betaFeedback

WorksMathematics

Non-Intrusive Data-Free Parametric Reduced Order Model for Geometrically Nonlinear Structures

Abstract

We present a fully non-intrusive parametric reduced-order modeling (PROM) framework for geometrically nonlinear structures subject to geometric variations. The method builds upon equation-driven Galerkin ROMs constructed from vibration modes and modal-derivative companion vectors, while nonlinear reduced tensors are identified from standard finite element outputs. A database of such ROMs is generated over a set of training samples, and all reduced operators-including the linear stiffness matrix, the quadratic and cubic nonlinear tensors, the Rayleigh damping parameters, and the reduction basis-are interpolated using Radial Basis Functions (RBFs). A global reduced basis is obtained through a two-level POD compression, combined with a MAC-guided reordering strategy to ensure parametric smoothness. The resulting PROM preserves the symmetry and polynomial structure of the reduced equations, enabling robust and efficient adaptation to new parameter values. Analytical parameter sensitivities follow directly from the interpolation model. The approach is demonstrated on a parametrically curved panel and a wing-box with geometric variations, showing excellent agreement with high-fidelity simulations and enabling substantial reductions in computational cost for parametric analyses.

Introduction

Dynamic simulations of lightweight components play a pivotal role in the design of aerospace structures, allowing optimal design and fatigue life prediction [1, 2, 3, 4, 5]. At the foundation of these simulations lies the FE method that is used to derive a set of time-dependent ordinary differential equations whose time integration provides the prediction of structural displacements and stresses to the external loads. The solved equations are nonlinear, as they are derived assuming a nonlinear strain-displacements relationship. This relationship is either based on the Green-Lagrange strain tensor or on the Von Kármán strain-displacement law [6]. Moreover, the discretized geometries are usually complex, and they require fine meshing to correctly capture the stress gradients. This leads to high dimensional models, referred to as High Fidelity Models (HFMs). The combination of high dimensionality and nonlinearity results in large solution times that can hinder the design of aerospace components, especially when repeated analysis are required.

Equation-based Reduced Order Models (ROMs) have been introduced in the few last decades to cut down the computational costs associated to dynamic analysis of structures. The most classic approaches for ROM construction rely on the Galerkin Projection (GP) method, whereby the solution to the HFM equations is searched in a linear subspace spanned by few vectors stored in a Reduction Basis (RB). Then, the HFM equations are projected on the RB. In equation-driven ROMs, the RB is assumed based on the physics of the problem. For linear dynamical systems, a set of low frequency Vibration Modes (VMs) constitutes an optimal RB, as proven by Modal Analysis theory [7]. In fact, the typical external excitation usually excites only low-frequency modes. A good RB for nonlinear systems can be constructed by complementing the RB of the underlying linear systems with additional vectors that capture the bending-stretching coupling

Email address: [email protected] (Alexander Saccani) triggered by geometric nonlinearities in thin-walled structures. Among these approaches, the method of Static Modal Derivatives (SMDs) [8] [9] and of Dual Modes (DMs) [10] have gained popularity and have been demonstrated to reliably approximate the response of geometrically nonlinear structures subjected to dynamic loading.

After RB identification, the construction of the ROM is completed by computing the ROM operators associated to the projection of the HFM equations. While the reduced mass and stiffness matrices can be swiftly computed from their HFM counterpart through projection on the RB, the projection of the nonlinear forces is usually approximated with a cubic polynomial in the generalized reduced coordinates [5]. The coefficients of this cubic polynomial expansion are stored in third and fourth order tensors.

If the FE method is based on the Total Lagrangian Formulation [6] and if the adopted material is linear elastic, this approximation becomes exact. This could theoretically allow for the computation of the reduced tensors via direct projection of their HFM counterpart [5], following a so called direct-method. However, in most commercial FE programs, the FE implementation is based on the Updated Lagrangian formulation [6] or on Corotational formulations, and the decomposition of internal elastic forces in their quadratic and cubic components is thus not accessible. For this reason, reduced nonlinear tensors are identified using the FE program as a black box. A set of equations in the unknown nonlinear tensors entries is formed by reading the nonlinear restoring forces or tangent stiffness matrices corresponding to a set of imposed displacements. The method using the nonlinear force goes under the name of Enforced Displacement (ED) method [11], while the one exploiting the tangent stiffness matrix is known as Enhanced Enforced Displacement (EED) method [12]. Since tensors are identified only by using common input-output quantities, both these methods are said to be non-intrusive [5].

Unlike methods where the RB is constructed from data compression of HFM simulations, using for example, Proper Orthogonal Decomposition (POD), equation-driven ROMs are relatively computationally cheap. However, the computational cost for model construction can still be significant, and is usually dominated by the cost for reduced tensor identification. Moreover, in structural optimization [13, 14, 15], uncertainty quantification [16], or model adaptation scenarios, this overhead has to be sustained multiple times. A way to reduced the computational cost for ROM construction is to resort to Parametric ROMs (PROMs). These techniques exploit an offline-online decomposition. In this setting, significant computational resources are used in the offline stage to identify a ROM that ideally works for any parameter realization in a prescribed parameter set. Then, ROM adaptation to any parameter realization is performed on the fly in the online stage. The initial overhead for PROM construction is usually larger than that for constructing a ROM for a prescribed parameter value. However, for multiple evaluations, the PROM strategy pays off, as ROM adaptation is extremely fast.

Parametric data-driven ROMs techniques have reached a high level of maturity, while equation-driven ROMs lag behind. A pioneering method for PROM construction for structural dynamics was put out in [17] and extended in [18]. There, the authors devised a way to interpolate ROMs constructed with a RB of VMs while preserving important properties of the ROM operators, such as symmetry and positive definiteness. The method, however, was applied only to linear dynamical systems.

In [15] the authors presented a parametric model for investigating the variability of the response of geometrically-nonlinear structures subjected to uncertainty in the geometry and boundary conditions. The reduction strategy pursued in the paper is the Implicit Condensation and Expansion (ICE) [19]. In [20] and [16] the authors presented a parametric ROM constructed with a RB of VMs and MDs to swiftly adapt the ROM to geometric imperfections. This model was shown to capture the strong dependence of the response to the geometrical parameters. However, it is intrusive in nature, as it requires the modification of the FE formulation at the element level.

In this work we present a fully non-intrusive Parametric Reduced Order Model (PROM) for predicting the nonlinear dynamic response of structures undergoing geometric variations. Differently from previous approaches such as [20] and [16], the proposed method does not require access to element-level quantities and can therefore be implemented directly on top of commercial FE software (e.g., ABAQUS, ANSYS, NAS-TRAN). This non-intrusive character is essential for industrial applications, where the FE formulation is typically inaccessible. The construction of the PROM proceeds in two stages. First, a database of equation-driven ROMs is generated at selected parameter samples using vibration modes and modal derivatives to capture nonlinear geometric coupling. Second, all reduced operators—including the reduction basis, the linear stiffness matrix, the quadratic and cubic nonlinear tensors, and the Rayleigh damping parameters—are interpolated across the parameter domain using Radial Basis Functions (RBFs). This interpolation strategy enables instantaneous adaptation of the ROM to any new parameter value and provides analytical parameter sensitivities through direct differentiation of the interpolants. To the best of our knowledge, this is the first non-intrusive PROM for geometrically nonlinear structures that interpolates all reduced operators, including the nonlinear reduced tensors, while preserving their intrinsic symmetry and the polynomial structure of the reduced equations. Additionally, analytical gradients of the ROM in the parameters can be easily obtained by differentiation of the interpolation formula. These features make the proposed PROM particularly attractive for uncertainty quantification, structural optimization, and parametric model adaptation.

We structure the presentation of our method as follows. In § 2, we set out the theory underlying equation-based ROMs for geometrically nonlinear structures. In § 3, we describe the construction of the PROM. In § 4, we eventually apply the PROM to two examples, namely a curved panel and a wing box with geometrical shape variations.

Preliminaries

Problem Statement

The starting point for the construction of the PROM herein proposed are the equations of motion for FE models of thin-walled structures with geometric nonlinearities and linear elastic material, which read

M(p)q¨+C(p)q˙+f(q,p)=f(e)(t,p),(1)\mathbf{M}(\mathbf{p})\ddot{\mathbf{q}}+\mathbf{C}(\mathbf{p})\dot{\mathbf{q}}+\mathbf{f}(\mathbf{q},\mathbf{p})=\mathbf{f}^{(e)}(t,\mathbf{p}), \tag*{(1)}

where q\mathbf{q}, q˙\dot{\mathbf{q}}, q¨∈Rn\ddot{\mathbf{q}} \in\mathbb{R}^{n} are the vectors of nodal displacements, velocities and accelerations, M,C∈Rn×n\mathbf{M},\mathbf{C} \in\mathbb{R}^{n \times n} are the mass and viscous damping matrices, f(q)\mathbf{f}(\mathbf{q}), f(e)∈Rn\mathbf{f}^{(e)} \in\mathbb{R}^{n} are the vectors of internal and external forces respectively, and p∈Rnp\mathbf{p} \in\mathbb{R}^{n_p} is the vector of parameters that is used to account for variations in the geometry, material, and boundary conditions.

The elastic internal forces are derived by assuming a nonlinear strain-displacement relation based on the Green-Lagrange strain tensor or on the approximate Von Kármán strains [6]. This leads to a cubic formulation that writes:

f(q,p)=K(1)(p)⋅q+K(2)(p):(q⊗q)+K(3)(p):(q⊗q⊗q)⟺fi=Kij(1)qj+Kijk(2)qjqk+Kijkl(3)qjqkql,(2)\mathbf{f}(\mathbf{q},\mathbf{p})=\mathbf{K}^{(1)}(\mathbf{p})\cdot\mathbf{q}+\mathbf{K}^{(2)}(\mathbf{p}):(\mathbf{q}\otimes\mathbf{q})+\mathbf{K}^{(3)}(\mathbf{p}):(\mathbf{q}\otimes\mathbf{q}\otimes\mathbf{q})\Longleftrightarrow f_i=K^{(1)}_{ij}q_j+K^{(2)}_{ijk}q_jq_k+K^{(3)}_{ijkl}q_jq_kq_l, \tag*{(2)}

where K(1)∈Rn×n\mathbf{K}^{(1)} \in\mathbb{R}^{n \times n} is the linear stiffness matrix, K(2)∈Rn×n×n\mathbf{K}^{(2)} \in\mathbb{R}^{n \times n \times n} and K(3)∈Rn×n×n×n\mathbf{K}^{(3)} \in\mathbb{R}^{n \times n \times n \times n} are respectively the tensors associated to the quadratic and cubic forces, "⊗\otimes" is the dyadic product operator, and ⋅\cdot, ::, ⋮\vdots, ∴\therefore denote respectively, single, double and triple contraction. The tensors K(2)(p)\mathbf{K}^{(2)}(\mathbf{p}) and K(3)(p)\mathbf{K}^{(3)}(\mathbf{p}) are commonly referred to as nonlinear stiffness operators, while K(1)(p)\mathbf{K}^{(1)}(\mathbf{p}) is the linear stiffness.

These operators are parameter dependent if geometrical and material variations are considered. Notably, the linear stiffness matrix and the nonlinear tensors are sparse and symmetric with respect to all axis, as they can be derived through partial differentiation of a quartic potential function [21]. For instance Kijk(2)=Kikj(2)=Kjik(2)=Kjki(2)=Kkij(2)=Kkji(2)K^{(2)}_{ijk}=K^{(2)}_{ikj}=K^{(2)}_{jik}=K^{(2)}_{jki}=K^{(2)}_{kij}=K^{(2)}_{kji}. Moreover, we assume that the linear stiffness matrix K(1)\mathbf{K}^{(1)} is positive definite. This is equivalent to assuming that the undeformed configuration is stable.

In the following, we also adopt the Raleigh Damping model [7], where the damping matrix C\mathbf{C} is defined as a linear combination of the mass and stiffness matrices, M\mathbf{M} and K(1)\mathbf{K}^{(1)}, as

C(p)=α(p)M(p)+β(p)K(1)(p),(3)\mathbf{C}(\mathbf{p})=\alpha(\mathbf{p})\mathbf{M}(\mathbf{p})+\beta(\mathbf{p})\mathbf{K}^{(1)}(\mathbf{p}), \tag*{(3)}

with parameter-dependent α,β∈R\alpha,\beta\in\mathbb{R}.

The objective of this work is the construction of a PROM for the investigation of the dynamic response of the HFM model, in Eq. (1), for parametric variations defined over a prescribed parameter set. Specifically, we define this set as

Sp:={p∈Rnp∣pimin⁡≤pi≤pimax⁡, i=1,…,np},(4)\mathcal{S}_{\mathbf{p}}:=\left\{\mathbf{p}\in\mathbb{R}^{n_p}\mid p_i^{\min}\le p_i\le p_i^{\max},\ i=1,\ldots,n_p\right\}, \tag*{(4)}

where pimin⁡∈Rp_i^{\min}\in\mathbb{R} and pimax⁡∈Rp_i^{\max}\in\mathbb{R} are respectively the minimum and maximum prescribed values for pip_i.

Equation Driven Galerkin ROMs for Geometrically Nonlinear Structures

In this subsection we provide the reader with a condensed outline on equation-driven Galerkin ROMs for systems with geometric nonlinearities, as this is the foundation for the construction of the PROM here presented.

