin{bmatrix} \Psi_I^{M T} & \Psi_C^{M T} & 0 & \Psi_C^{M T}P^T\\ 0 & 0 & \Psi_I^{S T} & 0 \end{bmatrix} \begin{bmatrix} K_{II}^M & K_{IC}^M & 0 & 0\\ K_{CI}^M & K_{CC}^M & 0 & 0\\ 0 & 0 & K_{II}^S & K_{IC}^S\\ 0 & 0 & K_{CI}^S & K_{CC}^S \end{bmatrix} \begin{bmatrix} \Psi_I^M & 0\\ \Psi_C^M & 0\\ 0 & 0\\ P\Psi_C^M & 0 \end{bmatrix} \tag*{(21)}

and the reduced condensed residual vector can be written as

G^cond=[ΨIMTΨCMT0ΨCMTPT00ΨIST0][GIMGCMGISGCS](22)\hat{G}_{\mathrm{cond}} = \begin{bmatrix} \Psi_I^{M T} & \Psi_C^{M T} & 0 & \Psi_C^{M T}P^T\\ 0 & 0 & \Psi_I^{S T} & 0 \end{bmatrix} \begin{bmatrix} G_I^M\\ G_C^M\\ G_I^S\\ G_C^S \end{bmatrix} \tag*{(22)}

Alternatively, the reduced quantities can be written as a sum over all elements ee of the system, which is the form that is used for the implementation of the hyperreduction method in the next section, i.e.

G^cond=∑e∈EΦeTGe(23)\hat{G}_{\mathrm{cond}} = \sum_{e\in\mathcal{E}} \Phi_e^T G_e \tag*{(23)}
K^cond=∑e∈EΦeTKeΦe(24)\hat{K}_{\mathrm{cond}}=\sum_{e\in\mathcal{E}}\Phi_e^T K_e\Phi_e \tag*{(24)}

Here, Φe\Phi_e, GeG_e, and KeK_e denote the parts of the global projection matrix Φ\Phi, residual vector GG, and stiffness matrix KK, associated with element ee. The set of all elements E\mathcal{E} is the union of the sets of elements of each substructure E=⋃insEi\mathcal{E}=\bigcup_{i}^{n_s}\mathcal{E}_i.

The reduced residual and reduced stiffness matrix contain internal forces, inertia forces. Additionally, the reduced residual contains external forces. Only the internal forces RR and the tangential stiffness matrix KK depend nonlinearly on the displacements and have to be updated in each Newton-Raphson iteration. We split the reduced residual into three parts

G^cond=∑e∈EΦeTRe+∑e∈EΦeTMρ,eΦeu¨−∑e∈EΦeTFext,e=R^cond+M^ρ,conda¨−F^ext,cond(25)\hat{G}_{\mathrm{cond}}=\sum_{e\in\mathcal{E}}\Phi_e^T R_e+\sum_{e\in\mathcal{E}}\Phi_e^T M_{\rho,e}\Phi_e\ddot{u}-\sum_{e\in\mathcal{E}}\Phi_e^T F_{\mathrm{ext},e}=\hat{R}_{\mathrm{cond}}+\hat{M}_{\rho,\mathrm{cond}}\ddot{a}-\hat{F}_{\mathrm{ext},\mathrm{cond}} \tag*{(25)}

The reduced internal force part R^cond\hat{R}_{\mathrm{cond}} is updated in each iteration of the Newton-Raphson method, while the reduced inertia part M^ρ,cond\hat{M}_{\rho,\mathrm{cond}} can be precomputed in the offline stage or at the beginning of the online stage. It is multiplied with the reduced acceleration vector a¨\ddot{a} to obtain the inertia forces in the reduced system. The reduced acceleration vector a¨\ddot{a} is computed by projecting the Newmark-β\beta equations Eq. (6) into the reduced space by the global projection matrix Φ\Phi. Analogously, the reduced stiffness matrix is also split into a stiffness part, related to the internal forces, and an inertia part, related to the reduced mass matrix

K^cond=∑e∈EΦeTKR,eΦe+1βΔt2∑e∈EΦeTMρ,eΦe=K^R,cond+1βΔt2M^ρ,cond.(26)\hat{K}_{\mathrm{cond}}=\sum_{e\in\mathcal{E}}\Phi_e^T K_{R,e}\Phi_e+\frac{1}{\beta\Delta t^2}\sum_{e\in\mathcal{E}}\Phi_e^T M_{\rho,e}\Phi_e=\hat{K}_{R,\mathrm{cond}}+\frac{1}{\beta\Delta t^2}\hat{M}_{\rho,\mathrm{cond}}. \tag*{(26)}

This split into internal forces and inertia forces is used for the implementation of the hyperreduction method, since only the part related to the internal forces has to be updated in each iteration of the Newton-Raphson method.

Component-wise hyperreduction by energy conserving sampling and weighting

The reduced system derived in section 2.2 still depends on the original size of the problem because all elements have to be evaluated to assemble the reduced internal force vector R^cond\hat{R}_{\mathrm{cond}} and the reduced tangential stiffness matrix K^cond\hat{K}_{\mathrm{cond}}. Hyperreduction techniques such as DEIM [6], ECSW [1, 12], or ECM [17] reduce this effort by evaluating only a reduced number of elements to approximate R^cond\hat{R}_{\mathrm{cond}} and K^cond\hat{K}_{\mathrm{cond}}. In this work we use ECSW for each component because in contrast to DEIM it does not require additional snapshots for the internal force vectors. DEIM does not preserve the structure of the equation system and can lead to numerical instabilities, especially for large rotations because of the split of the forces into linear and nonlinear parts Rutzmoser [2018]. Compared to ECM, ECSW was easier to implement into our existing finite element code, since it works with element level quantities, while ECM works with Gauss point level quantities. In preliminary investigations, ECM does not have significant advantages over ECSW in terms of accuracy and computational efficiency. Since ECSW has advantages in terms of implementation since it works with element level quantities and is widely used in the literature, ECSW is selected as the hyperreduction method for this work.

The component-wise ECSW approximation is not identical to applying ECSW to the assembled system, because the online projection operator contains interface contributions induced by the mortar mesh coupling. The approximation is expected to be accurate if the offline boundary parametrization spans the interface deformation states encountered online. This assumption is tested in the numerical examples (§ 5) by applying the same component weights to assemblies and loading conditions not used during training. In § 3.1 we discuss essential ingredients of the standard ECSW method (§ 3.1) and then present the component-wise extension in § 3.2.

Energy conserving sampling and weighting

In our implementation we follow the descriptions of Farhat et al. [2015] and Rutzmoser [2018]. In this section we give a brief overview of the method; for more details the reader is kindly referred to the original articles. The idea of the ECSW method is to approximate the virtual work of the internal forces in the reduced system by a weighted sum of the element virtual work contributions of a reduced set of elements E~\widetilde{\mathcal{E}}. The total virtual work of the reduced system is the sum of the element virtual work contributions over the whole set of elements E\mathcal{E}. This approximation is written as

δaTR~=δaT∑e∈EΨeTRe≈δaT∑e∈E~weΨeTRe,(27)\delta a^{T}\widetilde{R}=\delta a^{T}\sum_{e\in\mathcal{E}}\Psi_{e}^{T}R_{e}\approx\delta a^{T}\sum_{e\in\widetilde{\mathcal{E}}}w_{e}\Psi_{e}^{T}R_{e}, \tag*{(27)}

where wew_e are the element weights and Ψe\Psi_e is the part of the projection matrix Ψ\Psi associated with element ee. The virtual work equality Eq. (27) has to hold for arbitrary virtual reduced displacements δa\delta a, leading to

∑e∈EΨeTRe≈∑e∈E~weΨeTRe.(28)\sum_{e\in\mathcal{E}}\Psi_{e}^{T}R_{e}\approx\sum_{e\in\widetilde{\mathcal{E}}}w_{e}\Psi_{e}^{T}R_{e}. \tag*{(28)}

The weights wew_e can be computed by enforcing Eq. (28) for a set of displacement snapshots SSnap=[U1,U2,…,Ul]S_{\mathrm{Snap}}=[U_1,U_2,\ldots,U_l]. In this work, we only use converged snapshots. Since we search for a sparse solution, we minimize the zero-norm of the weight vector ∥w∥0\lVert w\rVert_0, which leads to the constrained optimization problem

w∗=arg⁡min⁡∥w∥0, subject to ∥Yw−b∥2 and w≥0.(29)w^{*}=\arg\min\lVert w\rVert_{0},\ \text{subject to }\lVert Yw-b\rVert^{2}\ \text{and}\ w\geq0. \tag*{(29)}

An exact solution of this is computationally infeasible, instead we use a greedy algorithm to compute an approximate solution. This algorithm, the sparse non-negative least squares algorithm (sNNLS), is described in the work by Farhat et al. [ [12]. The algorithm successively adds promising candidate elements, recomputes weights, and terminates when the convergence criterion

∥Yw−b∥≤τ∥b∥(30)\lVert Yw-b\rVert\leq\tau\lVert b\rVert \tag*{(30)}

is reached. The accuracy of the solution can be controlled by the tolerance τ\tau. A smaller value of τ\tau leads to a larger number of nonzero entries in ww. Consequently, a higher number of elements has to be evaluated in the online stage.

