Handout # 13 (MEEN 617)

Numerical Integration to Find Time

Response of MDOF mechanical system
The EOMS for a linear mechanical system are
 + D U
MU  +K U =F (1)
(t ) (t )
where U, U, and U  are the vectors of generalized displacement,
velocity and acceleration, respectively; and F( t ) is the vector of
generalized (external forces) acting on the system.
M, D,K represent the matrices of inertia, viscous damping and
stiffness coefficients, respectively1. The solution of Eq. (1) is
unique with specified initial conditions U o , U o .

Numerical solution methods

• direct numerical integration
• modal analysis (mode displacement or mode
acceleration)

(Undamped and Damped) Mode superposition methods are

discussed earlier, see Lecture notes 7,8 and 11. The accuracy of
these methods is determined by the number of modes selected. It is
also important the time spent in performing the eigenvector
analysis. Modal superposition methods are most general.
elegant, and very powerful.

1
The matrices are square with n-rows = n columns, while the vectors are n-
rows.

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 1
Direct numerical integration implies a marching in time,
step-by-step procedure, solving the algebraic form of Eq. (1) at
specific time values. The direct qualification implies that, prior to
the numerical integration; no transformation of the EOM (1) is
carried out.

In a numerical integration method, EOM (1) is satisfied at

discrete time intervals, Δ t apart. In essence, a sort of quasi-static
equilibrium is sought at each time step.

Recall that a reliable numerical integration scheme should

a) reproduce EOM as time step Δ t → 0
b) provide, as with physical model, bounded solutions for any
size of time step, i.e. method should be stable
c) reproduce the physical response with fidelity and accuracy.

Numerical integration methods are usually divided into two

categories, implicit and explicit.
Consider the ODEs x = f ( x,t ) (2)

In an explicit numerical scheme, the ODEs are represented in

terms of known values at a prior time step, i.e.

xi+1 = xi + Δ t f ( xi , ti ) , (3)

while in an implicit numerical scheme

xi+1 = xi + Δ t f ( xi+1 , ti ) (4)

Explicit numerical schemes are conditionally stable. That is,

they provide bounded numerical solutions only for (very) small
time steps. For example,

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 2
 Tn
Δ t ≤τ crit = (5)
π
2π
where Tn = and ωn = K
M are the natural period and natural
ωn
frequency of the system, respectively. The restriction on the time
step is too severe when analyzing stiff systems, i.e. those with
large natural frequencies.

Note that in a MDOF system, the condition defined in Eq.

(5) is too restrictive since Tn refers to the smallest period of
natural motion (largest natural frequency). More often than
not, this critical time is NOT known unless the
eigenanalysis is completed. In nonlinear systems, the
critical time step limit worsens.

Implicit numerical schemes are unconditionally stable, i.e.

do not impose a restriction on the size of the time step Δ t .
(However, accuracy may be compromised if too large time steps
are used).

An integration method is unconditionally stable if the

solution for any initial condition does not grow without
bound for any time step Δ t , in particular when is large Δ t Tn .
The method is only conditionally stable if the above holds
provided that Δ t Tn is smaller than a certain value, usually
called the stability limit.

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 3
The Wilson-θ method2
It is an extension of the linear-acceleration method. That is,
within a time step, as shown in Figure 1, the acceleration vector is
proportional to time. Over the time interval 0 ≤τ ≤ θ Δt ,

U  t + τ ⎡ U
 t +τ = U  t +θΔt − U
 t ⎤
θ Δt ⎣ ⎦ (6)

with θ ≥ 1.

Wilson-θ linear acceleration method

Acceleration velocity
Xi +θ X i +θ
X (τ ) = a + bτ
Xi +1 Xi +1

Xi X i
τ 2
X (τ ) = X i + aτ + b
2
t
τ t+θΔτ
t
τ t+θΔτ
ti
tι+1=tι ti ti+1
+Δτ
X : component of vector U
displacement
X i +θ

θ >1 X i +1
τ τ 2 3
X (τ ) = X i + X iτ + a + b
Xi 2 6

t
τ t+θΔτ
ti ti+1

Fig 1. Depiction of components of acceleration, velocity and

displacement for numerical integration – Wilson-θ method

Integration of Eq. (6) gives the vector of velocity as

2
This numerical method is extremely popular among the structural dynamics
community.

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 4
  t +τ = U
 t +τ U
 t + τ 2
 t +θΔt − U
 t ⎤
U ⎡⎣ U ⎦ (7)
2θ Δt