Galerkin-Projection

The GP is arguably the most classical and established technique for the construction a ROM. It consists in restricting the solution space of the HFM to a linear subspace spanned by a RB, V∈Rn×mV \in\mathbb{R}^{n \times m} (with m≪nm \ll n), and in solving the projected HFM equations for the reduced coordinate vector, η(t)∈Rm\eta(t) \in\mathbb{R}^{m}:

q(t)≈Vη(t),(5)q(t) \approx V\eta(t), \tag*{(5)}
VTMVη¨+VTCVη˙+VTf(Vη)=VTf(e).(6)V^{T}MV\ddot{\eta} + V^{T}CV\dot{\eta} + V^{T}f(V\eta) = V^{T}f^{(e)}. \tag*{(6)}

The ROMs constructed using GP preserve the Lagrangian Structure [22] associated to the HFM, a desirable feature that guarantees the ROM to inherit the same stability characteristics of the time integration scheme applied to the HFM equations [22].

Equation-Driven Reduction Basis

The RB used in Eq. (6) dictates accuracy and efficiency of the ensuing ROM: the larger the RB, the higher the expected level of detail, yet the lower the computational gain. In this work, the RB is identified following an equation-driven approach, where all the information required for RB construction is inferred from the physics modeled in the equations. In this way, the RB construction process is extremely efficient compared to data-driven methods, such as POD, that rely on HFM simulations.

Instead, a set of low frequency VMs ϕi∈Rn\phi_i \in\mathbb{R}^{n} are computed by solving the generalized eigenvalue problem defined as

(−Mωi2+K(1))ϕi=0,for i=1,…,nϕ,(7)(-M\omega_i^2 + K^{(1)})\phi_i = 0,\quad\text{for } i = 1,\ldots,n_{\phi}, \tag*{(7)}

where ωi\omega_i is the ithi^{\mathrm{th}} angular eigenfrequency. Of these vectors, only a few are selected based on the Modal Participation Factor (MPF) of the external load [7] and the load frequency bandwidth, and are used to form the RB. Throughout the paper, we identify the selected modes with their mode number, which we collect in set Sϕ⊂NS_{\phi} \subset\mathbb{N}.

A RB of VMs is expected to capture the response of the structure as it vibrates about its equilibrium position in a linear regime. However, as the displacements vibration amplitudes increase to a point where geometric nonlinearities are activated, the ROM with RB of VMs only predicts usually an overly stiff response. To correct for this effect, additional vectors are added to the RB, serving as additional DOFs to capture the hallmarks of the nonlinear response. For geometrically nonlinear structures, SMDs and DMs represent a good choice of companion vectors.

The SMDs are defined from the pairs of VMs retained in the RB. Specifically, the SMD θij∈Rn\theta_{ij} \in\mathbb{R}^{n} stemming from VMs ϕi\phi_i and ϕj\phi_j is defined as

K(1)θij=−∂2f(q)∂ϵi∂ϵj∣ϵi,ϵj=0,(8)K^{(1)}\theta_{ij} = -\left.\frac{\partial^2 f(\mathbf{q})}{\partial\epsilon_i \partial\epsilon_j}\right|_{\epsilon_i,\epsilon_j=0}, \tag*{(8)}

where q=ϕiϵi+ϕjϵj\mathbf{q} = \phi_i\epsilon_i + \phi_j\epsilon_j. In this way, SMDs are retrieved by solving all the linear systems of equations for all the different combinations of pairs of VMs in the RB. It is worth mentioning that the r.h.s. in Eq. (8) can be approximated using finite differences [23]. Namely, if a central difference scheme is employed, then

∂2f(q)∂ϵj∂ϵi∣ϵi,ϵj=0=∂KT(q)∂ϵi∣ϵi,ϵj=0ϕj≈KT(hϕi)−KT(−hϕi)2hϕj,(9)\left.\frac{\partial^2 f(\mathbf{q})}{\partial\epsilon_j \partial\epsilon_i}\right|_{\epsilon_i,\epsilon_j=0} = \left.\frac{\partial\mathbf{K}^{T}(\mathbf{q})}{\partial\epsilon_i}\right|_{\epsilon_i,\epsilon_j=0}\phi_j \approx\frac{\mathbf{K}^{T}(h\phi_i)-\mathbf{K}^{T}(-h\phi_i)}{2h}\phi_j, \tag*{(9)}

where KT∈Rn×n\mathbf{K}^{T} \in\mathbb{R}^{n \times n} is the tangent stiffness matrix, and hh is a user-defined scalar perturbation parameter. As a result, the SMDs can be computed non-intrusively from commercial FE programs. A major limitation in the use of SMDs as companion vectors to VMs arises from their quadratic increase in number with the number of VMs retained in the RB, leading to potentially inefficient ROMs. In fact, if nϕn_{\phi} is the number of VMs in the RB, the total number of ensuing SMDs is nϕ/2+nϕ2/2n_{\phi}/2 + n_{\phi}^{2}/2. Selection strategies for SMDs have been proposed in [24], reducing in this way the total number of vectors in the RB. Although derived from the different perspective [9, 8] of quasi-static modal analysis, it can be easily seen that SMDs appear in the second order Taylor series expansion of the nonlinear static solution problem

f(q∗)−∑i∈SϕϵiK(1)ϕi=0.(10)f(\mathbf{q}^{*})-\sum_{i\in S_{\phi}}\epsilon_i\mathbf{K}^{(1)}\phi_i=0. \tag*{(10)}

Interestingly, the same equations lie at the foundation of the DMs approach [25, 26, 10]. In fact, the DMs are obtained by solving the equations in (10) with a numerical scheme, for imposed loads in the directions of single VMs and linear combinations of pairs of VMs. Proper Orthogonal Decomposition is then used to identify the optimal linear subspace that spans the set of computed nonlinear solutions, retrieving the DMs [10].

Both DMs and SMDs are capable of capturing the nonlinear effects triggered by the nonlinear coupling between the VMs in the RB and the ones left out. In fact, this coupling is exercised in the construction of the RB (Eq. (10)) using loads that are orthogonal to all VMs not included in the RB. The contribution of the VMs left out from the RB, to the static solutions of Eq. (10), differs from zero only because of the nonlinearity. In this way, when using DMs or MDs as additional companion vectors to low frequency VMs, it is implicitly assumed that the modal dynamic coupling is closely related to the modal static coupling. The successful employment of SMDs in [20, 16, 23, 27, 28, 29, 30] and DMs in [26, 10, 5, 12] confirms the validity of this assumption.

In this way, the RB generally writes

V=[Φ,Θ]∈Rn×m,(11)\mathbf{V}=[\mathbf{\Phi},\mathbf{\Theta}]\in\mathbb{R}^{n\times m}, \tag*{(11)}

where Φ∈Rn×nϕ\mathbf{\Phi}\in\mathbb{R}^{n\times n_{\phi}} and Θ∈Rn×nθ\mathbf{\Theta}\in\mathbb{R}^{n\times n_{\theta}} are the matrices containing, respectively, the low frequency VMs, and the SMDs or DMs. While the first part of the RB is mass orthogonal by construction, the second part is mass orthogonalized to guarantee good numerical conditioning. For this purpose, the Gram-Schmidt algorithm [23] can be used.

Tensorial Formulation of ROM

Time integration of the ROM in the form written in Eq. (6) usually leads to small, if any, computational gains compared to the HFM. This is attributable to the expensive evaluations of the nonlinear internal forces and Jacobians required for the reconstruction of their reduced counterpart. In fact, the required number of operations scales with the size of the underlying FE model. Significant computational gains are achievable by leveraging an offline-online decomposition, where the ROM operators are computed once and for all in the offline model construction stage and used online for efficient evaluations [21, 31]. In this setting, the reduced mass and damping matrices are precomputed by projecting their HFM counterpart on the RB, while the reduced elastic forces are formulated as a cubic polynomial of the reduced coordinates:

f~(η)≜VTf(Vη)=K~(1)η+K~(2):(η⊗η)+K~(3):(η⊗η⊗η)⟺f~i=K~ij(1)ηj+K~ijk(2)ηjηk+K~ijkl(3)ηjηkηl,(12)\tilde{\mathbf{f}}(\boldsymbol{\eta})\triangleq\mathbf{V}^{T}\mathbf{f}(\mathbf{V}\boldsymbol{\eta})=\tilde{\mathbf{K}}^{(1)}\boldsymbol{\eta}+\tilde{\mathbf{K}}^{(2)}:(\boldsymbol{\eta}\otimes\boldsymbol{\eta})+\tilde{\mathbf{K}}^{(3)}:(\boldsymbol{\eta}\otimes\boldsymbol{\eta}\otimes\boldsymbol{\eta})\Longleftrightarrow\tilde{f}_i=\tilde{K}^{(1)}_{ij}\eta_j+\tilde{K}^{(2)}_{ijk}\eta_j\eta_k+\tilde{K}^{(3)}_{ijkl}\eta_j\eta_k\eta_l, \tag*{(12)}

Here, the reduced linear stiffness matrix, K~(1)∈Rm×m\tilde{\mathbf{K}}^{(1)}\in\mathbb{R}^{m\times m}, and the reduced nonlinear stiffness tensors, K~(2)∈Rm×m×m\tilde{\mathbf{K}}^{(2)}\in\mathbb{R}^{m\times m\times m} and K~(3)∈Rm×m×m×m\tilde{\mathbf{K}}^{(3)}\in\mathbb{R}^{m\times m\times m\times m}, are defined as

K~(1)=VTK(1)VK~ij(1)=VliKlk(1)Vkj(13)\tilde{\mathbf{K}}^{(1)}=\mathbf{V}^{T}\mathbf{K}^{(1)}\mathbf{V} \qquad\qquad\qquad\qquad \tilde{K}^{(1)}_{ij}=V_{li}K^{(1)}_{lk}V_{kj} \tag*{(13)}
K~(2)=(VTK(2)⋅V)⋅21V⟺K~ijk(2)=VliKlrp(2)VpkVrj(14)\tilde{\mathbf{K}}^{(2)}=\left(\mathbf{V}^{T}\mathbf{K}^{(2)}\cdot\mathbf{V}\right)_{\cdot21}\mathbf{V} \qquad\Longleftrightarrow\qquad \tilde{K}^{(2)}_{ijk}=V_{li}K^{(2)}_{lrp}V_{pk}V_{rj} \tag*{(14)}
K~(3)=((VTK(3)⋅V)⋅31V)⋅21VK~ijkl(3)=VpiKpqrs(3)VqjVrkVsl.(15)\tilde{\mathbf{K}}^{(3)}=\left(\left(\mathbf{V}^{T}\mathbf{K}^{(3)}\cdot\mathbf{V}\right)_{\cdot31}\mathbf{V}\right)_{\cdot21}\mathbf{V} \qquad\qquad \tilde{K}^{(3)}_{ijkl}=V_{pi}K^{(3)}_{pqrs}V_{qj}V_{rk}V_{sl}. \tag*{(15)}

Notice that the reduced order tensors are generally not sparse, as their full-order counterparts, but they are still symmetric. This feature allows for improved efficiency in reduced forces evaluations and limited storage requirements [21]. Similarly, the reduced tangent stiffness matrix, which is used in the assembly of the Jacobian of the residual in time integration, can be obtained by differentiating Eq. (12). This yields an expression that is a quadratic polynomial in the reduced coordinates, as reported in Appendix.

The computation of reduced tensors using Eq. (13) is possible only if the FE full order tensors are available. This is usually not the case when commercial FE programs are used. Consequently, the reduced order tensors are identified using non-intrusive techniques (a.k.a. indirect methods) that leverage standard outputs from commercial FE software. The ED [11] and EED [12] are the most popular approaches for tensor identification. While in the ED method the tensors are identified by computing the internal forces to a set of imposed displacements, in the EED method tensor identification is carried out from evaluations of the tangent stiffness matrix, thus improving efficiency [12].

Non intrusive tensor identification is generally an expensive procedure, and usually dominates the cost for the construction of equation based ROMs, especially when the number of the vectors in the RB is large. In [32] it was shown that decent speed-ups can be achieved by combining EED with hyperreduction techniques.

Non-Intrusive Parametric Reduced Order Model

Before proceeding with the description of the method for PROM construction, we introduce the normalized parameter vector p^\hat{\mathbf p}. Normalization is convenient, since the entries in the parameter vector p\mathbf p, defined in Eq. (4), may potentially differ by dimensions and order of magnitude. This makes the comparison between different parameter configurations challenging.

For this reason, we define the normalized parameter vector p^∈Rnp\hat{\mathbf p} \in\mathbb{R}^{n_p} from p\mathbf p as

p^i=pi−pimin⁡pimax⁡−pimin⁡,for i=1,…,np,(16)\hat{p}_i = \frac{p_i-p_i^{\min}}{p_i^{\max}-p_i^{\min}}, \qquad\text{for } i=1,\ldots,n_p, \tag*{(16)}

After normalization, the entries of the parameter vector can vary from zero to one, and the set of interest Sp\mathcal{S}_{\mathbf p} (defined in Eq. (4)) becomes a unitary hypercube centered at the origin. In the following, we denote this set by Sp^\mathcal{S}_{\hat{\mathbf p}}.