The quantities YY, ww and bb are

Y=[Ψ1TR1(ΩU1)⋯ΨnETRnE(ΩU1)⋮⋱⋮Ψ1TR1(ΩUl)⋯ΨnETRnE(ΩUl)](31)Y=\begin{bmatrix} \Psi_{1}^{T}R_{1}(\Omega U_{1}) & \cdots& \Psi_{n_{\mathcal{E}}}^{T}R_{n_{\mathcal{E}}}(\Omega U_{1})\\ \vdots& \ddots& \vdots\\ \Psi_{1}^{T}R_{1}(\Omega U_{l}) & \cdots& \Psi_{n_{\mathcal{E}}}^{T}R_{n_{\mathcal{E}}}(\Omega U_{l}) \end{bmatrix} \tag*{(31)}
w=[w1⋮wnE](32)w=\begin{bmatrix} w_{1}\\ \vdots\\ w_{n_{\mathcal{E}}} \end{bmatrix} \tag*{(32)}
b=[∑e∈EΨeTRe(ΩU1)⋮∑e∈EΨeTRe(ΩUl)](33)b=\begin{bmatrix} \sum_{e\in\mathcal{E}}\Psi_{e}^{T}R_{e}(\Omega U_{1})\\ \vdots\\ \sum_{e\in\mathcal{E}}\Psi_{e}^{T}R_{e}(\Omega U_{l}) \end{bmatrix} \tag*{(33)}

The operator Ω\Omega is defined as Ω=ΨΨT\Omega=\Psi\Psi^{T}. It computes the orthogonal projection of a snapshot UU onto the subspace spanned by Ψ\Psi. In each row of the matrix YY, the element contributions ΨeTRe\Psi_{e}^{T}R_{e} of all elements ee are evaluated for one snapshot. In the vector bb the sum of all element contributions ∑e∈EΨeTRe\sum_{e\in\mathcal{E}}\Psi_{e}^{T}R_{e} is evaluated for each snapshot. The weight vector ww is computed by the sNNLS such that the weighted sum of the element contributions in YY approximates the sum of all element contributions in bb for all snapshots. Each row in the equation system Yw=bYw=b is Eq. (28) evaluated for one snapshot.

With the weights ww, the hyperreduced internal force vector R~\widetilde{R} and the hyperreduced tangential stiffness matrix K~R\widetilde{K}_{R} can be constructed

R~≈∑e∈E~weΨeTRe(34)\widetilde{R}\approx\sum_{e\in\widetilde{\mathcal{E}}}w_{e}\Psi_{e}^{T}R_{e} \tag*{(34)}
K~R≈∑e∈E~weΨeTKeΨe(35)\widetilde{K}_{R}\approx\sum_{e\in\widetilde{\mathcal{E}}}w_{e}\Psi_{e}^{T}K_{e}\Psi_{e} \tag*{(35)}

The above equations are weighted sums over the reduced set of elements E~\widetilde{\mathcal{E}}, which contains all elements with nonzero weights: E~={e∈E∣we>0}\widetilde{\mathcal{E}}=\{e\in\mathcal{E}\mid w_e>0\}. In the following, hyperreduced quantities that are approximated by ECSW are denoted by a tilde ⋅~\widetilde{\cdot}.

Application to the component-wise MOR method

In the standard ECSW method, the projection matrix Ψ\boldsymbol{\Psi}, the weights w\boldsymbol{w}, and the reduced set of elements Eˉ\bar{\mathcal{E}} are computed for the whole system. In the component-wise model order reduction method, this cannot be done, since the tied-contact conditions and the Dirichlet conditions applied to the assembled system are not known beforehand. Additionally, there exist no snapshots for the whole system, which makes offline weight computation for the whole system impossible. We propose in this paper to compute the projection matrix, the weights, and the reduced sets of elements for each substructure separately from the substructure snapshots. The weights and the reduced sets of elements can be combined to get the weights and the reduced set of elements for the whole system. The latter are then used for the rapid online evaluation of the reduced internal force vector and the reduced tangential stiffness matrix.

For each substructure ii, we compute a projection matrix Ψi\boldsymbol{\Psi}^{i}, the ECSW weights wi\boldsymbol{w}^{i} and the reduced set of elements Eˉi\bar{\mathcal{E}}^{i} from the substructure snapshots. The reduced sets of elements of the substructures are then combined to get the reduced set of elements for the whole system E~=⋃i=1nsEˉi\tilde{\mathcal{E}}=\bigcup_{i=1}^{n_s}\bar{\mathcal{E}}^{i} and the weights are combined to get the weights for the whole system w=[w1w2…wns]\boldsymbol{w}=\left[\boldsymbol{w}^{1}\quad\boldsymbol{w}^{2}\quad\ldots\quad\boldsymbol{w}^{n_s}\right]. With these weights, the reduced internal force vector, defined in Eq. (36) and the reduced tangential stiffness matrix, defined in Eq. (37) can be approximated by a weighted sum over the reduced set of elements E~\tilde{\mathcal{E}} as described in the previous section

R~cond≈∑e∈E~weΦeTRe(36)\widetilde{\boldsymbol{R}}_{\mathrm{cond}}\approx\sum_{e\in\tilde{\mathcal{E}}}w_e\boldsymbol{\Phi}_e^{T}\boldsymbol{R}_e \tag*{(36)}
K~R,cond≈∑e∈E~weΦeTKeΦe(37)\widetilde{\boldsymbol{K}}_{R,\mathrm{cond}}\approx\sum_{e\in\tilde{\mathcal{E}}}w_e\boldsymbol{\Phi}_e^{T}\boldsymbol{K}_e\boldsymbol{\Phi}_e \tag*{(37)}

This is an approximation of the reduced internal forces and the reduced tangential stiffness parts derived in Eq. (25) and Eq. (26) by a weighted sum over the combined reduced set of elements E~\tilde{\mathcal{E}}. The projection matrix Φ\boldsymbol{\Phi} is the global projection matrix defined in Eq. (20) and contains the projection matrices of all substructures as well as the tied contact conditions. As a consequence, the slave side contact DOFs of all elements in the reduced set of elements are projected by the projection matrix of the master side.

Computation of snapshots

Snapshot generation

In this work, we compute the snapshots of the substructures by performing simulations of single substructures with parametrized Dirichlet boundary conditions at the possible interfaces and boundary faces. The snapshots are computed by a Latin Hypercube Sampling (LHS) (cf. [27, 37]) of the parameter space of the boundary conditions. This approach was also used by Ritzert et al. [ [32].

In other works, the snapshots are computed by simulating a small system of a few substructures with parametrized boundary conditions. This is sometimes called oversampling strategy, see e.g. [9, 10]. The advantage of the oversampling strategy is that the snapshots are computed in a realistic setting. The interface deformations are influenced by the neighboring substructures, which is also the case in the systems used in the online phase. The training systems have to be chosen such that different coupling scenarios are covered. The disadvantage is that the snapshot generation is more expensive compared to computing snapshots of single substructures. Computing snapshots of single substructures is computationally cheaper than the oversampling strategy, but the boundary conditions have to be parametrized to span the linear subspace in which the possible coupling scenarios lie.

In an assembled system, interface deformations are dominated by translations and rotations of the interfaces. Additionally, the interfaces undergo deformations in the normal- and tangential directions. In this work, we compute the displacement boundary conditions by applying parametrized translations and rotations to the possible interfaces of the substructure. To allow for normal and tangential deformations, we use two different strategies. The first strategy is applied for two-dimensional problems, it is used later in § 5.1.1. There we use three different sets of boundary conditions for each sample. The first set suppresses both normal deformation and tangential deformations. The second set permits normal deformation, while the third set permits tangential deformations. For three-dimensional problems we use the second strategy (see § 5.1.2), where elastic blocks are attached to the interfaces of the substructure. The displacement boundary conditions are applied to the outer edges of these blocks instead of directly to the interfaces. As a result, the interfaces can deform in the normal direction and contract in tangential directions. For three-dimensional problems we have two tangential and one normal boundary condition direction that can all be turned on and off. The first strategy leads to seven different sets of boundary conditions for each sample, because all possible combinations of activated and inactivated boundary conditions are used. This would significantly increase the computational cost of the snapshot generation.

A challenge is to determine the range of the parameters for the boundary conditions. The ranges can be chosen by testing a small assembled system, that contains different coupling scenarios. In nonlinear problems, not just the shape of the deformation but also the magnitude of the deformation has to be represented by the snapshots. There is no clear rule to choose the parameter ranges for the boundary conditions, but some rules of thumb can be given. In many systems that experience large deformations, the biggest part of the displacement is caused by rigid body motions of the substructures. The substructure deformations are often dominated by tension while shearing is small. The reason for this is that at free surfaces, no shear reactions are possible. This leads to a moment reaction that rotates the substructure towards tensile configurations. If the substructures are slender or have free surfaces the parametrization should represent this effect. Tangential displacements and rotations of the interfaces of the substructures should be smaller than the tensile normal displacements. Choosing a larger parameter space leads to a higher number of substructure POD modes, that often do not represent the deformations that occur in the online phase.

