Professional Documents
Culture Documents
AnexplicitintegrationmethodR2 Cleanversion
AnexplicitintegrationmethodR2 Cleanversion
net/publication/364818678
CITATIONS READS
7 206
5 authors, including:
Peng Yuan
The Hong Kong Polytechnic University
37 PUBLICATIONS 285 CITATIONS
SEE PROFILE
All content following this page was uploaded by Peng Yuan on 28 October 2022.
7 Hannover, Germany.
c
8 Institute for Risk and Uncertainty and School of Engineering, University of Liverpool,
13 PR China.
19
20
1
21 Abstract: Computational efficiency and accuracy are two of the most important properties
22 of integration methods. In this study, an explicit single-step integration method with third-order
23 accuracy is proposed to improve computational efficiency and accuracy. The proposed method
25 responses without physical damping. The algorithmic dissipation property is also improved to
26 filter out high-frequency spurious responses on the premise of sufficient accuracy in the low-
27 frequency response domain. Moreover, since the proposed method can advance the calculation
28 step-by-step via the known displacement and velocity items, the time-consuming factorization
29 of an effective stiffness matrix is not required in the calculation for the structure with a lumped
30 mass matrix. In other words, high computational efficiency is ensured in the proposed method.
31 The properties of the proposed method (i.e., the accuracy, convergence, dissipation, efficiency,
32 and effectiveness of the nonlinearity) are demonstrated by using four representative examples.
34 is adopted to comparatively evaluate the accuracy and convergence of the proposed method, a
35 typical nonlinear dynamic system is used to investigate the effectiveness of the nonlinear
36 calculation; a linear Howe truss model is employed to explore the effectiveness and efficiency
38 demonstrate the potential to solve the large-scale nonlinear system. Results show that the
39 proposed method can obtain more accurate results in the nonlinear dynamic calculations; and
40 its time consumption is around 65% of that of the Yuan method, 38% of that of the Kim method,
2
42 Keywords: Explicit integration method; Prediction-correction; High efficiency and accuracy;
3
44 1. Introduction
46 Computational efficiency and accuracy are two crucial properties of integration methods. An
47 admirable integration method should be able to achieve high computational efficiency and
48 accuracy simultaneously. Therefore, many research efforts have been made in the step-by-step
49 integration methods to achieve high efficiency and accuracy in dynamic calculations [3][4].
50 The step-by-step integration method consists of two categories, i.e., the implicit
51 integration method and the explicit integration method [5]. The implicit integration method is
52 unconditionally stable, but its computational efficiency is relatively low due to the time-
54 are required in the solution of complicated nonlinear dynamic problems, which makes the
55 implicit method less efficient in comparison with the explicit integration method [7]. The
56 explicit integration method is conditionally stable, and its computational time step is subject to
57 the critical time step size. The critical time step size is determined by the highest frequency of
59 is generally not required in the explicit calculation for dynamic problems with a diagonal mass
60 matrix, which widely exist in practical engineering [6]. Therefore, the high-efficiency
61 calculation can be realized in the analyses by using explicit integration methods. It is worth
62 noting that for transient problems, e.g., crash simulations and seismic analyses, a small-time
64 Therefore, the efficiency and accuracy of the explicit method have been the main goals pursued
4
65 by researchers in transient dynamic calculations rather than the limitation of the critical time
67 The central difference (CD) method with second-order accuracy is a well-known single-
68 step explicit integration method, and it is usually treated as a reference method because the
69 maximum critical time step is attained in explicit methods [9][10]. However, the CD method
70 cannot provide the algorithmic damping/dissipation that can be used to filter out high-
71 frequency spurious vibrations [11][12]. It is worth noting that spurious vibrations are a kind of
72 response that only exists in mathematical results, not in practical engineering. Moreover, even
73 if the diagonal mass matrix is used in the calculation, the factorization of the equivalent
74 stiffness matrix is required for nonlinear problems when the CD method is used in the
75 calculation [13]. The Verlet method, a single-step explicit integration method, was proposed to
76 evade the factorization of the CD method under nonlinear scenarios [14], but only first-order
77 accuracy was achieved in results when physical damping is taken into consideration in the
78 calculation. The Chung and Lee (CL) method improved the Verlet method [15], and second-
79 order accuracy can be ensured in results for the dynamic system with the physical damping,
80 but the computational accuracy can be enhanced further. Kim developed an explicit integration
81 method with second-order accuracy [16], and the algorithmic damping in the high-frequency
82 domain was optimized. The computational accuracy can be further improved, and Kim’s
83 method will be compared in this study. Therefore, it is essential to simultaneously improve the
85 Except for two-time-step explicit methods, continuous three-time steps are also employed
5
86 to develop explicit methods to improve computational efficiency and accuracy. For instance,
87 Zhai developed a single-step explicit integration method with first-order accuracy to decouple
88 a large dynamic system, and the decoupling property had been proved in the example of the
89 train-bridge coupling system [17]. However, Zhai’s method failed to acquire high accuracy in
90 the low-frequency response domain, and a small-time step must be adopted for satisfying
91 results. Yuan et al. (simplified as the Yuan method) proposed an explicit integration method
92 with second-order accuracy by introducing the displacement and velocity items at the previous
93 time [18]. Furthermore, Yuan et al. developed a new explicit integration method with third-
94 order accuracy by selecting appropriately known quantities and algorithmic parameters [19].
95 Although the accuracy has improved, its computational efficiency is relatively low, and an
96 additional starting procedure is required at the first calculational step. Therefore, it is important
98 In addition, explicit integration methods with multiple sub-steps were proposed to obtain
99 higher accuracy in the numerical analysis [20], such as the fourth-order Runge-Kutta method
100 [21], the Noh and Bathe method [22], the Kim and Lee method [23], the Kim method [24], the
101 Rezaiee-Pajand and Karimi-Rad method [25], and the Liu and Guo method [26]. Although high
102 accuracy was realized in these methods, a complex calculation procedure would be performed.
103 These methods need to solve the equations of motion at least twice during a time step to obtain
104 response results. For instance, four dynamic calculations are conducted in a time step for the
106 compared with the single-step method, which has been proven in [27]. Thus, these methods
6
107 with multiple sub-steps would not be discussed in this study.