The equation driven PROM that we herein present leverages an offline-online decomposition, where in the offline stage we construct a database of ROMs that are subsequently interpolated online using RBFs. More formally, for a given realization of the parameter vector belonging to the parametric set of interest p^∈Sp^\hat{\mathbf p} \in\mathcal{S}_{\hat{\mathbf p}}, we define the corresponding ROM as the collection of six objects:

ROM⁡(p^):={V(p^),M~(p^),C~(p^),K~(1)(p^),K~(2)(p^),K~(3)(p^)},(17)\operatorname{ROM}(\hat{\mathbf p}) := \left\{\mathbf V(\hat{\mathbf p}),\widetilde{\mathbf M}(\hat{\mathbf p}),\widetilde{\mathbf C}(\hat{\mathbf p}),\widetilde{\mathbf K}^{(1)}(\hat{\mathbf p}),\widetilde{\mathbf K}^{(2)}(\hat{\mathbf p}),\widetilde{\mathbf K}^{(3)}(\hat{\mathbf p})\right\}, \tag*{(17)}

In this definition, V(p^)∈Rm\mathbf V(\hat{\mathbf p}) \in\mathbb{R}^{m} is the RB constructed at p^\hat{\mathbf p}, and the reduced order model operators M~(p^)\widetilde{\mathbf M}(\hat{\mathbf p}), C~(p^)\widetilde{\mathbf C}(\hat{\mathbf p}), K~(1)(p^)\widetilde{\mathbf K}^{(1)}(\hat{\mathbf p}), K~(2)(p^)\widetilde{\mathbf K}^{(2)}(\hat{\mathbf p}), K~(3)(p^)\widetilde{\mathbf K}^{(3)}(\hat{\mathbf p}) are defined as the projection of their HFM parametric-dependent counterpart on the parametric RB. In this way, the reduced mass, stiffness, and damping matrices read

M~(p^)=VT(p^)M(p^)V(p^),K~(1)(p^)=VT(p^)K(1)(p^)V(p^),C~(p^)=VT(p^)C(p^)V(p^),(18)\widetilde{\mathbf M}(\hat{\mathbf p}) = \mathbf V^{T}(\hat{\mathbf p})\mathbf M(\hat{\mathbf p})\mathbf V(\hat{\mathbf p}), \qquad\widetilde{\mathbf K}^{(1)}(\hat{\mathbf p}) = \mathbf V^{T}(\hat{\mathbf p})\mathbf K^{(1)}(\hat{\mathbf p})\mathbf V(\hat{\mathbf p}), \qquad\widetilde{\mathbf C}(\hat{\mathbf p}) = \mathbf V^{T}(\hat{\mathbf p})\mathbf C(\hat{\mathbf p})\mathbf V(\hat{\mathbf p}), \tag*{(18)}

while the nonlinear tensors, K~(2)(p^)\widetilde{\mathbf K}^{(2)}(\hat{\mathbf p}), K~(3)(p^)\widetilde{\mathbf K}^{(3)}(\hat{\mathbf p}), are defined through Eqs. (13).

In this work, we generate some training parameter samples and collect them in a training set Sp^train:={p^i}i=1Nt\mathcal{S}_{\hat{\mathbf p}}^{\mathrm{train}} := \left\{\hat{\mathbf p}^{i}\right\}_{i=1}^{N_t} with p^i∈Sp^\hat{\mathbf p}^{i} \in\mathcal{S}_{\hat{\mathbf p}}. Then, for each of the parameter samples we construct a ROM that we store in a ROM database, defined as

SROMtrain={ROM⁡(p^i)∣p^i∈Sp^train}i=1Nt.(19)\mathcal{S}_{\mathrm{ROM}}^{\mathrm{train}} = \left\{\operatorname{ROM}(\hat{\mathbf p}^{i}) \mid\hat{\mathbf p}^{i} \in\mathcal{S}_{\hat{\mathbf p}}^{\mathrm{train}}\right\}_{i=1}^{N_t}. \tag*{(19)}

This set of ROMs is used to train a RBF interpolation model that allows us to swiftly compute new ROMs, for not-sampled parameter values. Specifically, we construct RBF interpolants for all the individual ROM operators (i.e. RB, reduced mass and stiffness matrices, reduced tensors of nonlinear forces, and reduced damping matrix).

In the following, we split the exposition into three parts: first, we set forth the construction of the ROM database; second, we illustrate the interpolation scheme; third we present the properties of the proposed interpolation method.

Reduced Order Models Database Construction

Smoothness Requirement for ROM

Before proceeding with the construction of the ROMs database, we should guarantee that the way we define the single ROMs allows for interpolation, i.e. their variation with the parameters is continuous. Indeed, any attempt in interpolating a discontinuous function with continuous interpolants might potentially yield rather poor results. To ensure continuity of the ROM with p\mathbf{p}, we need to guarantee that the HFM operators (mass matrix, damping matrix and nonlinear forces) and the RB are continuously varying in p\mathbf{p}. Under these conditions, the ROM obtained by GP is continuous by construction.

For FE models of structures with geometric nonlinearities and linear elastic material, the continuity of the HFM is ensured if variations in p∈Sp\mathbf{p} \in\mathbf{S}_{\mathbf{p}} result in continuous variations in the mesh nodal coordinates, boundary conditions, or material properties. Conversely, the RB varies continuously across different parameters realizations only for some careful choice and consistent ordering of its vectors. Based on these observations, we construct ROMs using a global RB that becomes parameter-dependent after mass orthogonalization. As shown in the following, this RB meets the continuity requirement.

ROM Database Construction Algorithm

The construction of the ROM training database is presented in the following seven points, summarized in Algorithm 1.

  1. The first step towards ROM construction consists of sampling the admissible parameter space Sp^\mathbf{S}_{\hat{\mathbf{p}}}, generating a user-defined number NtN_t of training points that are collected in the training set Sp^train\mathbf{S}_{\hat{\mathbf{p}}}^{\mathrm{train}}. The algorithm presented here for PROM generation can be virtually used with any sampling scheme, provided that good coverage of the parameter set in ensured. Both structured and scattered grids can be used, as the interpolation scheme that we adopt is based on RBFs, and can handle both cases. However, in the authors’ experience, scattered grids generated with quasi-random sampling schemes, such as Latin Hypercube Sampling (LHS) [33], provide the best compromise between low number of samples and accuracy of the interpolation model.

  1. As a second step, we construct a FE model for each point in Sp^train\mathbf{S}_{\hat{\mathbf{p}}}^{\mathrm{train}}, and compute all the VMs within the frequency range of interest. Then, for each constructed model, we select the set of VMs that are excited at linear level by the load. This selection is carried out based on the Modal Participation Factor (MPF)[7] and on the load frequency content, as detailed in [34, 32]. The selected VMs are collected in a matrix Φi∈Rn×nϕi\mathbf{\Phi}^{i} \in\mathbb{R}^{n \times n_{\phi}^{i}}, where the index ii refers to the training sample number, and nϕin_{\phi}^{i} is the number of selected VMs. Notice that in principle, nϕin_{\phi}^{i} can be different for the different training samples.

  1. The third step is devoted to the computation of RB vectors needed to tackle nonlinearities, namely Θ\mathbf{\Theta} in Eq. (11). Again, for each point ii in Sp^train\mathbf{S}_{\hat{\mathbf{p}}}^{\mathrm{train}}, we construct a different matrix Θi∈Rn×nθi\mathbf{\Theta}^{i} \in\mathbb{R}^{n \times n_{\theta}^{i}}, where nθin_{\theta}^{i} is the number of additional vectors. Both SMDs or DMs can be used, but it is important to mention that the proposed PROM construction method is not restricted to this choice. In fact, any other approach that allows one to identify a good RB for capturing nonlinearities could be equivalently employed at this stage.

  1. In the fourth step, the global RB is constructed from the set of previously computed local RB, Φi\mathbf{\Phi}^{i} and Θi\mathbf{\Theta}^{i}, for i=1,…,Nti = 1, \ldots, N_t. Specifically, two matrices ΦG\mathbf{\Phi}_{\mathrm{G}} and ΘG\mathbf{\Theta}_{\mathrm{G}} are constructed by collecting together all the column vectors in Φi\mathbf{\Phi}^{i} and Θi\mathbf{\Theta}^{i}, for all the training points:

ΦG=[Φ1,…,ΦNt],ΘG=[Θ1,…,ΘNt].(20)\mathbf{\Phi}_{\mathrm{G}} = [\mathbf{\Phi}^{1}, \ldots, \mathbf{\Phi}^{N_t}], \qquad\mathbf{\Theta}_{\mathrm{G}} = [\mathbf{\Theta}^{1}, \ldots, \mathbf{\Theta}^{N_t}]. \tag*{(20)}

Here, ΦG∈Rn×∑inϕi\mathbf{\Phi}_{\mathrm{G}} \in\mathbb{R}^{n \times\sum_i n_{\phi}^{i}}, and ΘG∈Rn×∑inθi\mathbf{\Theta}_{\mathrm{G}} \in\mathbb{R}^{n \times\sum_i n_{\theta}^{i}}. Then, the column vectors in ΦG\mathbf{\Phi}_{\mathrm{G}} and ΘG\mathbf{\Theta}_{\mathrm{G}} are unit normalized, before separately applying Singular Value Decomposition (SVD), and obtaining

ΦG=LΦΣΦRΦT,ΘG=LΘΣΘRΘT,(21)\mathbf{\Phi}_{\mathrm{G}} = \mathbf{L}_{\Phi}\mathbf{\Sigma}_{\Phi}\mathbf{R}_{\Phi}^{T}, \qquad\mathbf{\Theta}_{\mathrm{G}} = \mathbf{L}_{\Theta}\mathbf{\Sigma}_{\Theta}\mathbf{R}_{\Theta}^{T}, \tag*{(21)}

where LΦ\mathbf{L}_{\Phi} and LΘ\mathbf{L}_{\Theta} are the left singular vectors, RΦT\mathbf{R}_{\Phi}^{T} and RΘT\mathbf{R}_{\Theta}^{T} are the right singular vectors, and ΣΦ\mathbf{\Sigma}_{\Phi} and ΣΘ\mathbf{\Sigma}_{\Theta} are the diagonal matrices of singular values. Next, we define the global RB as

V=[L~Φ,L~Θ],(22)\mathbf{V} = [\widetilde{\mathbf{L}}_{\Phi}, \widetilde{\mathbf{L}}_{\Theta}], \tag*{(22)}

where L~Φ∈Rn×mϕ\widetilde{\mathbf{L}}_{\Phi} \in\mathbb{R}^{n \times m_{\phi}} and L~Θ∈Rn×mθ\widetilde{\mathbf{L}}_{\Theta} \in\mathbb{R}^{n \times m_{\theta}} are the snapshot matrices containing respectively mϕm_{\phi} and mθm_{\theta} left singular vectors, stored respectively in LΦ\mathbf{L}_{\Phi} and LΘ\mathbf{L}_{\Theta}, and corresponding to the largest singular values. A suitable choice of mϕm_{\phi} and mθm_{\theta} is usually made using a relative truncation energy criterion, based on which the number of retained modes is chosen so that the energy of the projected snapshot matrix is retained up to a defined threshold. For a generic snapshots matrix A∈Rn×Ns\mathbf{A} \in\mathbb{R}^{n \times N_s}, with Ns≤nN_s \le n, the fraction of the energy retained in its reconstruction with m≤Nsm \le N_s left singular vectors can be written as

em=∑i=1mσi2∑i=1Nsσi2,(23)e_m = \frac{\sum_{i=1}^{m} \sigma_i^2}{\sum_{i=1}^{N_s} \sigma_i^2}, \tag*{(23)}

where σi\sigma_i are the singular values.

In this setting, two different relative energy threshold eϕe_{\phi} and eθe_{\theta} can be defined, resulting potentially in to different number of vectors in L~Φ\widetilde{\mathbf{L}}_{\Phi} and L~Θ\widetilde{\mathbf{L}}_{\Theta}. Selecting appropriate energy truncation thresholds is critical for accuracy of the ensuing ROM and is left to the analyst. However, based on the authors’ experience, energy threshold values greater than 0.99 generally yield accurate predictions.

At this point, two remarks can be made. In principle, a global RB could be constructed by stacking together all the vectors in ΦG\mathbf{\Phi}_{G} and ΘG\mathbf{\Theta}_{G} in a single matrix. However, the PROM derived using this RB would be extremely inefficient, as its size would be equal to the sum of the degrees of freedom of the local RBs over the entire set of training points. Moreover, this RB would carry a lot of unnecessary repeated information, since the variations in Φi\mathbf{\Phi}^{i} and Θi\mathbf{\Theta}^{i} are expected to be relatively small across different parameter realizations. Based on these observations, we propose to compress the global RB with POD, applied separately to the ’linear part’ and to the ’nonlinear part’ of the RB. This latter choice is made to allow more freedom in the RB compression process as to what pertains the accuracy of the resulting global RB in approximating the linear and nonlinear components of the local RB (i.e. Φi\mathbf{\Phi}^{i} and Θi\mathbf{\Theta}^{i}). In fact, based on the authors’ experience, for the same global RB size, the best results are obtained using a higher energy threshold for the truncation of the left singular vectors of the linear part, rather than for the nonlinear part, i.e. eϕ≥eθe_{\phi} \ge e_{\theta}.

Additionally, the POD compression process, and consequently the extracted compressed RB, are significantly impacted by the scaling of the vectors in the snapshot matrix in Eq. (20). In contrast to snapshots matrices constructed from displacements of HFM simulations, where the relative scaling of the different snapshots carries a physical meaning, the scaling of the vectors in ΦG\mathbf{\Phi}_{G} and ΘG\mathbf{\Theta}_{G} lacks of physical interpretation. Indeed, both VMs and SMDs are defined up to a multiplicate constant. For this reason, we unit normalize the columns of the snapshots matrices prior to POD compression. By doing so, we aim for a compression that is optimal in a global sense, as POD weighs all the different vectors equally.