Rigid body modes

When multiple substructures are connected the deformation of one substructure can lead to rigid body translations and rotations of another substructure. The substructure modes need to include modes that can represent arbitrary rigid body motions. For small strains the rigid body motions can be represented by the six modes of the null space of the stiffness matrix. These modes are the three translation modes and the three linearized rotation modes. The linearized rotation modes contain displacements specific to a single rotation angle. If other rotations are approximated by this mode they would change the volume and shape of the body. This additional deformation has to be represented by additional modes.

Let us consider a 2D example to illustrate this. In Figure 1 a square with edge length 2 is rotated around the center by an angle of α=60∘\alpha= 60^\circ. In the left plot, one can see the square in its initial and rotated configuration. The plot in the center shows the fit of the rotation by a single linearized rotation mode around the angle αˉ=36.87∘\bar{\alpha} = 36.87^\circ. It can be seen that the rotation cannot be fitted and that the volume of the rectangle increases. The right plot shows the fit by two modes, where we used a volumetric expansion as the second mode. With these two modes we can represent the rotation with machine precision. The two red squares are the contributions of the rotation mode and the volumetric expansion mode. The rotation mode leads to an increase in area which is counteracted by the shrinkage of the volumetric mode. In 2D all rigid body deformations can be represented by two translation modes, one rotation mode and one volumetric expansion mode.

Two dimensional example of a rotation and the representation by modes

Figure 1. Two dimensional example of a rotation and the representation by modes. The left plot shows the initial and the rotated configuration. The center plot shows the fit by a single rotation mode, which leads to an increase in area. The right plot shows the fit by two modes, where the second mode is a volumetric expansion/contraction mode that counteracts the increase in area.

The same principle holds also for general rotations in 3D, the rotation modes introduce deformations of the body when used for different angles and rotation axes. These additional deformations have to be counteracted by a linear combination of deformation modes. For general rotations these deformation modes are not as easy to interpret as in the 2D case. They still contain volumetric expansion modes, but also other deformations like shearing and stretching.

The nine modes that can describe rotation can be derived from the general description of the rigid body rotation of a 3D body. The displacement resulting from a rigid body rotation can be written as

u(p,X)=R(p)X(38)u(\boldsymbol{p},\boldsymbol{X}) = R(\boldsymbol{p})\boldsymbol{X} \tag*{(38)}

where X\boldsymbol{X} is the position vector of a point in the body and RR is a matrix that maps the position vector to the displacement vector uu. The vector p\boldsymbol{p} contains the parameters for a general rotation, e.g. the rotation angle and the rotation axis or Euler angles. It is related to a rotation matrix Q(p)Q(\boldsymbol{p}) by R(p)=Q(p)−1R(\boldsymbol{p}) = Q(\boldsymbol{p}) - 1. For general rotations the matrix R(p)R(\boldsymbol{p}) has 9 independent entries, because each entry depends nonlinearly on the parameters in p\boldsymbol{p}. Therefore, we can define nine spatial modes that can express any rotational deformation.

These modes can be computed by expressing the displacement resulting from rotations using nine spatial modes

u(p,X)=∑i=19Vi(X)ai(p)(39)\boldsymbol{u}(\boldsymbol{p},\boldsymbol{X}) = \sum_{i=1}^{9} \boldsymbol{V}_i(\boldsymbol{X})a_i(\boldsymbol{p}) \tag*{(39)}

where Vi\boldsymbol{V}_i are the rotation modes and aia_i are the amplitudes of the modes depending on the rotation parameters p\boldsymbol{p}. The modes Vi\boldsymbol{V}_i depend only on the position X\boldsymbol{X} and not on the rotation parameters p\boldsymbol{p}. We choose the modes Vi\boldsymbol{V}_i to be linearly dependent on the position X\boldsymbol{X}: Vi(X)=AiX\boldsymbol{V}_i(\boldsymbol{X}) = \boldsymbol{A}_i\boldsymbol{X}, where Ai\boldsymbol{A}_i are constant matrices. By setting Eq. (38) and Eq. (39) equal to each other and by inserting the above definitions for Vi\boldsymbol{V}_i, we can derive

(R(p)−∑i=19Aiai(p))X=0.(40)\left(\boldsymbol{R}(\boldsymbol{p}) - \sum_{i=1}^{9}\boldsymbol{A}_i a_i(\boldsymbol{p})\right)\boldsymbol{X} = 0. \tag*{(40)}

The constant matrices Ai\boldsymbol{A}_i are the constant modal coefficients of an affine decomposition of R(p)\boldsymbol{R}(\boldsymbol{p}), with ai(p)a_i(\boldsymbol{p}) being the corresponding nonlinear coefficients. When we write the rotation matrix using index notation, we obtain

∑j=13∑k=13Rjk(p)ej⊗ek=∑i=19Aiai(p)(41)\sum_{j=1}^{3}\sum_{k=1}^{3}R_{jk}(\boldsymbol{p})\boldsymbol{e}_j\otimes\boldsymbol{e}_k = \sum_{i=1}^{9}\boldsymbol{A}_i a_i(\boldsymbol{p}) \tag*{(41)}

with ej\boldsymbol{e}_j being the unit vector in jj-th direction. From that equation we can directly read off the constant matrices Ai=ej⊗ek\boldsymbol{A}_i = \boldsymbol{e}_j\otimes\boldsymbol{e}_k and the amplitudes ai(p)=Rjk(p)a_i(\boldsymbol{p}) = R_{jk}(\boldsymbol{p}), with i=k+3(j−1)i = k + 3(j - 1). For each node of the mesh, the modes required to express an arbitrary rotation can be written as

Vi(X)=(ej⊗ek)X=Xkej,i=k+3(j−1).(42)\boldsymbol{V}_i(\boldsymbol{X}) = (\boldsymbol{e}_j\otimes\boldsymbol{e}_k)\boldsymbol{X} = X_k\boldsymbol{e}_j,\qquad i = k + 3(j - 1). \tag*{(42)}

For the 2D case, the same principle holds, but the rotation matrix has only two independent entries, which leads to two rotation modes. In 2D the rotation is described by the single rotation angle α\alpha. The displacements resulting from a rotation are

u(α,X)=R(α)X=[cos⁡(α)−1−sin⁡(α)sin⁡(α)cos⁡(α)−1][XY].(43)\boldsymbol{u}(\alpha,\boldsymbol{X}) = \boldsymbol{R}(\alpha)\boldsymbol{X} = \begin{bmatrix} \cos(\alpha)-1 & -\sin(\alpha) \\ \sin(\alpha) & \cos(\alpha)-1 \end{bmatrix} \begin{bmatrix} X\\ Y \end{bmatrix}. \tag*{(43)}

The same displacement can be represented as

u(α,X)=[1001]X(cos⁡(α)−1)+[0−110]Xsin⁡(α)(44)\boldsymbol{u}(\alpha,\boldsymbol{X}) = \begin{bmatrix} 1 & 0\\ 0 & 1 \end{bmatrix} \boldsymbol{X}(\cos(\alpha)-1) + \begin{bmatrix} 0 & -1\\ 1 & 0 \end{bmatrix} \boldsymbol{X}\sin(\alpha) \tag*{(44)}

From which we can read off the two rotation modes

V1(X)=[1001]X=[XY]andV2(X)=[0−110]X=[−YX].(45)\boldsymbol{V}_1(\boldsymbol{X}) = \begin{bmatrix} 1 & 0\\ 0 & 1 \end{bmatrix} \boldsymbol{X} = \begin{bmatrix} X\\ Y \end{bmatrix} \quad\text{and}\quad \boldsymbol{V}_2(\boldsymbol{X}) = \begin{bmatrix} 0 & -1\\ 1 & 0 \end{bmatrix} \boldsymbol{X} = \begin{bmatrix} -Y\\ X \end{bmatrix}. \tag*{(45)}

Additionally to the rotation modes, we need three translation modes in 3D and two translation modes in 2D to represent all rigid body motions. These can simply be defined as constant modes Vi=eiV_i=e_i with eie_i being the unit vector in ii-th direction.

In matrix notation, the full 3D rigid body projection matrix can be defined for each node XiX^i as

ΨRBMi(X)=[100Xi00Yi00Zi000100Xi00Yi00Zi000100Xi00Yi00Zi](46)\mathbf{\Psi}_{\mathrm{RBM}}^i(\mathbf{X}) = \begin{bmatrix} 1 & 0 & 0 & X^i & 0 & 0 & Y^i & 0 & 0 & Z^i & 0 & 0 \\ 0 & 1 & 0 & 0 & X^i & 0 & 0 & Y^i & 0 & 0 & Z^i & 0 \\ 0 & 0 & 1 & 0 & 0 & X^i & 0 & 0 & Y^i & 0 & 0 & Z^i \end{bmatrix} \tag*{(46)}

In 2D, the nodal rigid body projection matrix is

ΨRBMi(X)=[10Xi−Yi01YiXi](47)\mathbf{\Psi}_{\mathrm{RBM}}^i(\mathbf{X}) = \begin{bmatrix} 1 & 0 & X^i & -Y^i \\ 0 & 1 & Y^i & X^i \end{bmatrix} \tag*{(47)}

The full projection matrix ΨRBM=[(ΨRBM1)T  (ΨRBM2)T,…,(ΨRBMk)T]T\mathbf{\Psi}_{\mathrm{RBM}} = [(\mathbf{\Psi}_{\mathrm{RBM}}^1)^\mathsf{T} \; (\mathbf{\Psi}_{\mathrm{RBM}}^2)^\mathsf{T}, \ldots, (\mathbf{\Psi}_{\mathrm{RBM}}^k)^\mathsf{T}]^\mathsf{T} is obtained by stacking the nodal projection matrices ΨRBMi\mathbf{\Psi}_{\mathrm{RBM}}^i for all kk nodes and normalizing the columns. These projection matrix are not unique: any basis which spans the same linear subspace could be used instead. ONe possible choice might be a POD basis computed from snapshots of pure rigid body motions, which results in exactly 9 non-vanishing singular values in 3D and 4 non-vanishing singular values in 2D. For the illustrations in Figure 1, we used a different projection matrix.

