SiSc Laboratory Winter Semester 16/17

Sensitivity analysis of a Viscoelastic Material

Kishan Gireesan Rao & Jie Ruan

M.Sc. Simulation Sciences, RWTH Aachen university

Supervisor: Dipl.-Ing. Johannes Neumann

Institute of Applied Mechanics, RWTH Aachen university
Content
Abstract................................................................................................................. 1
Introduction ........................................................................................................... 1
Methods ................................................................................................................ 2
Results .................................................................................................................. 4
Conclusion .......................................................................................................... 14
Reference ........................................................................................................... 15
Abstract

This project aims at parameter sensitivity analysis of bituminous mortar, a typical

viscoelastic material widely used on German roads. For the constitutive law of this
material, we choose the generalized Maxwell model. A large number of
parameters conduct some difficulty for the parameters identification, then we use
parameters sensitivity analysis to obtain the effects of parameters on dynamic
modulus. Automatic differentiation methods are applied to get the derivative of
dynamic modulus to parameters, then we can analyse the parameter sensitivity.

Introduction

Fig 1. Viscoelastic model

One typical characteristic of viscoelastic material is that it shows both viscous and
elastic when deforming, thus we can describe the mathematic model of
viscoelastic material as Fig 1. The spring models elasticity while the dashpot
models viscosity.

Fig 2. Generalized Maxwell model

The generalized Maxwell model is shown in Fig 2, which is a number of Maxwell

model connected in parallel. The Maxwell model can be represented by a purely
viscous damper to model fluid and a purely elastic spring to model solid
connected in series. Similar to Hooke's Law, the relational expression of stress

1
and strain for elastic component is   E , where  is the stress, E is the
elastic modulus and  is the strain. Meanwhile, the stress-strain rate relationship
of viscous components obeys Newton`s law and can be given as    , where
 is viscosity coefficient and  is strain rate. Since for Maxwell model the spring
and dashpot is connected in series, which means the stress of elastic component
and viscous component are same while the strain of them are additive. Then we
can write the relation of stress and strain for viscoelastic material using Maxwell
model as:
 
   (1)
 E

Methods

Since Hooke`s law and Newton`s law both follow linear superposition, the
generalized Maxwell model also follows linear superposition, which means the
effectiveness of the Maxwell model is additive w.r.t its arms. Thus, we can simplify
this model as one single Maxwell model when doing parameters sensitivity
analysis. Besides, the generalized Maxwell model can also discrete relaxation
time spectrum by a series of discrete relaxation time sequence.

We use dynamic modulus, the modulus value of G, to represent the relations

between the stress and strain. The complex number G is defined as G  G 'iG ' ' ,
where G ' is storage modulus and G ' ' is loss modulus. The expression of G '
and G ' ' in generalized Maxwell model can be written as:
N
( ii ) 2
G ' ( )   g i (2)
i 1 1  ( ii ) 2
N
 i i
G' ' ( )   g i (3)
i 1 1  ( ii ) 2
Where N is the number of Maxwell models in parallel, g is the spring stiffness,

 is relaxation time (   ) and  is frequency.
E

Fig 3. Stress and strain for viscoelastic material

2
We can find that purely elastic materials have stress and strain in phase, while in
purely viscous materials strain lags stress by a phase lag. Viscoelastic materials
show both behaviors, which means exhibiting some lag in strain. Fig 3 shows the
definition of stress and strain. X axis is time and y axis is stress and strain,  0
and  0 means the amplitude of stress and strain respectively,  is the phase
2
angel. The period of stress and strain are both . Then the formulas

of stress and strain can be defined as:
   0 sin( t ) (4)
   0 sin( t   ) (5)

Fig 4. G in complex plane

Since there exist a phase lag between storage modulus and loss modulus, we
choose complex plane to express G as Fig 4. Dynamic modulus is defined as
below:

| G | G0  0  (G ' ) 2  (G ' ' ) 2 (6)
0
G'  G 0 cos  (7)
G' '  G 0 sin  (8)