108 Based on the discussion of the accuracy, dissipation, and efficiency above, this study
109 proposes a single-step explicit integration method with third-order accuracy, and the fourth-
110 order accuracy can be achieved in the absence of physical damping. An algorithmic dissipation
111 property is improved to filter out spurious response contents in the high-frequency response
112 domain on the premise of enough accuracy in the low-frequency response domain. It is worth
113 noting that it is very hard to balance the accuracy and dissipation properties of integration
114 methods. Moreover, the proposed method can be advanced step-by-step by only using previous
115 displacement and velocity items. To detail the derivative and demonstration process of the
118 demonstrated to obtain improved accuracy and dissipation properties. In Section 3, two
119 algorithmic indexes (i.e., stability and accuracy) are studied to demonstrate the algorithmic
120 properties of the proposed method. In Section 4, four representative examples are analyzed by
121 comparing state-of-the-art methods (e.g., the CD method, the Zhai method [17], the Kim
122 method [16], and the Yuan method [19]) to explore the algorithmic properties of the proposed
123 method (e.g., accuracy and efficiency). Some conclusions are summarized in Section 5.
125 A new prediction-correction integration method is proposed to improve the accuracy and
127 schemes for the displacement and velocity items are proposed, and their parameters are
7
128 identified based on the polynomial accuracy of the prediction scheme. Subsequently, the local
129 truncation error of the prediction-correction scheme is investigated to ensure the accuracy of
130 the overall scheme. Finally, the standard formulas with two parameters are formulated.
132 An SDOF system is often used to construct a new numerical integration method to obtain
133 specified algorithmic properties (e.g., high efficiency or accuracy) [28][29]. The governing
134 equation of motion and its initial conditions for the nonlinear SDOF system are, respectively,
138 where x(t), ẋ(t), and ẍ(t) are the displacement, velocity, and acceleration of the SDOF system,
139 respectively, the superposed dots represent the derivative with respect to time, f(t) is the
140 external force applied to the system, d0 and v0 are the initial displacement and velocity of the
141 system, respectively, and N(x(t), ẋ(t)) is the vector of nonlinear internal forces. For linear
144 where ω is the system angular frequency, and ξ is the corresponding damping ratio.
145 To solve the dynamical system by using the step-by-step integration method, the whole
146 computational time T = [t0, tm] is discretized into an n-step time series with a time interval Δt.
147 Three continuously computational time steps are presented in Fig. 1 to illustrate the
8
149 computations step by step (e.g., from the time step t to the time step t+Δt, as shown in Fig. 1),
150 known quantities (i.e., displacements, velocities, and accelerations) at the time steps t and t-Δt
151 are usually treated as initial conditions to calculate the responses of the unknown quantities at
152 the time step t+Δt [4][15]. Then, this procedure is repeated until the whole computation is
153 completed, and system responses within T can be obtained. In other words, the computational
154 procedure can be advanced directly by the explicit method. It is worth mentioning that both
155 matrix factorizations and nonlinear iterations are necessary for the implicit integration method.
156
158 To make full use of the known quantities at the time steps t-Δt and t, e.g., the displacement
159 and velocity items, the following schemes are proposed to predict the displacement and
162 b5 b
xt + t = x t − t + 6 x t +b7 xt − t + b8 xt (4)
t t
163 where the superposed wavy lines indicate prediction values of the displacement and velocity
164 at the time step t+Δt, and b1~8 are eight free parameters. The general integration expression is
165 used to determine free parameters in the displacement prediction schemes, and more
166 information on the general integration expression can be found in [32]. The following three
167 relationships involving parameters b1, b2, b3, and b4 are derived to ensure second-order
168 accuracy of the displacement prediction scheme (i.e., Eq. (3)), and the detailed derivation
9
170 b1 = 2b4 − 3 , b2 = 4 − 2b4 , and b3 = b4 − 2 (5 a-c)
171 The parameter b4 will be further determined according to the discussion of algorithmic
172 properties later. Similarly, to guarantee the second-order accuracy of the velocity prediction
173 scheme (i.e., Eq. (4)), the three relationships, involving parameters b5, b6, b7, and b8, can be
176 Only two parameters (i.e., b4 and b8) are retained in Eqs. (3) and (4). For the nonlinear
177 and linear scenarios, the following acceleration items at the time step t+Δt can be derived by
181 Based on the prediction acceleration at the time step t+Δt (i.e., xt +t ), the correction
182 schemes of the displacement and velocity items at the time step t+Δt are proposed as follows:
184 b14 b b
xt + t = x t − t + 15 x t + 16 x t + t +b17 xt − t + b18 xt (10)
t t t
185 where b9~18 are the ten parameters to be determined. It is worth mentioning that Eqs. (3), (4),
186 (9), and (10) are constructed following the Taylor series to make full use of known quantities.
187 Similarly, the following eight relationships can be identified, and two independent parameters
188 (b13 and b18) are retained in Eqs. (9) and (10), respectively.
189 b9 = 5 − 18b13 , b10 = 18b13 − 4 , b11 = 2 − 8b13 , and b12 = 4 − 10b13 (11 a-d)
10
8 + 5b18 8 − b18 2 + b18
190 b14 = , b15 = −4 − b18 , b16 = , and b17 = (12 a-d)
4 4 2
192 and it includes the prediction stage (i.e., Eqs. (3), (4), (7), and (8)) and correction stage (i.e.,
193 Eqs. (9) and (10)). Although the correction stage is added in the solution procedures in
194 comparison to the Zhai method and Yuan method, this stage is a highly efficient vector
195 operation with only a slight influence on computational efficiency, which will be demonstrated
196 later. Moreover, for the purpose of simplification, the four unknown parameters (i.e., b4, b8, b13,
197 and b18) are designated as α, β, γ, and φ, respectively, which will be further identified in the
200 Each equation of the proposed scheme is independent, and the local truncation error of the
201 proposed method is examined to ensure the second-order accuracy of the overall scheme in this
202 section [7][19][32][33]. Specifically, based on the proposed method (i.e., Eqs. (1) to (12)), the
203 overall scheme can be derived by assuming f(t) = 0 in Eqs. (1a) and (2). The following recursive
204 relationship between the displacement and velocity can be obtained, and a similar derivation
xt xt − t 0 0 1 0 xt − t
xt xt − t 0 0 0 1 xt − t
206
= A = (13)
31 32 33 34 xt
xt + t xt
x x 41 42 43 44 xt
t + t t
207 where A is the amplification matrix, and each element in the matrix (e.g., μ31) is given in
208 Appendix II. The characteristic equation of the matrix A is derived as follows:
11
209 q( ) = − det( A − I) = − 4 + A4 3 + A3 2 + A2 + A1 (14)
210 where λ denotes the eigenvalue of A, I is the identity matrix, and coefficients A1, A2, A3, and
211 A4 are given in Appendix III. Based on the characteristic equation, the difference equation of
1
213 (− xt + 2 t + A4 xt +t + A3 xt + A2 xt −t + A1 xt − 2 t ) = 0 (15)
t 2
214 Based on the difference equation (i.e., Eq. (15)), the local truncation error (LTE) can be
215 derived. The derivation process of LTE is that all items on the left side of Eq. (15) are first
216 expanded as Taylor series at the time step t, then, Taylor series of the displacement at the time
217 step t is subtracted by the expanded Eq. (15), which leads to the truncation terms (i.e., LTE),
220 where O ( t ) is the truncation error item of the overall integration scheme, and the coefficients
3 2
222 B0 =
2
( xt + 2 xt + xt ) = 0 (17 a)
3
224 B2 = ( − 2 ) 2 xt + ( − 4 − 3 ) xt(3) + ( − − 1) xt(4) (17 c)
2 2 8
226 where xt(3) , xt(4) , and xt(5) denote the third-order, fourth-order, and fifth-order derivatives with
227 respect to time t, respectively. Eqs. (17a) and (17b) are equal to zero for arbitrary parameters
228 α, β, γ, φ, and ξ. In other words, the proposed integration method can always keep second-order
232 The coefficient of Δt2 equals zero (i.e., LTE = B3t 3 + O(t 4 ) ), and the proposed integration
233 method has third-order accuracy at least. Two free parameters (i.e., α and φ) still exist in the
238 and fourth-order accuracy (i.e., LTE = O(t 4 ) ) can also be achieved in the proposed method. In
239 addition, the proposed method cannot maintain numerical stability when φ = 16/3, which will
242 Based on the accuracy analysis above, the proposed explicit integration method has a third-
243 order accuracy for arbitrary parameter values φ and α, and the standard computational
246 12 12
xt + t = x t −t − x +5xt −t + 8xt (20 b)
t t t
247 xt +t = f (t ) − N ( xt +t , xt +t ) (20 c)
13
8 + 5 4+ 8− 2+
249 xt + t = x t −t − xt+ x t + t + xt −t + xt (20 e)
4t t 4t 2
250 Since the quantities at the time step t-Δt are involved in the standard equations, the
252 1 2
x1 = x0 + tx0 + t x0 (21)
2
253 x1 = x0 + tx0 (22)
254 where x0, 𝑥̇ 0 , and 𝑥̈ 0 are the initial displacement, velocity, and acceleration, respectively, and
255 x1 and 𝑥̇ 1 are the displacement and velocity at the first step, respectively. The standard
256 computational procedure for a multi-DOF system is given in Appendix IV. Based on the
257 discussion above, some features of the proposed method are summarized as follows:
258 (1) The third-order accuracy is achieved in the proposed method at least, and the fourth-
260 (2) The displacement and velocity responses at the time step t+Δt can be computed only
261 by using the previous displacements and velocities at the time steps t and t-Δt;
262 (3) The proposed explicit method is a single-step integration method, and only vector
263 calculations are conducted in the calculation for the dynamic problem with a lumped mass
265 And (4) iterative calculations are not required in the solution of velocity-dependent
268 Stability and accuracy are two important indexes to measure the effectiveness of a new
269 integration method, and the characteristics of the stability and accuracy of the proposed method
14
270 are investigated by assuming the parameter α = (10+13φ)/(3φ).
272 The spectral radius ρ is widely adopted to measure the stability of a numerical integration
273 method, and the algorithmic stability can be ensured when ρ ≤ 1 [33]. For the proposed
274 integration method, the spectral radius ρ can be expressed as follows [34]:
276 where λ1, λ2, λ3, and λ4 are four eigenvalues of the amplification matrix A, which can be solved
277 by Eq. (14). Moreover, to ensure algorithmic stability without physical damping, the following
278 requirement can be derived according to the Routh-Hurwitz criteria, and a similar derivation
280 0 (24)
281 To explore the characteristics of the stability varying with the parameter φ, the curves of
282 the spectral radius considering different parameter values are plotted in Fig. 2. It is shown that
283 (1) the stability requirement (i.e., ρ ≤ 1) can be satisfied when φ < 0, (2) the abscissa values
284 of the spectral radius at the bifurcating points (as marked in Fig. 2) gradually increase with the
285 decreasing φ, and the ordinate values decrease first and then increase, (3) the critical ratio value
286 (i.e., the abscissa value) of Δtcr/T ((i.e., the critical time step Δtcr / the minimum period T) is
287 close to that of CD method (i.e., 1/π) when φ = -0.001, (4) the spectral radius (i.e., the ordinate
288 value) is close to one when φ = -0.001, and the stable regions of the proposed method decrease
289 gradually with the increase of damping ratios. Therefore, a large parameter value φ (i.e., a small
290 absolute value) should be adopted to obtain a relatively large critical time step size in the
15
291 computation. The suggested range of the parameter φ is from -10 to -0.01, which has a similar
292 critical time step size to CD method, and the stability area and critical step size at the suggested
294
295 Fig.2 Spectral radius curves of the proposed method under various parameters φ
296
297 Fig.3 Critical time step size of the proposed method for various parameters φ and ξ.
299 The numerical dissipation and dispersion are two indexes to measure the accuracy of the
300 numerical integration method [7][17], and they are, respectively, evaluated by using the
301 amplitude decay ξ̅ (AD) and the period elongation (T̅ -T)/T (PE), where ξ̅ and T̅ are the
302 algorithmic damping and period, and T is the exact period. Detailed information on the two
303 indexes can be found in [35][36][37]. Although the numerical dissipation represents the
304 computational error, it is useful for filtering out spurious content in the high-frequency response
305 domain. Spurious vibrations only exist in mathematical results, not in practical engineering.
306 One of the characteristics of an admirable integration method is that the proposed method can
307 obtain accurate responses in low-frequency response domains and adequately suppress
308 spurious vibration in high-frequency response domains simultaneously, which is usually not
310 To explore the characteristics of the accuracy of the proposed method, the curves of the
311 AD and PE for different damping scenarios are shown in Fig. 4 and Fig. 5, respectively. Four
312 typical explicit integration methods with a single time step (i.e., the CD method, Zhai method
16
313 [17], Kim method [16], and Yuan method [19]) are compared in the two figures to highlight the
314 superiority of the proposed method. It is important to note that to obtain the high accuracy in
315 the low-frequency domain and the high numerical dissipation in the high-frequency
316 simultaneously, the parameters are selected as φ = ψ = 0.6 for the Zhai method, β = 1/27 for
317 the Kim method, and (β = 0.2 and δ = 0.3) for the Yuan method, as marked in Fig. 5 and Fig.
318 6. Moreover, the Zhai method and the Yuan method are constructed by three continuous
319 computational time steps, and both the CD method and the Kim method are built by two
321 Fig. 4 shows that (1) the CD method has no numerical dissipation with the increase of
322 Δt/T, (2) the Zhai method presents a large numerical dissipation in the low-frequency response
323 domain (i.e., the small ratio stage of Δt/T), and a relatively small numerical dissipation is
324 observed in the high-frequency response domain (i.e., the large ratio stage of Δt/T), and (3) the
325 Yuan method seems to satisfy the condition of an admirable algorithm as it has an adequate
326 numerical dissipation in the high-frequency domain compared with the CD, Zhai, and Kim
327 methods. For the proposed integration method, the accuracy gradually increases in the low-
328 frequency response domain with the decreasing φ, and the algorithmic damping gradually
329 increases in the high-frequency response domain. In other words, high accuracy can be
330 achieved in the low-frequency domain, and large numerical dissipation can be obtained in the
331 high-frequency domain by adjusting parameters, e.g., φ = -2.5, as shown in Fig. 4. The AD
332 curves for different integration methods with ξ = 0.02 are shown in Fig. 4 (b). A similar trend
333 to the scenarios where ξ = 0 is captured. For instance, the high accuracy in the low-frequency
17
334 domain and the large numerical dissipation in the high-frequency domain are observed in the
335 proposed method, and both the accuracy and dissipation for the proposed method can be
337 The PE given in Fig. 5 shows that (1) a similar trend of the PE is captured in the scenarios
338 with or without physical damping (i.e., ξ = 0 in Fig. 5 (a) and ξ = 0.02 in Fig. 5 (b)), (2) the
339 value of the PE gradually increases with the increase of φ from positive values to negative
340 values in the high-frequency domain, (3) the Zhai method has the maximum PE compared with
341 other methods, and (4) the minimum PE in low-frequency parts can be obtained in the proposed
343
344 Fig.4 Comparison of the amplitude decay in the proposed integration method and the
346
347 Fig.5 Comparison of the period elongation in the proposed explicit integration method
350 Four representative examples are studied to demonstrate the algorithmic properties of the
351 proposed method by comparing with the existing state-of-the-art methods. Specifically, the
352 accuracy and convergence are evaluated by using an SDOF dynamic system with a theoretical
353 solution. The effectiveness of the nonlinear system calculation is investigated via a typical
354 nonlinear dynamic system. A linear Howe truss model subjected to impact and earthquake loads
355 is calculated to explore its effectiveness in calculating both high-frequency vibrations and
18
356 earthquake responses. Finally, a nonlinear wellbore model is used to study the solution ability
359 A typical SDOF linear dynamic system without damping is used to study the computational
360 accuracy theoretically. Namely, ξ = 0 and f(t) = 0 are set in Eq. (2). The equation of motion of
363 where ω = 2π rad/s, and T = 1 s. The initial conditions are x(0) = 1 m and ẋ(0) = 0 m/s. The
366 The time history curves of the displacement, velocity, and acceleration are depicted in Fig.
367 6, and responses of different integration methods are compared in the figure, e.g., the CD
368 method, the Zhai method (φ = ψ = 0.5), and Kim method (β = 1/27) with the second-order
369 accuracy, and the Yuan method with the third-order accuracy. Their stability limits Δtcr/T are
370 0.318 for the CD method, 0.318 for the Zhai method [17], 0.297 for the Kim method [16],
371 0.288 for the Yuan method [19], and 0.24 for the proposed method. A time step of 0.05 s is used
372 in the calculation for all integration methods. To observe the difference with the exact results
373 clearly, only the responses from 9 s to 10 s are shown in Fig. 6. It is shown that (1) similar
374 displacement responses are observed in the CD method and Zhai method, (2) the responses of
375 the proposed method well match up with the exact results compared with other methods, i.e.,
376 high accuracy is ensured in the low-frequency response domain (f = 1 Hz), and (3) the accuracy
19
377 of can be adjusted by using the parameters φ.
378
379 Fig. 6. Comparison of the accuracy in various integration methods and the exact
380 solution. (a) displacement, (b) velocity, and (c) acceleration.
382 The following global error defined in [12] is used to investigate the rate of convergence
(x i
p
− dip )2
384 Ep = i=0
n
(26)
(d
i=0
i
)
p 2
385 where p is the order of the derivative with respect to time, p = 0, 1, and 2 represent displacement,
386 velocity, and acceleration, respectively, 𝑥𝑖𝑝 and 𝑑𝑖𝑝 are the numerical and theoretical results at
387 the time step ti, respectively, and n is the number of the computational time step.
388 The rates of convergence for the displacement, velocity, and acceleration are presented in
389 Fig. 7, and the methods discussed in Fig. 6 are used to comparatively analyze the rate of
390 convergence. It is shown that (1) the CD method has the second-order accuracy in the
391 displacement responses, but it only maintains first-order accuracy in both the velocity and
392 acceleration responses, (2) both the Zhai method with φ = ψ = 0.5 and the Kim method with β
393 = 1/27 retain the second-order accuracy, (3) the Yuan method can keep the third-order accuracy
394 except for the second-order accuracy of velocity responses, (4) the proposed method has the
395 fastest rate of convergence and achieves the fourth-order accuracy in displacement, velocity,
396 and acceleration responses, and (5) the rate of convergence of the proposed method slightly
397 increases with the decreasing φ. Therefore, compared with the state-of-the-art integration
20
398 methods, the proposed method is characterized by a high rate of convergence and accuracy for
400
401 Fig. 7. Comparisons of the rate of convergence in the proposed method and typical
402 methods (T = 1). (a) displacement, (b) velocity, and (c) acceleration.
404 A typically nonlinear second-order differential equation (i.e., the well-known van der Pol’s
405 equation) is used to discuss the applicability of the proposed method in terms of nonlinear
408 where c0 = 2 is the damping coefficient. The initial conditions are x(0) = -0.02 and 𝑥̇ (0) = 0.
409 The computational time step is 0.2 s for the proposed method and the methods discussed
410 in Fig. 6, and the near-exact solution is calculated by using the Kim method with a time step of
411 0.002 s. Fig. 8 shows the dynamic responses of the nonlinear system. It is shown that (1) the
412 errors from the CD method, the Zhai method, the Kim method, and the Yuan method are
413 obvious compared with those of the proposed method, and (2) the results from the proposed
414 method overlap with the near-exact results when φ = -4. Therefore, compared with the
415 examined methods, the proposed method can more accurately compute the nonlinear dynamic
416 system.
417
418 Fig. 8 Dynamic responses of the nonlinear system calculated by various integration
21
420 4.3 A linear multi-DOF dynamic system
421 A linear Howe truss, as shown in Fig. 9, is used to discuss the algorithmic performance of the
422 proposed integration methods in terms of complex structures and loads [41][42]. The accuracy,
423 the dissipation, and the efficiency are investigated successively. The periodic impact load is set
424 as F(t) = Picos(ωt), in which Pi represents external loads acting on nodes, as shown in Fig. 9.
425 Two loading frequencies of ω = 0 and 100 π are employed to generate spurious responses in
426 the analysis, and the duration of the impact loads is 0.1 s. The material properties of the Howe
427 truss model are given in Fig. 9. Structural damping is ignored in the calculation, and the element
429
430 Fig. 9. A Howe truss under periodic impacts
432 The horizontal responses of Node 5 are shown in Fig. 10 when the loading frequency is
433 ω = 0. The near-exact solution is obtained by using the CD method with a time step of 0.00001
434 s, and a time step of 0.001 s is employed in the calculation of other methods. Only the proposed
435 method can achieve the approved results, and large errors are observed in the results from other
436 methods. In addition, more accurate solutions can be obtained by adopting smaller time steps
437 for all these methods, but the computational cost will remarkably increase. Thus, the proposed
438 method outperforms in terms of balancing accuracy and efficiency when linear multi-DOF
440
441 Fig. 10 Responses of Node 5 in the horizontal direction when ω = 0. (a) displacement, (b)
22
442 velocity, and (c) acceleration.
444 To generate the spurious vibrations, the structural stiffness is magnified 55 times, and the
445 loading frequency is 100 π. The computational results show that the response frequency ranges
446 from 94 to 697 Hz. The maximum vibration frequency exists in the horizontal vibration of
447 Node 5, whose responses are used to discuss the dissipation property. The response with full
448 frequencies is calculated by using the CD method with a time step of 0.00001 s, as shown in
449 Fig 11. The time steps used in other methods are 0.0003 s. The displacement responses are
450 given in Fig. 11 (a), and its enlarged view from 0.085 s to 0.093 s is depicted in Fig. 11 (b) to
451 observe spurious vibrations. Fig. 11 (a) shows that results in the low-frequency response
452 domain overlap with the full-frequency responses for all integration methods. In other words,
453 accurate results can be captured in the low-frequency domain by all methods. Moreover, the
454 spurious high-frequency vibrations, as marked in Fig. 11 (a), can be efficiently filtered out by
455 adjusting the parameter φ of the proposed method, which cannot be realized in the CD method.
456 Therefore, the proposed method is featured with the superior numerical dissipation property.
457
458 Fig. 11 Responses of Node 5 in the horizontal direction when ω = 100 π. (a) displacement
461 To explore the potential and computational accuracy under complex loadings, the Howe
462 truss under seismic impacts, as shown in Fig. 12, is calculated. A time step of 0.002 s is used
463 in the calculation for all the methods. The near-exact solution is solved by using a time step of
23
464 0.0001 s, and responses of 50 s are calculated in the analysis. The responses of Node 5 are
465 plotted in Fig. 13. We find that (1) the methods (i.e., the CD method, the Zhai method, and the
466 Kim method) have relatively poor results, (2) the Yuan method makes more progress in the
467 computing accuracy in terms of the displacement responses, as shown in Fig.13 (a), and (3)
468 satisfying results are observed in all responses by using the proposed method, including
469 displacement, velocity, and acceleration responses. Therefore, the proposed method is suitable
470 for solving the responses of multi-DOF systems under complex impacts.
471
473
474 Fig. 13 Earthquake responses of Node 5 in the horizontal direction. (a) displacement, (b)
475 velocity, and (c) acceleration.
477 To compare the computational efficiency of different integration methods, the same
478 accuracy is ensured in all methods. More specifically, considering the accuracy of integration
479 methods comprehensively (i.e., including the AD and PE), different ratios of Δt/TMin are
480 selected to obtain the same global errors of the acceleration (e.g., Log(E2) = 0.01), as shown in
481 Fig. 7 (c). The obtained ratios and time step sizes of the investigated methods are listed in Table
482 1. Fig. 14 gives the average time consumption of 10 calculations for the Howe truss subjected
483 to earthquake excitations with a duration of 50 s. All calculations are conducted by a computer
484 with an Intel(R) Core (TM) i7-8700. We find that (1) the largest time consumption is observed
485 in the CD method due to the time-consuming factorization of an effective stiffness matrix, and
24
486 (2) the proposed method has the shortest time consumption, which is around 65% of that of the
487 Yuan method, 38% of that of the Kim method, and 30% of that of the Zhai method. Therefore,
489 Table 1. The ratios of t / T and time step sizes for various methods.
490
491
494 A nonlinear plate model with 3600 DOFs, which is employed to simulate the wellbore
495 shown in Fig. 15, is studied to evaluate the computational accuracy of large-scale nonlinear
496 systems based on the plane strain assumption. The inner and outer boundaries and the thickness
497 of plating (d = 0.1 m) of the wellbore are shown in Fig. 15 (a), and more information can be
498 accessed in [43][44]. Considering the symmetry, only Pi/4 rad of the plate model is built to
499 simplify the modeling process, and the relevant parameters are given in Table. 2. Fig. 15 (a)
500 depicts the inner and outer wall dimensions. The periodic radial force applied to the inner wall
501 is set as F(t) = r0sin(ωt), in which r0 = 104 N and ω = 10π rad/s. The lump mass matrix is used
502 in the calculation and the damping ratio is 0.01. The minimum and maximum frequencies of
503 the model are 30.93 Hz and 16072 Hz, respectively. Considering the nonlinear damping with
504 the combined viscous and quadratic damping [45], the governing equation of the wellbore plate
506 Mb X + Cb X + Db X X + K b X = Fb (28 a)
25
507 Cb = a0Mb + a1K b , Db = a0Mb (28 b,c)
508 where Mb, Kb, and Cb are the mass matrix, the stiffness matrix, and the Rayleigh damping
509 matrix, respectively, X, Ẋ and Ẍ are the displacement, velocity and acceleration vectors,
510 respectively, Fb is the loading vector, Db is the constant matrix related to the mass matrix, and
511 a0 and a1 are the coefficients of the mass matrix and stiffness matrix, respectively.
512 Only pure vector operations are conducted in the calculation for the Kim method, the Zhai
513 method, and the Yuan method. Hence, they are selected in the comparative analysis. The CD
514 method is excluded in the comparison due to the factorization of the equivalent stiffness matrix
515 in nonlinear calculations [13]. The horizontal and vertical acceleration responses of the two
516 representative points P1 and P2 (as marked in Fig. 15 (b)) are presented in Figs. 16 and 17,
517 respectively, to explore the response properties. The near-exact solution is obtained by using
518 the Kim method with a time step of 2×10-7 s, and the time step of other methods is 2×10-6s.
519 Considering the critical time steps of integration methods, the algorithmic parameters are,
520 respectively, selected as β = 0.2 for the Kim method, φ = 0.5 and ψ = 0.5 for the Zhai method,
521 β = 0.2 and δ = 0.25 for the Yuan method, and φ = 0.01 for the proposed method, as shown in
522 Figs. 16 and 17. It is shown that only the response curves of the proposed method overlap with
523 the near-exact solution, and results from other methods are gradually divergent, as marked in
524 the locally enlarged views. Note that the Kim method, Zhai method, and Yuan method can also
525 obtain near-exact solutions if they adopt smaller time steps. Therefore, the proposed method
526 can be used to solve complex nonlinear dynamic systems, and the proposed method has a better
527 rate of convergence when compared with the Kim method, the Zhai method, and the Yuan
26
528 method.
529
530 Fig. 15 The nonlinear wellbore model with (a) the cross-section and (b) the simplified model.
531
532 Table 2. Main parameters of the wellbore model
533
534
535 Fig. 16 The dynamic responses of the point P1 close to the inner boundary. (a) horizontal
537
538 Fig. 17 The dynamic responses of the point P2 close to the out boundary. (a) horizontal
540 5. Conclusions
541 In this study, an explicit integration method is proposed, and two parameters φ and α are
542 introduced to adjust its stability and accuracy properties. A systematical theoretical and
543 numerical demonstration of the algorithmic properties of the proposed method is conducted,
544 and several typical integration methods, e.g., the CD method, the Kim method, the Yuan
545 method, and the Zhai method, are compared in the analysis. Four representative examples are
546 used to investigate the properties of the proposed method, including the accuracy, efficiency,
547 dissipation, and effectiveness of the nonlinear calculation. Some conclusions are summarized
548 as follows:
549 • The third-order accuracy can be realized in the proposed method at least, and the
551 • The proposed method is featured with desirable accuracy properties. Specifically,
27
552 high accuracy can be ensured in the low-frequency response domain, and
554 domain. Therefore, the proposed method can be used to efficiently filter out
557 • The proposed explicit method is a single-step integration method, and it can
558 advance the solution step-by-step via the obtained displacements and velocities.
560 for the structure with a lumped mass matrix. Namely, pure vector calculations are
561 conducted for the dynamic analysis. Therefore, high computational efficiency is
564 characterized by a high rate of convergence and accuracy for the linear dynamic
565 calculation, and more accurate results can be obtained in the nonlinear dynamic
566 calculation.
567 The limitations of the proposed method and future work are given as follows:
568 1. The proposed method cannot achieve self-starting, and the information from the first
569 step needs to be given in the calculation. Self-starting is possible when the displacement and
570 velocity at the time step (t0-Δt) are obtained by using initial conditions (t0).
571 2. The applicability of the proposed explicit integration method in more complex
28
573 References
574 [1] He JJ, Lu ZZ, Liu YM. New method for concurrent dynamic analysis and fatigue damage
576 [2] Yu Y, Yao H, Liu YM. Aircraft dynamics simulation using a novel physics-based learning
578 [3] Hughes Thomas JR. The finite element method: linear static and dynamic finite element
580 [4] Chopra AK. Dynamics of structures. New Jersey: Prentice Hall, 2011.
581 [5] Dokainish MA, Subbaraj K. A survey of direct time-integration methods in computational
583 [6] Yuan P, Cai CS, Li DJ, Xu GJ, Li CX. New method for concurrent dynamic analysis and
585 [7] Bathe KJ. Finite element procedures. New Jersey: Prentice Hall, 1996.
586 [8] Bathe KJ, Noh GW. Insight into an implicit time integration scheme for structural dynamics.
588 [9] Hulbert GM, Chung J. Explicit time integration algorithms for structural dynamics with
589 optimal numerical dissipation. Comput Methods Appl Mech Eng 1996;137(2):175-188.
590 [10] Kim W. An accurate two-stage explicit time integration scheme for structural dynamics
591 and various dynamic problems. Int J Numer Methods Eng 2019;120(1):1-28.
592 [11] Liu W, Guo WH. A Non-Iterative Integration Scheme Enriching the Solution to the
593 Coupled Maglev Vehicle-Bridge System. Int J Struct Stab Dyn 2021;21(04):2150052.
594 [12] Wen WB, Wei K, Lei HS, Duan SY, Fang DN. A novel sub-step composite implicit time
29
595 integration scheme for structural dynamics. Comput Struct 2017;182:176-186.
596 [13] Park KC. An improved stiffly stable method for direct integration of nonlinear structural
598 [14] Fung TC. Numerical dissipation in time-step integration algorithms for structural dynamic
600 [15] Chung JT, Lee J. A new family of explicit time integration methods for linear and non-
602 [16] Kim W. A simple explicit single step time integration algorithm for structural dynamics.
604 [17] Zhai WM. Two simple fast integration methods for large-scale dynamic problems in
606 [18] Yuan P, Li DJ, Cai CS, Xu GJ. An efficient decoupling dynamic algorithm for coupled
608 [19] Yuan P, Li DJ, Cai CS, Xu GJ. A Novel Decoupling Dynamic Method with Third-order
610 [20] Zhu ZH, Gong W, Wang LD, Li Q, Yu ZW, Harik IE. An efficient multi-time-step method
612 [21] Dormand JR, Prince PJ. A family of embedded Runge-Kutta formulae. J Comput Appl
614 [22] Noh G, Bathe KJ. An explicit time integration scheme for the analysis of wave
616 [23] Kim W, Lee J. An improved explicit time integration method for linear and nonlinear
618 [24] Kim W. A new family of two-stage explicit time integration methods with dissipation
620 [25] Rezaiee-Pajand M, Karimi-Rad M. A family of second-order fully explicit time integration
622 [26] Liu W, Guo WH. A novel predictor-corrector explicit integration scheme for structural
30
623 dynamics. Structures 2021;34:2735-2745.
624 [27] Braekhus J, Aasen J. Experiments with direct integration algorithms for ordinary
626 [28] Diogo FR, Walnório GF, Webe JM, Adenilcia Fernanda Grobério Calenzania. A review of
627 automatic time-stepping strategies on numerical time integration for structural dynamics
629 [29] Kim W, Reddy JN. Novel explicit time integration schemes for efficient transient analyses
631 [30] Delfim S, Georg G. Nonlinear structural dynamic analysis by a stabilized central
633 [31] Idesman AV. A new high-order accurate continuous Galerkin method for linear
635 [32] Cheng FY. Matrix analysis of structural dynamics: applications and earthquake
637 [33] Liu TH, Huang FL, Wen WB, He XH, Duan SY, Fang DN. Further insights of a composite
638 implicit time integration scheme and its performance on linear seismic response analysis.
640 [34] Rezaiee-Pajand M, Esfehani SAH, Karimi-Rad M. Highly accurate family of time
642 [35] Hilber HM, Hughes TJR. Collocation, dissipation and [overshoot] for time integration
644 [36] Chang SY. A family of noniterative integration methods with desired numerical dissipation.
646 [37] Bazzi G, Anderheggen E. The ρ-family of algorithms for time-step integration with
648 [38] Kougioumtzoglou IA, Spanos PD. Nonlinear MDOF system stochastic response
649 determination via a dimension reduction approach. Comput Struct 2013;126: 135-148.
650 [39] Kougioumtzoglou IA. Stochastic joint time--frequency response analysis of nonlinear
31
651 structural systems. J Sound Vib 2013;332(26):7153-7173.
652 [40] Semler C, Gentleman WC, Paı Doussis MP. Numerical solutions of second order implicit
654 [41] Yuan P, Li DJ, Cai CS, Xu GJ. Time integration method with high accuracy and efficiency
656 [42] Guo T, Frangopol DM, Chen YW. Fatigue reliability assessment of steel bridge details
657 integrating weigh-in-motion data and probabilistic finite element analysis. Comput Struct
658 2012;112:245-257.
659 [43] Hodge MO. Avoiding Borehole Failure by Time-dependent Stability Analysis of Stressed
661 [44] Yuan P, Dong Y. High-efficient decoupling method for coupling systems with multiple
663 [45] Franchetti P, Modena C, Feng MQ. Nonlinear damping identification in precast prestressed
665 [46] Wang CH, Xiao JZ, Wang CQ, Zhang CZ. Nonlinear damping and nonlinear responses of
666 recycled aggregate concrete frames under earthquake loading. Eng Struct
667 2019;201:109575.
668
670 The displacement prediction scheme of the proposed method can be written as follows:
672 Eq. A1 is similar to the general integration expression and detailed information on the
673 general integration expression can be found in [32]. By selecting suitable parameter values, Eq.
674 A1 can be exact in the special case where x is a polynomial of order m-1, i.e., x = tm-1, and the
675 second-order accuracy is obtained in the displacement prediction scheme, i.e., x = 1, t, t2.
681 b3 + b4 − b1 = 1 (A4)
686 Therefore, the following three relationships between parameters b1, b2, b3, and b4 can be
690 Using Eqs. (3) - (12), the following integration approximation operator can be obtained. A
0 0 1 0
0 0 0 1
692 Α= (B1)
31 32 33 34
41 42 43 44
693 where the coefficients in the matrix are expressed as follows:
698 t t 2t 4
(B6)
t 2
− 4t − 2 − 8 + +
2
t 2 2
42 = 2 − 16 + 4t 2 2 − − 2t 2 2 + 12t
699 2
(B7)
t
2 2
3t t
+ − − 4t + +5
4 2 2
36 12 9
43 = − − 16 − − 8t 2 + t 2
700 t t 2t
(B8)
t 2
+ 4t 2 + 2 + 8 − −
2
9
A2 = − 6 − 4t 2 2 − 3t 2 2 − 16t
704 2
(C2)
3t 2 2 3t
+ − 10t + +4
4 2
7t 2 2
A3 = 6 − 6 + 7t 2 2 + + 20t
705 2
(C3)
3t 2 2 13t 3t
− + − −6
4 2 2
707 Appendix IV. The standard computational procedure for a multi-DOF system
708 The standard computational procedure for multi-DOF systems can be written as follows:
34
709 Xt +t = (2 − 3) X t −t +(4 − 2 ) X t +( − 2)tXt −t + tXt (D1)
710 12 12
Xt + t = X t −t − X +5Xt −t + 8Xt (D2)
t t t
711 Xt + t = M −1(F t + t −N(Xt + t , Xt + t )) (D3)
8 + 5 4 + 8 − 2 +
713 Xt +t = X t −t − Xt + X t +t + Xt −t + Xt (D5)
4 t t 4 t 2
714 where Xt-Δt and Xt −t denote the vectors of the displacement and velocity at t-Δt, respectively,
715 Xt and Xt represent the vectors of displacement and velocity at t, respectively, Xt +t , Xt +t ,
716 and Xt +t are the vectors of the prediction displacement, velocity, and acceleration at t+Δt,
717 respectively, Xt+Δt and Xt +t are the vectors of the displacement and velocity at t+Δt,
718 respectively, Ft+Δt and N( Xt +t , Xt +t ) are the vectors of the external and internal forces at t+Δt,
719 respectively, and M is the mass matrix of dynamic systems. The computational flowchart of
720 the proposed method is shown in Fig. 1. Furthermore, computational procedures for numerical
722
724
35
725 Captions of table and figures
726 Table 1. The ratios of t / T and time step sizes for various methods.
727 Table 2. Main parameters of the wellbore model
731 Fig. 4. Comparison of the amplitude decay in the proposed integration method and the typical
732 integration methods. (a) ξ = 0 and (b) ξ = 0.02.
733 Fig. 5 Comparisons of the period elongation in the proposed explicit integration method and
734 typical methods. (a) ξ = 0 and (b) ξ = 0.02.
735 Fig.5 Comparisons of the period elongation in the proposed explicit integration method and
736 typical methods. (a) ξ = 0 and (b) ξ = 0.02.
737 Fig. 6. Comparison of the accuracy in various integration methods and the exact solution. (a)
738 displacement, (b) velocity, and (c) acceleration.
739 Fig. 7. Comparisons of the rate of convergence in the proposed method and typical methods (T
740 = 1). (a) displacement, (b) velocity, and (c) acceleration.
741 Fig. 8 Dynamic responses of the nonlinear system calculated by various integration methods.
742 (a) displacement, (b) velocity, and (c) acceleration.
743 Fig. 9. A Howe truss under periodic impacts
744 Fig. 10 Responses of Node 5 in the horizontal direction when ω = 0. (a) displacement, (b)
745 velocity, and (c) acceleration.
746 Fig. 11 Responses of Node 5 in the horizontal direction when ω = 100 π. (a) displacement and
747 (b) enlarged view.
748 Fig. 12 Vertical acceleration of ground motion
749 Fig. 13 Earthquake responses of Node 5 in the horizontal direction. (a) displacement, (b)
750 velocity, and (c) acceleration.
751 Fig. 14 Time consumptions for various methods
752 Fig. 15 The nonlinear wellbore model with (a) the cross-section and (b) the simplified model.
753 Fig. 16 The dynamic responses of the investigated point P1. (a) horizontal acceleration and (b)
754 vertical acceleration
36
755 Fig. 17 The dynamic responses of the investigated point P2. (a) horizontal acceleration and (b)
756 vertical acceleration
757 Fig. 18 The computational flowchart of the proposed method
37
758 Table 1. The ratios of t / T and time step sizes for various methods.
Methods Algorithm parameters Δt/TMin (×10-2s) Δt (×10-4s)
38
761 Table 2. Main parameters of the wellbore model
Parameters Unit Value
Modulus of elasticity N/m2 2×109
Poisson’s ratio - 0.3
Density kg/m3 2500
Damping ratio - 0.02
Outer diameter m 5
Inner diameter m 0.5
Angle rad π/4
762
763
39
764
40
767
768 Fig.2 Spectral radius curves of the proposed method under various parameters φ
769
41
770
771 Fig.3 Critical time step size of the proposed method for various parameters φ and ξ.
772
42
773
774
775 Fig.4 Comparison of the amplitude decay in the proposed integration method and the
43
777
778
779 Fig.5 Comparisons of the period elongation in the proposed explicit integration method
44
781
782
45
783
784 Fig. 6. Comparison of the accuracy between various integration methods and the exact
46
787
788
47
789
790 Fig. 7. Comparisons of the rate of convergence between the proposed method and
791 typical methods (T = 1). (a) Displacement, (b) velocity, and (c) acceleration.
792
48
793
794
49
795
796 Fig. 8 Dynamic responses of the nonlinear system calculated by various integration
50
799
800 Fig. 9. A Howe truss under impact
801
51
802
803
52
804
805 Fig. 10 Responses of Node 5 in the horizontal direction when ω = 0. (a) displacement, (b)
53
808
809
810 Fig. 11 Responses of Node 5 in the horizontal direction when ω = 100 π. (a) displacement
54
813
815
55
816
817
56
818
819 Fig. 13 Earthquake responses of Node 5 in the horizontal direction. (a) displacement, (b)
820 velocity, and (c) acceleration.
821
57
822
823 Fig. 14 Time consumptions for various methods
824
58
(a)
825
(b)
826
827 Fig. 15 The wellbore model with (a) the cross-section and (b) the simplified model.
59
828
829
830 Fig. 16 The dynamic responses of the investigated point P1 close to the inner boundary. (a)
831 horizontal acceleration and (b) vertical acceleration
832
833
60
834
835
836 Fig. 17 The dynamic responses of the investigated point P2 close to the out boundary. (a)
837 horizontal acceleration and (b) vertical acceleration
838
61
839
62