You are on page 1of 32

CONCRETE CREEP PREDICTION:

CALIBRATION USING A KELVIN-VOIGT MODEL




CEE 499H FINAL REPORT
HONORS SENIOR PROJECT








by
Carson Baker and Prof. John Stanton


UNIVERSITY OF WASHINGTON
DEPARTMENT OF CIVIL ENGINEERING
SEATTLE, WASHINGON 98195

XX August, 2014
Baker | ii

TABLE OF CONTENTS
List of Figures .............................................................................................................................................. iii
List of Tables ............................................................................................................................................... iii
1 Introduction ........................................................................................................................................... 1
1.1 Currently Accepted Models .......................................................................................................... 1
1.2 Kelvin-Voigt Model Overview ..................................................................................................... 3
1.3 Elastic elements ............................................................................................................................ 5
1.3.1 Constant Property Behavior .................................................................................................. 5
1.3.2 Time-Varying Behavior ........................................................................................................ 5
1.4 Kelvin-Voigt Elements ................................................................................................................. 7
1.4.1 Constant Property Behavior .................................................................................................. 7
1.4.2 Time-Varying Behavior ........................................................................................................ 8
1.5 Model advantages ....................................................................................................................... 12
2 Model Calibration ............................................................................................................................... 13
2.1 Introduction ................................................................................................................................. 13
2.2 Model Capabilities ...................................................................................................................... 13
2.2.1 General Calibration ............................................................................................................. 13
2.2.2 Individual Parameter-Specific Calibration .......................................................................... 18
2.3 Reduced parameter set ................................................................................................................ 19
2.3.1 ACI Creep Coefficient - C
cu
................................................................................................ 19
2.3.2 Time Scales and ........................................................................................................ 21
3 Conclusions ......................................................................................................................................... 25
3.1 Linked parameters to reduce complexity .................................................................................... 25
3.2 Future Work ................................................................................................................................ 25
Notation (Not updated) ............................................................................................................................... 29


Baker | iii

LIST OF FIGURES
Figure 1-A: Comparison of Current Strain Models ...................................................................................... 3
Figure 1-B: Schematic Diagram of a Singe Kelvin-Voigt Link.................................................................... 4
Figure 1-C: Schematic diagram of the Kelvin-Voigt Model used in Creep Analysis ................................... 4
Figure 1-D: Effect of ACI Constant on Concrete Strength ..................................................................... 7
Figure 1-E: Effect of Constant on Kelvin-Voigt Element Property Variance ........................................ 9
Figure 1-F: Effect of Time Intervals on Strain Approximation .................................................................. 11
Figure 1-G: Effect of Time Intervals on Strain Approximation Error ........................................................ 11
Figure 2-A: An example calibration fit to the GL2000 model. ................................................................... 14
Figure 2-B: An example calibration fit to the GL2000 model. ................................................................... 14
Figure 2-C: Calibration curves for each different model to 100 time steps. ............................................... 15
Figure 2-D: Kelvin-Voigt Plots from Input Parameters specified in Table 2-B. ........................................ 17
Figure 2-E: Comparison of calibrations for Volume to Surface ratios (update with legend) ..................... 18
Figure 2-F: Calibration of ACI and Exponential Models ........................................................................... 22
Figure 2-G: Effects of time constant T on the standard exponential model ............................................. 24



LIST OF TABLES
Table 1-A: Comparison of Factors Considered by Current Creep Models ................................................... 2
Table 1-B: Values of the constant for use in Equation (1-3) ..................................................................... 6
Table 1-C: Steps used in computing Figure 1-G. ........................................................................................ 11
Table 1-D: Parameters for Each Kelvin-Voigt Link ................................................................................... 12
Table 1-E: Parameters for Elastic and Kelvin-Voigt Links ........................................................................ 12
Table 2-A: Calibrated Kelvin model parameters for various creep models with 100 time steps. ............... 16
Table 2-B: Comparison of Different Kelvin-Voigt Input Parameters......................................................... 17
Table 2-C: Comparison of Input Parameters given varying Volume to Surface Ratios ............................. 19
Table 2-D: Relationship between the time-constants and ............................................................... 24
Baker | 1

1 INTRODUCTION
Concrete girders are widely used in various applications including highway bridges and in supporting
floor slabs. An understanding of their propensity to deflect under stress is critical in determining their
suitability in design. When concrete is subjected to stress it exhibits both an immediate elastic strain and a
time-dependent creep strain. Several models exist that are able to describe concrete strain under constant
load. However the majority of applications which utilize concrete girders subject them to variable stress,
the effect of which is not reflected in currently accepted models for analyzing concrete strain.
The purpose of this report is to further the development of a (Viscoelastic Projection) model created by
Prof. John Stanton and Bill Davison of the University of Washington. The model seeks to allow a user to
input a history of varying stress patterns which, when also given appropriate material parameters, would
allow the model to more accurately predict creep strains. While the model is able to be calibrated in order
to match the given output from several current creep models, it is currently unable to predict creep strains
given information describing a concrete samples material parameters and load history.
(Model Name ideas)
UW Model
Kelvin Voigt (KV) Model
Kelvin Voigt 2014 (KV14) Model
Variable Parameter Viscous (VPV) Model
Variable Kelvin Voigt (VKV) Model
Variable Viscous (VV) Model
Viscoelastic Model
Stanton-Davison Model
Variable Viscoelastic (VV) Model
Viscoelastic Projection (VP) Model Carson preferred.
Viscoelastic Prediction (VP) Model
Davison Projection Model

