State-space based discretize-then-differentiate adjoint sensitivity method for transient responses of non-viscously damped systems
Introduction
Design sensitivity analysis (DSA) is used to quantify the influence of measurable outputs with respect to a set of design variables [1]. It plays a vital role in a wide variety of fields, such as model updating [2], gradient-based optimization [3], damage detection [4] and structural reliability [5]. In practice, engineering structures are often subjected to dynamic loadings, such as impact, shock and seismic excitations, which may cause severe undesired noises and vibrations. Therefore, efficient and accurate DSA methods must be developed for transient response problems. In principle, methods for the DSA can be divided into three broad categories [6]: finite difference method (FDM), direct differentiate method (DDM) and adjoint variable method (AVM). The FDM is easy to implement, but it suffers from severe computational burden and numerical error associated with the step size. Therefore, it is typically used for comparison and verification of the sensitivity results obtained from new methodologies. The AVM is more efficient than the DDM when the number of design variables is large and the number of active constraints is small [7], e.g., in topology optimization that typically involves thousands or millions of design variables, and vice versa.
In the AVM, the sensitivity is computed by first adding some residual equations to the response function to built an augmented function and subsequently by differentiating the resulting augmented function. In general, there are two ways to perform the AVM according to the order of discretization and the differentiation, namely differentiate-then-discretize method and discretize-then-differentiate method, respectively. The differentiate-then-discretize method first differentiates the augmented response function and then transforms the adjoint variable problem into a terminal value problem which is similar to the prime problem. And finally the prime and terminal problems are calculated at each discretized time point. On the contrary, the discretize-then-differentiate method consists in differentiating directly the discretized version of the augmented response functions for each time point [8]. The literatures of the gradient-based optimization methods that rely on the differentiate-then-discretize method are well-developed. Some dealt with the design of frequency response problems [9], [10], [11] which aim to improve the performances of the structural system under harmonic excitations. Others are related to the transient response problems [12], [13], [14], [15], which are used to minimizing the dynamic responses, such as displacement, velocity, acceleration and stress, generated by the structure under external loads.
While the focus on differentiate-then-discretize AVM continues to dominate the literature [16], there have been some notable exceptions, particularly in the areas of sensitivity analyses of transient responses. One of the earliest papers to explore the use of discretize-then-differentiate AVM in combination with topology optimization was a 2012 paper by Le et al. [17]. Here the authors used this algorithm to avoid numerical inconsistencies and optimized material microstructural to achieve desired energy propagation in a two-phase composite plate. Since then, a number of authors have also investigated the performances of the discretize-then-differentiate approach on sensitivity analysis of transient responses. Jensen et al. [18] pointed out that, for viscously damped systems, the traditional differentiate-then-discretize AVM may lead to inconsistent sensitivities. It was found that, besides the appropriate choice of time step and summation rule for the discrete approximation, the more cumbersome discretize-then-differentiate AVM is also an effective way to resolve the inconsistent sensitivities. Later on, Nakshatrala and Tortorelli [19] presented a transient dynamic topology optimization framework for rate independent small deformation elastoplastic materials systems and derived the adjoint sensitivity expression based on the discretize-then-differentiate approach. A year later, they further extended the discretize-then-differentiate method to iteratively updated the micro-structural design parameters and developed a multiscale topology optimization framework for nonlinear structural systems. Other DSA methods for the nonlinear transient problems based on the discretize-then-differentiate AVM were also studied [8], [20].
Mathematically, the damping force can not be ignored in the DSA of transient responses [21]. However, most of the above mentioned studies are developed in assumption of undamped or viscously damped models. In practice, viscoelastic materials have been extensively applied in controlling vibration and noise of engineering structures. The viscoelastic effects are of great importance to get a feasible design in structural optimization algorithms [22]. The non-viscous damping model, which assumes that the dissipative forces depend on the past history of motion via convolution integrals over suitable kernel functions [23], provides an effective way to accurately describe the dissipative forces of viscoelastic materials. It has been considered to be the most general damping model within the scope of linear systems [24]. Currently, most of the studies on the DSA for non-viscously damped systems are focused on the eigensolution problems [25], [26], [27] and frequency response problems [28], [29], [30]. As to the problems of transient responses, Li et al. [31] extended the discrete Fourier transform and inverse discrete Fourier transform algorithms to non-viscously damped systems and presented the AVM and DDM to calculate the sensitivity of transient response. James and Waisman [32] derived a novel discretize-then-differentiate time-dependent AVM and incorporated into a topology optimization framework accounting for viscoelastic creep responses. Yun and Youn [33] developed a Newmark- based discretize-then-differentiate AVM on calculating the transient response sensitivity for non-viscously damped systems and then utilized it to optimize the microstructural of viscoelastically damped structures subjected to dynamic loads. Recently, Ding et al. [34] proposed a DSA method of transient responses for non-viscously damped systems based on the DDM to obtain both first- and second-order derivatives.
For viscously damped systems, the differentiate-then-discretize AVM may lead to inconsistency errors [35] and the alternative discretize-then-differentiate AVM can resolve this issue [18]. The effect of the order of the discretization and the differentiation on the accuracy of DSA is still uncertain for non-viscously damped systems. Since it is difficult to directly differentiate the equations of motion of non-viscously damped systems with some convolution integral terms, most of the published studies are based on the discretize-then-differentiate method, which firstly discretize the convolution integral terms and then approximate them using certain summation scheme. However, this approximation may introduce large numerical error. Ding et al. [36] developed a state-space based differentiate-then-discretize AVM on DSA of transient responses for non-viscously damped systems. The equations of motion of non-viscously damped systems are equivalently transformed into a state-space form without approximation and a modified precise integration method (MPIM) [37] is adopted to derive the adjoint variables. It is proved that, compared with the Newmark- based discretize-then-differentiate AVM [33], the state-space based differentiate-then-discretize AVM shows obvious advantages on the computational accuracy, efficiency and the implementation effort [36]. But the comparison is somewhat unfair, because the time integration methods are different and the MPIM is much more accurate and efficient than the Newmark- method on calculating the transient responses of non-viscously damped systems.
