Improved Algorithm For Efficient and Realistic Creep Analysis of Large Creep-Sensitive Concrete Structures
Improved Algorithm For Efficient and Realistic Creep Analysis of Large Creep-Sensitive Concrete Structures
Recent compilation of data on numerous large-span prestressed creep effects were also documented for four large segmen-
segmentally erected box girder bridges revealed gross underesti- tally erected box girders in Japan.3,6,7 A subsequent search of
mation of their multi-decade deflections. The main cause has been data on segmentally erected prestressed box girder bridges,
identified as incorrect and obsolete creep prediction models in undertaken under the auspices of the recently established
various existing standard recommendations and is being addressed RILEM Committee TC-MDC (Multi-Decade Creep), led to
in a separate study. However, previous analyses of the excessive
a list of 69 large spans (in 10 countries) that suffered grossly
deflections of the Koror-Babeldaob (KB) Bridge in Palau and of
four Japanese bridges have shown that a more accurate method excessive long-time deflections.8-10 It is likely that many
of multi-decade creep analysis is required. The objective of this more cases of excessive deflections exist worldwide. Clearly,
paper is to provide a systematic and comprehensive presentation, the problem is widespread and its cause must be systematic.
appropriate not only for bridges but also for any large creep- Analyses of the data on the KB Bridge and the Japanese
sensitive structure. For each time step, the solution is reduced to bridges3,6,7 with Model B3 (calibrated by multi-decade labo-
an elastic structural analysis with generally orthotropic elastic ratory tests11-13) showed that the main causes of the gross
moduli and eigenstrains. This analysis should normally be three- underestimation are twofold:
dimensional (3-D). It can be accomplished with a commercial finite 1. Incorrect and obsolete material models for concrete
element code such as ABAQUS. Based on the Kelvin chain model, creep surviving in standard design recommendations of ACI,
the integral-type creep law is converted to a rate-type form with CEB-fib, and other engineering societies; and
internal variables, which account for the previous history. For time 2. An oversimplified method of creep structural analysis.
steps short enough to render aging during each step to be negli-
gible, a unique continuous retardation spectrum for each step is
The former is being addressed in a separate study.6-8 This
obtained by Laplace transform inversion using simple Widder’s article will focus attention on the latter. Assembling previ-
formula. Discretization of the spectrum then yields the current ously published results, it will give a comprehensive presen-
Kelvin chain moduli. The rate-type creep analysis is computation- tation of the method of multi-decade creep analysis, which
ally more efficient than the classical integral-type analysis. More can use a general-purpose finite element code such as
importantly, though, it makes it possible to take into account the ABAQUS and has led to excellent agreement with the obser-
evolution of various inelastic and nonlinear phenomena such as vations on the KB Bridge and four Japanese bridges.
tensile cracking, cyclic creep, and stress relaxation in prestressing For the purpose of creep analysis, the prestressed concrete
tendons at variable strain, as well as the effects of humidity and box girders have traditionally been analyzed as one-dimen-
temperature variations, and the effect of wall thickness variation sional (1-D) beams, sometimes using, for the top slab,
on drying creep and shrinkage. Finally, the advantages compared approximate correction formulas for the shear lag.14-16 Such
to the existing commercial programs, based on step-by-step inte-
analysis, however, still popular in commercial design soft-
gration of memory integrals, are pointed out and illustrated by a
simple example.
ware, can have large errors, mainly for two reasons:
1. It cannot realistically capture the shear lag effects,
Keywords: aging; bridge deflections; commercial programs; continuous which are not only elastic but also aging viscoelastic, occur
retardation spectrum; cracking; creep and shrinkage; cyclic creep; Kelvin not only in the top slab but also in the vertical walls and
chain model; rate-type algorithm; steel relaxation; three-dimensional the bottom slab, and are caused by both the vertical shear
analysis; variable environment. forces at the piers and by the concentrated loads from tendon
anchors; and
INTRODUCTION 2. It cannot account for the differences in drying creep
The severity of the effects of multi-decade creep of properties and in shrinkage caused by the differences in the
concrete has long been underestimated. In 2008, however, drying rates of slabs, resulting from different thicknesses
the release of detailed technical data on the Koror-Babeldaob and environmental exposure.
(KB) Bridge in Palau, which collapsed in 19961,2 as a result Properly, the box girders should be analyzed as thick
of a retrofit, provided a wakeup call. When built in 1977, this shells, for which a three-dimensional (3-D) finite element
prestressed segmentally erected box girder had the world- analysis is necessary. Such analysis is, of course, also
record span of 241 m (791 ft). Within 18 years, its midspan necessary for other creep-sensitive structures such as super-
deflection (compared to the design camber) reached 1.61 m tall buildings.
(5.3 ft), which was approximately three times larger than
the deflection calculated in design.3 The 18-year prestress
loss was measured to be approximately 50% (as an average ACI Structural Journal, V. 109, No. 5, September-October 2012.
of nine readings), which was more than the double of the MS No. S-2010-366.R2 received September 13, 2011, and reviewed under Institute
publication policies. Copyright © 2012, American Concrete Institute. All rights
normally predicted loss4,5 (refer to Fig. 1). reserved, including the making of copies unless permission is obtained from the
Thanks to other data released in 2008 by Shimizu copyright proprietors. Pertinent discussion including author’s closure, if any, will be
published in the July-August 2013 ACI Structural Journal if the discussion is received
Construction, Tokyo, similar underestimations of long-term by March 1, 2013.
Fig. 1—Excessive deflection recorded in KB Bridge in Palau. (Note: 1 m = 3.28 ft; 1 MPa
= 145 psi.)
In real design of large creep-sensitive structures, the inte- the nonlinear phenomena, the rate-type stress-strain rela-
gral-type creep law arising from the principle of superposi- tion ceases to obey the principle of superposition. A further
tion must be converted to an equivalent rate-type creep law highly nonlinear phenomenon is the stress relaxation of steel
with internal variables whose current values account for the tendons, which is not viscoelastic. The evolution of steel
previous history of viscoelastic strain. There are basically relaxation may be significantly influenced by the strain vari-
two reasons for choosing the rate-type approach: ation in concrete and thus needs to be calculated as part of
1. One is computational: The strain history need not be structural creep analysis to take into account the time varia-
stored and no sums over all the previous time steps need to be tion of prestress loss due to the combined effects of concrete
computed to approximate the history integrals. Thus, the rate- creep, shrinkage, and steel relaxation (note that the usual
type approach makes creep computations far more efficient. assumption of either a time-constant prestress loss or time-
For the design of very large structures, the rate-type form of variable steel relaxation estimated in advance is generally
creep law is almost inevitable. It also makes it easier to simu- insufficient for creep-sensitive structures).
late the effects of numerous changes in the structural system, The key property enabling the rate-type analysis is the fact
as in segmental construction with sequential prestressing of that any realistic integral-type stress-strain relation of aging
many segments or in super-tall building erection. viscoelasticity can be approximated with any desired accu-
2. Another is the necessity to capture the phenomena racy by a rate-type creep law visualized by the Kelvin chain
causing deviations from the principle of superposition in model. This model consists of a series of Kelvin units m =
time.19-21 These phenomena include the cracking and other 1, 2, 3, ···, N (Fig. 2), each of which involves a spring of
damage of concrete, variations of humidity and tempera- stiffness Em(t) coupled in parallel with a dashpot of viscosity
ture with the corresponding changes in the rate of aging (or hm(t) = Em(t)tm, where tm are the suitably chosen retardation
hydration), bond slip, and cyclic creep. Upon inclusion of times. In step-by-step analysis, one takes advantage of the
( ) (1 – 2n)/(2(1 – n)).
m
L tm = − k →∞
(1)
( k − 1)!
1 n / (1 − n) n / (1 − n) 0 0 0
where C(k) is the k-th order derivative on time t of the creep 1 n / (1 − n) 0 0 0
part C(t, tn – 1/2) of the compliance function J(t, tn – 1/2); C(t, t′) 1− n 1 0 0 0
D= (5)
= J(t, t′) – 1/E0; and E0 is the instantaneous (or short-time)
elastic modulus. Note that, generally, L(tm) is different for
(1 + n)(1 − 2 n) n* 0 0
n* 0
each integration point of each finite element in each time
step. The effects of thickness tb, current local humidity h, n*
current local temperature T, and so on are introduced into the
spectrum through the compliance curve J(t, t′), in which t′ is
constant and represents the age of concrete in the middle of The usual algorithms for numerical integration of first-
the current time step. order ordinary differential equations, such as the central or
In practice, the limit in Eq. (1) need not be calculated backward difference methods or the Runge-Kutta method,
and it suffices to use k = 3. A discrete approximation of the fail by numerical instability. The reason is that these methods
continuous spectrum gives the discrete spectrum are stable only if Dt << t1 while, to reach long times, the
time steps Dt need to be increased to values larger (in fact,
( )
A tm =
1
( ) ( )
= L t m ln10 D log t m = L t m ln10 (2)( ) many orders of magnitude larger) than the shortest retarda-
( )
E tm tion time tm = t1. An unconditionally stable algorithm, called
the exponential algorithm, was devised to overcome this
which corresponds to a finite (or discrete) Kelvin chain and problem.20,23,29,30 In this algorithm, one calculates (for each
is required for numerical computations. integration point of each finite element, in each time step)
Note that, in the original rate-type creep analysis,29 the
discrete spectrum A(tm) was calculated by least-square fitting bm = e
− Dt / t m
( )
, l m = t m 1 − b m / Dt (6)
( ) ( )
Nb
NUMERICAL IMPLEMENTATION AND ALGORITHM De 1′′ = ∑ 1 / t n −1/ 2 + q3 / q2 × 1 − b m g m( n −1)
All the creep and shrinkage models predict the average (or m =1
(15)
effective) properties of creep and shrinkage over the thick-
( )
Nd
( n −1)
ness of a slab of the cross section directly from the envi- + ∑ 1 − bm g m
m =1
ronmental humidity, taking into account the slab thickness.
This simplification is quite poor for cross sections subject
to pure bending, partly because the creep prediction model
has been calibrated by the tests under axial load. However, De 2′′ = q4s ( n −1) Dt / t n −1/ 2 (16)
in flanged cross sections such as those of box girders, the
distribution of longitudinal normal stress across the wall is
nearly uniform—that is, the bending component within the where Ab(tmb) is the discretized spectrum for the non-aging
wall alone is relatively small and the force resultant is nearly constituent in Model B3; e″1 is the 6 × 1 column matrix
centric within the wall thickness. In that case, the aforemen- (vector) of the inelastic strain increments in the current
tioned simplification works much better. It suffices to use volume fraction of solids; and e″2 is the 6 × 1 column matrix
only one finite element through the slab thickness, except of inelastic strain increment of viscous flow.
in the top slab because that slab is subjected to transverse More realistic is an analysis in which the relative humidity
bending from roadway loads. hp in the capillary pores of concrete at various points and
A computer program based on the algorithm described in times is computed from the nonlinear diffusion equation for
the following was written and applied to the KB Bridge in concrete drying,38,39 for which a 1-D version for transverse
Palau.3,6,7 This program (which can be freely downloaded from moisture transport is generally sufficient. This approach
www.civil.northwestern.edu/people/Bažant/PDFs/Papers) further has the advantage that the creep and shrinkage law
can be adapted to other segmental bridges, as well as other for a material point is simpler and more accurate than it is for
types of concrete structures. On a state-of-the-art personal the cross-section average.
computer (PC) (for example, a quad-core, 4 GB memory), Model B3 includes an alternative that allows such point-
the entire 3-D finite element creep analysis of the KB Bridge wise calculation of creep and shrinkage. In this case, the
ran for approximately 50 minutes. The flowchart of the algo- thickness of the slab needs to be subdivided into at least
rithm used in this program is shown in Fig. 2. six finite elements, which increases the number of all finite
Note that the number N of internal variables could actu- elements in the structure approximately six times. The inte-
ally be reduced to approximately five, but then the first L(tm) gration of the diffusion equation for drying (and wetting) of
(m = 1) would have to be computed as the integrated area concrete38 is done numerically, simultaneously with creep
under the spectrum up to –∞ in the log-time scale (the reason structural analysis. This more refined approach is a straight-
forward generalization of the present approach and will not
is that Kelvin units with tm << Dt behave as springs, and the
be discussed further.
compliances of springs coupled in series can be combined
into one compliance). Using N = 22 for Model B3 and N = PRESTRESSING STEEL RELAXATION AT
13 for other models increases the demand on computer time VARIABLE STRAIN
and storage but is simpler to program, which is important So far, the practice has been to calculate the stress relax-
for the user. The spectral values L(tm) for tm >> t (the current ation in prestressing tendons from simple formulas that give
f ( t , t ′ ) = 2.35g c
(t − t ′)0.6 and e = 0.000129 for t = 10,000 days. The ACI formula gives
(18) almost the same results; refer to Fig. 3.
10 + ( t − t ′ )
0.6
the creep is modeled by linear aging viscoelasticity, which ACI Committee 209 and since 1990 by the CEB Model
is implemented in the primitive form of an integral-type Code, yet no update has been made in SOFiSTiK (nor
creep law (as introduced for nuclear reactor structures RSTAB/RFEM or InfoCAD).
before 1970). The memory integrals are numerically inte- By virtue of taking the aging into account, the errors of
grated step by step from the complete stress history of each the AAEM for loads applied at a young age (which is a
beam element, which increases the demands on computer typical situation for box girder segments) are significantly
time and storage. This fact, and the need to investigate many smaller, compared to the exact linear viscoelastic solutions,
load cases and combinations, are probably the reasons why, than the errors of the Trost method.24,52 The AAEM often
despite the availability of two-dimensional (2-D) and 3-D gives surprisingly good estimates of the linear aging visco-
finite elements, SOFiSTiK does not use the memory inte- elastic solutions in simple beam structures. However, even
grals in combination with these elements. Rather, in the case if SOFiSTiK switched to AAEM, much larger errors due to
of 2-D or 3-D elements, SOFiSTiK uses a simple quasi- the use of quasi-elastic analysis for multi-dimensional finite
elastic algebraic analysis based on a one-step incremental elements and of pure linear viscoelasticity for 1-D beam
elastic relation of the following uniaxial form elements would still remain.
Although the major source of error in a program such as
SOFiSTiK is the embedded CEB creep model, one must
s ( t ) − s (t1 ) = E ′′ (t , t1 ) e ( t ) − f (t , t1 ) e (t1 ) − e sh (t , t1 ) (20) accept this model to isolate other errors through numerical
examples. The first example is a freestanding plain concrete
column under uniaxial compression. The relevant concrete
where s, e is axial stress and strain; t1 is the age of concrete properties are fc28 = 30 MPa (4350 psi) (which gives E28 =
at load application; t is current age; e(t1) is initial elastic 31,008 MPa [4496.2 ksi] according to the CEB formula) and
strain at t1; f(t, t1) and esh are creep coefficient and shrinkage an environmental humidity h = 70%. The first compressive
strain, respectively, as specified by the 1990 CEB Model stress s (1) = 1 MPa (145 psi) is applied at t0 = 7 days and
subsequently additional compressive stresses s (i) = 1 MPa
Code; and E″(t, t1) is the incremental Young’s modulus for
(145 psi) (i = 2, 3, ···, 12) are applied every 7 days. It is
the entire period from t1 to t. According to the 1967 Trost
found that, in this example, the integral-type algorithm of
method,50 E″ = E28/(1 + rf), where r is Trost’s so-called
SOFiSTiK predicts the creep deformations accurately
relaxation coefficient, typically taken as 0.8.
enough. The deviations from the exact analytical solution
But r and E28 do not take into account the creep aging.
and the present algorithm are barely distinguishable (Fig. 3).
This effect can be captured by the age-adjusted effective The second example is the ill-fated KB Bridge in Palau.
modulus (AAEM) method,20,23,24,51 in which E28 is replaced The information about this bridge can be found in investiga-
with E(t) and r with aging coefficient c, as shown in 1972 by tion reports and recent studies.1-3,6,7,18,53 To explain the unex-
Bažant.51 The AAEM has been endorsed since 1982 by pected huge deflection, a comprehensive investigation of this
bridge was carried out at Northwestern University3,6,7; refer and shrinkage analysis, the age at start of drying is taken as
also to Fig. 4. By applying the present rate-type algorithm to t0 = 7 days, which is also the assumed segmental erection
3-D analysis, the long-term deflections have been computed cycle. The average environmental humidity is h = 0.70. In
for different creep and shrinkage models, as plotted in each erection cycle, one segment, including its dead load, is
Fig. 4 with respect to the end of construction. As seen in added and prestressed at the end of the 7-day period.
the figure, such analysis based on Model B3 can match the The linear viscoelastic analysis of the KB Bridge in
measured deflection very well, provided the creep parame- SOFiSTiK (modeled as beam) predicts substantially less
ters are calibrated by the long-term creep tests of Brooks11,12; deflection than the recorded measurements, which are
refer to B3 (Set 2) in Fig. 4. represented by the diamonds in Fig. 4. Furthermore, when
Unlike the rate-type analysis with ABAQUS, SOFiSTiK compared with the rate-type 3-D analysis, which uses the
uses a simplified approach with 45 nodes and 44 beam same CEB model, SOFiSTiK also substantially underpre-
elements corresponding to the segmental construction2,3,6,7; dicts the creep deflection. The integral-type algorithm used
refer to Fig. 4. The prestressing system is modeled in this calculation gives a deflection of only approximately
by 44 tendon groups consisting of six to 40 individual 32 mm 79% of the value predicted by the CEB Model Code when
(1.25 in.) diameter threaded alloy bars with 1030 MPa applied in the ABAQUS 3-D analysis, which itself underes-
(150 ksi) nominal tensile strength, which are assumed to timates the observed creep deflections by a factor of approxi-
be located in the center of the top slab. The jacking force mately 1/3; refer to Fig. 4.
Even if the viscoelastic analysis and finite element simu-
for each tendon is 600 kN (135 kips) with 5% of initial
lation were perfect, SOFiSTiK and similar design software
prestress loss. According to the design specifications, the
programs cannot correctly simulate the nonlinear evolution
28-day concrete properties are fc′ = 35.9 MPa (5207 psi) of viscoplastic steel relaxation at variable strain, the varia-
and Ec = 28.3 GPa (4105 ksi). The pier and foundation are tions of temperature and humidity, and the nonlinear effects
substituted by an idealized vertical spring of kz = 1.42 × of cracking and cyclic loading (although a correction for
108 kN/m (9.73 × 106 kip/ft) and a rotational spring of km,y cracking can, of course, be made in a time step; this would
= 1.42 × 109 kN-m/rad (1.05 × 109 kip-ft/rad). The CEB be incorrect in combination with memory integrals for creep
Model 199054 is used not only for concrete creep but also for because the cracking has no delayed memory). The effects
steel relaxation. The steel relaxation evolution is handled in of wall thickness differences among the top slab, walls, and
SOFiSTiK either by a fixed 1000-hour relaxation factor or, bottom slab on the drying creep and shrinkage can only be
as chosen in this example, by the stress-dependent quadratic captured by more refined models using a rate-type creep law
function according to the CEB Model Code 1990,54 inter- and 2-D or 3-D finite elements.
polated from three points: 2% at 0.60fpk, 4% at 0.70fpk,
SUMMARY AND CONCLUSIONS
and 6.67% at 0.80fpk, where fpk is the characteristic tensile 1. Although many ingredients of the present algorithm
strength of the prestressing tendons. Although the actual have been developed separately—some of them long
calculations in SOFiSTiK are a black box, the specified ago19,22,24,27,29-31,33,34,55—this paper combines them into one
quadratic CEB relaxation function in itself is not a good comprehensive algorithm, gives the previously missing
approximation for multi-decade relaxation.40 For the creep details needed for programming, and incorporates into the
lς (1+1/ c)
(A2)
44. ACI Committee 209, “Prediction of Creep, Shrinkage and Temperature
Effects in Concrete Structures,” Designing for Effects of Creep, Shrinkage
and Temperature in Concrete Structures, SP-27, American Concrete Insti- where
tute, Farmington Hills, MI, 1971 (Reapproved 2008), pp. 51-93.
45. ACI Committee 209, “Guide for Modeling and Calculating Shrinkage
and Creep in Hardened Concrete (ACI 209.2R-08),” American Concrete F ( e ) − g f y′
Institute, Farmington Hills, MI, 2008, 45 pp. ς= , r = r0 ehx (A3)
46. SOFiSTiK AG, “AQB Design of Cross Sections and of Prestressed s p − g f y′
Concrete and Composite Cross Sections v13.64,” Software Manual, 2010.
47. SCIA, “SCIA Engineer. Reference Guide v2010.1,” Software
Manual, 2010, www.scia-online.com. (last accessed Apr. 26, 2011) Herein, k, c, r0, and h are positive empirical constants for
48. Infograph GmbH, “InfoCAD 10.4 Manual,” Software Manual, 2010,
www.infograph.eu. (last accessed Apr. 26, 2011) the given steel; and r0 is an empirical function of e(t). For oil-
49. Dlubal Engineering Software, “RSTAB 7. Structural Analysis for tempered wire (OT Series in Magura et al.41), r0 = 0.34 and
General Frameworks. Program Description,” Software Manual, 1999,
www.dlubal.com. (last accessed Apr. 26, 2011) h = 0.01. At constant strain, F(e) = s0 = initial prestress; T(t)
50. Trost, H., “Auswirkungen des Superpositionsprinzip auf Kriech- und is absolute temperature; T0 = 298 K; and Q/kB ≈ 14,600 K43
Relaxations-Probleme bei Beton und Spannbeton,” Beton- und Stahlbet-
onbau, V. 62, 1967, pp. 230-238, 261-269.
(kB = Boltzmann constant). Furthermore, Et = tangential
51. Bažant, Z. P., “Prediction of Concrete Creep Effects Using Age- modulus = initial E-modulus if the prestress is within linear
Adjusted Effective Modulus Method,” ACI JOURNAL, Proceedings V. 69, range; sp = F(e) = short-time stress-strain curve and, for
No. 4, Apr. 1972, pp. 212-217.
52. Bažant, Z. P., and Najjar, L. J., “Comparison of Approximate Linear linear range, F(e) = Ee; f y′ is the yield strength (1% offset);
Methods for Concrete Creep,” Journal of the Structural Division, ASCE, g f y′ is the threshold below which there is no relaxation
V. 99, No. ST9, 1973, pp. 1851-1874.
53. Burgoyne, C., and Scantlebury, R., “Why Did Palau Bridge Collapse?” (safely g = 0.45, although often g = 0.55); x = F(e)/fy′ – g ;
The Structural Engineer, 2006, pp. 30-37. k, c are the empirical exponents40; and the initial relaxation
54. CEB-FIP Model Code, Model Code for Concrete Structures, Comité
euro-international du béton (CEB), Bulletin d’Information No. 213 and curve s0 – s ∝ tk (k = 0.08, but the CEB Model Code gives
214, Lausanne, Switzerland, 1990, 437 pp. different values for different steels; c ≈ 2).