  1. In the fifth step, we mass orthogonalize the global RB in Eq. (22), for all points in Sp^train\mathbf{S}_{\hat{\mathbf{p}}}^{\mathrm{train}}. This operation is needed to guarantee a well conditioned ROM. Mass orthogonalization is performed by computing the VMs of the different ROMs in the training set, obtained by projecting the parametric mass and stiffness FE matrices onto the global RB. Specifically, for each realization p^i∈Sp^train\hat{\mathbf{p}}^{i} \in\mathbf{S}_{\hat{\mathbf{p}}}^{\mathrm{train}}, we solve the generalized eigenvalue problem

−VTM(p^i)Vωj2+VTK(1)(p^i)Vϕ~ji=0,for j=1,…,m,m=nϕ+nθ,(24)-\mathbf{V}^{T}\mathbf{M}(\hat{\mathbf{p}}^{i})\mathbf{V}\omega_{j}^{2}+\mathbf{V}^{T}\mathbf{K}^{(1)}(\hat{\mathbf{p}}^{i})\mathbf{V}\tilde{\boldsymbol{\phi}}_{j}^{i}=0,\quad\text{for } j=1,\ldots,m,\quad m=n_{\phi}+n_{\theta}, \tag*{(24)}

where ϕ~ji∈Rn\tilde{\boldsymbol{\phi}}_{j}^{i} \in\mathbb{R}^{n}, and ωj\omega_j are respectively the jjth VM and angular frequency of the ROM in reduced coordinates. Using the VMs, the mass-orthogonalized RBs are then retrieved by performing the change of variables

Vi=VΦ~i,(25)\mathbf{V}^{i}=\mathbf{V}\widetilde{\mathbf{\Phi}}^{i}, \tag*{(25)}

with

Φ~i=[ϕ~1i,…,ϕ~mi],(26)\widetilde{\mathbf{\Phi}}^{i}=\left[\tilde{\boldsymbol{\phi}}_{1}^{i},\ldots,\tilde{\boldsymbol{\phi}}_{m}^{i}\right], \tag*{(26)}

where again index ii refers to p^i∈Sp^train\hat{\mathbf{p}}^{i} \in\mathbf{S}_{\hat{\mathbf{p}}}^{\mathrm{train}}.

After mass-orthogonalization, the global RB inherits local features, since, for each parameter p^i∈Sp^train\hat{\mathbf{p}}^{i} \in\mathbf{S}_{\hat{\mathbf{p}}}^{\mathrm{train}}, it is transformed using the local parameter-dependent mass and stiffness matrices. However, the range of Vi\mathbf{V}^{i}, corresponding to different p^i\hat{\mathbf{p}}^{i}, remains the same, as the column vectors are just recombined to diagonalize the linear part of the different ROMs.

  1. The vectors in the different RBs Vi\mathbf{V}^{i} vary continuously across the parameter samples in the training set, if consistently ordered. In fact, if M(p^)\mathbf{M}(\hat{\mathbf{p}}) and K(1)(p^)\mathbf{K}^{(1)}(\hat{\mathbf{p}}) vary continuously with p^\hat{\mathbf{p}}, as a consequence, also the VMs of the ROM, obtained by solving the eigenvalue problem in Eq. (24), vary continuously with p^\hat{\mathbf{p}} [35]. However, the ordering of the column vectors in each RB plays a crucial role when it comes to guarantee continuity, column by column. Ordering vectors based on the corresponding ROM angular frequencies (see Eq. (24)) is unwarranted and can lead to discontinuities in the RBs across the parameter set. In fact, mode veering and mode crossing phenomena can potentially appear [36, 37, 35], affecting the ROM VMs. When mode veering occurs, the eigenfrequencies corresponding to pairs of different VMs approach each other, as the parameters of the systems are varied, and then veer away without crossing. Close to the veering point, the shapes of the two involved VMs change smoothly, and the modes morph into one another [38, 37]. In contrast, in mode crossing the eigenfrequencies of two different VMs coalesce for a critical value of the parameters and then separate again. In this case, the VMs retain their shape before and after the crossing, without interacting [38]. At the crossing point, the space associated to the two crossing VMs becomes degenerate, and every combination of the two VMs is also a VM of the system [7].

Reordering algorithms have been employed in the literature to track branches of VMs, both in numerical and experimental procedures [39, 40, 35, 40, 38]. These algorithms rely on the Modal Assurance Criterion (MAC) [41], a normalized index that expresses the similarity of different mode shapes. In this work, we use a mass weighted MAC number to perform consistent reordering of the vectors in the set of RBs, SV:={Vi}i=1Nt\mathrm{S}_{\mathrm{V}} := \{\mathbf{V}^{i}\}_{i=1}^{N_t}, with Vi\mathbf{V}^{i} defined in Eq. (25). Specifically, consider two RBs, VI\mathbf{V}^{I} and VJ\mathbf{V}^{J}, corresponding to the pair of parameter realizations p^I,p^J∈Sp^\hat{\mathbf{p}}^{I},\hat{\mathbf{p}}^{J} \in\mathrm{S}_{\hat{\mathbf{p}}}. For each pair of column vectors viI\mathbf{v}^{I}_{i} in VI\mathbf{V}^{I} and vjJ\mathbf{v}^{J}_{j} in VJ\mathbf{V}^{J}, we define the mass weighted MAC number as

MACijIJ=(viITM(I)vjJ)2viITM(I)viI⋅vjJTM(I)vjJ,for i,j=1,…,m,(27)MAC^{IJ}_{ij} = \frac{\left(\mathbf{v}^{I}_{i}{}^{\mathsf{T}}\mathbf{M}^{(I)}\mathbf{v}^{J}_{j}\right)^{2}}{\mathbf{v}^{I}_{i}{}^{\mathsf{T}}\mathbf{M}^{(I)}\mathbf{v}^{I}_{i}\cdot\mathbf{v}^{J}_{j}{}^{\mathsf{T}}\mathbf{M}^{(I)}\mathbf{v}^{J}_{j}}, \qquad\text{for } i,j=1,\ldots,m, \tag*{(27)}

where M(I)=M(p^I)\mathbf{M}^{(I)}=\mathbf{M}(\hat{\mathbf{p}}^{I}) is the mass matrix computed for parameter p^I\hat{\mathbf{p}}^{I}.2 The MAC number can vary from zero to one, indicating low modal correlation for values near zero and high modal correlation for values close to one. The mass weighted MAC number is preferred to the standard MAC number, as it is used to reorder mass orthogonal vectors - see the definition of local RBs in Eqs. (25) and (24). Based on the authors’ experience, this choice results in a more robust reordering scheme compared to when the standard MAC number is used, as the difference in MAC numbers between correlated and uncorrelated pairs of vectors increases.

Reordering is performed one RB at a time. A reference RB VI\mathbf{V}^{I} is chosen from set SV\mathrm{S}_{\mathrm{V}}. Then, its vectors are related to the vectors of the RB to be reordered, here referred to as VJ∈SV\mathbf{V}^{J} \in\mathrm{S}_{\mathrm{V}}. Specifically, the MAC number matrix is computed from the vectors in the two RBs using Eq. (27). Then, vectors in VI\mathbf{V}^{I} are associated to the vectors of VJ\mathbf{V}^{J} that maximize the MAC number. Formally, we can define a reordering vector bIJ∈Rm\mathbf{b}^{IJ} \in\mathbb{R}^{m} as

biIJ=arg⁡max⁡1≤j≤m  MACijIJ,for i=1,…,m,(28)b^{IJ}_{i} = \underset{1\leq j\leq m}{\arg\max}\; MAC^{IJ}_{ij}, \qquad\text{for } i=1,\ldots,m, \tag*{(28)}

and the reordered RB V~J\widetilde{\mathbf{V}}^{J} is obtained by reordering the columns of VJ\mathbf{V}^{J} as

V~J(:,i)=VJ(:,biIJ),(29)\widetilde{\mathbf{V}}^{J}(:,i)=\mathbf{V}^{J}(:,b^{IJ}_{i}), \tag*{(29)}

where we adopt index notation.

In the present Algorithm, vectors in different RBs are ordered using different reference RBs. This choice is motivated by the fact that RB vectors might exhibit strong variations with the parameters and using a constant reference might result in poor tracking of VMs braches. In fact, if the reference point far is too far away from the reorder point in the parameter space, correlation between the vectors in the two RBs might be lost. Consequently, for each RB that we order, we choose as reference RB, the one corresponding to the sample point in the parameter space that is the closest to the reordering point. The adopted distance metric is the ll-2 norm in the normalized parameter space:

d(p^i,p^j)=∥p^i−p^j∥2,(30)d(\hat{\mathbf{p}}^i,\hat{\mathbf{p}}^j)=\lVert\hat{\mathbf{p}}^i-\hat{\mathbf{p}}^j\rVert_2, \tag*{(30)}

where p^i\hat{\mathbf{p}}^i and p^j\hat{\mathbf{p}}^j are two realizations of the normalized parameter vector.

We then define two different sets: one containing reordered RBs, referred to as S~V\widetilde{\mathbf{S}}_{\mathrm{V}}, and one containing the RBs that still need to be reordered, namely SV\mathbf{S}_{\mathrm{V}}. The idea is to subsequently draw RBs from the set of not ordered RBs SV\mathbf{S}_{\mathrm{V}}, reordering them and add them to the set of reordered RBs S~V\widetilde{\mathbf{S}}_{\mathrm{V}}, until all RBs are reordered. At each iteration, the selection of the RB to order from SV\mathbf{S}_{\mathrm{V}}, and of the reference RB from S~V\widetilde{\mathbf{S}}_{\mathrm{V}} is performed based on the distances between all the parameter values corresponding to the RB in the two sets. These distances are computed using Eq. (30) and stored in matrix D∈R∣SV∣×∣S~V∣\mathbf{D}\in\mathbb{R}^{|\mathbf{S}_{\mathrm{V}}|\times|\widetilde{\mathbf{S}}_{\mathrm{V}}|} as

Dij=d(p^i,p^j).(31)D_{ij}=d(\hat{\mathbf{p}}^i,\hat{\mathbf{p}}^j). \tag*{(31)}

Here, p^i\hat{\mathbf{p}}^i and p^j\hat{\mathbf{p}}^j are parameter realizations whose associated RBs are respectively in SV\mathbf{S}_{\mathrm{V}} and S~V\widetilde{\mathbf{S}}_{\mathrm{V}}. We then compare all the values in the distance matrix and select the RBs in SV\mathbf{S}_{\mathrm{V}} and S~V\widetilde{\mathbf{S}}_{\mathrm{V}} that correspond to the minimum distance in the parameter space.

This translates in reordering RB SV(i∗)\mathbf{S}_{\mathrm{V}}(i^*) using the reference RB S~V(j∗)\widetilde{\mathbf{S}}_{\mathrm{V}}(j^*), where

(i∗,j∗)=arg⁡min⁡1≤i≤∣SV∣,1≤j≤∣S~V∣ Dij.(32)(i^*,j^*)=\underset{1\leq i\leq|\mathbf{S}_{\mathrm{V}}|,1\leq j\leq|\widetilde{\mathbf{S}}_{\mathrm{V}}|}{\arg\min}\ D_{ij}. \tag*{(32)}

Loosely speaking, we define a starting reference point in the parameter space, and progressively move away from it, while reordering the nearby RBs by referring them to the closest RB in the reordered set, as shown in Figure 1. The reader should note that an initial reference RB must be defined for the first reordering iteration. This choice is arbitrary and is expected to affect the final ordering in the vectors of the different RBs but not reordering consistency.

Figure 1. Schematic illustration of the reordering scheme used for RB reordering for a 2D parameter space, for the first three iterations. At each iteration, we denote in magenta and red, respectively, the parameter points corresponding to the reference RB and to the RB that is reordered at the current iteration.

A limitation of RB reordering using the present approach is that two or more distinct vectors in the RB to be reordered, VJ\mathbf{V}^{J}, may be associated to the same vector of the reference RB, VI\mathbf{V}^{I}. This situation arises when repeated values appear in the reordering vector bJI\mathbf{b}^{I}_{J}, defined in Eq. (28). As a result, the reordered RB becomes rank deficient and the ensuing ROM ill-defined. This issue highlights the inability of the MAC to distinguish different modes and is usually a symptom of an overly coarse parameter sampling, which causes too large variations in the shape of neighboring RBs vectors. An effective remedy is to refine the parameter sampling to reduce the distance between training points, thereby increasing the spatial correlation across RBs.

  1. The last step in the ROMs database construction consists of computing the reduced order operators from the full order operators using the orthogonalized, reordered RBs. In particular, reduced mass, stiffness, and damping matrices are obtained using Eq. (18), while the reduced nonlinear stiffness tensors are identified using either the ED or the EED method.

Following the aforementioned procedure, we are able to construct a ROM database in the form presented in Eq. (19). Moreover, the ROM operators are guaranteed to vary continuously across the parameter set of interest, allowing for interpolation.