The aim of this work is to develop a novel state-space based discretize-then-differentiate AVM on DSA of transient responses for non-viscously damped systems. Firstly, the equations of motion of non-viscously damped systems are transformed into a state-space formulation to eliminate the convolution integral terms. Secondly, the transient responses are obtained by the MPIM and the residual equations are discretized at each time point. Then, the augmented response functions are differentiated and the adjoint variables are calculated to deduce the transient response sensitivities for the non-viscously damped systems. Finally, theoretical and numerical comparisons are made to investigate the performances of the proposed state-space based discretize-then-differentiate AVM with the existing methods. It is hoped that this study will clarify the effects of the order of the differentiation and the discretization on the computational considerations of the state-space based AVMs for transient response sensitivities of non-viscously damped systems. The paper is organized as follows: in Section 2, the problem of DSA is formulated by introducing some theoretical backgrounds on the non-viscously damped systems and inconsistent sensitivities. In Section 3, the state-space based differentiate-then-discretize AVM is reviewed covering the introduction of the MPIM. A novel state-space based discretize-then-differentiate AVM for transient responses of non-viscously damped systems is presented in Section 4, which is also the main contribution of this paper. In Section 5, the flowchart and some computational considerations including the numerical error, implementation effort and computational complexity are studied and compared. The proposed approach together with two other methods are applied on two numerical examples to compare the performances in Section 6. Finally, important concluding remarks are drawn in Section 7.
Section snippets
Theoretical background and problem description
In this section, theoretical background on non-viscously damped systems is firstly introduced. Then, the problem of design sensitivity analysis (DSA) is formulated by using the adjoint variable method (AVM). Specifically, the methodology of two AVM approaches, namely differentiate-then-discretize method and discretize-then-differentiate method, are presented. Finally, two types of errors occur in the DSA are defined.
Review of state-space based differentiate-then-discretize AVM for transient responses of non-viscously damped systems
In order for better comparison among the proposed method and the existing methods, the state-space based differentiate-then-discretize method will be reviewed in detail. In this section, the Newmark- based discretize-then-differentiate method is not shown due to its cumbersome numerical implementation (see Ref. [33] for details).
State-space based discretize-then-differentiate AVM for transient responses of non-viscously damped systems
In this section, a new state-space based DSA method for transient responses of non-viscously damped systems is proposed by using the discretize-then-differentiate AVM. The calculations on the transient responses of the proposed method are the same as the differentiate-then-discretize AVM reviewed in the previous section. Therefore, this section will directly start with the DSA by first discretizing and then differentiating the transient response functions.
Computational considerations of the proposed and the comparative methods
The computational considerations of the proposed state-space based discretize-then-differentiate AVM are studied and compared with the Newmark- based discretize-then-differentiate AVM [33] and the sate-space based differentiate-then-discretized AVM [36]. van Keulen et al. [39] provided several criteria for choosing an appropriate DSA method: accuracy, consistency, computational cost and implementation effort. The above mentioned approaches will be discussed according to these criteria.
Numerical examples and discussions
In this section, the performances of the proposed state-space based discretize-then-differentiate AVM will be studied and compared with the other two methods via two numerical examples on various conditions. All computations are performed on a laptop computer with a Windows 10, 64 bit operating system and a Intel Core i7-8565U CPU (2.00 GHz) with 8.00 GB random-access memory.
Conclusions
This paper develops a new design sensitivity analysis (DSA) method for transient responses of non-viscously damped systems. The energy dissipation of the non-viscous damping systems is modeled by the past history of motion via convolution integrals over kernel functions. The adjoint variable method (AVM) is adopted to derive the DSA. It is verified by previous studies that, for viscously damped systems, the discretize-then-differentiate AVM can improve the consistency of the sensitivity
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by the National Natural Science Foundation of China (Grant No. 51805383), the Hong Kong Scholars Program (Grant No. XJ2019019) and the Hubei Provincial Natural Science Foundation of China (Grant No. 2018CFB345).
References (43)
- et al.
Review and unification of methods for computing derivatives of multidisciplinary computational models
AIAA J
(2013) - et al.
Estimation of beam material random field properties via sensitivity-based model updating using experimental frequency response functions
Mech Syst Signal Pr
(2018) - et al.
A new data-driven topology optimization framework for structural optimization
Comput Struct
(2020) - et al.
Nonlinear breathing crack identification from time-domain sensitivity analysis
Appl Math Model
(2020) - et al.
Sampling-based system reliability-based design optimization using composite active learning kriging
Comput Struct
(2020) - et al.
A review of optimization of structures subjected to transient loads
Struct Multidiscip Optim
(2006) - et al.
A unified framework for nonlinear path-dependent sensitivity analysis in topology optimization
Int J Numer Meth Eng
(2018) - et al.
Adjoint sensitivity analysis and optimization of hysteretic dynamic systems with nonlinear viscous dampers
Struct Multidiscip Optim
(2018) - et al.
A method for topology optimization of structures under harmonic excitations
Struct Multidiscip Optim
(2018) - et al.
An improved parametric level set method for structural frequency response optimization problems
Adv Eng Soft
(2018)