Snapshot centering

As we will show in the numerical examples, it can be an advantage to remove the rigid body mode part of the snapshots before computing the POD basis. The substructure snapshots often contain rigid body motions. Each snapshot UiU_i can be expressed as the sum of the deformation part and the rigid body contribution Ui=UiDefo+UiRBMU_i = U_i^{\mathrm{Defo}} + U_i^{\mathrm{RBM}}. In order to get pure deformation snapshots without rigid body motions the rigid body motion contributions need to be subtracted

UiDefo=Ui−UiRBM.(48)U_i^{\mathrm{Defo}} = U_i - U_i^{\mathrm{RBM}}. \tag*{(48)}

The rigid body part can be expressed with the rigid body motion POD basis shown in § 4.2

UiRBM≈ΨRBMaiRBM,(49)U_i^{\mathrm{RBM}} \approx\mathbf{\Psi}_{\mathrm{RBM}} a_i^{\mathrm{RBM}}, \tag*{(49)}

where aiRBMa_i^{\mathrm{RBM}} is the amplitude vector of the rigid body motion modes. The amplitudes can be computed by a linear least squares fit min⁡aiRBM∈R9∥ΨRBMaiRBM−Ui∥2\min_{a_i^{\mathrm{RBM}}\in\mathbb{R}^9}\lVert\mathbf{\Psi}_{\mathrm{RBM}}a_i^{\mathrm{RBM}}-U_i\rVert^2. The solution of this minimization is given by the normal equation ΨRBMTΨRBMaiRBM=ΨRBMTUi\mathbf{\Psi}_{\mathrm{RBM}}^\mathsf{T}\mathbf{\Psi}_{\mathrm{RBM}}a_i^{\mathrm{RBM}}=\mathbf{\Psi}_{\mathrm{RBM}}^\mathsf{T}U_i. Since ΨRBM\mathbf{\Psi}_{\mathrm{RBM}} is an orthonormal basis, the product ΨRBMTΨRBM=1∈R9\mathbf{\Psi}_{\mathrm{RBM}}^\mathsf{T}\mathbf{\Psi}_{\mathrm{RBM}}=1\in\mathbb{R}^9 gives the identity. The rigid body motion displacement vector UiRBMU_i^{\mathrm{RBM}} can directly be computed

UiRBM=ΨRBMΨRBMTUi(50)U_i^{\mathrm{RBM}} = \mathbf{\Psi}_{\mathrm{RBM}}\mathbf{\Psi}_{\mathrm{RBM}}^\mathsf{T}U_i \tag*{(50)}

This gives the deformation part of the snapshot

UiDefo=Ui−ΨRBMΨRBMTUi.(51)U_i^{\mathrm{Defo}} = U_i - \mathbf{\Psi}_{\mathrm{RBM}}\mathbf{\Psi}_{\mathrm{RBM}}^\mathsf{T}U_i. \tag*{(51)}

The POD basis is than computed from a modified snapshot matrix

SDefo=[U1DefoU2Defo…UlDefoSRBM],(52)S_{\mathrm{Defo}} = \left[ U_1^{\mathrm{Defo}} \quad U_2^{\mathrm{Defo}} \quad\ldots\quad U_l^{\mathrm{Defo}} \quad S_{\mathrm{RBM}} \right], \tag*{(52)}

that contains all deformation contributions of the snapshot and the pure rigid body motion snapshots SRBMS_{\mathrm{RBM}}.

Alternatively, the rigid body part of each snapshot could be computed by a Kabsch algorithm [20], which computes the optimal rotation and translation to align two point clouds. But since we use modes to describe the finite rigid body motions in the online phase, we use the same modes to compute the rigid body part of the snapshots. It should be noted that UiRBMU_i^{\mathrm{RBM}} contains not only the rigid body motion but also deformations.

Numerical examples

In this article, we test different aspects of the component-wise hyperreduction method using two main numerical examples. First, we investigate the effect of the snapshot centering explained in § 4.3 and the accuracy of the component-wise ECSW method (§ 3.2) for 2D and 3D systems. Then we demonstrate the performance of the method using larger systems with many substructures, where we show its predictive capabilities for quasi-static and dynamic problems. In § 6.1.1, we show that we can use quasi-static elastic snapshots to predict the behavior of a nonlinear hyper-viscoelastic dynamic system. In all numerical simulations two different types of substructures are used. In the first subsection we explain the boundary conditions and the parameter space for the snapshot sampling of both substructures.

Snapshot sampling

In this section, we use two different substructures, which were already used in our previous publication Ritzert et al. [ [32]. In the following, the boundary conditions and parameter spaces for the sampling process of both substructures are explained. For all snapshots a compressible Neo-Hookean material model with the Lame constants λ=14907 MPa\lambda= 14907\ \mathrm{MPa}, μ=34783 MPa\mu= 34783\ \mathrm{MPa} is used. The Helmholtz free energy of the Neo-Hookean material model is

ψ(C)=μ2(tr⁡(C)−3−ln⁡(det⁡(C)))+λ4(det⁡(C)−1−ln⁡(det⁡(C)))(53)\psi(C) = \frac{\mu}{2}\left(\operatorname{tr}(C) - 3 - \ln(\det(C))\right) + \frac{\lambda}{4}\left(\det(C) - 1 - \ln(\det(C))\right) \tag*{(53)}

Substructure A

The first substructure is shown in Figure 2. The edges marked in blue are the possible interfaces for boundary conditions, and tied contact interfaces. The same substructure with the same snapshot sampling procedure was already used in our work Ritzert et al. [ [32]. We will briefly summarize the parametrization and boundary conditions of the sampling procedure.

Figure 2. Illustration of the snapshot parametrization. Figure adapted from Ritzert et al. [ [32].

In the snapshot sampling, we apply Dirichlet boundary conditions on those edges. These are parametrized by three parameters on each edge: The parameter dxd_x or dyd_y describes the displacement perpendicular to the edge, the parameter pylp_y^l, pyrp_y^r, pxtp_x^t or pybp_y^b is responsible for the displacement parallel to the edge, and the parameter φzl\varphi_z^l, φzr\varphi_z^r, φzt\varphi_z^t, φzb\varphi_z^b parametrizes xx- and yy-displacements coming from a rotation of the edge. The parameters used in the sampling differ slightly from the ones used in Ritzert et al. [ [32]. Here, we use

dx∈(−1.2 mm,4 mm)dy∈(−1.2 mm,4 mm)pxl∈(−4 mm,4 mm)φzl∈(−15∘,15∘)pxr∈(−4 mm,4 mm)φzr∈(−15∘,15∘)pyt∈(−4 mm,4 mm)φzt∈(−15∘,15∘)pyb∈(−4 mm,4 mm)φzb∈(−15∘,15∘)\begin{aligned} d_x &\in(-1.2\ \mathrm{mm}, 4\ \mathrm{mm}) & d_y &\in(-1.2\ \mathrm{mm}, 4\ \mathrm{mm}) \\ p_x^l &\in(-4\ \mathrm{mm}, 4\ \mathrm{mm}) & \varphi_z^l &\in(-15^\circ, 15^\circ) \\ p_x^r &\in(-4\ \mathrm{mm}, 4\ \mathrm{mm}) & \varphi_z^r &\in(-15^\circ, 15^\circ) \\ p_y^t &\in(-4\ \mathrm{mm}, 4\ \mathrm{mm}) & \varphi_z^t &\in(-15^\circ, 15^\circ) \\ p_y^b &\in(-4\ \mathrm{mm}, 4\ \mathrm{mm}) & \varphi_z^b &\in(-15^\circ, 15^\circ) \end{aligned}

As described in § 4.1, we consider three different boundary conditions cases. In the first case, we apply Dirichlet boundary conditions in tangential and normal direction of each edge. In the second and third case, only the tangential DOFs or the normal DOFs are prescribed. This strategy allows for non-planar deformations of the interfaces. For a more detailed explanation of this choice the reader is referred to Ritzert et al. [ [32].

We generat 50 samples using LGS in the 10-dimensional parameter space. Two load steps per sample and the three boundary condition cases yield 300 snapshots in total.