One more integration delivers the vector of displacements as

t+ U τ 2
 t + τ 3
 t +θΔt − U
 t ⎤ (8)
U t +τ = U t +τ U ⎡⎣ U ⎦
2 6θ Δt
Specification of the vectors at time τ = θ Δt
(θ Δ )
2
 t +θΔt = U t + θΔt U
 t + t  t +θΔt − U  t ⎤
U ⎡⎣ U ⎦
2θ Δt (9)
 t + θΔt ⎡ U
=U  
⎣ t +θΔt + U t ⎤⎦

(θΔt ) (θΔt )
2 3
t
U t +θΔt = U t + θΔt U +  t
U +  t +θΔt − U
⎡⎣ U  t ⎤
2 6θ Δt ⎦
(10)
 t + (θΔt ) ⎡ U
2

= U t + θΔt U  t +θΔt + 2U
 t ⎤
6 ⎣ ⎦

From Eqs. (9) and (10), solve for the acceleration and velocity at
the end of the interval, i.e.
6 6 
 t +θΔt =
U [ U t +θΔt − U ] − Ut − 2 U t (11)
(θΔt ) θΔt
2 t

 t +θΔt = 3 [ U t +θΔt − U t ] − 2 U
U  t − θΔt U
 t (12)
θΔt 2

Now, state EOM (1) at the end of the time interval ( t + θ Δt )


at t + θ Δt ⇒ M U 
t +θ Δt + D U t +θ Δt + K U t +θ Δt = Ft +θ Δt (13)

where Ft +θ Δt = Ft + θ [ Ft +Δt − Ft ] (14)

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 5
is a (time) projection of the external force vector.

Substitute Eqs. (11) and (12) into (13) to determine an equation

from which U t +θΔt can be obtained.

Next, substitute this vector into Eq (12) and (11) to obtain

 t +θΔt and U
U  t +θΔt .

Lastly, use these vectors to determine in Eqs. (8) and (9) the
 t +Δt and U t +Δt .
displacement and velocity vectors at τ = Δt , i.e. U

The resulting system of equations for solution at each

time step is
ˆ U t +θΔt = Fˆ t +θΔt
K (15)

where
ˆ = K + ao M + a1D
K (16)

is the effective stiffness matrix, and

Fˆ t +θ Δt = Ft + θ [ Ft +Δt − Ft ] +
M ( ao U t + a2 U t
 ) +
 + 2U
t (17)

D ( a1 U t + 2 U
 +a U
t 3
 )
t

is the effective load vector at t + θ Δt . Above,

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 6
 6 3 θ Δt ao
ao = , a = , a = 2 a , a = , a = ,
(θ Δt ) θ Δt θ
2 1 2 1 3 4
2
(18)
a2 3 Δt Δt 2
a5 = − , a6 = 1 − , a7 = , a8 =
θ θ 2 6
The effective stiffness matrix can be easily decomposed (once) as
ˆ = L Dg LT If M, K, D are symmetric.
K

Otherwise,
ˆ =LU
K Δ , i.e. the product of a lower triangular
and upper triangular form matrices.

Hence, the solution of Eq. (15) proceeds step by step in a process

of backward and forward substitutions, i.e.

L Y = Fˆ t +θΔt
U Δ U t +θΔt = Y (19)

Solution of Eq. (19) gives Ut +θΔt . Next, calculate the

displacements, velocity, and acceleration vectors at ( t + Δt ) using

 t +Δt = a4 [ U t +θΔt − U t ] + a5 U
U  t + a6 U
 t
 t +Δt = U
 t + a7 ⎡ U
 
U ⎣ t +Δt + U t ⎤⎦ (20)
 t + a8 ⎡ U
U t +Δt = U t + Δt U  
⎣ t +Δt + 2U t ⎤⎦

Analysis shows that θ =1.37-1.40 for unconditional stability.

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 7
The Newmark-β method
This method is also an extension of the linear-acceleration
method. The velocity and displacement vectors at the end of the
time interval ( t + Δt ) are, as shown in Fig 2:

U  t + ⎡(1 − δ ) U
 t +Δt = U  t + δ U
 t +Δt ⎤ Δt \
⎣ ⎦
 t + ⎡( 1 2 − α ) U
Ut +Δt = Ut + Δt U  t + α U
 t +Δt ⎤ Δt 2
⎣ ⎦ (21)

where (α , δ ) are parameters selected to promote numerical

stability and/or gain in accuracy.

Linear acceleration method

Acceleration velocity

Xi +1 X i +1
Xi X (τ ) = a + bτ X i τ2
X (τ ) = X i + a τ + b
2

τ τ
ti ti+1 t t
ti ti+1

X : component of vector U
displacement

X i +1
τ τ
2 3
X (τ ) = X i + X iτ + a + b
Xi 2 6

τ
ti ti+1 t

Fig 2. Depiction of components of acceleration, velocity and

displacement for numerical integration – Newmark method

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 8
δ = 1 2 ,α = 16 , Eq. (21) represents the linear acceleration method
(θ=1 in Wilson’s method).