We would like to stress that with the aforementioned procedure for ROM construction, we end up with a database of ROM whose linear part is diagonalized. This proves to be a useful feature for ROM interpolation, as we will show in § 3.2.3. Furthermore, if the ROM VMs (obtained using Eq. (24)) are mass-normalized, the ROM mass matrix M~\tilde{\mathbf{M}} is the identity matrix, while the reduced stiffness matrix K~(1)\tilde{\mathbf{K}}^{(1)} is diagonal and contains the square of the ROM angular frequencies [7]. Specifically,

M~jk(p^i)=δjk,K~jk(1)(p^i)=ωj(i)2δjk,for j,k=1,…,m;(33)\tilde{M}_{jk}(\hat{\mathbf{p}}^{i}) = \delta_{jk}, \quad\tilde{K}_{jk}^{(1)}(\hat{\mathbf{p}}^{i}) = \omega_{j}^{(i)2}\delta_{jk}, \quad\text{for } j,k=1,\ldots,m; \tag*{(33)}

with p^i∈Sp^train\hat{\mathbf{p}}^{i} \in S_{\hat{\mathbf{p}}}^{\mathrm{train}}, and δij\delta_{ij} being the Kronecker delta. As a result, the Rayleigh damping reduced matrix is also diagonal and writes

C~jk(p^i)=α(p^i)δjk+β(p^i)K~jk(1)(p^i)δjk,for j,k=1,…,m.(34)\tilde{C}_{jk}(\hat{\mathbf{p}}^{i}) = \alpha(\hat{\mathbf{p}}^{i})\delta_{jk} + \beta(\hat{\mathbf{p}}^{i})\tilde{K}_{jk}^{(1)}(\hat{\mathbf{p}}^{i})\delta_{jk}, \quad\text{for } j,k=1,\ldots,m. \tag*{(34)}

Interpolation scheme for Parametric ROM

RBF interpolant computation

The interpolation of the ROM database is carried out using RBF. Let Q\mathbf{Q} be any of the ROM operators V\mathbf{V}, C~\tilde{\mathbf{C}}, K~(1)\tilde{\mathbf{K}}^{(1)}, K~(2)\tilde{\mathbf{K}}^{(2)}, and K~(3)\tilde{\mathbf{K}}^{(3)}, 3^3 and NeN_e be the total number of non-zero entries in Q\mathbf{Q} that we collect here in vector g∈RNe\mathbf{g} \in\mathbb{R}^{N_e}. Any generic jjth entry in Q\mathbf{Q} is a real-valued function of the ROM parameter vector, gj:p^∈RNp↦Rg_j:\hat{\mathbf{p}} \in\mathbb{R}^{N_p} \mapsto\mathbb{R}, and can be approximated using RBFs as

gj(p^)≈∑i=1NtWjiγ(∥p^−p^i∥2),(35)g_j(\hat{\mathbf{p}}) \approx\sum_{i=1}^{N_t} W_{ji}\gamma(\lVert\hat{\mathbf{p}}-\hat{\mathbf{p}}^{i}\rVert_2), \tag*{(35)}

where p^i∈Sp^train\hat{\mathbf{p}}^{i} \in S_{\hat{\mathbf{p}}}^{\mathrm{train}} are the centers of the RBF interpolant, WjiW_{ji} the associated weights, and γ:R↦R\gamma:\mathbb{R}\mapsto\mathbb{R} is the RBF kernel function. This equation can also be written in matrix form as

g(p^)≈W⋅γ(p^),(36)\mathbf{g}(\hat{\mathbf{p}}) \approx\mathbf{W}\cdot\boldsymbol{\gamma}(\hat{\mathbf{p}}), \tag*{(36)}

where W∈RNe×Nt\mathbf{W}\in\mathbb{R}^{N_e\times N_t} is the matrix of weights and γ∈RNt\boldsymbol{\gamma}\in\mathbb{R}^{N_t} with γi=γ(∥p^−p^i∥2)\gamma_i=\gamma(\lVert\hat{\mathbf{p}}-\hat{\mathbf{p}}^i\rVert_2). The RBF weights are computed by evaluating the ROM operator vector g\mathbf{g} at the NtN_t training points in Sp^trainS_{\hat{\mathbf{p}}}^{\mathrm{train}} and solving the linear systems of equations for the matrix of weights W∈RNe×Nt\mathbf{W}\in\mathbb{R}^{N_e\times N_t} that reads

GT=ΓT⋅WT.(37)\mathbf{G}^{T}=\boldsymbol{\Gamma}^{T}\cdot\mathbf{W}^{T}. \tag*{(37)}

Here, G∈RNe×Nt\mathbf{G}\in\mathbb{R}^{N_e\times N_t} and Γ∈RNt×Nt\boldsymbol{\Gamma}\in\mathbb{R}^{N_t\times N_t} are defined as

G=[g(p^1),…,g(p^Nt)],Γ=[γ(∥p^1−p^1∥2)γ(∥p^1−p^2∥2)⋯γ(∥p^1−p^Nt∥2)γ(∥p^2−p^1∥2)γ(∥p^2−p^2∥2)⋯γ(∥p^2−p^Nt∥2)⋮⋮⋱⋮γ(∥p^Nt−p^1∥2)γ(∥p^Nt−p^2∥2)⋯γ(∥p^Nt−p^Nt∥2)].(38)\mathbf{G}=[\mathbf{g}(\hat{\mathbf{p}}^1),\ldots,\mathbf{g}(\hat{\mathbf{p}}^{N_t})],\quad \boldsymbol{\Gamma}= \begin{bmatrix} \gamma(\lVert\hat{\mathbf{p}}^1-\hat{\mathbf{p}}^1\rVert_2) & \gamma(\lVert\hat{\mathbf{p}}^1-\hat{\mathbf{p}}^2\rVert_2) & \cdots& \gamma(\lVert\hat{\mathbf{p}}^1-\hat{\mathbf{p}}^{N_t}\rVert_2) \\ \gamma(\lVert\hat{\mathbf{p}}^2-\hat{\mathbf{p}}^1\rVert_2) & \gamma(\lVert\hat{\mathbf{p}}^2-\hat{\mathbf{p}}^2\rVert_2) & \cdots& \gamma(\lVert\hat{\mathbf{p}}^2-\hat{\mathbf{p}}^{N_t}\rVert_2) \\ \vdots& \vdots& \ddots& \vdots\\ \gamma(\lVert\hat{\mathbf{p}}^{N_t}-\hat{\mathbf{p}}^1\rVert_2) & \gamma(\lVert\hat{\mathbf{p}}^{N_t}-\hat{\mathbf{p}}^2\rVert_2) & \cdots& \gamma(\lVert\hat{\mathbf{p}}^{N_t}-\hat{\mathbf{p}}^{N_t}\rVert_2) \end{bmatrix}. \tag*{(38)}

The reader should note that the solution of these linear systems of equations is usually not computationally intensive, as it requires only one factorization of matrix Γ\boldsymbol{\Gamma}, which is usually small. Moreover, interpolating ROM operators based on Eq. (35) is much faster than constructing a new ROM as described in § 2.2. In fact, the RBF model evaluation requires only NtN_t evaluations of the RBF kernel γ(p^)\gamma(\hat{\mathbf{p}}) and one matrix-vector multiplication. Then, after interpolation, the entries in the interpolated vector g\mathbf{g} are rearranged in the interpolated ROM operators C~(p^)\tilde{\mathbf{C}}(\hat{\mathbf{p}}), K~(1)(p^)\tilde{\mathbf{K}}^{(1)}(\hat{\mathbf{p}}), K~(2)(p^)\tilde{\mathbf{K}}^{(2)}(\hat{\mathbf{p}}), K~(3)(p^)\tilde{\mathbf{K}}^{(3)}(\hat{\mathbf{p}}), and V(p^)\mathbf{V}(\hat{\mathbf{p}}), retrieving ROM(p^)ROM(\hat{\mathbf{p}}) in its canonical form.

Validation of RBF Kernel

In RBF interpolation, different kernels can be used, among which popular choices are Gaussian, multiquadratic, and inverse multi-quadratic. The kernel functions may feature a dependency on few hyperparameters that govern their shape. In the following, we denote the RBF parameter vector with ϵ\epsilon. For example, the Gaussian kernel RBF, γg(x,ϵ)=e−(ϵ∥x−xi∥2)2\gamma_g(\boldsymbol{x},\epsilon)=e^{-(\epsilon\lVert\boldsymbol{x}-\boldsymbol{x}_i\rVert_2)^2}, is a function of the shape parameter ϵ\epsilon that governs its peakedness.

The correct choice of RBF hyperparameters plays a crucial role on interpolation accuracy and ad-hoc tuning can be made using a validation set. Specifically, different weights are computed using different RBF hyperparameters, resulting in different interpolants. Then, accuracy of the different interpolants, in approximating the function evaluations in the validation set, is compared, and the interpolation model associated with the lowest error is selected.

Following this scheme, we first construct different interpolants of the ROM operators, from the parameter samples in Sp^train\mathbf{S}_{\hat{p}}^{\mathrm{train}}, using different values of the hyperparameter vectors ϵ\epsilon. This is done by repeatedly solving the linear systems in Eq. (37), obtaining ϵ\epsilon-dependent weight matrices W(ϵ)\mathbf{W}(\epsilon). Then, we construct a validation parameter set Sp^val:={p^i}i=1Nv\mathbf{S}_{\hat{p}}^{\mathrm{val}}:=\{\hat{\boldsymbol{p}}_i\}_{i=1}^{N_v}, using the same sampling scheme adopted for the construction of the training set, and its associated ROM database SROMval={ROM(p^i)∣p^i∈Sp^val}i=1Nv\mathbf{S}_{\mathrm{ROM}}^{\mathrm{val}}=\{\mathrm{ROM}(\hat{\boldsymbol{p}}^i)\mid\hat{\boldsymbol{p}}^i\in\mathbf{S}_{\hat{p}}^{\mathrm{val}}\}_{i=1}^{N_v}. This is done by following the procedure detailed in § 3.1. Different approximations to the ROM operators are obtained as shown in Eq. (36) for all p^i∈Sp^val\hat{\boldsymbol{p}}^i\in\mathbf{S}_{\hat{p}}^{\mathrm{val}}, using weights W(ϵ)\mathbf{W}(\epsilon) corresponding to the different ϵ\epsilon values. An average relative error measure is defined as

erel(ϵ)=∑i=1Nv∥g(p^i)−W(ϵ)γϵ(p^i)∥2∥g(p^i)∥2,with p^i∈Sp^val,(39)e_{\mathrm{rel}}(\epsilon)=\sqrt{\sum_{i=1}^{N_v}\frac{\lVert\boldsymbol{g}(\hat{\boldsymbol{p}}^i)-\mathbf{W}(\epsilon)\boldsymbol{\gamma}_{\epsilon}(\hat{\boldsymbol{p}}^i)\rVert_2}{\lVert\boldsymbol{g}(\hat{\boldsymbol{p}}^i)\rVert_2}},\qquad\text{with }\hat{\boldsymbol{p}}^i\in\mathbf{S}_{\hat{p}}^{\mathrm{val}}, \tag*{(39)}

where in vector g\boldsymbol{g} we collect the values of the exact entries of the ROM operators stored in SROMval\mathbf{S}_{\mathrm{ROM}}^{\mathrm{val}}. Based on this error metric, we select the interpolation model corresponding to the value of ϵ\epsilon for which the error is the lowest. It is worth stressing that different RBF interpolation models are constructed separately for the entries of the different ROM operators, resulting in different optimal values of ϵ\epsilon. This additional flexibility in the choice of the RBF hyperparameters allows us to improve the fit, as compared to the case where the same value of ϵ\epsilon is used for all operators.

Moreover, we remind the reader that even if our interpolation model is validated here only with respect to RBF hyperparameters, in principle, one could use the validation set to optimally select the best RBF kernel function. This could be achieved following the same procedure adopted for RBF hyperparameters validation, paying the cost of additional solutions of Eq. (37).

Algorithm 1 ROM Database Construction

Input: constructFeModel, Sp\mathbf{S}_p, NtN_t

Output: ROM database SROMtrain\mathbf{S}_{\mathrm{ROM}}^{\mathrm{train}}

Generation of training set

  1. Sptrain←\mathbf{S}_{p}^{\mathrm{train}} \leftarrow generateParamSamples(Sp,Nt)(\mathbf{S}_p,N_t)4

  2. Sp^train←\mathbf{S}_{\hat{p}}^{\mathrm{train}} \leftarrow normalizeParamSamples(Sptrain,pbounds)(\mathbf{S}_{p}^{\mathrm{train}},\mathbf{p}_{\mathrm{bounds}})5

Construction of the local RBs

  1. for i=1,…,Nti=1,\ldots,N_t do

  2. p^i←Sp^train(i)\hat{\mathbf{p}}^i \leftarrow\mathbf{S}_{\hat{p}}^{\mathrm{train}}(i)

  3. FeModeli←\mathrm{FeModel}^i \leftarrow constructFEModel(p^i)(\hat{\mathbf{p}}^i)6

  4. Mi←FeModel.MassMatrix()\mathbf{M}^i \leftarrow\mathrm{FeModel}.\mathrm{MassMatrix}(), K(1),i←FeModel.LinStiffMatrix()\mathbf{K}^{(1),i} \leftarrow\mathrm{FeModel}.\mathrm{LinStiffMatrix}()

  5. Φi←\mathbf{\Phi}^i \leftarrow computeAndSelectVMs(Mi,K(1),i,f(e)(t))(\mathbf{M}^i,\mathbf{K}^{(1),i},f^{(e)}(t))7

  6. Θi←\mathbf{\Theta}^i \leftarrow computeBasisNL(FeModel,Φi)(\mathrm{FeModel},\mathbf{\Phi}^i)8

  7. end for

Construction of the global RB

  1. ΦG←[Φ1,…,ΦNt]\mathbf{\Phi}_{G}\leftarrow[\mathbf{\Phi}^{1},\ldots,\mathbf{\Phi}^{N_t}], ΘG←[Θ1,…,ΘNt]\mathbf{\Theta}_{G}\leftarrow[\mathbf{\Theta}^{1},\ldots,\mathbf{\Theta}^{N_t}]

  2. [LΦΣΦRΦT]←SVD(ΦG)[\mathbf{L}_{\Phi}\mathbf{\Sigma}_{\Phi}\mathbf{R}_{\Phi}^{T}]\leftarrow\mathrm{SVD}(\mathbf{\Phi}_{G}), [LΘΣΘRΘT]←SVD(ΘG)[\mathbf{L}_{\Theta}\mathbf{\Sigma}_{\Theta}\mathbf{R}_{\Theta}^{T}]\leftarrow\mathrm{SVD}(\mathbf{\Theta}_{G})9

  3. V←[L~Φ(:,1:mϕ),L~Θ(:,1:mθ)]\mathbf{V}\leftarrow[\tilde{\mathbf{L}}_{\Phi}(:,1:m_{\phi}),\tilde{\mathbf{L}}_{\Theta}(:,1:m_{\theta})]10