Fig 5. Linear fit complex modulus

3
Fig 5 shows the dynamic modulus data from experiments and linear fit by formula.
The crosses represent experimental data, red line represents dynamic modulus
by linear fit, blue dashed line represents loss modulus and green dashed line
represents storage modulus. Both X axis and Y axis are in log scale. From the
definition formula of loss modulus, we can find that, the peak value of loss
modulus appears when relaxation time  multiply by frequency  equals to 1,
that means relaxation time is inversely proportional to characteristic frequency.
For Maxwell model with high characteristic frequency, the material has low
relaxation time. However, for Maxwell model with low characteristic frequency, the
material has high relaxation time.

In order to perform the parameter sensitivity analysis, we need to get the partial
derivatives of the dynamic, storage and loss moduli with respect to g and  .
Since classical methods are slow for partial derivatives with respect to many
inputs, we choose automatic differentiation method using the DCO toolbox. This
method arrives at derivatives by the use of chain rule based on machine precision,
and using optimization to cheap gradients result in low computation cost.

Furthermore, the values of the derivatives with respect to a particular parameter

shows the extent of influence that particular parameter has on the result. Using
this knowledge and the DCO toolbox, the following graphs have been obtained
which are used for our analysis.

Lastly, the derivatives have been obtained for only one Maxwell arm and the
analysis is done for the parameters. Since a Visco-elastic material can essentially
be modelled as a number of Maxwell arms connected in parallel, we can say that
the behavior of the resulting material is the additive effect of the individual
Maxwell arms.

Results

The following graphs were obtained on a linear space grid with values of g and
 varying from 0 to 10, i.e.:
0 ≤ g ≤ 10
0 ≤  ≤ 10

The attached source code which uses the DCO toolbox in order to calculate the
derivatives was modified accordingly to obtain the required values of the
derivatives and the values of Storage and Loss Moduli. The derivatives were
obtained for an omega value of 0.5 in all the cases.
ω = 0.5

These values were then plotted using the surface plot in MATLAB.

4
Storage Modulus:

Fig 6. Storage Modulus vs. Spring stiffness and Relaxation time

Figure 6 shows the variation of the storage modulus as g and  vary in a

3-Dimensional graph on a linear space. We see that the storage modulus
increases asymptotically as the value of relaxation time increases but increases
linearly with the increase in the value of spring stiffness. This also shows that the
increase in value of omega has an influence on the value of the storage modulus.

Fig 7. Storage Modulus vs. Relaxation time Fig 8. Storage Modulus vs. Spring stiffness

5
Influence of the parameters on the Storage Modulus:
When the derivative of the Storage modulus is taken with respect to the relaxation
time, we obtain:

Fig 9. Derivative of Storage Modulus w.r.t  vs. Spring stiffness and Relaxation time

6
 Relaxation time (  )
Figure 9 shows the 3-Dimensional view of the derivative of the Storage modulus
with respect to the relaxation time. We observe that the peak is reached very
close to the numerical value of omega. This shows that the influence of the
relaxation time is strongest when it’s value is close to the value of induced
frequency. We can also observe that the spring stiffness only induces a linear
scaling and thus does not have a deep impact on the Storage modulus.

Spring Stiffness ( g )

When the derivative of the Storage modulus is taken with respect to the spring
stiffness, we obtain:

We immediately observe that when the derivative is taken with respect to the
spring stiffness, the g on the R.H.S Vanishes and it becomes independent of the
value of spring stiffness. This is also confirmed in the following graphs shown in
figure 10.

7
Fig 10. Derivative of Storage Modulus w.r.t g vs. Spring stiffness and Relaxation time

Relaxation time (  ) Spring Stiffness ( g )

8
Loss Modulus:

Fig 11. Loss Modulus vs. Spring stiffness and Relaxation time

Figure 11 shows the variation of the Loss modulus as g and  vary in a

3-Dimensional graph on a linear space. We see that the storage modulus peaks
at certain values with increase in the value of relaxation time increases but
increases linearly with the increase in the value of spring stiffness. This also
shows that the increase in value of omega has an influence on the value of the
storage modulus.

Fig 12. Loss Modulus vs. Relaxation time Fig 13. Loss Modulus vs. Spring stiffness

9
Influence of the parameters on the Loss Modulus:
When the derivative of the Loss modulus is taken with respect to the relaxation
time, we obtain:

Fig 14. Derivative of Loss Modulus w.r.t  vs. Spring stiffness and Relaxation time

GL
oss

Relaxation time (  )

10
Figure 14 shows the 3-Dimensional view of the derivative of the Loss modulus
with respect to the relaxation time. We observe that the derivative cuts the X-axis
at a value that is very close to the numerical value of the reciprocal of omega. This
shows that the influence of the relaxation time is strongest when it’s value is close
to the value of induced frequency. From figure we can also observe that the spring
stiffness only induces a linear scaling and thus does not have a deep impact on
the Loss modulus.

Spring Stiffness ( g )

When the derivative of the Loss modulus is taken with respect to the spring
stiffness, we obtain:

Just like the in the case of the storage modulus, we immediately observe that
when the derivative is taken with respect to the spring stiffness, the g on the
R.H.S Vanishes and it becomes independent of the value of spring stiffness. This
is also confirmed in the following graphs shown in figure.

11
Fig 15. Derivative of Storage Modulus w.r.t g vs. Spring stiffness and Relaxation time

Relaxation time (  ) Spring Stiffness ( g )

12
In order to see the influence of the parameters – Spring stiffness and the
relaxation time when they both vary, a superposition can be made by calculating
the norm of the partial differentials.

Norm of the partial differentials of Storage modulus:

2 2
] +[ ] }

Relaxation time (  )

Fig 16. Norm of the partial differentials of

Storage Modulus w.r.t g vs. Spring
stiffness and Relaxation time

Spring Stiffness ( g )

Norm of the partial differentials of Loss modulus:

2 2
] +[ ] }

Relaxation time (  )

Fig 17. Norm of the partial differentials

of Loss Modulus w.r.t g vs. Spring
stiffness and Relaxation time

Spring Stiffness ( g )

13
Norm of the partial differentials of Storage modulus:
2 2
] +[ ] }

Relaxation time (  )

Fig 18. Norm of the partial differentials

of Dynamic Modulus w.r.t g vs.
Spring stiffness and Relaxation time

Spring Stiffness ( g )

In the figures 16-18, we observe the same trends as in were in the individual
components of the Dynamic modulus. Thus we can conclude the analysis in the
following section.

Conclusion

 Influence of the spring stiffness on the dynamic modulus is purely linear in

nature. Therefore, it is always predictable and error in measurement of
data has the same effect at any value in the spectrum.
 Influence of the relaxation time is larger close to the characteristic
frequency (induced frequency) and reduces drastically at values further
away from it. This means that any error in the measurement of values close
to the characteristic frequency leads to a larger error in the measurement
of the dynamic modulus.
 This analysis can further be extended to a visco-elastic material comprising
of a number of Maxwell arms with the influence of the parameters being
additive in nature.

14
References

 https://en.wikipedia.org/wiki/Viscoelasticity
 https://www.nag.com/content/nag-and-algorithmic-differentiation
 “Multiscale modelling of Asphalt concrete”, Johannes Neumann,
Jaan-Willem Simon, Stefanie Reese, ECCOMAS Congress 2016, Crete
(Greece), June 9, 2016
 “The Art of Differentiating Computer Programs, An Introduction to
Algortihmic Differentiation”, Number 24 in Software, Environments and
Tools, SIAM 2012.

15