δ = 1 2 ,α = 1 4 , brings unconditional stability, and Eq. (21)

represents the average acceleration method discussed for SDOF
systems. See Lecture Notes 6.

The method solves the algebraic form of EOM (1) at the end of the
time interval

t + Δt ⇒ M U 
t +Δt + D U t +Δt + K U t +Δt = Ft +Δt (22)

Substitution of Eqs. (21) above leads to the algebraic equation:

ˆ U t +Δt = Fˆ t +Δt
K (23)

where ˆ = K + ao M + a1D
K (24)

is the effective stiffness matrix, and

Fˆ t +Δt = Ft +Δt + M ( ao U t + a2 U
 +a U
t 3
 ) +
t
(25)
D ( a1 U t + a4 U
 +a U
t 5
 )
t

is the effective load vector at t + Δt . Above,

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 9
 6 δ 1 1 δ
ao = , a = , a = , a = − 1, a = − 1,
α Δt 2 1
α Δt 2
α Δt 3
2α
4
α
(26)
Δt ⎛ δ ⎞
a5 = ⎜ − 2 ⎟ , a6 = Δt (1 − α ) , a7 = δΔt
2 ⎝α ⎠

In general, it is recommended that δ ≥ 2 ,α ≥ 4 ( 2 + δ )

2
1 1 1 (27)

The effective stiffness matrix is easily decomposed (once) as

ˆ = L U Δ for solution of Eq. (23), i.e. to determine U .
K t +Δt

The vectors of acceleration and velocity follow from

 t +Δt = a0 [ U t +Δt − U t ] − a2 U
U  t + a3 U
 t
(28)
 t +Δt = U
U  t + a6 U t + a7 U
 t +Δt

Stability of Numerical Method

In general, the numerical method solves
ˆ Ut = Fˆ t and K
K ˆ Ut +Δt = Fˆ t +Δt in two consecutive time steps.

For a sufficiently small time step,

ˆFt +Δt = Fˆ t + ∂ ˆ
F ∂ 2 ˆ
F Δt 2
∂ Fˆ
Δt + 2 + .... ≈ K U t +
ˆ Δt (29)
∂t ∂t 2 ∂t
Hence
∂ Fˆ
K U t +Δt = K U t +
ˆ ˆ Δt (30)
∂t

After some algebraic manipulations, U t +Δt = A U t + L F (31)

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 10
Hence, assume LF=0 for stability analysis. A is known as the
convergence matrix, Then,

U1 = A U 0 ,
 
U 2 = A U1 = A 2 U 0
(32)
.....
 n
U n = A U n−1 = A U 0

The numerical solution is bounded, i.e. stable, if the spectral radius

of the convergence matrix is less than one,

( )
ρ A = λmax <1 (33)

That is, none of its eigenvalues is greater or equal to unit.

Numerical Solution for Nonlinear

Systems
The EOMS for a non-linear mechanical system are
 + F = F
MU (34)
NL( t ) (t )

(
 ,U
where FNL( t ) = f NL U, U )
 is a nonlinear function of
 and U
U, U,  , i.e. the vectors of generalized displacement, velocity
and acceleration. Specification of Eq. (34) at two times,
sufficiently close to each other, gives:

 + F = F
MU (35a)
t NL( ) (t ) t

 + F
MU = F(t +Δt ) (35b)
t +Δt NL( t +Δt )

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 11
Note that in (35a) all information is known. Subtracting (b) from
(a) gives:

 + ΔF = ΔF
M ΔU t NL( ) (t ) t
(36)

Now, let
ΔFNL( ) = FNL (t +Δt ) − FNL (Δt ) 
Δt

⎛ ∂FNL ⎞ ⎛ ∂FNL ⎞  ⎛ ∂FNL ⎞  (37)

⎜ ⎟ ΔU t + ⎜  ⎟ ΔU t + ⎜  ⎟ ΔU t
⎝ ∂ U ⎠t ⎝ ∂ U ⎠t ⎝ ∂ U ⎠t
where
 = ( U
t +Δt − U t ) , ΔU t = ( U t +Δt − U t ) , ΔU t = ( U t +Δt − U t )
ΔU     
t

(38)