Mass-orthogonalize RB

  1. for i=1,…,Nti=1,\ldots,N_t do

  2. M~i←VTMiV\tilde{\mathbf{M}}^i\leftarrow\mathbf{V}^{T}\mathbf{M}^i\mathbf{V}, K~(1),i←VTK(1),iV\tilde{\mathbf{K}}^{(1),i}\leftarrow\mathbf{V}^{T}\mathbf{K}^{(1),i}\mathbf{V}

  3. Φ~i←VMs(M~i,K~(1),i)\tilde{\mathbf{\Phi}}^i\leftarrow\mathrm{VMs}(\tilde{\mathbf{M}}^i,\tilde{\mathbf{K}}^{(1),i})11

  4. Vi←VΦ~i\mathbf{V}^i\leftarrow\mathbf{V}\tilde{\mathbf{\Phi}}^i

  5. end for

Reorder RBs

  1. [V1,…,VNt]←[\mathbf{V}^{1},\ldots,\mathbf{V}^{N_t}]\leftarrow reorderRBs([V1,…,VNt],Sp^train)([\mathbf{V}^{1},\ldots,\mathbf{V}^{N_t}],\mathbf{S}_{\hat{p}}^{\mathrm{train}})12

Construct ROMs

  1. for i=1,…,Nti=1,\ldots,N_t do

  2. M~i←ViTMiVi\tilde{\mathbf{M}}^i\leftarrow\mathbf{V}^{iT}\mathbf{M}^i\mathbf{V}^i, K~(1),i←ViTK(1),iVi\tilde{\mathbf{K}}^{(1),i}\leftarrow\mathbf{V}^{iT}\mathbf{K}^{(1),i}\mathbf{V}^i, C~i←ViTCiVi\tilde{\mathbf{C}}^i\leftarrow\mathbf{V}^{iT}\mathbf{C}^i\mathbf{V}^i

  3. [K~(2),i,K~(3),i]←[\tilde{\mathbf{K}}^{(2),i},\tilde{\mathbf{K}}^{(3),i}]\leftarrow identifyNonlinearTensors(Vi,FeModeli)(\mathbf{V}^i,\mathrm{FeModel}^i)13

  4. SROMtrain(i)←{Vi,M~i,K~(1),i,C~i,K~(2),i,K~(3),i}\mathbf{S}_{\mathrm{ROM}}^{\mathrm{train}}(i)\leftarrow\{\mathbf{V}^i,\tilde{\mathbf{M}}^i,\tilde{\mathbf{K}}^{(1),i},\tilde{\mathbf{C}}^i,\tilde{\mathbf{K}}^{(2),i},\tilde{\mathbf{K}}^{(3),i}\}

  5. end for

Properties of PROM interpolation

In this Subsection we present the properties of the interpolation model herein proposed. As reported in the literature, any good technique for constructing a ROM for structural dynamics should preserve the physics of the problem, to avoid incurring significant loss in accuracy, or worse, numerical issues pertaining to ROM time integration [22, 18, 17]. Such properties are related to the Lagrangian Structure inherent to the HFM, based on which:

(i) The mass matrix mass M\mathbf{M} is symmetric-positive definite, as the kinetic energy is always positive; (ii) The linear stiffness matrix K(1)\mathbf{K}^{(1)} and the nonlinear stiffness tensors K(2),K(3)\mathbf{K}^{(2)},\mathbf{K}^{(3)} are symmetric about all axes, as detailed in § 2.1, since they are obtained by differentiation of an elastic potential function;

(iii) The linear stiffness matrix K(1)\mathbf{K}^{(1)} is positive definite, as the undeformed configuration is stable;

(iv) The damping matrix C\mathbf{C} is positive definite, as the system is overall dissipative.

A ROM is said to preserve the Lagrangian structure when the aforementioned properties are carried over from the HFM operators to their ROM counterparts. Based on this definition, Galerkin Projection ROMs are structure preserving, as shown in [22]. The interpolation technique herein presented is also structure preserving, with some caveats. In fact:

(i) The reduced matrix mass matrix M~\widetilde{\mathbf{M}} is not interpolated, as it is constant and equal to the identity. Thus its symmetry and positive definiteness is guaranteed;

(ii) The reduced linear stiffness matrix is diagonal K~(1)\widetilde{\mathbf{K}}^{(1)} (see Eq. (33)) and only its diagonal entries are interpolated, preserving symmetry. In contrast, its positive definiteness is not necessarily preserved. Based on the authors’ experience, if a positive definite RBF kernel is used and if the training grid is fine enough, the interpolated entries of K~(1)\widetilde{\mathbf{K}}^{(1)} are positive. However, to avoid issues, we always check the sign of the entries in K~(1)\widetilde{\mathbf{K}}^{(1)} after interpolation and before time integration. Since K~(1)\widetilde{\mathbf{K}}^{(1)} is diagonal, if its entries are positive, then the matrix is positive definite. This allows us to confirm that this property is indeed preserved;

(iii) The nonlinear stiffness tensors of the interpolated ROM, K~(1)\widetilde{\mathbf{K}}^{(1)} and K~(3)\widetilde{\mathbf{K}}^{(3)}, are symmetric by construction since, for each pair or triplet of symmetric entries, we interpolate their value and then store it in the tensors to respect symmetries.

(iv) The reduced damping matrix C~\widetilde{\mathbf{C}} is obtained using Eq. (34). Specifically, instead of interpolating it entry by entry, we interpolate the values of α(p^)\alpha(\hat{p}) and β(p^)\beta(\hat{p}). Eventually, we reconstruct C~\widetilde{\mathbf{C}} using those interpolated values and the interpolated K~(1)\widetilde{\mathbf{K}}^{(1)}. Again, an easy check on the positive definiteness of C~\widetilde{\mathbf{C}} can be made by looking at the signs of the interpolated α\alpha and β\beta damping parameters.

Once more, the positive definiteness of K~(1)\widetilde{\mathbf{K}}^{(1)} and C~\widetilde{\mathbf{C}} is not enforced as in [18, 17], but comes as a consequence of accurate interpolation. In the investigated examples, these properties have always been satisfied; thus, there was no need to use a different, more involved, interpolation scheme. However, we remind the reader that in critical cases where positive definiteness of K~(1)\widetilde{\mathbf{K}}^{(1)} and C~\widetilde{\mathbf{C}} is difficult to attain, one could, in principle, use the interpolation strategy presented in [17] for the linear part of the ROM, while keeping the ROM database construction procedure and the nonlinear tensor interpolation unmodified.

Applications

In this section we present the construction of the parametric interpolated ROM for two different structures. The first is a canonical curved panel, while the second is the box of a NACA wing. In both cases, parametric geometrical variations are considered. Accuracy and computational performance of the parametric ROMs will be benchmarked against HFM simulations. In this work, we used our Matlab based in-house FE software YetAnotherFEcode [42] as HFM and for constructing the parametric ROM. Although our FE code allows for an intrusive construction of the ROM, we used it in a non-intrusive fashion to demonstrate that the presented methodology can in principle be applied using commercial FE programs, as demonstrated in [32, 23].

Curved Panel

Model description

The rectangular curved panel under investigation is depicted in Figure 2a. It has dimensions Lx=40L_x = 40 mm, Ly=20L_y = 20 mm, thickness t=0.8t = 0.8 mm, and is clamped at all four edges. The material is linear elastic with Young’s modulus E=70E = 70 GPa, Poisson’s ratio ν=0.33\nu= 0.33, and density ρ=2700\rho= 2700 Kg/m3^3. Three geometrical parameters, acting on the curvature of the panel, were considered. Parametric variations can be used to model uncertainties introduced in the manufacturing process or, alternatively, to explore new designs.

Figure 2. In (a) geometry of the flat panel. In (b) the parameters-dependent shape functions used to define the height of the plate. In (c), realization of the mesh for p1=1.3tp_1 = 1.3t, p2=0.89tp_2 = 0.89t, and p3=1.6p_3 = 1.6 (In the illustration, the out of plane zz coordinates are scaled by a factor of 10).

In this test case, the final parameter-dependent geometry was obtained by superposing out-of-plane geometrical shapes to the flat rectangular panel, as shown in Fig.. Specifically, we constructed two out-of-plane geometrical shapes z1(x,y,p1)z_1(x,y,p_1) and z2(x,y,p2)z_2(x,y,p_2) by extruding two arch of circumferences passing through the edges of the plate along yy and xx axis, respectively. Here, p1p_1 and p2p_2 are defined as the height of the midspan for each of the geometrical shapes. The idea was then to add z1(x,y,p1)z_1(x,y,p_1) to z2(x,y,p2)z_2(x,y,p_2) to obtain the parametric out-of-plane shape. In order to make the problem more challenging, we decided to introduce an additional parameter p3p_3 that skews the geometry towards one side of the panel, breaking symmetry. For this purpose, we defined another shape function z3z_3 as z3(x,y,p3)=p3(1−2Lx)+2Lxxz_3(x,y,p_3)=p_3\left(1-\frac{2}{L_x}\right)+\frac{2}{L_x}x. Here, parameter p3p_3 dictates the inclination of this plane that is forced to be equal to unity for x=Lx/2x=L_x/2, for any value of p3p_3. In this way, we defined the final out-of-plane shape of the panel as

z(x,y,p1,p2,p3)=z1(x,y,p1)z3(x,y,p3)+z2(x,y,p2).(40)z(x,y,p_1,p_2,p_3)=z_1(x,y,p_1)z_3(x,y,p_3)+z_2(x,y,p_2). \tag*{(40)}

The parameter bounds considered here are shown in Table 1. As we show in the following, these parameter variations are enough to trigger significant variations in the panel response.

pminp_{min}pmaxp_{max}
p1p_10.250t0.250t1.500t1.500t
p2p_20.620t0.620t1.875t1.875t
p3p_31122

Table 1. Parametric bounds for the curved plate

Rayleigh damping model was assumed, where the mass and stiffness proportionality coefficients of the damping matrix, α\alpha and β\beta, were obtained by imposing a modal damping ratio of 0.01 to the first and second linear VMs.

Parametric ROM construction

The training set of parameters was constructed by drawing Nt=14N_t=14 quasi-random samples using the Latin Hypercube Sampling scheme. This sampling strategy maximizes the distance between points for given number of samples, ensuring a good coverage of the set of interest. The generated samples are shown in Figure 3. Then, the database of ROM was constructed following the procedure described in § 3.1. For each of the points in the training set, a FE model was constructed, using 600 elements with quadratic shape functions, based on the Von-Kármán strain-displacement relationship [6]. All assemblies had the same topological equivalent mesh (i.e. same element connectivity) with 8505 DOFs.

Figure 3. Training, validation and test samples of the curved panel.

Then, for each FE model, the first 12 VMs were computed and selected based on the load Modal Participation Factor. In this example, we considered the transient response of the panel to a uniform pressure load. Thus, we expected the response to be dominated at linear level by low frequency VMs. For each assembly instance, we retained all the linear VMs in the frequency bandwidth [0, 630 Hz] with non-zero MPF. This led to the inclusion of VMs 1, 2, 3, 8, 10, 11 etc (after reordering had been applied). The reader should note that the choice of the maximum frequency of interest was based on the analyst experience.

In this test case, we used the Static Modal Derivatives as complementary vectors used to capture nonlinearities. Their number grows almost quadratically with the number of retained VMs, hence it is essential to include only the most relevant one, to avoid undermining computational efficiency of the resulting ROM. The Static Modal Derivatives were selected using the criteria proposed in [32], which consists in ranking them based on the cross multiplication of the MPF of the different VMs. This is a variant of the more established selection criterion proposed in [24], and has the advantage that Modal Derivatives selection comes with no additional computational cost (linear runs are not needed). For each of the assemblies, the selected SMDs were those coupling VMs 1 – 1, 1 – 2, 1 – 3, 2 – 2, 2 – 3, 3 – 3. The SMDs were computed according to Eqs. (8) and (9), where the tangent stiffness derivative was obtained with finite differences, using a perturbation step h=10−8h=10^{-8}.

The global RB was then constructed by unit normalizing all the VMs and Static Modal Derivatives, and stacking them in two different snapshots matrices, ΦG\boldsymbol{\Phi}_{G} and ΘG\boldsymbol{\Theta}_{G}, respectively. Singular Value Decomposition was then applied separately to the two snapshot matrices, and 11 and 12 left singular vectors were selected from ΦG\boldsymbol{\Phi}_{G} and ΘG\boldsymbol{\Theta}_{G}. This corresponds to the relative truncation energy thresholds ϵΦ=0.997\epsilon_{\Phi}=0.997 and

eθ=0.980e_\theta= 0.980. The energy lost in compression 15 is plotted in Figure 4 for the two parts of the RB, ΦG\Phi_G and ΘG\Theta_G, as a function of the number of retained POD modes. As can be seen from Figure 4 the graph has a kink in both two different plots in correspondence of the sixth POD mode. This is because for each instance in the training set, six VMs and SMDs were chosen. The first six POD modes could be interpreted as an average of the first six modes throughout the training set. Instead, the higher order POD modes are needed to capture parametric variations of these mode shapes across the training set.

Figure 4. Energy lost in SVD data compression using different number of left singular vectors for ΦG\Phi_G (a) and ΘG\Theta_G (b).

After the global basis was constructed, Nt=14N_t = 14 different RB were obtained through mass normalization using the different mass matrices corresponding to the different FE instances in the training set. Then using the different RBs, the reduced order mass, stiffness and damping matrices were obtained by projection of their parameter-dependent HFM counterpart, while the nonlinear tensors are identified with EED. This required 299 evaluations of the nonlinear tangent stiffness matrix.