1.1 CURRENTLY ACCEPTED MODELS
There exist several recognized models for predicting concrete creep deformations. ACI Report 209.2R-08
details several models including their formulations and a summary of their calibrations. The models listed
in the report include:
ACI 209R-08 (ACI Committee 209, 2008)
Bazant-Baweja B3 (Bazant and Baweja, 1995)
CEB MC90-99 (Muller and Hilsdorf, 1990)
GL2000 (Gardner and Lockwood, 2004)
In addition to the models listed in the ACI Report, the AASHTO LRFD manual provides an additional
model intended specifically for use in highway bridge girders.

Each model is derived from the assumption that a unit element of concrete is placed under constant stress
and experiences both elastic and creep strains. However the parameters that are used to determine the
Baker | 2

ultimate strain of the concrete vary between models. Table 1-A compares the effects considered across
these four models.

Table 1-A: Comparison of Factors Considered by Current Creep Models
FACTORS
MODEL
Figure 1-A
Values
ACI
209R-08
Bazant-
Baweja B3
CEB
MC90-99
GL2000 AASHTO
L
o
a
d
i
n
g

H
i
s
t
o
r
y

Age of concrete at
time of loading (days)
Considered Considered Considered Considered Considered 7
Age of concrete when
drying begins (days)
Considered Considered Considered Considered Considered 7
Age of steam curing
before loading (days)
1 to 3 days - - - 1 to 3 days N/A
Age of moist curing
before loading (days)
1 day 1 day 14 days - 1 day 7
G
e
o
m
e
t
r
y

&

E
n
v
i
r
o
n
m
e
n
t

Relative humidity, % 40 to 100 40 to 100 40 to 100 Considered 35 to 100 80
Volume-surface ratio,
(in)
Considered Considered Considered Considered Considered 2
Specimen shape - Considered - - - Unit Cube
C
o
n
c
r
e
t
e

T
y
p
e

, psi -
2,500 to
10,000
2,900 to
13,000
Up to
11,890
Up to
15,000
9350
Rate of strength gain - - - - Considered N/A
Cement type I, III I, II, III I, II, III I, II, III I, II, III III
Coarse to Fine Ratio
(coarse/total agg, %)
Considered - - - - 63.38
Aggregate to Cement
Ratio,
- 2.5 to 13.5 - - - 4.327
Water to Cement
Ratio,
-
0.35 to
0.85
- - - 0.3985
Unit Weight (pcf) Considered - - - - 150
Cement content,
pound per cubic yard
Considered
270 to
1215
- - - 660
Air Content (%) Considered - - - - 2
Slump (in) Considered - - - - 3

As current models utilize different parameters in their formulation, each predicts a different value of
strain for a given concrete specimen. This can be seen be plotting the strain vs. time resulting from a 1 ksi
load using the same parameter set for each model, as shown in Figure 1-A. The variations evident in the
curves are a result of both the differences in formulation and differences in parameter utilization.
Conclusions should not be drawn from the figure concerning which model is more conservative as each
model weighs parameters differently.
Baker | 3



Figure 1-A: Comparison of Current Strain Models
The models listed fall into two categories based on their formulation. The first category describes an
output which asymptotically reaches an ultimate creep strain; these models include the ACI, CEB, and
AASHTO models. The second category contains models which continue to strain over time to no limit,
albeit more slowly with time; these models include the B3 and GL2000 models. While the prediction of
unbounded creep strains is conservative for design use, little evidence exists to support this behavior.
A perfect model does not exist to predict creep strains, and no model is more suitable in all situations. The
irregularity of available creep data creates further uncertainty concerning the reliability of the models.
Additionally, the inability to account for variable stresses lessens the utility of the models for design use,
as the effects of additional creep strains and creep recovery due to changes in stress cannot be described.
1.2 KELVIN-VOIGT MODEL OVERVIEW
As concrete girders do not typically experience constant stress loading, it is desirable to use a model that
allows for time-varying stresses in order to more accurately compute long-term deflections. A Kelvin-
Voigt model is a viscoelastic model which may be used to account for the rheological behavior exhibited
in concrete creep. The standard model is schematically represented in Figure 1-B.
Baker | 4


Figure 1-B: Schematic Diagram of a Singe Kelvin-Voigt Link
A single Kelvin-Voigt element is comprised of a linear elastic spring and a linear viscous dashpot linked
in parallel. Multiple Kelvin-Voigt links may be added in series to allow for increased versatility of the
model as shown in Figure 1-C. Additionally an elastic spring may be added in series to account for elastic
strains.



Figure 1-C: Schematic diagram of the Kelvin-Voigt Model used in Creep Analysis
The notation used in the model represents the components of concrete creep:
where:

elastic modulus of concrete.


Kelvin unit spring elastic stiffness.


Kelvin unit dashpot viscosity.
This model may be used to describe the strains within a concrete element. The purely elastic deformations
due to the elastic modulus of concrete are represented by a single linear elastic spring, while the creep
deformations are modeled by the Kelvin-Voigt components. Immediately after a load is introduced to the
system, the dashpots do not move, and the only deformation in the model occurs in the elastic spring. If
this load remains constant over time, the dashpots deform and simulate creep strains, while the elastic
deformations remain unchanged.
Elastic Viscoelastic
Baker | 5

1.3 ELASTIC ELEMENTS
The model may be analyzed by considering the purely elastic components representing the elastic
modulus of concrete separately from the viscoelastic components which are used to determine the creep
strain.
1.3.1 Constant Property Behavior
The basic constitutive equation for a linear elastic spring is given by:

(1-1)

where:

the stress of the linear elastic spring.


the elastic modulus of a linear spring.


the strain of the linear elastic spring.


When the stress

and elastic modulus

are both known for the model, the resulting elastic strains


may be easily computed. These strains occur immediately with changing stress, and the system operates
the same regardless of tensile or compressive stresses. This formulation of the model implies that the
spring completely returns to its undeformed condition when stress is removed. This behavior however is
not observed in concrete, and thus the model has been modified to account for the time-varying properties
of the concrete.
1.3.2 Time-Varying Behavior
Concrete is a complicated composite of several materials, and does not deform in a linear elastic manner.
However an estimate of the elastic modulus is useful for determining stresses when the strains are small.
As the elastic modulus is a function of the modulus of elasticity of the aggregates as well as the cement
binder and their relative proportions, the modulus of elasticity for concrete

varies significantly
between concrete mixes. The elastic modulus may however be approximated through relation to the mean
compressive strength. The American Concrete Institute (ACI) report 209R-08 allows the modulus of
elasticity to be calculated according to the following equation:


(1-2)

where:

the secant modulus of elasticity of concrete at days after casting.


the unit weight of concrete in

the design concrete compressive strength at days after casting.


This equation defines the elastic modulus over time by correlation to the mean compressive strength over
time. According to this formulation, as the concrete continues to cure and increase in strength the concrete
grows stiffer. However while the concrete becomes more resistant to deformational change over time, this
effect does not lend the spring additional energy to return completely to its previously unstrained state if
the stresses were removed. Instead the spring stiffens in place with time as the concrete cures.
Baker | 6

The ACI 209R-08 report also provides a general equation in fractional polynomial form for computing
the mean compressive strength

at any time as:

(1-3)

where: the time after the casting of the concrete measured in days.
ACI regulated constant measured in days.

ACI regulated constant measured in days.

the design concrete compressive strength at 28 days in psi.


The constant is a function of both the type of cement used and the method of curing used. The ratio


describes the age of concrete in days when the one half of the ultimate design compression strength,


is reached. ACI 209R-08 provides recommended values for as reproduced in Table 1-B.
Table 1-B: Values of the constant for use in Equation (1-3)
Type of cement Moist-cured concrete Steam-cured concrete
I 4.0 1.0
III 2.3 0.7

Varying the value of modifies both the rate of strength gained and the ultimate concrete strength, but
does not affect the 28-day concrete strength. The effect of the value used for is displayed in Figure 1-D.

Baker | 7


Figure 1-D: Effect of ACI Constant on Concrete Strength
By substituting the value of

found through Equation (1-3) into Equation (1-2), the value of

at
any time may be determined. These equations are entirely empirical and are not based on
thermodynamic theory. Rather they are a generally accepted method for approximating a workable elastic
modulus using readily available information. The variables required to determine the elastic strains are
the mean compressive strength of the concrete at 28 days

, the unit weight of the concrete, the type of


cement used and the method of curing. Each of these are standardized quantities and are easily
determined in practice.
1.4 KELVIN-VOIGT ELEMENTS
The Kelvin-Voigt elements are used to model viscoelastic creep strains. As multiple model components
interact together to affect the creep strain, their analysis requires a more complicated formulation than
those describing the elastic strains. The classical Kelvin-Voigt model uses constant property parameters
and does not make allowance for these properties to vary with time. The behavior of the standard Kelvin-
Voigt model is discussed in Section 1.4.1, while the modified Kelvin-Voigt model with time-varying
parameters is formulated in Section 1.4.2.
1.4.1 Constant Property Behavior
At any given point in time, the strains of the two Kelvin-Voigt components

and are identical. Thus


the total strain of a single Kelvin-Voigt link is given by:

(1-4)

Baker | 8

where:

the strain of a Kelvin-Voigt link.


the strain of a Kelvin-Voigt spring.


the strain of a Kelvin-Voigt dashpot.


Additionally the total stress of a single Kelvin-Voigt link is the sum of the component stresses:

(1-5)

where:

the stress of a Kelvin-Voigt link.


the stress of a Kelvin-Voigt spring.


the stress of a Kelvin-Voigt dashpot.


The general constitutive equation for a Kelvin-Voigt link is:

(1-6)

In this standard form of the Kelvin-Voigt model, the spring stiffness

and dashpot stiffness are


constant. However this leads to behavior which causes the system to return asymptotically to its original
undeformed configuration when the load is removed. As with elastic strains, this behavior is not observed
in concrete. Thus in a fashion similar to the elastic strain formulation, the model has been modified to
emulate the time-varying properties of concrete.
1.4.2 Time-Varying Behavior
Each of the two Kelvin-Voigt components are given time-varying properties. The time dependence of the
elements

and take the same form:


) (1-7)

where: the property value (Either

or ).
the time since the casting of the concrete.
a constant parameter which controls how rapidly the Kelvin-Voigt element
properties change with time.
The selection of the Equation (1-7) for use in describing the time dependence of the parameters is
arbitrary. There exists no standard for describing

and , but this formulation is a versatile and


intuitive means of modifying the parameter to reflect initial and ultimate values. While modifications to
this formulation are possible, the viscoelastic components of the (Viscoelastic Projection) model have
been programmed in this fashion. Varying the value of modifies the rate at which a parameter changes
Baker | 9

stiffness, but does not affect the ultimate value of the parameter. The effect of the value used for is
displayed in Figure 1-E. The value for

was set to 1.0, and the value for

was set to 2.0.




Figure 1-E: Effect of Constant on Kelvin-Voigt Element Property Variance
To determine the rate of creep strain, the relation to the rate of change of the stress of a Kelvin-Voigt link
must be known. By differentiating the general constitutive equation given in Equation (1-6) and
accounting for the time-dependence of each component, the stress is found to vary with time by:

(1-8)

The

term represents the additional change in stress induced by the change in stiffness of the spring
over time. If this term is non-zero, additional spring energy arises in the system as the spring
incrementally stiffens. However in order to maintain a thermodynamically consistent system, the spring
may not acquire additional energy on its own. While the spring does become more resistant to
deformational change over time, this effect does not lend the spring additional energy to return to its
previously unstrained state. As with the spring used to describe elastic strains, the Kelvin-Voigt spring
stiffens in place with time. The

term is then set to zero, and equation (1-8) may be simplified to:

(1-9)
Baker | 10


where:

the effective stiffness, given by


It is important to note that this formulation assumes that creep deformation occurs in the same manner for
both compressive and tensile stresses, which is in a similar fashion to the elastic strain modeling
described in Section 1.3.2.
To determine the rate of strain within a Kelvin-Voigt link, it would be desirable to solve equation (1-9)
for

as a function of time. However no closed form solution exists when

and have arbitrary


forms. In order to determine the rate of strain, a numerical approach must then be utilized. The equation
may be solved incrementally over short time steps using a constant value for

but still allowing the


properties

and to vary with time. This reduces the problem to the solution of an algebraic equation:

) (1-10)

where:

the initial strain rate of the Kelvin-Voigt link at the beginning of the time interval
considered.

the rate of relaxation of the Kelvin-Voigt link


With Equation (1-10) the rate of change of strain is related to the stiffness properties

and . When
gets very large, the equation is simplified to the rate of change in stress divided by the effective stiffness.
As the time steps used to perform the numerical analysis are made shorter, the approximation approaches
a smooth curve. The effect of using different length time steps are displayed in Figure 1-F. Leaving all
other parameters unchanged, the curves are computed using varying amounts of time steps. As the
number of steps increase, the precision of the curves increases, and the plot becomes smoother. While
additional precision may be achieved with additional steps, operating with more than 1000 steps is
computationally expensive, and the precision gained is minimal. This demonstrated in Figure 1-G. While
the amount of error saved by utilizing more steps is initially significant, as the number of steps increases
the error difference between a stepped function and a truly smooth curve quickly converges to zero. The
number of steps used in Figure 1-G are given in Table 1-C.
Baker | 11


Figure 1-F: Effect of Time Intervals on Strain Approximation

Figure 1-G: Effect of Time Intervals on Strain Approximation Error
Table 1-C: Steps used in computing Figure 1-G.
Steps: 10 12 15 25 50 110 260 640 1590 3985 10000
Baker | 12

Based on the above equations, both Kelvin-Voigt components have three associated parameters which
control their behavior: an initial value, a final value, and an associated time constant. These parameters
are summarized in Table 1-D.
Table 1-D: Parameters for Each Kelvin-Voigt Link
Kelvin Spring, Kelvin Dashpot,
Initial Value


Final Value


Time Constant



When an elastic link as described in Section 1.3 is added in series with the Kelvin-Voigt components as in
the case of the Model Name, a total of eight input parameters are defined. These are provided in Table
1-E. While the Initial Value for the Elastic Spring under ACI is not truly an initial value and is instead
defined at 28 days, it may be understood as a reference point analogous to the initial value used in Kelvin-
Voigt formulation.
Table 1-E: Parameters for Elastic and Kelvin-Voigt Links
Elastic Parameters Kelvin-Voigt Parameters
Elastic Spring, Kelvin Spring, Kelvin Dashpot,
Initial Value


Final Value


Time Constant



1.5 MODEL ADVANTAGES
By formulating stress in this manner, the stress of the model at any point can be computed by summing
the strain values over a series of short time intervals in order to compute total strain at any time in the
concretes life. This methodology provides several advantages:
The stress is allowed to dynamically vary throughout the timespan of the analysis. Current
models rely on constant-stress conditions for analysis. The change in stress is accounted for by
selecting a different value for

and

at any time t. This is a much more elegant method


than, for example, superposition of several creep curves for each change in stress.
The model may be used to describe both creep gain when loads are applied and creep recovery
when loads are removed.
The model properties may be more easily related to the chemical properties of the concrete, rather
than based off of empirical data alone.
Baker | 13

2 MODEL CALIBRATION
2.1 INTRODUCTION
While the flexibility of a Kelvin-Voigt model may allow creep strains to be more accurately modeled due
to variable stresses, there is no inherent relation between the model parameters given in Table 1-D and the
material factors known to effect creep listed in Table 1-A. In order to use the Model Name to predict
creep strains, rather than describing them after the fact, it is necessary to calibrate the model parameters.
As a result of the complicated nature of creep effects, available creep data is sporadic. This makes
calibration difficult and ill-defined as there is significant uncertainty in determining the accuracy of the
predictions. However it is possible to choose stiffness values for the Kelvin spring and dashpot elements
which create plots closely mimicking the results of the standard models described in Section 1.1. As these
models are already sufficiently calibrated to available creep data, calibration of the Model Name to the
output of these standard models effectively results in calibration to creep data. Since the models account
for material factors including relative humidity and the volume to surface ratio, by generating a curve
using the Model name which adequately matches these standard models given a variety of input
parameters, it is demonstrated that the model has the capability to change its output to adjust for these
factors in a similar fashion to the standard models.
The process of calibration is a function of minimizing the error between the output of the Kelvin-Voigt
model and the output of the desired model. The function values for each timestep are computed for both
models, and their differences are recorded. The mean squared error of the models is then computed. As
the Kelvin parameters are modified, the mean squared error is varied. Through iteration, an idealized
parameter set is determined which minimizes the mean squared error between the models.
(Prof Stanton: In the MATLAB program the mean squared error is divided by the maximum strain to get
the relative error, and this is what is minimized. What difference does this make? Why not just compare
the mean squared error?)
While it is possible to choose stiffness values for the Kelvin spring and dashpot elements which create
plots closely mimicking the results of standard models, this only attests to the mathematical capabilities
of the model, and does nothing for predicting creep strains of a given a concrete element.
2.2 MODEL CAPABILITIES
In order to determine to what degree the Model Name could be made to correspond to the standard
models, calibrations were made by varying the input parameters listed in Table 1-D. To simplify the
calibration process, a single Kelvin link was first used to minimize the number of parameters required.
2.2.1 General Calibration
Calibrations were run for each standard creep model and using the same parameters as those listed in
Table 1-A. An example calibration set to the GL 2000 model found after with 1000 data points and 200
iterations is depicted in Figure 2-A and Figure 2-B.
Baker | 14


Figure 2-A: An example calibration fit to the GL2000 model.

Figure 2-B: An example calibration fit to the GL2000 model.

Baker | 15

Similar processes were performed to calibrate the Model Name to each of the other standard models, the
results of which are shown in Figure 2-C.

Figure 2-C: Calibration curves for each different model to 100 time steps.

As the amount of uncertainty in the accuracy of creep prediction well exceeds the precision to which the
Model Name is able to match the standard models, it was determined that one Kelvin link provided
sufficient flexibility based on the control of the model output.
Calibration based on minimizing error alone is unintuitive and complicated. By increasing the timesteps
available to the model, the error in strain approximation may be minimized as demonstrated in Figure
1-G. However the model parameters do not approach a single optimum set as the number of time steps
increase. Additionally, calibration can return negative values for spring stiffness and dash-pot viscosity,
which does not correlate to an easily understood material response. Results of calibration of Model Name
to the standard models given the inputs specified in Table 1-A using varying amounts of time steps are
listed in Table 2-A.

Baker | 16

Table 2-A: Calibrated Kelvin model parameters for various creep models with 100 time steps.

100 Time Steps 200 Time Steps 1000 Time Steps

Fit Model
Element
Type
Initial
Value
Ratio of
Final to
Initial
Value
Intrinsic
Time
Constant
Initial
Value
Ratio of
Final to
Initial
Value
Intrinsic
Time
Constant
Initial
Value
Ratio of
Final to
Initial
Value
Intrinsic
Time
Constant
Initial
Values
Spring
Stiffness
7000 0.3 20.0 7000 0.3 20.0 7000 0.3 20.0
Dash-Pot
Viscosity
1000 6000 900 1000 6000 900 1000 6000 900
AASHTO
Model
Spring
Stiffness
-7159.0 -1.2 2.4 2164.1 0.4 2.0 7805.2 0.5 15.8
Dash-Pot
Viscosity
92.0 93425.0 558.0 94 44666 344.0 1120.6 3443.7 486.1
ACI
Model
Spring
Stiffness
4689.7 1.3 8.7 7249.1 0.6 2.0 1277.8 2.8 10.4
Dash-Pot
Viscosity
-211.0 -65010.0 1008.0 1506.7 3848.2 578.2 888.0 11106.0 1194.0
B3
Model
Spring
Stiffness
3919.7 0.9 10.7 5229.8 0.4 11.4 2054.4 0.9 26.9
Dash-Pot
Viscosity
1096.4 5421.2 384.3 1239.8 669.1 747.6 1399.4 5831.1 838.1
CEB
Model
Spring
Stiffness
-6671.9 -1.0 2.3 12441.0 0.4 7.0 8909.0 0.4 3.2
Dash-Pot
Viscosity
-1266 -23917 1548 594 13669 622.0 1045.5 6615.9 628.1
GL2000
Model
Spring
Stiffness
8621.9 0.3 19.1 5571.0 0.3 25.5 3821.5 0.4 17.2
Dash-Pot
Viscosity
803.8 8197.1 880.4 1193.7 4343.2 1024.7 1339.2 3644.7 1113.8

As suggested by the results given above, different combinations of the input parameters may produce
numerically similar outputs. Table 2-B provides two substantially different input parameters, while Figure
2-D plots the similar outputs using these parameters.
Baker | 17

Table 2-B: Comparison of Different Kelvin-Voigt Input Parameters

100 Time Steps 200 Time Steps

Fit Model
Element
Type
Initial
Value
Ratio of
Final to
Initial
Value
Intrinsic
Time
Constant
Initial
Value
Ratio of
Final to
Initial
Value
Intrinsic
Time
Constant
GL2000
Spring
Stiffness
1202 2.7 37.0 7258 0.2 20.4
Dash-Pot
Viscosity
947.0 15995 20.4 1075.2 5410.5 1004.6

(Edit plot and table to use data from Table 1-A)

Figure 2-D: Kelvin-Voigt Plots from Input Parameters specified in Table 2-B.
As a result, when performing calibrations, multiple sets of input parameters may prove to be acceptable
given a desired level or error. It is unclear which set is a more suitable selection. Additionally, providing
the optimization algorithm with different initial values leads to a different set of final values which have
similar levels of error.
Baker | 18

(To do this, use Individual Calibration Carson and vary increment5_Ek and increment5_eta. Using
values of 1.5 for each of these produces a great fit each time, but with wildly different values, as given in
the table. However, running the model multiple times with the same starting point results in the same
answer. So at least there is some degree of consistency there.)

2.2.2 Individual Parameter-Specific Calibration
In order to use the Model Name to effectively predict the creep strains which would result from a given
stress history and mix of concrete, it must be understood how to vary the model parameters based on
material conditions. To understand how the Kelvin-Voigt parameters may relate to specific material
properties, individual material properties were varied with the intention of determining a relation with the
Kelvin parameters. One example of an investigated parameter was the volume to surface ratio. The GL
2000 model and the parameters given in Table 1-A were used, and the values for the volume to surface
ratio were set at 2 (red), 3, and 4. Calibrations to each plot were performed. The plots are given in Figure
2-E and the input parameters in Table 2-C.

Figure 2-E: Comparison of calibrations for Volume to Surface ratios (updatewith legend)

Baker | 19

Table 2-C: Comparison of Input Parameters given varying Volume to Surface Ratios

(V/S) = 2 (V/S) = 3 (V/S) = 4

Fit Model
Element
Type
Initial
Value
Ratio of
Final to
Initial
Value
Intrinsic
Time
Constant
Initial
Value
Ratio of
Final to
Initial
Value
Intrinsic
Time
Constant
Initial
Value
Ratio of
Final to
Initial
Value
Intrinsic
Time
Constant
GL 2000
Spring
Stiffness
9085.8 0.2 14.2 7319.2 0.2 24.6 8991.2 0.2 21.8
Dash-Pot
Viscosity
864.6 6828.4 903.8 1955.0 2604.7 767.5 697.0 6595.77 702.0

No logical relation between parameters could be determined, as each seemed to vary significantly and
independently. As a result it was concluded that overall the model was too flexible. While the Model
Name can match many given creep strain curves well, it outputs parameters inconsistently. The goal of
the Model Name is not to match an existing model, but to predict creep strains. If there is no level of
consistency in the model parameters selected, then it is not able to be used in prediction. It is then
apparent the model requires additional constraints be placed upon it to remove the irregularities.

2.3 REDUCED PARAMETER SET
The Model Name relies on eight parameters as compared in Table 1-E. In order to reduce the flexibility of
the model, it was determined that the parameters ought to be defined in terms of one another to limit the
possible values of the model. One effort to accomplish this was through parameter relation, whereby
model parameters were formulated based on known material properties of the concrete.
2.3.1 ACI Creep Coefficient - Ccu
Specifically, it was investigated whether the model parameters could be described based on the
commonly used creep coefficient C
cu
defined by ACI 209R-08. This term acts as the ratio of the creep
strain to the initial elastic strain as given in Equation (2-1).


(2-1)

where:

the ACI creep coefficient.


the strain induced by creep.


the initial elastic strains.


Using the creep coefficient, the ultimate strain, including both elastic and creep strains, may be described
by Equation (2-2).

(2-2)
Baker | 20


ACI provides further guidance on determining the creep coefficient as a function of time through a
fractional polynomial, and a means to adjust the coefficient to account for additional effects including
curing method and duration (moist or steam), relative humidity, member size, and volume to surface ratio.
While values for Ccu may range from 1.0 for high-strength concretes to 3.0 for low strength concretes,
the recommended average value is 2.35. It is important to note that this coefficient is only a numerical
approximation, and is far from accurately describing creep. The additional models described in Section
1.1 in large part were developed to resolve the shortcomings of the ACI C
cu
coefficient, however the
coefficient continues to be in common use.
The parameters within the (Viscoelastic Projection) model may be described through relation to the ACI
C
cu
coefficient. When a stress is applied to the model, the elastic portion deforms immediately, and an
elastic strain is determined according to Equation (1-1). This equation may be rewritten as


(2-3)

When the viscoelastic elements stop deforming, (namely when

and

are zero) a creep strain may be


determined as


(2-4)

As C
cu
is defined as the ratio between creep strain and elastic strain, the following relation is found:

)
(

(2-5)

When more than one Kelvin-Voigt link is used, the equation takes the following form:

(2-6)

It is important to note that this relation holds true regardless of applied stress, which may be constant or
variable. Additionally, this relationship may be applied even as

and

vary with time. (Questionable...


Can C
cu
be considered a meaningful term when stress is variable? If it is defined by ACI for creep under
constant stress loading, are we using this improperly?) As long as enough time is permitted to pass such
that

and

are zero, this relation is always valid.



Baker | 21

2.3.2 Time Scales and
The elastic and viscoelastic components operate on different time scales according to their formulation. It
was attempted to set the parameters on the same time scale, so that the elastic and viscoelastic
components would behave in a similar fashion. The two components are formulated differently, as the
elastic components are derived from the ACI fractional polynomial in Equation (1-3) and the viscoelastic
components are formulated from an exponential model in Equation (1-7). No closed form solution exists
to match the two models together. However it was discovered that a relation could be formulated to have
the two time functions mimic each other as close as possible. The two models have been reproduced
below in their simplest formats.
The ACI Model:
[


] (2-7)

The exponential model:

) (2-8)

Rather than calibrating the model by varying parameters until a minimized mean squared error was found,
calibration was built upon ensuring fundamental key characteristics of the curve were maintained.
Specifically, both curves were designed to pass through zero at , both curves were to pass through
1.0 at , and both curves were to approach the same asymptote at . This design ensures that
the curves can be used to fit readily available information, namely the 28 day mix properties.
As the exponential model approaches

with time, the value of

must match the asymptote from the


ACI model, namely . Determining the value for T is more complicated. The curves have different
characteristics and are not as easily compared. The exponential model does not readily pass through the
origin, but instead has an initial value at , as shown in Figure 2-F. In order to force the curve
through zero, the function was shifted to pass through

at days, and then traced backwards to


pass through zero. This produced a curve that was characteristically similar and fit available data.

Baker | 22


Figure 2-F: Calibration of ACI and Exponential Models

To achieve this numerically, it was first determined what values would send the function through zero.
(Add more here. What we are really doing is setting the curve to pass through 28 days since we are
shifting it over)

)
(2-9)

(2-10)

(

(2-11)

(

(2-12)

(2-13)

)

(2-14)

This function is true for , but we now shift the curve to the right so that this function launches at
. Thus to make the graph coincide with , we must substitute in for , as shown below.
Baker | 23

)

(2-15)

)

(2-16)

And when

is substituted for the behavior expected by the ACI model, the final equation becomes of the
form:

)

(2-17)

)

(2-18)


)

(2-19)
Or alternatively,

)
)

(2-20)

)

(2-21)

This defines a function for T which depends solely on a, thereby relating the two constants together. This
relationship ensures that for any value of a, there exists a T which causes
Baker | 24


Figure 2-G: Effects of time constant T on the standard exponential model
Values for T in relation to common values for are given in Table 2-D below:
Table 2-D: Relationship between the time-constants and
Type of
cement
Moist-cured concrete Steam-cured concrete
(days) (days)
I 4.0 14.4 1.0 8.4
III 2.3 11.2 0.7 7.6

The variable denotes the time at which the function has reached

of its final value.


In terms of the physical Kelvin-Voigt model, this would mean that the dashpot has reached more than half
of its stiffness in days, and that shortly thereafter, the viscoelastic elements are relying almost entirely
on the elastic spring for their stiffness. Unfortunately the values produced for the variable are rather low,
with most values under two weeks time. This is not intuitively what is understood about concrete creep
behavior. In order for the model to accurately account for creeps long-term effects, the dashpot must not
have reached its final position. However by relating to in this fashion, using in just
days, the dashpot has reached

of its final value. This also has unfortunate


implications for modeling creep recovery. In this case if all loads were removed at days, only
of creep would be available to be recovered.
Baker | 25

3 CONCLUSIONS
The viscoelastic model is a powerful and flexible model which may be used to model concrete creep. The
model is able to be calibrated to match all modern creep models in used in common practice. However,
the model remains unable to predict concrete creep due to complications of matching the model
parameters to material properties of the concrete. Calibrations are made difficult for two primary reasons.
Firstly, the model is too flexible as multiple combinations of model parameters can return equivalent
outputs. Secondly, the relationship between specific material properties, such as the volume to surface
ratio, to specific model parameters, such as dashpot stiffness, is not yet understood.
3.1 LINKED PARAMETERS TO REDUCE COMPLEXITY
In order to limit the flexibility of the model, relationships were investigated between parameters which
would limit the inherent flexibility by placing additional constraints on the model and aid in linking the
model parameters to material properties. To this effect, the spring stiffnesses

and

were related to
the ACI creep coefficient in the following manner.

(3-1)

It is important that this does relation does not describe directly any material property of the concrete. The
ACI creep coefficient is only an approximation tool for describing creep, and does not describe an
intrinsic property of the concrete. However, it is a commonly used method for determining creep.
The two element types of the Viscoelastic Projection model are based upon two different mathematical
models in their formulation of their time scales. The elastic spring element is based on a fractional
polynomial model according to ACI recommendations. The viscoelastic element is built on an
exponential model. These two models rely on different time scales based on the time constants and .
In order to reduce the flexibility of the model, constraints were placed on the time constants so that the
time scales would behave as similarly as possible. To this effect, the following relation was developed:

)

(3-2)

While the relationship eliminates a variable by relating the two parameters together, this constraint
produces rather small values for . This is not desirable, as larger values of allow for effects of creep
recovery due to variable stresses to be more accurately modeled.

3.2 FUTURE WORK
While the Viscoelastic Projection model has been shown to be powerful enough to describe creep strains
by matching the output of standard creep models, the model remains uncalibrated and unable to predict
creep strains given material properties of a concrete specimen. The model as currently formulated does
not produce unique outputs, rather multiple inputs can produce an equivalent result. This creates difficulty
in calibration, as assigning relationships between material properties and model parameters is unclear. In
Baker | 26

order to aid in calibration efforts, it is recommended that the number of model parameters be reduced by
placing additional constraints on the model, thereby limiting its flexibility.
As currently designed, two formulations are used for computing the time-dependence of elements; the
elastic elements are fractional polynomials, and the viscoelastic elements are exponentials. To mitigate
the complexity this introduces, it was investigated whether or not the time constants could be defined in
terms of each other. While a relationship was developed, its results are not favorable. It is thus
recommended that one formulation type be used for both elements. While ACI suggests use of a
fractional polynomial for computation of elastic strains, an exponential model may be better suited as it is
more powerful in that multiple parameters may be specified, namely

, and , rather than the single


parameter in the fractional polynomial model. However, it may be preferred to use multiple fractional
polynomials to describe creep strains, as their formulation is significantly simpler, and matching material
properties to a single time constant may be more intuitive. Additionally, as a fractional polynomial is a
gentler curve than an exponential curve, a fractional polynomial may be found to be more representative
of concrete behavior.
It is also recommended that additional data be obtained on creep recovery, as this may influence current
understandings of creep behavior and provide guidance for what model parameters are appropriate.
Once the model is refined to the point at which relationships between material properties and model
parameters may be more easily developed, it is recommended that efforts be made to establish them. This
may be achieved in several ways. It may be appropriate that each material property effects a single model
parameter. This would be most likely in the case of multiple fractional polynomial links where each link
could more easily describe a material property independently. In the case of exponential links, it would be
more difficult to assign a single material property to each link as three variables would need to be defined,
and the complexity of the model would grow significantly with each link. Alternatively, if exponential
links are preferred, a series of relationships could be developed describing how multiple material
properties effect a single model parameter. This is most similar to current creep models, where a series of
computations based on numerous inputs defines a single variable.









Baker | 27


Creep Recovery:



10
0
10
1
10
2
10
3
10
4
0
1
2
3
x 10
-4
Creep Recovery Curve Fits
Time (days)
S
t
r
a
i
n


AASTO/Yue Model Prediciton
Kelvin Model Fit
Baker | 28




WEBSITE
Photo of creep rig
Kelvin model basic explanation. No equations, just dashpot stuff.
Plot how E elastic from ACI and the Ek exponential model can be overlayed to look pretty close.
Show a and b formula.
Also show some plots of matching ACI curves. I think I should show how we changed the GL
2000 model and the Kelvin model was able to match it. I have a good picture of that.



Baker | 29

NOTATION (NOT UPDATED)

= elastic modulus of concrete

= the effective stiffness, given by

= Kelvin unit spring elastic stiffness


= the time since the casting of the concrete.
T = a constant which controls how rapidly the Kelvin-Voigt element properties change with time.

= the strain of a single Kelvin-Voigt link

= the strain of a Kelvin-Voigt spring

= the strain of a Kelvin-Voigt dashpot

= the initial strain of the Kelvin link


= Kelvin unit dashpot viscosity

= the stress of a Kelvin-Voigt link

= the stress of a Kelvin-Voigt spring

= the stress of a Kelvin-Voigt dashpot


the decay time of the unit

You might also like