And define the following matrices, evaluated at each time step,

⎛ ∂F ⎞ ⎛ ∂F ⎞ ⎛ ∂F ⎞
K t = ⎜ NL ⎟ ; Dt = ⎜ NL ⎟ ; M t = ⎜ NL ⎟ (39)3
 
⎝ ∂ U ⎠t ⎝ ∂ U ⎠t ⎝ ∂ U ⎠t

These matrices represent linearized stiffness (K), viscous

damping (D) and inertia (M) coefficients. Substitution of Eq.
(39) into (37) and into Eq. (36) gives:

( M + M t ) ΔU t + Dt ΔU t + K t ΔUt = ΔFt (40)

3
The elements of the stiffness matrix are, for example,
∂ FNLb
K a ,b = , a, b = 1, 2,...n
Ua

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 12
 The solution of this algebraic equation proceeds in the same
form as for the SDOF system, see Lecture Notes 6.

Thus, the numerical treatment is similar, except that at each time

step, linearized stiffness, damping and inertia coefficients need be
calculated.

Although the recipe is identical; however with the apparent

nonlinearity, the method does not guarantee stability for (too) large
time steps.

References
Bathe, J-K, 1982 (1st ed), 2007 latest, “Finite Element Procedures,”
Prentice Hall.

Press, W.H., Flannery, B.P., Teukolsky, S.A., and Vetterling, W.T., 1986
(1st edition), “Numerical Recipes, The Art of Scientific Computing,”
Cambridge University Press, UK.

Other
San Andrés, L, 2008, “Numerical Integration to Find Time Response of
SDOF mechanical system,” MEEN 617, Lecture Notes 6,
http://phn.tamu.edu/me617

Piche, R., and P. Nevalainen, 1999, “Variable Step Rosenbrock Algorithm

for Transient Response of Damped Structures,” Proc. IMechE, Vol. 213, part
C, Paper C05097.

MEEN 617 HD 13 Numerical Integration for Time Response: MDOF system L. San Andrés © 2008 13
 Numerical Integration of MDOF Linear systems with
Viscous Damping - NEWMARK METHOD
Original by Dr. Luis San Andres for MEEN 617 class / 2008

The equations of motion are: M d2U/dt2 + C dU/dt + K U = P(t) (1)

where M,C,K are nxn matrices of inertia, viscous damping and stiffness coefficients, and U,
V=dU/dt, W=d2U/dt2,and are the nx1 vectors of displacements, velocity and accelerations, resp,
and P(t) is the nx1 vector of generalized forces.
Eq. (1) is solved numerically with initial conditions, at t=0, Uo,Vo=dU/dt
=====================================================================================
1. Define elements of inertia, stiffness, and damping matrices:
n := 3 # of DOF
⎛⎜ 100 0 0 ⎟⎞ ⎛⎜ 2⋅ 107 −1⋅ 107 0 ⎟⎞
⎛⎜ 5000 0 0 ⎞
⎟
K := ⎜ −1⋅ 107 2.5⋅ 107 −.5⋅ 107 ⎟
M := ⎜ 0 100 0 ⎟
⎜ 0 0 50 ⎟ ⎜ ⎟ C := ⎜ 0 2500 −1000 ⎟ ⋅ .0
⎝ ⎠ ⎜ 0 ⎜ 0 −1000 1000 ⎟
−.5⋅ 10 .5⋅ 10 ⎟⎠ ⎝ ⎠
7 7
⎝
kg N/m N.s/m

⎛⎜ 0 ⎟⎞ ⎛⎜ 0 ⎟⎞
Input vectors of initial conditions: Fmax := 10000 N
U0 := ⎜ 0 ⎟ V0 := ⎜ 0 ⎟
⎜0 ⎟ ⎜0 ⎟
m ⎝ ⎠ m/s ⎝ ⎠ Timpulse := 0.01

natural frequencies
smallest natural period Tmin = 0.011
⎛⎜ 0.026 ⎟⎞
Tmin 1
Set sampling time: Δt := −3 = ⎜ 0.016 ⎟
Δt = 5.377 × 10 should be less than
2 f ⎜ 0.011 ⎟
10
smallest period ⎝ ⎠
Number of time steps: Ntimes := 2 3
Ntimes = 1.024 × 10
Pulse load

Thalf := 0.5⋅ Timpulse EXAMPLE

t
F ( t) := Fmax⋅ if t < Thalf
Thalf

⎡ ( t − Timpulse) ⎤ if T ≤ t < T
⎢ −Fmax⋅ ⎥ half impulse
⎣ Thalf ⎦ f 1 = 61.64
0 otherwise

Build time and force vectors: Timpulse

= 1.86
Δt
i := 0 .. Ntimes − 1

ti := i⋅ Δt Fimp := F ( ti) Note that impact must be well captured

i
 i
Pulse load

1 .10
4
1 .10
4
Timpulse
= 1.86
Δt
Fimp
i 5000 F ( tt) 5000

0
0 0.1 0
0 0.1
ti
tt

Numerical sampling vs. actual function

Set parameters for Newmark integration method:

α := 0.25⋅ ( 0.5 + δ )
2
δ := 0.5

δ
a3 := ⎛⎜
1 1 1 ⎞
a0 := a1 := a2 := ⎟−1 α = 0.25
α ⋅ Δt
2
α ⋅ Δt
1
α ⋅ Δt
1 ⎝ 2⋅ α ⎠

⎛δ⎞
a4 := ⎜ ⎟ − 1 a5 := ⎜
⎛ Δt ⎞ ⋅ ⎛ δ − 2⎞ a6 := Δt⋅ ( 1 − δ )
⎟⎜ ⎟ a7 := δ ⋅ Δt
⎝α⎠ ⎝ 2 ⎠ ⎝α ⎠

Form effective stiffness matrix: Ke := K + a0⋅ M + a1⋅ C

Triangularize effective K matrix, since n=3 (small), best to find inverse, Be := Ke− 1
(effective flexibility matrix)

Calculate response for impact exerted at one of DOFs ⎛⎜ 0 ⎟⎞

Force vector
g ( t) := ⎜ 0 ⎟
⎜ F ( t) ⎟
⎝ ⎠
numerical procedure

Find initial Acceleration (W) from EOM & include external force at t=0, i.e initial time
⋅ ( K⋅ U0 + C⋅ V0 − g ( t0) )
−1
W0 := −M t0 := 0
T
initial accel. W0 = ( 0 0 0 ) U: displacement,
V: velocity
Step by step numerical integration W: acceleration
〈〉
NumInt ( g) := u 0 ← U0
〈〉
v 0 ← V0 Set initial conditions
〈〉
w 0 ← W0
for i ∈ 1 .. Ntimes − 1
P ← g ( ti)

(
〈 〉 〈 〉 〈 〉
)〈 〉
( 〈 〉 〈 〉
P ← P + M⋅ a0⋅ u i−1 + a2⋅ v i−1 + a3⋅ w i−1 + C⋅ a1⋅ u i−1 + a4⋅ v i−1 + a5⋅ w i−1 )
〈〉
u i ← Be⋅ P
〈〉 〈〉 ( 〈 〉 )
〈 〉 〈 〉
w i ← a0⋅ u i − u i−1 − a2⋅ v i−1 − a3⋅ w i−1 set (new)
acceleration (W),
〈〉 〈 〉 〈 〉 〈〉
v i ← v i−1 + a6⋅ w i−1 + a7⋅ w i velocity (V), and
displacement (U)
return u

find numerical response z := NumInt ( g)

rows ( z) = 3 3
cols ( z) = 1.024 × 10

Extract time responses obtained from numerical integration

i := 0 .. Ntimes − 1

U1 := z0 , i
i U2 := z1 , i
i U3 := z2 , i
i

numerical procedure

Tmax_plot := .22
RESULTS of numerical response: Timpulse = 0.01
UNDAMPED NATURAL FREQS.
⎛⎜ 0.026 ⎟⎞ ⎛⎜ 38.52 ⎞⎟
T = ⎜ 0.016 ⎟ f = ⎜ 61.64 ⎟
⎜ 0.011 ⎟ ⎜ 92.996 ⎟
⎝ ⎠ ⎝ ⎠
=============================== VERY LARGE TIME STEP

0.004

U1 0.002
i

U2
i 0

U3
i
0.002

0.004
0 0.022 0.044 0.066 0.088 0.11 0.13 0.15 0.18 0.2 0.22
ti
U1
U2
U3

max( t) = 5.5 −3 Timpulse

Δt = 5.377 × 10 = 1.86
Δt

=============================== LARGE TIME STEP

0.004

U1 0.002
i

U2
i 0

U3
i
0.002

0.004
0 0.022 0.044 0.066 0.088 0.11 0.13 0.15 0.18 0.2 0.22
ti
U1
U2
U3

tlast( t) = 2.75 −3 Timpulse

Δt = 2.688 × 10 = 3.72
Δt
max( t) = 2.75
 ⎝ ⎠ ⎝ ⎠
=============================== SMALL TIME STEP

0.004

U1 0.002
i

U2
i 0

U3
i
0.002

0.004
0 0.022 0.044 0.066 0.088 0.11 0.13 0.15 0.18 0.2 0.22
ti
U1
U2
U3

max( t) = 0.275 −4 Timpulse

Δt = 2.688 × 10 = 37.199
Δt