Substructure B

In Figure 3, we show the dimensions of the substructure as well as the boundary value problem used for the snapshot sampling. In this example, a parametric displacement UBC=UBC(dx,dy,dz,φx,φy,φz)U_{BC} = U_{BC}(d_x, d_y, d_z, \varphi_x, \varphi_y, \varphi_z) is prescribed on both surfaces Γuˉ1\Gamma_{\bar{u}}^{1} and Γuˉ2\Gamma_{\bar{u}}^{2}, where dxd_x, dyd_y, dzd_z define translational components along the respective coordinate axes, and φx\varphi_x, φy\varphi_y, φz\varphi_z describe rotations around them. For boundary conditions on the surface Γuˉ1\Gamma_{\bar{u}}^{1}, the parameter ranges used in the LHS are

Boundary value problem for the snapshot computation

Figure 3. Boundary value problem for the snapshot computation, consisting of three substructures. The snapshots are collected for the central substructure. Figure adapted from Ritzert et al. [ [32]

dx∈(−36 mm,84 mm)φx∈(−70∘,70∘)dy∈(−44 mm,36 mm)φy∈(−70∘,70∘)dz∈(−150 mm,150 mm)φz∈(−40∘,40∘)\begin{aligned} d_x &\in(-36\ \mathrm{mm},84\ \mathrm{mm}) & \varphi_x &\in(-70^\circ,70^\circ) \\ d_y &\in(-44\ \mathrm{mm},36\ \mathrm{mm}) & \varphi_y &\in(-70^\circ,70^\circ) \\ d_z &\in(-150\ \mathrm{mm},150\ \mathrm{mm}) & \varphi_z &\in(-40^\circ,40^\circ) \end{aligned}

For each sampled parameter set, two simulations are performed: one with the displacement applied to Γuˉ1\Gamma_{\bar{u}}^{1} while Γuˉ2\Gamma_{\bar{u}}^{2} remains fixed, and another with the roles reversed. When the boundary conditions are applied on Γuˉ2\Gamma_{\bar{u}}^{2}, the parameters are dxd_x and dyd_y are swapped, while the other parameters remain unchanged.

Comparison between snapshots with and without snapshot centering

3D example

In this numerical example, we use substructure B in two different systems. The boundary value problems are illustrated in Figure 4. On the left, four components are assembled into a ring structure with a cut, with 15876 DOFs in total. The other system consists of five substructures with 19845 DOFs, forming an "Omega"-like shape. The displacement boundary conditions can be seen in Figure 4. In the following, we use the term "ring"-structure for the example on the left and the term "omega" for the structure shown on the right.

Illustration of the geometry and boundary value problems for the 3D examples, showing the “ring” and “omega” systems

Figure 4. Illustration of the geometry, and the boundary value problems for the 3D examples. Figure adapted from Ritzert et al. [ [32]

In Figure 5 it can be seem, that the accuracy of the solution is improved when the snapshots are centered according to § 4.3. In the centering the rigid body motion part computed from the rigid body modes are subtracted from each snapshot. The modes are computed from the pure deformation snapshots and the rigid body motion snapshots. For the "ring" system, the difference between the two methods becomes small for a higher number of modes per substructure. For 60 modes per substructure centering reduces the maximum error from 1.7 % to 0.24 %. For 80 modes per substructure the error is only reduced from 0.07 % to 0.06 %. For the "omega" system, the difference is larger. For 80 modes per substructure the error is reduced from 1.9 % to 0.3 %. It can be concluded that the centering of the snapshots can increase the accuracy of the solution for small numbers of modes per substructure. If the number of modes is large enough to produce accurate results, the centering has only a small influence on the accuracy of the solution.

Relative displacement error over applied displacement for the “ring” and “omega” systems, for normal and centered snapshots

Figure 5. This plot shows the influence of the centering of the snapshots (see § 4.3) on the accuracy of the solution for the "ring" system on the left and the "omega" system on the right. The relative displacement error is plotted over the applied displacement for different numbers of modes per substructure.

In Figure 6 and Figure 7 we show the mean displacement error and the element ratio of evaluated elements to the total number of elements over the tolerance τ\tau for both systems ("ring" and "omega"). The tolerance τ\tau is the parameter controlling the convergence criterion (Eq. (30)) of the weight computation of the ECSW method. The mean error is computed as the mean of the relative displacement error over all load steps

Influence of the tolerance on the accuracy and element ratio for the ring system

Figure 6. Influence of the tolerance τ\tau on the accuracy of the solution of the "ring" system. On the left, the mean displacement error (Eq. (54)) is plotted over τ\tau. The dashed lines are the mean displacement error without ECSW. On the right, the increase of the ratio of evaluated elements is plotted.

Influence of the tolerance on the accuracy and element ratio for the omega system

Figure 7. Influence of the tolerance τ\tau on the accuracy of the solution of the "omega" system. On the left, the mean displacement error (Eq. (54)) is plotted over τ\tau. The dashed lines are the mean displacement error without ECSW. On the right, the increase of the ratio of evaluated elements is plotted.

e=1Nload∑i=1Nload∥UMORi−URefi∥2∥URefi∥2.(54)e = \frac{1}{N_{\mathrm{load}}} \sum_{i=1}^{N_{\mathrm{load}}} \frac{\lVert U_{\mathrm{MOR}}^{i} - U_{\mathrm{Ref}}^{i} \rVert_{2}}{\lVert U_{\mathrm{Ref}}^{i} \rVert_{2}} . \tag*{(54)}

Here, NloadN_{\mathrm{load}} is the number of load steps and UMORU_{\mathrm{MOR}} is either the POD solution or the ECSW solution. For both systems the errors converge to the POD reduced solution. It can be seen, that smaller numbers of POD modes per substructure require smaller values of the tolerance τ\tau to reconstruct the POD solution. For the "ring" example, τ=0.025\tau= 0.025 for 60 modes per substructure gives accurate results with a difference to the POD mean error of 1.7⋅10−31.7 \cdot10^{-3}. For this hyperreduced computation 9.5% of all elements have to be evaluated. The difference of the ECSW solution for the "omega" structure is 2.3⋅10−32.3 \cdot10^{-3}, where 70 modes per substructure and a tolerance of τ=0.025\tau= 0.025 were used. This simulation required the evaluation of 11.0% of all elements. The selected elements with their corresponding weights are visualized in Figure 8 for three different values of τ\tau. The weights where computed with 70 POD modes per substructure. It can be seen, that the selected elements are located predominantly at the interfaces of the substructure. In the regions with many ECSW elements, the weights are small compared to the highest weights. The elements at the boundary are neccessary, to ensure that the mesh tying conditions can be enforced.

Visualization of ECSW weights for the substructure used in the “ring” and “omega” systems

Figure 8. Visualization of the ECSW weights for the substructure used in the "ring" and "omega" systems. The colorbar shows the weights of the elements. The three subfigures show the weights for different tolerances τ\tau computed with the 70 POD modes.

2D example

The mesh and the boundary value problem are shown in Figure 9. The structure is fixed on the left side in horizontal and vertical direction and a horizontal displacement of ux=30 mmu_x = 30\,\mathrm{mm} is applied on the right side of the substructure on the bottom right. This substructure has a finer mesh because it undergoes the largest strains.

Figure 9. Geometry, mesh and boundary conditions for the 2D example with 2×32 \times3 substructures. Figure adapted from Ritzert et al. [ [32]

First, the influence of the centering of the snapshots is investigated. Figure 10 shows that the influence is very small. For all numbers of modes the accuracy is in a similar range. The reason is that the snapshots here do not contain rigid body motions because the boundary conditions in the sampling are applied directly to the component. In the centering also volumetric deformations are removed because ΨRBM\boldsymbol{\Psi}_{\mathrm{RBM}} contains volumetric deformation modes, as explained in § 4.2. This leads to the small differences between the modes of the centered and normal snapshots.

Influence of the centering of the snapshots on the accuracy of the solution for the 2 × 3 system

Figure 10. Influence of the centering of the snapshots (see § 4.3) on the accuracy of the solution for the 2×32 \times3 system. The displacement error is plotted over the applied displacement for different numbers of modes per substructure.

Figure 11 shows that the errors converge to the errors of the POD solution if the tolerance τ\tau is decreased. The dashed lines are the mean displacement error over all time steps of the reduced system without ECSW. The right plot shows the growing amount of evaluated elements for a decreasing τ\tau. For sufficiently converged results, a tolerance of τ=10−2\tau= 10^{-2} is necessary, but with a tolerance of τ=0.05\tau= 0.05 errors below 1% can already be achieved. For τ=10−2\tau= 10^{-2}, 15.9% of the elements need to be evaluated when 35 POD modes are used per substructure. In Figure 12 the selected elements and their weights are visualized for three different values of τ\tau. Similar to the 3D examples, the selected elements are located predominantly at the interfaces of the substructure as τ\tau decreases. These elements at the boundary have comparatively small weights. The elements with the highest weights are located in the center of the substructure.

Influence of the tolerance on the accuracy of the solution of the 2 × 3 system

Figure 11. Influence of the tolerance τ\tau on the accuracy of the solution of the 2×32 \times3 system. On the left, the mean displacement error is plotted over τ\tau. The dashed lines are the mean displacement error without ECSW. On the right, the increase of the ratio of evaluated elements is plotted.

Visualization of the ECSW weights for the substructure used in the 2 × 3 system

Figure 12. Visualization of the ECSW weights for the substructure used in the 2×32 \times3 system. The colorbar shows the weights of the elements. The three subfigures show the weights for different tolerances τ\tau.

Larger examples

In this section, the method is applied to structures composed of many substructures. First, we will show the predictive capabilities for a quasi static case and then for a finite-strain dynamic example.