Interpolation of the ROM database was carried out in the normalized parameter space using RBF with inverse multiquadratic kernel function [43], which writes

γ(δ,ϵ)=11+(ϵδ)2.(41)\gamma(\delta,\epsilon) = \frac{1}{\sqrt{1+(\epsilon\delta)^2}}. \tag*{(41)}

Here, δ=∥p^−p^i∥2\delta= \lVert\hat{p} - \hat{p}_i \rVert_2 is the l2l_2-norm distance and ϵ\epsilon is a tuning parameter governing the peakedness of the distribution. The best ϵ\epsilon value was chosen with the aid of a validation set. Specifically, Nv=3N_v = 3 validation samples were used, and the validation ROM database was constructed following the same procedure for the construction of the training ROM database. Different weights were computed for 50 equally spaced ϵ\epsilon values in a logarithmic scale, ranging from 10−210^{-2} to 10. Approximation accuracy, in reconstructing the ROM operators in the validation set, was compared for the different interpolants. Based on this analysis, we selected the optimal ϵ\epsilon values that we reported in Table 2.

K~(1)\widetilde{\mathbf{K}}^{(1)}K~(2)\widetilde{\mathbf{K}}^{(2)}K~(3)\widetilde{\mathbf{K}}^{(3)}Vα\alphaβ\beta
ϵ\epsilon0.0120.5342.8480.2010.2010.231

Table 2. Optimal ϵ\epsilon values used in the interpolation of ROM operators of the curved panel.

Parametric ROM interpolation performance

The accuracy of the interpolated ROM was evaluated using a test set containing five parameter samples that had not been used for the construction of the ROM. These samples are shown in Figure 3 and listed in Table 3. The test load was a short pulse modeled as a uniform pressure in space, acting in the positive $z direction, varying in time as

pip^ip1p_1p2p_2p3p_3
p1p^11.09t0.79t1.04
p2p^20.47t1.61t1.32
p3p^30.82t0.97t1.41
p4p^41.33t1.84t1.77
p5p^50.54t1.19t1.81

Table 3. Values of the parameter samples in the test set.

p(t)=a⋅sin⁡(πTpulset)⋅H(Tpulse−t),(42)p(t) = a \cdot\sin\left(\frac{\pi}{T_{\mathrm{pulse}}}t\right) \cdot H(T_{\mathrm{pulse}} - t), \tag*{(42)}

where HH is the Heaviside, Tpulse=1/900T_{\mathrm{pulse}} = 1/900 s. and a=1.962a = 1.962 kPa. The time variation of the load is shown in Figure 5 (a).

Figure 5. In (a) uniform pressure load time history, in (b) out-of-plane displacement time history of the midspan (normalized to the panel thickness) for the individual parameter realizations in the test set.

For each of the five test points, we run time integration over the time span [0,0.06 s][0, 0.06\ \mathrm{s}], using five different models:

  • (i) the HFM, which is here the benchmark;

  • (ii) the interpolated parametric ROM evaluated at the test points,

  • (iii) the ROM in the training database corresponding to the training point that is the closest to the test point (we used as distance metric the l2l_2 norm in the normalized parameter space);

  • (iv) the ROM obtained by projecting the global RB on the full order model evaluated at the test point, here referred to as recomputed ROM;

  • (v) the linearized model corresponding to the test point.

In all cases, time integration was carried out using the Newmark-β\beta integration scheme [7], with a constant time step dt=10−4dt = 10^{-4} s. The panel responses for the out-of-plane degree of freedom (along zz axes) at the midspan obtained with the full model, for the different test parameter realizations, are shown in Figure 5.

A large variability of the transient response across the parameter set is observed, both in terms of frequency content and amplitudes of vibration. This is also confirmed by the variation of the system first natural frequency, which ranges from a minimum value of 147.9 Hz to a maximum value of 241.1 Hz.

A comparison of the results obtained from time integration of the different models is presented in Figure 6 and Figure 7. In Figure 6 we plot the time history of the out-of-plane displacement for the node in the middle of the panel (same as shown in Figure 5), while in Figure 7 the time history of the in-plane degree of freedom (along xx axes) of the node located at x=0.35Lxx = 0.35L_x, y=0.35Lyy = 0.35L_y. For all points in the test set, it is possible to observe a good overlap of the time histories obtained with the HFM and the interpolated ROMs, both for out-of-plane and in-plane degrees of freedom. The difference between the HFM with the linearized model confirms that the structure is vibrating in a nonlinear, as displacements in the order of the thickness also suggest. For all test cases (Figure 6 a - Figure 6e), the characteristic period of the linear unforced vibrations appears to be smaller than the one of the nonlinear models, thus indicating a softening behavior of the nonlinear structure. The comparison with the closest ROM in the training set is presented here to highlight the need for interpolation. Even if for some parameter realizations the interpolated ROM and the closest ROM show a similar behavior (Figure 6 a and Figure 6e), there are instances (Figure 6 b,Figure 6c, and Figure 6d) where the difference is more marked, and using the closest ROM instead of the interpolated ROM leads to erroneous results. Moreover, the reader should note that the recomputed ROM provides the best results, as the interpolated ROM is just its approximation. However, this comes at the price of reconstructing the ROM tensors, for each new parameter evaluation.

Figure 6. Out-of-plane displacement time history at panel midspan for different geometrical configurations in the test set.

Figure 7. In plane displacement time history at panel midspan for different geometrical configurations in the test set.

The average computational time to run time integration using the HFM and the interpolated ROM was, respectively, 848 s and 23 s, corresponding to a speed-up of 36.9. The reported computational times refer to a local machine equipped with 16 GB RAM and processor Intel Core i7-12700 H operating at 2.3 GHz. The modest computational gain is owing to the fact that the starting FE model is relatively small, especially compared to real-life industrial test cases, where the ROMs are clearly computationally advantageous.

On ROM construction time

Interpolating the database of ROMs to obtain a new adapted ROM is an extremely fast operation, compared to the computation of the tensors. In fact, a new ROM can be constructed in less than a second once the interpolation weights are known, while the non-intrusive identification of tensors from the RB using the EED method requires on average around 1830 s. This second cost dominates the overall cost for model construction, as the SMDs and the VMs were computed in less than 10 s, for each instance in the parameter set. Tensor identification can be in principle parallelized, allowing for a major reduction in the construction time of the ROM database. In this setting, the identification of tensors corresponding to different parameter instances can be dispatched to different workers. The reader should note that the interpolation approach pays off only if multiple runs are needed, such as, for example, in uncertainty quantification or structural optimization settings. In this second scenario, the interpolated ROM is also quite attractive because the gradients of the model with respect to the parameters, required for the search of the optimal parameter values, can be efficiently computed through analytical differentiation of the interpolation Equation (36). If standard ROM techniques are employed, such gradients might be computed with Finite Differences or automatic differentiation techniques. The first ones are computationally intensive, as they require multiple ROM constructions for values of the parameters perturbed about a reference point. The second ones are usually more computationally affordable and accurate, yet they require specific code and cannot be integrated in a non-intrusive ROM construction framework.

NACA Wing Box

Model description

As a second test case we investigate the response of an aircraft wing with parametric geometry. The set of different geometrical configurations was defined starting from a nominal rectangular wing box with NACA 0012 profile. The wing consists of a skin panel stiffened with ribs along the longitudinal and lateral directions. A model with the same geometry was considered in [44, 30]. Then the nominal wing geometry was morphed into different new geometric shapes by introducing three geometric parameters as shown in Figure 8 a.

Figure 8. In (a) parametrized wing geometry. In (b) finite element model for zero tapering, twist and sweep angles. In (c) load time variation used for dynamic analysis.

The first parameter, p1p_1, is the taper ratio and is defined as the ratio between the chord dimensions of the profiles at the tip and at the root of the wing. The profile sections across the span are rescaled by the taper ratio while preserving the proportions of the section.

The second parameter, p2p_2, is the angle of rotation of the tip wing cross section about the wing longitudinal axis passing from the origin of the centroid of the root section. The third parameter, p3p_3, is the sweep angle which results in a inclination of the wing profiles towards the trailing edge of the wing.

The validity ranges for the parameters in the model are shown in Table 4. The wing material is linear elastic with Young’s modulus E=70E = 70 GPa, Poisson’s ratio ν=0.3\nu= 0.3, and density ρ=2700\rho= 2700 Kg/m3^3. All sections in the model have a uniform thickness t=2t = 2 mm.

pmin⁡p_{\min}pmax⁡p_{\max}
p1p_10.751
p2p_2−3-3 deg+3+3 deg
p3p_30+14.4+14.4 deg

Table 4. Investigated parametric variations for the wing model

The FE model was constructed using 59040 triangular shell elements with linear shape functions, featuring 3 nodes per element and 6 dofs per node. The employed elements are based on the Von Karman strain-displacement kinematic relation. Clamped boundary conditions constraining all the displacements and rotations at the root of the wing were applied. This resulted in a model with 134195 variables. Meshes

with same topology (same number of nodes, elements and connectivity) were used to model the different geometrical instances considered herein.

Objective of the study is the prediction of the displacement level of structures with different geometries to an external pulse pressure acting on the nodes highlighted in red in Figure 8b, along the negative zz direction. The load varies in time as in Eq. (42) with Tpulse=0.05T_{\mathrm{pulse}} = 0.05 s and its time history is plotted in Figure 8c. We assume Rayleigh damping, with α\alpha and β\beta obtained by imposing a damping ratio of 1% to the first two structural VMs.

Parametric ROM construction

Two different sets of 12 and 4 parameter samples were generated using Latin Hypercube Sampling and used as training and validation sets, respectively. Then, different FE models, corresponding to the different sampled parametric variations, were obtained by moving the nodes of the FE mesh corresponding to a wing with zero sweep and twist angles and null taper ratio. For each FE model we computed the VMs and selected them based on the load Modal Participation Factor. Based on this analysis we decided to construct a single-mode ROM, employing only the first VM, as the other modes excited by the load are much stiffer.

To prove the concept that the presented methodology can in principle be used with a generic RB for the nonlinear part (provided that it is representative of the nonlinear solution), we decided to use Dual Modes instead of SMDs. Specifically, for each ROM we solved two nonlinear static problems for imposed loads in the direction of the first VMs, of the form ±K(1)ϕ1s1\pm K^{(1)}\phi_1s_1, where the s1s_1 is a scaling factor that is used to rescale the VM so that the maximum displacement of the mode at the tip of the wing is 3.5 cm. This value was found to be sufficient to trigger the non-linearity while remaining in the neighborhood of the equilibrium point.

Then, VMs and Dual Modes corresponding to different models are unit normalized and stacked respectively in two different snapshot matrices ΦG\Phi_{\mathrm{G}} and ΘG\Theta_{\mathrm{G}}, respectively. Singular Valued Decomposition was applied separately to ΦG\Phi_{\mathrm{G}} and ΘG\Theta_{\mathrm{G}}, and 5 left singular vectors were extracted from each of the two matrices to form the RB for the parametric ROM. This corresponds to a relative retained truncation energy threshold of 0.998 for the VMs, and 0.99 for the Dual Modes. Energy decay of VMs and DMs with POD mode number are shown in Figure 9. Then, this RB was mass-orthogonalized using the different mass matrices, corresponding to the different training samples, and the RB vectors reordering procedure was carried out. Eventually, mass and linear stiffness matrices were computed by projection of their HFM counterparts, whereas nonlinear stiffness tensors were identified non-intrusively using the EED method. This last step required 65 evaluations of the tangent stiffness matrix for each assembly in the training and validation sets.

Figure 9. Energy lost in SVD data compression using different number of left singular vectors for ΦG\boldsymbol{\Phi}_{G} (a) and ΘG\boldsymbol{\Theta}_{G} (b), for the wing test case.

Once the ROM database was constructed, RBF interpolation weights for the ROMs operators were computed for the inverse multiquadratic kernel in Eq. (41). The RBF shape parameter ϵ\epsilon was tuned using the validation set. The ϵ\epsilon values resulting from this procedure are reported in Table 5.

K~(1)\widetilde{\mathbf{K}}^{(1)}K~(2)\widetilde{\mathbf{K}}^{(2)}K~(3)\widetilde{\mathbf{K}}^{(3)}V\mathbf{V}α\alphaβ\beta
ϵ\epsilon0.00850.00770.00850.13850.00770.3854

Table 5. Optimal ϵ\epsilon values used in the interpolation of ROM operators of the NACA wing.

ROM interpolation performance

The parametric ROM performance is assessed by a direct comparison with HFM simulations for different random realizations of the parameter vector. Specifically, a test set with five samples was constructed by sampling the feasible parameter space with LHS. The values of the sampled parameters are reported in Table 6, while the corresponding geometrical configurations are shown in Figure 10. For these configurations,

pip^{i}p1p_{1} [deg]p2p_{2} [deg]p3p_{3} [−-]
p1\mathbf{p}^{1}0.92−2.22-2.220.61
p2\mathbf{p}^{2}0.801.734.68
p3\mathbf{p}^{3}0.86−1.33-1.336.04
p4\mathbf{p}^{4}0.972.8511.15
p5\mathbf{p}^{5}0.81−0.31-0.3111.80

Table 6. Values of the parameter samples in the test set for the wing.

Geometrical configurations in the test set

Figure 10. Geometrical configurations in the test set.

time integration for the applied impulse load was performed using the Newmark scheme for:

  1. the HFM, which is here the benchmark;

  2. the interpolated parametric ROM evaluated at the test points;

  3. the linearized model corresponding to the test point.

The displacements of a node on the left edge of the node tip section, obtained by time integrating the HFM for the parameter values in the test set are shown in Figure 11. The reader should note that the variability in the parameters affects the characteristic frequency of vibration as well as the amplitude. This variability is well captured by the interpolated ROM, as shown in Figure 12 and Figure 13. The difference between the linear response and the nonlinear prediction shows hardening behavior.