Quasi static simulations

In Figure 13 we show three different assemblies under different loading conditions. For all systems, we show the deformed configuration with the substructure boundaries. The undeformed configuration is displayed in the background. We consider a compressible Neo-Hookean material model with λ=3727\lambda= 3727 MPa and μ=8696\mu= 8696 MPa. These parameters are different from the parameters used in the snapshot sampling (cf. § 5.1). We use 35 POD modes per substructure and a tolerance of τ=0.05\tau= 0.05, leading to 59 elements per substructure that need to be evaluated. This number is the same for the coarse and the fine mesh used in the three systems.

Sketch of three different assemblies and their corresponding boundary conditions

Figure 13. Sketch of three different assemblies and their corresponding boundary conditions in the undeformed and deformed configuration.

The results of the simulations are summarized in Table 1. The differences in the computational time are due to the different amount of fine and coarse meshes in the three systems. When the mesh is refined, the speed-up of the ECSW method increases because the number of evaluated elements is independent of the mesh. For the system (c) we can achieve the highest speed-up of tref/tECSW=12.536t_{\mathrm{ref}}/t_{\mathrm{ECSW}} = 12.536 because 32 of the 100 substructures are discretized using the finer mesh.

SystemDOFsreduced DOFselement ratiodisplacement errorspeed-uptime ratio
(a)43449035000.0890.00949.6670.103
(b)32880028000.0950.00868.410.119
(c)56133635000.0680.005112.5360.080

Table 1. Summary of the results for the three different assemblies from Figure 13.

It can be seen that the solution of all three systems could be predicted with an error below 1 %. When finer meshes were used, the speed-ups became higher because the number of evaluated elements in the reduced mesh is almost completely independent of the fidelity of the mesh. In general, we observed that the number of ECSW elements depends mostly on the number of modes and the tolerance τ\tau. Mesh refinement only leads to small differences in the number of elements that need to be evaluated.

Finite-strain dynamic example

With this numerical example, we want to show that the method can also be applied to nonlinear dynamic substructuring problems. Additionally, we show that we can predict finite-strain hyper-viscoelastic material behavior with POD modes and ECSW weights and elements that were computed with elastic material behavior.

Figure 14 shows the boundary value problem of the forced vibration simulation. On the left side, we apply a time-dependent displacement in zz-direction and fix the displacement in xx-direction. On the right side, we measure the zz-displacement at the tip. The input function is the damped sine wave

Figure 14. Sketch of the mesh and the boundary value problem of the forced vibration example. The time-dependent zz-displacements on the left side are given by the function Eq. (55). On the right side we measure the zz-displacements at the tip.

uˉz(t)=−32e−5tsin⁡(5πt) 50 mm.(55)\bar{u}_z(t)=-\frac{3}{2}e^{-5t}\sin(5\pi t)\,50\ \mathrm{mm}. \tag*{(55)}

This function is also shown in Figure 15.

Image not extracted

Figure 15. Plots of the tip displacement uztipu_z^{\mathrm{tip}} over time for two different sets of material parameters.

The general hyper-viscoelastic material model used here, is derived in Appendix § 11.§ 1. This general model is specified by choosing suitable functions for the Helmholtz free energy ψ(C,Cˉe)\psi(C,\bar{C}_e) and the dissipation potential g(Zˉ)g(\bar{Z}). The Helmholtz free energy consists of an equilibrium and a non-equilibrium part

ψ=ψeq(C)+ψneq(Cˉe),(56)\psi=\psi_{eq}(C)+\psi_{neq}(\bar{C}_e), \tag*{(56)}

where both energies are of Neo-Hookean type

ψeq(C)=μeq2(tr⁡(C)−3−ln⁡(det⁡(C)))+λeq4(det⁡(C)−1−ln⁡(det⁡(C)))(57)\psi_{eq}(C)=\frac{\mu_{eq}}{2}\left(\operatorname{tr}(C)-3-\ln(\det(C))\right)+\frac{\lambda_{eq}}{4}\left(\det(C)-1-\ln(\det(C))\right) \tag*{(57)}
ψneq(Cˉe)=μneq2(tr⁡(Cˉe)−3−ln⁡(det⁡(Cˉe)))+λneq4(det⁡(Cˉe)−1−ln⁡(det⁡(Cˉe))).(58)\psi_{neq}(\bar{C}_e)=\frac{\mu_{neq}}{2}\left(\operatorname{tr}(\bar{C}_e)-3-\ln(\det(\bar{C}_e))\right)+\frac{\lambda_{neq}}{4}\left(\det(\bar{C}_e)-1-\ln(\det(\bar{C}_e))\right). \tag*{(58)}

The dissipation potential is chosen as

g(Zˉ)=12μneqtr⁡(dev⁡(Zˉ)2)+19κtr⁡(Zˉ)2,(59)g(\bar{Z})=\frac{1}{2\mu_{neq}}\operatorname{tr}\left(\operatorname{dev}(\bar{Z})^2\right)+\frac{1}{9\kappa}\operatorname{tr}(\bar{Z})^2, \tag*{(59)}

with κ=λneq+2/3μneq\kappa=\lambda_{neq}+2/3\mu_{neq}.

To investigate if the POD bases shown in § 5.1.1 can be used for different material behavior, we consider two different sets of material parameters These different parameter sets result in different damping behavior. The parameters are listed in Table 2.

Material AMaterial Bunit
λeq\lambda_{\mathrm{eq}}932932MPa
μeq\mu_{\mathrm{eq}}21742174MPa
λneq\lambda_{\mathrm{neq}}37261864MPa
μneq\mu_{\mathrm{neq}}86964348MPa
η\eta421/s

Table 2. Material parameters for the finite strain hyper-viscoelastic material model.

Evaluation of the reduced simulation The system is solved with 40 modes per substructure and an ECSW tolerance of τ=0.05\tau= 0.05. The ratio of the unreduced degrees of freedom nn and the reduced degrees of freedom mm is n/m=137394/1200=114.5n/m = 137394/1200 = 114.5. For the solution we evaluate 59 elements per substructure for both the fine and coarse discretized substructures. In this system 3 of the 30 substructures have a finer mesh. The element ratio of reduced mesh to the full mesh is ∥Eˉ∥/∥E∥=0.084\lVert\bar{\mathcal{E}}\rVert/\lVert\mathcal{E}\rVert= 0.084.

For the analysis of the accuracy we use two different error measures. The first error measure is a cross-correlation error of the tip displacement

e=uztip,MOR⋅uztip,Refuztip,Ref⋅uztip,Ref−1.(60)e = \frac{u_z^{\mathrm{tip,MOR}}\cdot u_z^{\mathrm{tip,Ref}}}{u_z^{\mathrm{tip,Ref}}\cdot u_z^{\mathrm{tip,Ref}}} - 1. \tag*{(60)}

This error can capture if the shape of the tip displacement is predicted correctly. Slight phase shifts of the tip displacement can lead to large errors in the mean squared error, but the cross-correlation error is not affected by this. Additionally, we compute the mean squared error over all time steps for the full field

MSE=1nsteps∑t=1nsteps(UtMOR−UtRef)2.(61)\mathrm{MSE} = \frac{1}{n_{\mathrm{steps}}}\sum_{t=1}^{n_{\mathrm{steps}}}\left(U_t^{\mathrm{MOR}}-U_t^{\mathrm{Ref}}\right)^2. \tag*{(61)}

For material A, the cross-correlation error is e=1.8 10−4e = 1.8\,10^{-4} and the mean squared error of the full displacement field is MSE=5.1 10−4\mathrm{MSE} = 5.1\,10^{-4}. For material B the errors are roughly twice as large because the deformations are also larger. Here, the cross-correlation error is e=2.9 10−4e = 2.9\,10^{-4} and the mean squared error of the full displacement field is MSE=2.5 10−3\mathrm{MSE} = 2.5\,10^{-3}. In general the errors are very small and the tip displacement can be predicted with high accuracy. In Figure 15 the tip displacement over time is plotted for both materials. The reference solution is plotted as a dashed line and the solution of the reduced system is plotted as a solid line. Visually, the solutions are indistinguishable.

Conclusions and outlook

We developed a component-wise hyperreduction method for nonlinear solid mechanics problems, where the hyperreduced components can be assembled into different systems. The component-wise hyperreduction is based on POD and ECSW on the component level and the substructures are connected by mortar mesh tying.

The projection matrices are computed offline using single component simulations with parametrized boundary conditions on possible interfaces. The snapshots data matrix contains twelve additional modes to model finite rotations and translations. The ECSW weights and elements are computed from the same snapshots in the offline phase using the full substructure projection matrix. In the online phase, the POD basis, the weights, and the ECSW elements are reused for different assemblies and boundary conditions. For the mesh tying, the slave side interface degrees of freedom are projected onto the POD space of the master side.

In the numerical examples, we could show that the substructure POD bases and ECSW meshes can be used for different assemblies and boundary conditions. They are robust under changes in material parameters and can be used also for finite strain hyper-viscoelasticity, even though the snapshots were generated with a purely hyperelastic Neo-Hookean material. We showed that finite rotations and translations can be described by adding three translation and nine rotation modes to the snapshot matrix. We showed how the rotation modes can be derived from the general description of finite rotations and how they can be used to improve the accuracy of the reduced order model. The advantage of this

approach is its simplicity. With only a few additional modes, the reduced order model can be used for large rotations and translations. The results show that the component-wise hyperreduction can significantly reduce the computational cost, while maintaining good accuracy. The hyperreduced simulations converge toward the POD solution with increasing number of ECSW elements. The key result is not only that ECSW accelerates a reduced model, but that the ECSW-selected component meshes remain reusable after assembly. This suggests that hyperreduced components can serve as transferable simulation building blocks for modular nonlinear solid mechanics.

In future work, we will investigate more efficient hyperreduction methods for the substructures and apply the method to more complex inelastic problems. The assembly of the hyperreduced components is currently very intrusive and requires access to the solver. We plan work on making the method non-intrusive and implement it in a commercial solver.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Data availability

Data will be made available on request.

Acknowledgements

The authors gratefully acknowledge the funding granted by the German Research Foundation (DFG). The results presented here were developed within the subproject A01 of the Transregional Collaborative Research Center (CRC) Transregio (TRR) 280 with project number 444299690. Furthermore, T. Brepols, J. Kehls, and S. Reese gratefully acknowledge the funding that was granted within the subproject B05 "Coupling of intrusive and non-intrusive locally decomposed model order reduction techniques for rapid simulations of road systems" of the DFG CRC/TRR 339 with the project number 496338782, that was strongly involved in the origin of the paper.

Appendix

Finite strain hyper-viscoelastic material model

For the derivation of the hyper-viscoelastic material model, we follow the method described by Holthusen et al. [ [19] and define the inelastic quantities in the co-rotated intermediate configuration. It is derived from a Helmholtz free energy depending on the total strain and the elastic strain and an evolution equation describing the viscous strains.

The kinematics of the hyper-viscoelastic material are described by the multiplicative split of the deformation gradient F=FeFiF = F_{\mathrm{e}}F_{\mathrm{i}} into elastic FeF_{\mathrm{e}} and inelastic parts FiF_{\mathrm{i}}. The two tensors can be split by a polar decomposition Fe=VeReF_{\mathrm{e}} = V_{\mathrm{e}}R_{\mathrm{e}}, Fi=RiUiF_{\mathrm{i}} = R_{\mathrm{i}}U_{\mathrm{i}} into non-unique rotational parts Re,RiR_{\mathrm{e}},R_{\mathrm{i}} and unique stretch tensors Ve,UiV_{\mathrm{e}}, U_{\mathrm{i}}. The material model is formulated using uniquely defined strain tensors in the reference configuration and in the co-rotated intermediate configuration. In the reference configuration we use the right Cauchy-Green strain C=FTFC = F^{T}F and the inelastic Cauchy-Green strain, which is defined as Cˉi=UiT\bar{C}_{\mathrm{i}} = U_{\mathrm{i}}^{T}. The co-rotated elastic Cauchy-Green strain can be computed by a partial push forward Cˉe=Ui−1CUi−1\bar{C}_{\mathrm{e}} = U_{\mathrm{i}}^{-1}CU_{\mathrm{i}}^{-1} by the inelastic stretch tensor UiU_{\mathrm{i}}. All of these strain tensors are not affected by the rotational non-uniqueness. Based on this stretch, we can further define the co-rotated rate of deformation tensor

Dˉi=sym⁡(U˙iUi−1)=12Ui−1Cˉ˙iUi−1.(62)\bar{D}_{\mathrm{i}} = \operatorname{sym}(\dot{U}_{\mathrm{i}}U_{\mathrm{i}}^{-1}) = \frac{1}{2}U_{\mathrm{i}}^{-1}\dot{\bar{C}}_{\mathrm{i}}U_{\mathrm{i}}^{-1}. \tag*{(62)}

The stress and the thermodynamic driving force are derived from the Helmholtz free energy ψ=ψ(C,Cˉe)\psi=\psi(C,\bar{C}_e), depending on the right Cauchy-Green strain CC and the co-rotated elastic Cauchy-Green strain Cˉe\bar{C}_e. The second Piola Kirchhoff stress is derived from the Clausius-Duhem inequality by using the standard Coleman-Noll procedure, leading to

S=2∂ψ∂C+Ui−1∂ψ∂CˉeUi−1.(63)S=2\frac{\partial\psi}{\partial C}+U_i^{-1}\frac{\partial\psi}{\partial\bar{C}_e}U_i^{-1}. \tag*{(63)}

This stress definition leads to the remaining dissipation inequality Zˉ⋅Dˉ˙i≥0\bar{Z}\cdot\dot{\bar{D}}_i\geq0, where the thermodynamic driving force Zˉ\bar{Z} is defined as

Zˉ=2Cˉe∂ψ∂Cˉe.(64)\bar{Z}=2\bar{C}_e\frac{\partial\psi}{\partial\bar{C}_e}. \tag*{(64)}

The remaining dissipation inequality can be fulfilled if the evolution equation of Dˉi\bar{D}_i is derived from a non-negative and convex potential g(Zˉ)g(\bar{Z}) that is assumed to be an isotropic function of Zˉ\bar{Z}. Using Eq. (62), we can write the evolution equation in terms of the inelastic right Cauchy-Green strain as

C˙i=2ηUi∂g(Zˉ)∂ZˉUi.(65)\dot{C}_i=2\eta U_i\frac{\partial g(\bar{Z})}{\partial\bar{Z}}U_i. \tag*{(65)}

Here, η\eta is a material parameter controlling the rate of dissipation. In a time discretized setting on the interval t∈[tn,tn+1]t\in[t_n,t_{n+1}] with Δt=tn+1−tn\Delta t=t_{n+1}-t_n, we solve the evolution equation by an exponential map algorithm (see e.g. [8, 40, 7]). After some derivation steps shown in the work Holthusen et al. [ [19], one arrives at the residual equation that is solved at each integration point

r(Ui)=Uiexp⁡(−2Δt η∂g(Zˉ)∂Zˉ)Ui−Uin2=0,(66)r(U_i)=U_i\exp\left(-2\Delta t\,\eta\frac{\partial g(\bar{Z})}{\partial\bar{Z}}\right)U_i-U_{in}^{2}=0, \tag*{(66)}

where we dropped the index n+1n+1. It can be seen that the internal variable in this material model is the inelastic stretch UiU_i. This residual equation is solved by a Newton-Raphson procedure.

References

  1. [1]Steven S An, Theodore Kim, and Doug L James. Optimizing cubature for efficient integration of subspace deformations. ACM transactions on graphics (TOG), 27(5):1–10, 2008.
  2. [2]Jernej Barbić and Yili Zhao. Real-time large-deformation substructuring. ACM Transactions on Graphics, 30(4):1–8, July 2011. ISSN 0730-0301, 1557-7368. doi: 10.1145/2010324.1964986.DOI
  3. [3]Diana Bonomi, Andrea Manzoni, and Alfio Quarteroni. A matrix deim technique for model reduction of nonlinear parametrized problems in cardiac mechanics. Computer Methods in Applied Mechanics and Engineering, 324: 300–326, 2017.DOI
  4. [4]Tuan Anh Bui, Junyoung Park, and Jun-Sik Kim. A reduced-order model for geometrically nonlinear curved beam structures with substructuring techniques. International Journal of Non-Linear Mechanics, 162:104724, June 2024. ISSN 0020-7462. doi: 10.1016/j.ijnonlinmec.2024.104724.DOI
  5. [5]Anindya Chatterjee. An introduction to the proper orthogonal decomposition. Current science, pages 808–817, 2000.
  6. [6]Saifon Chaturantabut and Danny C. Sorensen. Nonlinear Model Reduction via Discrete Empirical Interpolation. SIAM Journal on Scientific Computing, 32(5):2737–2764, January 2010. ISSN 1064-8275. doi: 10.1137/090766498. Publisher: Society for Industrial and Applied Mathematics.DOI
  7. [7]Daniel Christ and Stefanie Reese. A finite element model for shape memory alloys considering thermomechanical couplings at large strains. International Journal of solids and Structures, 46(20):3694–3709, 2009.
  8. [8]Wulf Dettmer and Stefanie Reese. On the theoretical and numerical modelling of armstrong–frederick kinematic hardening in the finite strain regime. Computer Methods in Applied Mechanics and Engineering, 193(1-2):87–116, 2004.DOI
  9. [9]Philipp Diercks, Karen Veroy, Annika Robens-Rademacher, and Jörg F Unger. Multiscale modeling of linear elastic heterogeneous structures via localized model order reduction. International Journal for Numerical Methods in Engineering, 124(20):4580–4602, 2023.arxiv.org/abs/2201.10374
  10. [10]Mehran Ebrahimi and Masayuki Yano. A hyperreduced reduced basis element method for reduced-order modeling of component-based nonlinear systems. Computer Methods in Applied Mechanics and Engineering, 431:117254, November 2024. ISSN 0045-7825. doi: 10.1016/j.cma.2024.117254.DOI
  11. [11]Mehran Ebrahimi and Masayuki Yano. An online-adaptive hyperreduced reduced basis element method for parameterized component-based nonlinear systems using hierarchical error estimation. Computer Methods in Applied Mechanics and Engineering, 449:118590, February 2026. ISSN 00457825. doi: 10.1016/j.cma.2025.118590.DOI
  12. [12]Charbel Farhat, Todd Chapman, and Philip Avery. Structure-preserving, stability, and accuracy properties of the energy-conserving sampling and weighting method for the hyper reduction of nonlinear finite element dynamic models. International Journal for Numerical Methods in Engineering, 102(5):1077–1110, 2015. ISSN 1097-0207. doi: 10.1002/nme.4820.DOI
  13. [13]Erik Faust and Lisa Scheunemann. Empirical material sampling and linearisation–a simple and efficient strain-space model order reduction approach for computational homogenisation in large-deformation hyperelasticity. arXiv preprint arXiv:2604.27179, 2026.arxiv.org/abs/2604.27179
  14. [14]Roy Featherstone. Robot dynamics algorithms. Annexe thesis digitisation project 2016 block 5, 1984.
  15. [15]Felix Fritzen and Matthias Leuschner. Nonlinear reduced order homogenization of materials including cohesive interfaces. Computational Mechanics, 56(1):131–151, 2015.DOI
  16. [16]Fariborz Ghavamian, Paolo Tiso, and Angelo Simone. Pod–deim model order reduction for strain-softening viscoplasticity. Computer Methods in Applied Mechanics and Engineering, 317:458–479, 2017.DOI
  17. [17]J. A. Hernández, M. A. Caicedo, and A. Ferrer. Dimensional hyper-reduction of nonlinear finite element models via empirical cubature. Computer Methods in Applied Mechanics and Engineering, 313:687–722, January 2017. ISSN 0045-7825. doi: 10.1016/j.cma.2016.10.022.DOI
  18. [18]J. A. Hernández, J. R. Bravo, and S. Ares de Parga. CECM: A continuous empirical cubature method with application to the dimensional hyperreduction of parameterized finite element models. Computer Methods in Applied Mechanics and Engineering, 418:116552, January 2024. ISSN 0045-7825. doi: 10.1016/j.cma.2023.116552.DOI
  19. [19]Hagen Holthusen, Christiane Rothkranz, Lukas Lamm, Tim Brepols, and Stefanie Reese. Inelastic material formulations based on a co-rotated intermediate configuration—application to bioengineered tissues. Journal of the Mechanics and Physics of Solids, 172:105174, March 2023. ISSN 00225096. doi: 10.1016/j.jmps.2022.105174.DOI
  20. [20]Wolfgang Kabsch. A solution for the best rotation to relate two sets of vectors. Foundations of Crystallography, 32(5):922–923, 1976.
  21. [21]Jannick Kehls, Stephan Ritzert, Tim Brepols, and Stefanie Reese. Hyper-reduction of mechanical problems with plasticity and damage via discrete empirical interpolation. In Advances and Challenges in Computational Mechanics, pages 187–199. Springer, 2026a.DOI
  22. [22]Jannick Kehls, Stephan Ritzert, Lars Breuer, Qinghua Zhang, Stefanie Reese, and Tim Brepols. Multi-field decomposed hyper-reduced order modeling of damage-plasticity simulations. Computer Methods in Applied Mechanics and Engineering, 451:118702, April 2026b. ISSN 0045-7825. doi: 10.1016/j.cma.2025.118702.DOI
  23. [23]Theodore Kim and Doug L. James. Physics-based character skinning using multidomain subspace deformations. In IEEE Trans Visual Comput Graphics, volume 18, pages 1228–1240, 2012. doi: 10.1109/TVCG.2012.78.DOI
  24. [24]Robert J. Kuether, Matthew S. Allen, and Joseph J. Hollkamp. Modal Substructuring of Geometrically Nonlinear Finite-Element Models. AIAA Journal, 54(2):691–702, February 2016. ISSN 0001-1452, 1533-385X. doi: 10.2514/1.J054036.DOI
  25. [25]Robert J. Kuether, Matthew S. Allen, and Joseph J. Hollkamp. Modal Substructuring of Geometrically Nonlinear Finite Element Models with Interface Reduction. AIAA Journal, 55(5):1695–1706, May 2017. ISSN 0001-1452, 1533-385X. doi: 10.2514/1.J055215.DOI
  26. [26]Tod A. Laursen, Michael A. Puso, and Jessica Sanders. Mortar contact formulations for deformable–deformable contact: Past contributions and new extensions for enriched and embedded interface formulations. Computer Methods in Applied Mechanics and Engineering, 205-208:3–15, January 2012. ISSN 00457825. doi: 10.1016/j.cma.2010.09.006.
  27. [27]Michael D McKay, Richard J Beckman, and William J Conover. A comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics, 42(1):55–61, 2000.
  28. [28]Alexander Popp and Peter Wriggers, editors. Contact Modeling for Solids and Particles, volume 585 of CISM International Centre for Mechanical Sciences. Springer International Publishing, Cham, 2018. ISBN 978-3-319-90154-1 978-3-319-90155-8. doi: 10.1007/978-3-319-90155-8.DOI
  29. [29]Michael A. Puso. A 3D mortar method for solid mechanics. International Journal for Numerical Methods in Engineering, 59(3):315–336, January 2004. ISSN 0029-5981, 1097-0207. doi: 10.1002/nme.865.DOI
  30. [30]Annika Radermacher and Stefanie Reese. POD-based model reduction with empirical interpolation applied to nonlinear elasticity. International Journal for Numerical Methods in Engineering, 107(6):477–495, 2016. ISSN 1097-0207. doi: 10.1002/nme.5177. _eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/nme.5177.DOI
  31. [31]Annika Radermacher, Brett A Bednarczyk, Bertram Stier, Jaan Simon, Lei Zhou, and Stefanie Reese. Displacement-based multiscale modeling of fiber-reinforced composites by means of proper orthogonal decomposition. Advanced Modeling and Simulation in Engineering Sciences, 3(1):29, 2016.DOI
  32. [32]Stephan Ritzert, Jannick Kehls, Stefanie Reese, and Tim Brepols. Component-based model-order reduction with mortar tied contact for nonlinear quasi-static mechanical problems. International Journal for Numerical Methods in Engineering, 126(8):e70041, 2025.
  33. [33]Johannes Rutzmoser. Model Order Reduction for Nonlinear Structural Dynamics. PhD thesis, Technische Universität München, 2018.
  34. [34]David Ryckelynck. Hyper-reduction of mechanical models involving internal variables. International Journal for numerical methods in engineering, 77(1):75–89, 2009.DOI
  35. [35]Alexander Saccani and Paolo Tiso. Accelerating construction of nonintrusive nonlinear structural dynamics reduced-order models through hyperreduction. AIAA Journal, 64(3):1643–1660, 2026.arxiv.org/abs/2411.14262
  36. [36]Lisa Scheunemann and Erik Faust. A manifold learning approach to nonlinear model order reduction of quasi-static problems in solid mechanics: proof of concept and parameter study. Computational Mechanics, 77(4):983–1023, 2026.
  37. [37]Michael Stein. Large sample properties of simulations using latin hypercube sampling. Technometrics, 29(2):143–151, 1987.DOI
  38. [38]Paolo Tiso and Daniel J Rixen. Discrete empirical interpolation method for finite element structural dynamics. In Topics in Nonlinear Dynamics, Volume 1: Proceedings of the 31st IMAC, A Conference on Structural Dynamics, 2013, pages 203–212. Springer, 2013.DOI
  39. [39]Cyril Touzé, Alessandra Vizzaccaro, and Olivier Thomas. Model order reduction methods for geometrically nonlinear structures: a review of nonlinear techniques. Nonlinear Dynamics, 105(2):1141–1190, 2021.DOI
  40. [40]Ivaylo N Vladimirov, Michael P Pietryga, and Stefanie Reese. On the modelling of non-linear kinematic hardening at finite strains with application to springback—comparison of time integration algorithms. International Journal for Numerical Methods in Engineering, 75(1):1–28, 2008.DOI
  41. [41]Barbara I. Wohlmuth. Discretization Methods and Iterative Solvers Based on Domain Decomposition, volume 17 of Lecture Notes in Computational Science and Engineering. Springer Berlin Heidelberg, Berlin, Heidelberg, 2001. ISBN 978-3-540-41083-6 978-3-642-56767-4. doi: 10.1007/978-3-642-56767-4.DOI
  42. [42]Stephan Wulfinghoff. Computational crystal plasticity homogenization using empirically corrected cluster cubature (e3c) hyper-reduction. Computer Methods in Applied Mechanics and Engineering, 452:118681, 2026.arxiv.org/abs/2510.11187
  43. [43]Stephan Wulfinghoff and Jan Hauck. E3c hyper-reduction for computational homogenization in nonlinear mechanics. Computational Mechanics, 76(6):1563–1572, 2025.DOI
  44. [44]Masayuki Yano and Anthony T Patera. An lp empirical quadrature procedure for reduced basis treatment of parametrized nonlinear pdes. Computer Methods in Applied Mechanics and Engineering, 344:1104–1123, 2019.DOI
  45. [45]Matthew J Zahr, Philip Avery, and Charbel Farhat. A multilevel projection-based model order reduction framework for nonlinear dynamic multiscale problems in structural and solid mechanics. International Journal for Numerical Methods in Engineering, 112(8):855–881, 2017.DOI
  46. [46]Lei Zhou, Jaan-Willem Simon, and Stefanie Reese. Proper orthogonal decomposition for substructures in nonlinear finite element analysis: coupling by means of tied contact. Archive of Applied Mechanics, 88(11):1975–2001, November 2018. ISSN 1432-0681. doi: 10.1007/s00419-018-1427-1.DOI

Paper details

vlnc:W_dake4wyq
Version
v1 · 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