Figure 11. Wing out-of-plane (top) and in-plane (bottom) displacements of a node on the tip cross section for the different parameter realizations in the test set.

Figure 12. Out of plane response of the wing subjected to impulse loading at a node at the wing tip section. Comparison between HFM, interpolated ROM, and Linear model for the different parameter realizations in the test set, p1p^{1} (a), p2p^{2} (b), p3p^{3} (c), p4p^{4} (d), and p5p^{5} (e).

Figure 13. In plane response of the wing subjected to impulse loading at a node at the wing tip section. Comparison between HFM, interpolated ROM, and Linear model for the different parameter realizations in the test set, p1p^{1} (a), p2p^{2} (b), p3p^{3} (c), p4p^{4} (d), and p5p^{5} (e).

The HFMs were time integrated in parallel on the Euler cluster of ETH Zürich, using 5 cores (one for different test assembly) and 4 Gb RAM per core. No parallelization was used for the assembly of the internal forces and tangent stiffness matrices required within the time integration routine. A constant time step of 1.2⋅10−3 s1.2 \cdot10^{-3}\,\mathrm{s} was used. This resulted in an average total computation time of 4.789 h. In contrast, the ROMs

were time integrated using a larger constant time step h=5⋅10−3h = 5 \cdot10^{-3} s and the total average solution time was 4 s. This corresponds to a speed-up of the ROM with respect to the HFM of around 4310. The use of a larger time step to integrate the ROM compared to the HFM was possible as the high-frequency vibration effects, intrinsic to the HFMs, are not present in the ROMs.

On ROM construction time

The construction of the ROM was carried out in the Euler Cluster ETH Zürich, using 16 cores (one for different assembly in the training-validation sets) and 7 Gb RAM per core. The most computationally intensive procedures in the construction of the database of ROMs needed for PROM construction were the following:

(i) solving the nonlinear static problems needed in the Dual Modes computation, which required on average 62 s;

(ii) identifying the nonlinear tensors using EED, which was completed on average in 120 s;

For this specific test case, the non-intrusive construction of the ROM is relatively cheap. This owes to the fact that only few vectors are included in the RB.

Conclusion

In this work we presented a technique for constructing a PROM for structures with geometric nonlinearities, following an equation-based approach for the construction of the RB. The PROM can handle parametric variations in the geometry, and potentially in the material and boundary conditions. Thus, it can be used for model updating, uncertainty quantification, and structural optimization.

The PROM is constructed in two steps. A database of ROMs is constructed for random parameters samples drawn from the parameter set of interest. Then, a RBF interpolant is fitted to the entries of the ROM operators to obtain the PROM. he numerical examples demonstrate that the PROM reduces the computational effort for model adaptation to virtually zero, while still capturing substantial variations in the dynamic response across the parameter domain. Furthermore, analytical parameter sensitivities follow directly from differentiation of the PROM, a capability of particular relevance for gradient-based optimization.

Two limitations should be acknowledged. The first concerns the selection of an adequate number of parameter samples: too few lead to poor accuracy, while too many substantially increase the offline construction cost. Greedy sampling strategies [45] may provide an efficient remedy. The second limitation is inherent to the current geometric parametrization, which is restricted to topology-preserving mesh morphing. For example, in the context of stiffened panels, the method can account for variations in skin or stiffener thickness, but not for changes in the stiffener layout.

References

  1. [1]R. W. Gordon, J. J. Hollkamp, Reduced-order models for acoustic response prediction of a curved panel, 2011. doi:10.2514/6.2011-2081.
  2. [2]S. Spottswood, J. Hollkamp, T. Eason, On the use of reduced-order models for a shallow curved beam under combined loading, American Institute of Aeronautics and Astronautics (AIAA), 2008. doi:10.2514/6.2008-2235.DOI
  3. [3]A. Przekop, S. A. Rizzi, Dynamic snap-through of thin-walled structures by a reduced-order method, AIAA Journal 45 (2007) 2510–2519. doi:10.2514/1.26351.DOI
  4. [4]A. Przekop, S. A. Rizzi, D. S. Groen, Nonlinear acoustic response of an aircraft fuselage sidewall structure by a reduced-order analysis, in: Ninth International Conference on Recent Advances in Structural Dynamics, no. Paper-135, 2006.
  5. [5]M. P. Mignolet, A. Przekop, S. A. Rizzi, S. M. Spottswood, A review of indirect/non-intrusive reduced order modeling of nonlinear geometric structures, Journal of Sound and Vibration 332 (10) (2013) 2437–2460.DOI
  6. [6]M. Crisfield, Nonlinear Finite Element Analysis of Solids and Structures, Wiley, 1991.
  7. [7]M. Géradin, D. J. Rixen, Mechanical vibrations: theory and application to structural dynamics, John Wiley & Sons, 2015.
  8. [8]S. R. Idelsohn, A. Cardona, A load-dependent basis for reduced nonlinear structural dynamics, Computers & Structures 20 (1-3) (1985) 203–210.
  9. [9]S. R. Idelsohn, A. Cardona, A reduction method for nonlinear structural dynamic analysis, Computer Methods in Applied Mechanics and Engineering 49 (1985) 253–279. doi:10.1016/0045-7825(85)90125-2.DOI
  10. [10]K. Kim, A. G. Radu, X. Q. Wang, M. P. Mignolet, Nonlinear reduced order modeling of isotropic and functionally graded plates, International Journal of Non-Linear Mechanics 49 (2013) 100–110. doi:10.1016/j.ijnonlinmec.2012.07.008.DOI
  11. [11]A. A. Muravyov, S. A. Rizzi, Determination of nonlinear stiffness with application to random vibration of geometrically nonlinear structures, Computers and Structures 81 (2003) 1513–1523. doi:10.1016/S0045-7949(03)00145-7.DOI
  12. [12]R. Perez, X. Q. Wang, M. P. Mignolet, Nonintrusive structural dynamic reduced order modeling for large deformations: Enhancements for complex structures, Journal of Computational and Nonlinear Dynamics 9 (7 2014). doi:10.1115/1.4026155.DOI
  13. [13]M. Song, L. Renson, J. P. Noël, B. Moaveni, G. Kerschen, Bayesian model updating of nonlinear systems using nonlinear normal modes, Structural Control and Health Monitoring 25 (12 2018). doi:10.1002/stc.2258.DOI
  14. [14]T. L. Hill, P. L. Green, A. Cammarano, S. A. Neild, Fast bayesian identification of a class of elastic weakly nonlinear systems using backbone curves, Journal of Sound and Vibration 360 (2016) 156–170. doi:10.1016/j.jsv.2015.09.007.DOI
  15. [15]K. Park, M. S. Allen, A gaussian process regression reduced order model for geometrically nonlinear structures, Mechanical Systems and Signal Processing 184 (2 2023). doi:10.1016/j.ymssp.2022.109720.DOI
  16. [16]J. Marconi, P. Tiso, D. E. Quadrelli, F. Braghin, A higher-order parametric nonlinear reduced-order model for imperfect structures using neuman expansion, Nonlinear Dynamics 104 (2021) 3039–3063. doi:10.1007/s11071-021-06496-y.DOI
  17. [17]D. Amsallem, J. Cortial, K. Carlberg, C. Farhat, A method for interpolating on manifolds structural dynamics reduced-order models (2009).DOI
  18. [18]D. Amsallem, C. Farhat, An online method for interpolating linear parametric reduced-order models, Vol. 33, Society for Industrial and Applied Mathematics Publications, 2011, pp. 2169–2198. doi:10.1137/100813051.DOI
  19. [19]M. I. McEwan, J. R. Wright, J. E. Cooper, A. Y. Leung, A finite element/modal technique for nonlinear plate and stiffened panel response prediction, American Institute of Aeronautics and Astronautics Inc., 2001. doi:10.2514/6.2001-1595.DOI
  20. [20]J. Marconi, P. Tiso, F. Braghin, A nonlinear reduced order model with parametrized shape defects, Computer Methods in Applied Mechanics and Engineering 360 (3 2020). doi:10.1016/j.cma.2019.112785.DOI
  21. [21]J. Rutzmoser, Model order reduction for nonlinear structural dynamics, Ph.D. thesis, Technische Universität München (2018).
  22. [22]C. Farhat, T. Chapman, P. 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 (2015) 1077–1110. doi:10.1002/nme.4820.DOI
  23. [23]M. K. Mahdiabadi, P. Tiso, A. Brandt, D. J. Rixen, A non-intrusive model-order reduction of geometrically nonlinear structural dynamics using modal derivatives, Mechanical Systems and Signal Processing 147 (1 2021). doi:10.1016/j.ymssp.2020.107126.DOI
  24. [24]P. Tiso, Optimal second order reduction basis selection for nonlinear transient analysis, in: Modal Analysis Topics, Volume 3: Proceedings of the 29th IMAC, A Conference on Structural Dynamics, 2011, Springer, 2011, pp. 27–39.DOI
  25. [25]J. J. Hollkamp, R. W. Gordon, S. M. Spottswood, Nonlinear modal models for sonic fatigue response prediction: A comparison of methods, Journal of Sound and Vibration 284 (2005) 1145–1163. doi:10.1016/j.jsv.2004.08.036.DOI
  26. [26]A. G. Radu, M. P. Mignolet, A. G. Radu, X. Gao, Validation of reduced order modeling for the prediction of the response and fatigue life of panels subjected to thermo-acoustic effects (2003). URL https://www.researchgate.net/publication/262933555
  27. [27]C. S. Sombroek, P. Tiso, L. Renson, G. Kerschen, Numerical computation of nonlinear normal modes in a modal derivative subspace, Computers and Structures 195 (2018) 34–46. doi:10.1016/j.compstruc.2017.08.016.
  28. [28]L. Wu, P. Tiso, Nonlinear model order reduction for flexible multibody dynamics: a modal derivatives approach, Multibody System Dynamics 36 (2016) 405–425. doi:10.1007/s11044-015-9476-5.DOI
  29. [29]L. Andersson, P. Persson, K. Persson, Efficient nonlinear reduced order modeling for dynamic analysis of flat structures, Mechanical Systems and Signal Processing 191 (5 2023). doi:10.1016/j.ymssp.2023.110143.
  30. [30]S. Jain, P. Tiso, Simulation-free hyper-reduction for geometrically nonlinear structural dynamics: A quadratic manifold lifting approach, Journal of Computational and Nonlinear Dynamics 13 (7 2018). doi:10.1115/1.4040021.DOI
  31. [31]J. B. Rutzmoser, D. J. Rixen, A lean and efficient snapshot generation technique for the hyper-reduction of nonlinear structural dynamics, Computer Methods in Applied Mechanics and Engineering 325 (2017) 330–349. doi:10.1016/j.cma.2017.06.009.DOI
  32. [32]A. Saccani, P. Tiso, Accelerating construction of non-intrusive nonlinear structural dynamics reduced order models through hyperreduction (11 2024). URL http://arxiv.org/abs/2411.14262
  33. [33]J. C. Helton, F. J. Davis, Latin hypercube sampling and the propagation of uncertainty in analyses of complex systems, Reliability Engineering & System Safety 81 (1) (2003) 23–69.
  34. [34]P. Tiso, M. Karamooz Mahdiabadi, J. Marconi, Modal methods for reduced order modeling, in: Model Order Reduction: Volume 1: System-and Data-Driven Methods and Algorithms, De Gruyter, 2021, pp. 97–138.
  35. [35]O. Brüls, P. Duysinx, J. C. Golinvál, The global modal parameterization for non-linear model-order reduction in flexible multibody dynamics, International Journal for Numerical Methods in Engineering 69 (2007) 948–977. doi:10.1002/nme.1795.DOI
  36. [36]N. G. Stephen, On veering of eigenvalue loci, Journal of Vibration and Acoustics 131 (2009) 0545011–0545015. doi:10.1115/1.3147130.DOI
  37. [37]G. H. Heirman, F. Naets, W. Desmet, A system-level model reduction technique for the efficient simulation of flexible multibody systems, International Journal for Numerical Methods in Engineering 85 (2011) 330–354. doi:10.1002/nme.2971.DOI
  38. [38]O. Giannini, A. Sestieri, Experimental characterization of veering crossing and lock-in in simple mechanical systems, Mechanical Systems and Signal Processing 72-73 (2016) 846–864. doi:10.1016/j.ymssp.2015.11.012.DOI
  39. [39]K. Y. Kim TS, Mac-based mode-tracking in structural topology optimization, Computers and Structures (2000).DOI
  40. [40]E. Bonisoli, G. Marcuccio, C. Rosso, Crossing and veering phenomena in crank mechanism dynamics, in: Topics in Model Validation and Uncertainty Quantification, Volume 5: Proceedings of the 31st IMAC, A Conference on Structural Dynamics, 2013, Springer, 2013, pp. 175–187.DOI
  41. [41]R. J. Allemand, The modal assurance criterion–twenty years of use and abuse, Sound and vibration 37 (8) (2003) 14–23.
  42. [42]S. Jain, J. Marconi, P. Tiso, Yetanotherfecode (2021).DOI
  43. [43]G. B. Wright, Radial basis function interpolation: numerical and analytical developments, University of Colorado at Boulder, 2003.
  44. [44]S. Jain, P. Tiso, J. B. Rutzmoser, D. J. Rixen, A quadratic manifold for model order reduction of nonlinear structural dynamics, Computers and Structures 188 (2017) 80–94. doi:10.1016/j.compstruc.2017.04.005.DOI
  45. [45]A. Paul-Dubois-Taine, D. Amsallem, An adaptive and efficient greedy procedure for the optimal training of parametric reduced-order models, International Journal for Numerical Methods in Engineering 102 (2015) 1262–1292. doi:10.1002/nme.4759.

Paper details

vlnc:W_ag7k7scf
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.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents