Guidelines for Fibre-based Inelasticity Modelling of Reinforced Concrete

Members
J.P. Almeida & S. Das
European Centre for Training and Research in Earthquake Engineering,
EUCENTRE, Pavia, Italy

R. Pinho
University of Pavia, Italy

ABSTRACT:
The last years have seen a generalization in the use of fibre-based distributed inelasticity frame formulations for
seismic analyses, both in research and specialized professional practice. Nevertheless, major vulnerabilities
affect the performance of these models, the most serious of which is perhaps the occurrence of non-objective
responses for softening sectional behaviour. To overcome these problems, the choice of appropriate modelling
assumptions cannot be independent of a set of parameters such as the axial load ratio, reinforcement detailing,
section geometry or mechanical material properties. Previous investigations have shown and studied the
mechanisms underlying the aforementioned numerical pathologies and recommended regularization techniques,
validated by comparison against experimental results. Based on these state-of-the-art proposals and the
application of an illustrative binomial logistic regression, the present work outlines a strategy for the
development of guidelines that can assist the analyst in the choice of suitable modelling decisions and output
interpretation.
Keywords: Distributed inelasticity; Logistic regression; Reinforced concrete (RC); Objectivity; Fibre elements.

1. INTRODUCTION
Although the numerically efficient lumped plasticity models are an interesting tool for prompt
structural assessment and confirmation of design assumptions, the increasing computational power
promoted the election of distributed inelasticity fibre models as the preferred analysis tool whenever
detailed earthquake engineering is involved. In fact, the vast majority of researchers or professional
experts that carry out meticulous frame analysis do come across with the need to use such models. The
considerable amount of specific and commercial finite element software packages that already include
such formulations is the most indicative sign of its growing importance.
Not surprisingly, specialized models come hand in hand with recurrent difficulties in their use and
output interpretation. Unawareness of common numerical phenomena that can be triggered for certain
loading cases and for particular combinations of geometrical and mechanical properties, associated to
a worrying lack of clear aid guidelines for the definition of the initial model features, further add to the
problem.
1.1. Fibre models and element formulations in earthquake engineering
The term distributed inelasticity, whilst arguably more correct from the theoretical viewpoint, comes
second to the most known term fibre models. From a rigorous perspective, the latter can be used in
association with lumped plasticity approaches. It is recalled that fibre discretized sections allow for a
direct modelling of the axial load bending moment interaction, as well as biaxial loading. However,
their relatively heavy computational cost makes them more suitable for distributed inelasticity type of
formulations, where the element response is obtained from the integration of the behaviour at different
sectional integration points (IPs). Different finite element formulations have been developed in the
recent decades, being the differences amongst them well known from the theoretical viewpoint

[Hjelmstad amd Taciroglu, 2005]. In particular, it is noted that the classical displacement-based (DB)
models are not suitable to represent the nonlinear curvature distribution along the member, whereas
the strict verification of equilibrium in force-based (FB) elements overcomes such difficulty. That is,
to model non-linear material response the latter approach is preferable since it allows for the use of a
coarse mesh (one element per member), whereas the former requires suitable mesh discretization.
1.2. The analyst modelling options
The researcher or engineer that wants to model a RC frame element has a substantial (although finite)
number of modelling options at his/her disposal. Let us consider the context of fibre models and
assume that shear deformations are sufficiently small so that they can be neglected. The occurrence
and modelling of bar pullout and strain penetration will be discussed ahead. At the element level, the
construction of a model consists basically in the definition of the following variables: type of
formulation (FB or DB), mesh characteristics (number of elements per structural member and
respective lengths), numerical integration properties (integration scheme and number of integration
points per element), regularization method (is it required?, and if so which one to adopt?) and, finally,
definition of the sectional constitutive model (number and distribution of fibres throughout the section,
uniaxial stress-strain model for cover / core concrete and longitudinal steel). Some of the more
relevant questions that naturally arise are: How to carry out the previous selection? What are the
relevant factors that should be considered? How to take them into account?
1.3. Case-study
The objective of the present paper is to establish a logical platform to help analysts that use specialized
structural analysis software to choose the more appropriate modelling options by answering the
questions posed at the end of the previous paragraph. Such choice will depend on different loading,
geometric and mechanical parameters of the problem under consideration. To illustrate the proposed
approach, an experimental test is used as a benchmark. The specimen is the column unit No. 7, from
the work of Tanaka and Park [1990]. This is a largely referred RC cantilever of square section (0.55m
0.55m) and shear span ratio of 3.0. The longitudinal steel ratio sl is 1.25%, while the transverse
steel ratio st is 2.08%. The specimen is loaded with an axial load ratio of 30% and cyclically tested by
an imposed displacement at the free node.
All the (monotonic) analyses of this study were performed with the OpenSees software [Opensees,
2010]. The transversal section has been discretized in 250 layers. For the concrete, the Kent and Park
[1971] model has been used. Regarding the concrete parameters, the cover concrete has been modelled
with fc = 32.1 MPa, 0 = 0.0024 mm/mm and 20 = 0.0248 mm/mm. The confinement effect on the
core concrete has been evaluated with the Mander model [Mander et al., 1988], resulting on a
confinement factor of kc = 1.215 with which the values fcc = 39 MPa, compression strength of the
confined concrete, and c,max = 0.0052 mm/mm, strain corresponding to the maximum concrete stress,
are obtained. For what concerns the reinforcement steel, a bilinear stress-strain relationship is used,
with elastic modulus E = 200.000 MPa, yield stress fy = 510 MPa and 1% strain-hardening ratio.
In order to estimate the strain limits for moment-curvature analysis, the proposals and formulas
indicated by Priestley et al. [2007] have been adopted. For the damage-control compression strain, the
confined concrete compression strain limit is taken to occur when fracture of the transverse
reinforcement confining the core happens. For the present case this condition results in a value of c,dc
= 0.0289. A steel tension strain limit of s,dc = 0.6su = 0.06 is also considered.

2. MODELLING OBJECTIVE AND NON-OBJECTIVE RESPONSES

A numerical response of a distributed inelasticity model is said to be objective if, as the number of
elements per member (in DB models) or the number of IPs per member (in FB models) increase, then
the model output both at the global level and at the local level converge to a specific value. It is
also well-known that the occurrence of this objectivity is connected to the sectional constitutive
behaviour: while the sectional response of the most strained section is in the hardening range, then the
response is objective; as soon as one of the sections depicts a negative stiffness, then global level
measures (top node horizontal displacements) and local level measures (curvatures and strains) start
being non-objective. The magnitude of such localized error depends on the magnitude of the excursion
into the sectional post-peak range, as well as on the steepness of the corresponding softening branch.
Analysts with the minimum experience know that: (i) for DB formulations it is not justified, in any
case, to use more than 2 IPs per element and other than a Gauss-Legendre integration scheme; (ii) for
FB formulations 1 element per member should be used, with Gauss-Lobatto quadrature rule (possible
exception when non-objective responses are expected, as it will be discussed below).
In a previous study by the authors [Calabrese et al., 2010], the above results were explicitly illustrated
for the case-study. Under a 3% axial load ratio, which induces a hardening type of sectional response,
it was additionally showed that converged results for global level responses required: (i) DB
formulations: 4 elements per member, preferably using shorter elements at the member extremities;
(ii) FB formulations: 4 IPs per element. At the local level, converged responses could be obtained
with 6 IPs per element in FB approaches, while for DB elements an interesting type of engineering
localization occurs, addressed in the following sub-section (2.1). Again, it is underlined that the
previous results were inferred from the analysis of a cantilever.
On the other hand, for the experimental axial load ratio (30%), the ensuing softening behaviour
induces, as expected, a non-objective response. In such cases a regularization technique has to be
adopted, either for FB or DB formulations, such that the model output is physically significant. In
particular, these methods have been an active field of recent research concerning FB elements, which
are more suitable for non-linear analysis than DB elements. Coleman and Spacone [2001a] were the
first to present a regularization method based on a constant fracture energy release. The difficult
application of their procedure to practical engineering analysis (since the uniaxial concrete material
properties in each section had to be adapted as a function of the number of IPs per element) motivated
further research, from which it is of reference: (i) the proposal by Scott and Fenves [2006], based on a
modified Gauss-Radau quadrature in the plastic hinge regions and a Gauss-Legendre integration rule
in the interior part of the element; (ii) the solution put forward by Addessi and Ciampi [2007], in
which Gauss-Lobatto schemes are used both in the plastic hinge and the interior regions. The major
criticism to the two previous integration methods is that although regularized outputs for softening
behaviour can be obtained, the numerical accuracy for hardening behaviour decreases significantly,
producing a too flexible response. Scott and Hamutuoglu [2008] developed a curious integration
scheme, featuring two additional IPs very close to the element ends, which yields convergent solutions
for both types of sectional constitutive behaviours. Its most relevant downside is the reduction in the
order of accuracy from 2N-3 to N-3 (N being the number of Gauss-Lobatto IPs of the underlying
quadrature rule), which implies that a large number of IPs have to be considered in order to obtain the
same degree of precision for hardening responses; such substantial increase in computational power is
not negligible. Additionally, the high sensitivity to the placement of the two extra IPs can be of
concern. More recently, Lee and Filippou [2009] presented an interesting beam-column element based
on the integration of the smoother section deformation rate (instead of the section deformations) over
the end inelastic zones. However, its current restriction to bilinear moment-curvature relations under
fixed axial force, as well as accuracy issues for single curvature conditions, deserve future research.
Every single one of the foregoing proposals requires the definition of a plastic hinge region, which is
typically identified with the plastic hinge length (obtained from one of the many existing empirical
formulae). This association deserves further attention and remarks, which are carried out in paragraph

2.2. Finally, a specific but meaningful comment regarding regularization techniques for DB meshes is
made on paragraph 2.3.
2.1. The engineering type of localization in DB meshes under hardening behaviour
In the previous lines no guidance was provided regarding the recommended number of DB elements
per member to attain a converged response of the local measures under hardening behaviour. The
reason for such intentional omission is the difficulty (or impossibility) to provide a clear indication. As
a matter of fact, Fig. 2.1.(a) [Calabrese et al., 2010] shows that for the present case-study a convergent
response can only be obtained for a mesh of above 40-50 elements. In other words, it is not feasible to
directly evaluate local measures from common engineering DB meshes. This phenomenon can be
much more pronounced than what was noted in a previous work by Neuenhofer and Filippou [1997].
Such particular type of localization is due to the fact that DB formulations only satisfy equilibrium in
the weak (or average) form, which is apparent from Fig. 2.1.(b).
# IPs, FB
2

10

0.16

450

0.14

400

Curvature [m ]

350

0.12

Force [kN]

-1

0.1

300
Axial force at the second controlling section
Axial force at the first controlling section
Weighted sum

250

0.08

DB formulation

0.06

200

FB formulation
150

0.04
1

# Elements, DB

100

0.00

0.01

(a)

0.02

0.03

0.04

Top node horizontal displacement [m]

(b)

Figure 2.1. (a) Curvature at the bottom integration section for 2.5% drift: comparison between a DB uniform
mesh and a FB element with Gauss-Lobatto IPs; (b) Axial forces at the two controlling sections of the bottom
DB element (2 DB elements mesh: 0.5m + 1.15m) [Calabrese et al., 2010].

2.2. The role of the plastic hinge length as a regularization parameter

All regularization procedures require, in one form or another, the estimation of a plastic hinge region.
The latter is in general evaluated with an empirical formula for the plastic hinge length (e.g., the wellknown expression given by Paulay and Priestley [1992]). The distinction between the terms (i) length
of plastification the actual physical length over which the real plastic curvatures extend and severe
damage concentrates, and (ii) plastic hinge length a conventional parameter that deals indirectly with
effects of shear, bond-slip, etc., is clearly referred by Fardis [2009]. Unfortunately, the former term is
still rarely used, which increases the chances for an interchangeable misuse of different expressions.
The concept of length of plastification should be the one to be used for regularization purposes, which
was never the case in the aforementioned publications. A clarifying work, wherein the effect of the
axial load level in the length of plastification is also accounted for, is that by Bae and Bayrak [2008].
2.3. A widespread but questionable belief
It is commonly believed that, for DB formulations, the use of displacement interpolation functions
force localization over the length of an entire element [e.g., Coleman and Spacone, 2001b; Scott and
Hamutuoglu, 2008]. Such underlying principle justifies the regularization criterion of defining the
length of the most strained elements (typically, those at the member ends) as equal to the plastic hinge

length (or more appropriately, to the length of plastification). However, it was shown by Calabrese et
al. [2010] that the localized response may not actually concentrate at the element level but at the IP
level, as in FB elements. That is, considering 2 Gauss-Legendre IPs per element, the regularized
response should be based on a mesh where the size of the frame extremities elements is set as twice
the length of plastification.
3. STRATEGY FOR THE DEVELOPMENT OF MODELLING GUIDELINES
3.1. Modelling guidelines
The conclusions and comments of the previous section can be summarised in Fig. 3.1., which intends
to serve as a first general scheme of modelling guidelines to assist analysts in the definition of their
finite element model. It should be explicitly stated that such proposal was derived from the study of a
cantilever (single curvature) under a finite number of load cases and thus it is not yet recommended
for generalized application, which is dependent on the execution of a forthcoming detailed parametric
study.

Analysts modelling
choices ?

- Gauss-Legendre / Gauss-Lobatto - Number of elements / member (DB)

- Number of integration points
- Mesh characteristics (DB)

Software dependent

Hardening
FB
(preferred!)

DB
(artificially
stiffer, average
equilibrium)

Fully
HARDENING
response
expected or
STRAINSOFTENING
behaviour is
possible?
(suggestion: use
logit regression
model below for
initial guess)

Softening

Hardening

Global
response

4 IPs / element

Local
response

6 IPs / element

- Gauss-Lobatto Global/Local
response
-1 elem./member

Apply regularization
method

- Gauss-Lobatto
- 1 elem./member

- Gauss-Legendre
-2 IPs / element

Softening
- Gauss-Legendre
-2 IPs / element

Global
response

4 elements / member
(refine at extremities)

Local
response

Not reliable for

engineering meshes
(or postprocess?)

Global/Local
response

Lbottom elem = 2 Lp

Figure 3.1. Trial draft for modelling guidelines, inferred from the analysis of a cantilever.

The majority of structural analysis software packages do not include regularization procedures.
Consequently, as depicted in the figure above, it is apparent that perhaps the single most important
unknown that should influence the engineers initial modelling options as well as the final
interpretation of results is the possible occurrence of a strain-softening response in the member for
the specific deformation demand (and before reaching the steel and concrete compression damagecontrol strain limits). Although this can only be confirmed at the end of the analysis, it would be
helpful, in order to avoid or reduce lengthy iterative procedures, to have an estimate of the potential
for such occurrence. This will be addressed in the next section.
3.2. Logistic regression model for softening / hardening initial guess
It is known that overreinforced or heavily loaded (with axial load) RC members are prone to develop
softening responses. This straightforward observation motivated the development of a trial model,
obtained via a logistic regression, which is function of the independent variables (axial load ratio)
and sl (longitudinal steel ratio).

It is assumed, for the current illustrative purposes, that the transverse steel ratio st is constant for all
the analyses (= 2.08%). Therefore, the confined concrete compression strain limit, as well as the steel
tension strain limit (referred to in subsection 1.3.), were also kept unchanged.
Combinations between 15 values of the axial load ratio , varying between 0.5% and 30% according
2
to the expression of axial load 50 + 15.05 ( i 1) [ kN ] wherein i = 1, 2, 3,..,15 , and 10 values of the
longitudinal steel ratio sl, varying uniformly between 0.45% and 3%, originated 150 cases of analysis.
The moment-curvature relation for each of these cases was then produced with Opensees [2010] and
the following procedure was applied: (i) identification of the damage-control curvature, taken as the
minimum between the steel tension limit curvature and the confined concrete compression damagecontrol curvature; (ii) verification of the relative position of the damage-control curvature and
associated moment: if the latter was smaller than 0.95 times the largest observed moment within the
range of damage-control curvature, the corresponding softening / hardening binary response variable
was assigned a value of zero; in other words, it was considered that the probability of occurrence of a
softening non-objective behaviour, with eventual impact on the physical validity of the structural
response, was non-negligible; for the remaining cases it was assumed that hardening was the only
possible occurrence and consequently the softening / hardening binary response variable was assigned
a value of one. From this set of data the logistic regression was carried out.
The logistic function is the following:

f (z ) =

1
1 + e z

(1.1)

The variable z , known as logit, represents the exposure to the axial load ratio and the longitudinal
steel ratio, selected as independent variables. The output f ( z ) is confined to values between 0 and 1,
representing the probability of hardening given that set of explanatory variables. In the present case,
an output below 0.5 stands for a softening type of behaviour while above 0.5 indicates a hardening
response. The variable z is defined as:
z = a 0 + a1 + a 2 sl

(1.2)

where a 0 is the constant (also known as intercept) and a1 and a 2 are the regression coefficients of
and sl, respectively. The application of the binary logistic regression was performed with an ad-hoc
program, yielding the following best fit:
z = 32 180 799 sl

(1.3)

This model can now be compared with the initial fitting data, which is performed in Fig. 4.1. The case
numbers correspond to 10 sequential blocks of increasing longitudinal steel ratio, each one containing
15 cases of increasing axial load ratio. One first pertinent comment is related to the observation that
the majority of the model predictions come close to either 0 or 1, which indicates the strong
dependence of f ( z ) from the independent variables chosen (that is, the latter were a good selection).
Secondly, it is observed that, out of the 150 cases, the model only failed to predict the type of
behaviour of 5, which can be considered a very satisfactory regression. As physically expectable, if the
percentage of longitudinal steel is low, then a softening type of response is only predicted for large
values of the axial load ratio. As the steel ratio increases, softening is attained gradually for lower
values of the axial load.

1
0.9

Response Identifier

0.8
0.7
0.6
Correct Model Prediction
Wrong Model Prediction

0.5
0.4
0.3
0.2
0.1
0

50

100

150

Case Number
Figure 4.1. Comparison between the binomial logistic regression model and the fitting data.

The good match depicted in the figure above shows that logistic regressions are a promising tool to
develop a more comprehensive model function of a broader range of geometrical and mechanical
properties that will be able to impart a truly general character to the present guidelines. The latter
will provide an estimate (prior to the analysis) of the possible member behaviour and, as a function of
the output of interest to the analyst, thus assist him/her in the selection of the most appropriate
modelling parameters.

4. CONCLUSIONS
The generalization of performance-based design and assessment methodologies, alongside with the
sustained increase in computational power, keeps promoting the use of fibre-based distributed
inelasticity models. This class of models is affected by a number of numerical issues that should be
thoroughly understood to select appropriate modelling parameters, as well as to construct a critical
viewpoint on the model outputs. The current paper intends to give a contribution towards the creation
of simple and easy-to-follow aid guidelines.
With such a goal in view, a concise summary on well-known and other less known behavioural
features was carried out. In particular, attention was paid to commonly accepted modelling rules,
advantages and drawbacks of the most recent regularization methods for FB elements, particularities
of the traditional DB mesh performances, and the role of the plastic hinge length within the context of
softening regularization techniques.
Taking into consideration the above findings, as well as the conclusions of a numerical case-study, a
tentative scheme for modelling guidelines was drafted. Such proposal included the development of an
illustrative model based on binomial logistic regression to assess the probability of occurrence of
hardening or softening responses as an initial guess. The latter was found promising for incoming
extensions of the guidelines.
However, and despite the many contributions on this field, critical modelling challenges are still
present. In particular, there is room for the development of a new adaptive model for FB approaches,
able to automatically adapt the numerical integration scheme as a function of the sectional constitutive
state. Finally, it is the authors conviction that bar pull-out and strain-penetration phenomena should

be taken into account with calibrated zero-length elements placed at the member ends, therefore
allowing for a subtraction of such deformation contributions from the more general and sometimes
dubious plastic hinge length concept.
REFERENCES
Addessi, D. and Ciampi, V. (2007). A regularized force-based beam element with a damage-plastic section
constitutive law. International Journal for Numerical Methods in Engineering 70:5, 610-629.
Bae, S. and Bayrak, O. (2008). Plastic hinge length of reinforced concrete columns. ACI Structural Journal
105:3, 290-300.
Calabrese, A., Almeida, J.P. and Pinho, R. (2010). Numerical issues in distributed inelasticity modeling of RC
frame elements for seismic analysis. Journal of Earthquake Engineering 14:1, 38-68.
Coleman, J. and Spacone, E. (2001a). Localization issues in nonlinear frame elements. Journal of Structural
Engineering 127:11, 1257-1265.
Coleman, J. and Spacone, E. (2001b). Localization issues in nonlinear frame elements. Modelling of Inelastic
Behaviour of RC Structures Under Seismic Loads, ed. P.B. Shing and T. Tanabe (ASCE, Reston, VA), 403419.
Fardis, M.N. (2009). Seismic Design, Assessment and Retrofitting of Concrete Buildings, Springer.
Hjelmstad, K.,D. and Taciroglu, E. (2005). Variational basis of nonlinear flexibility methods for structural
analysis of frames. Journal of Engineering Mechanics 131:11, 1157-1169.
Kent, D.C. and Park, R. (1971). Flexural members with confined concrete. Journal of Structural Division 97:7,
1964-1990.
Lee, C.-L. and Filippou, F.C. (2009). Efficient beam-column element with variable inelastic end zones. Journal
of Structural Engineering 135:11, 1310-1319.
Mander, J.B., Priestley, M.J.N. and Park, R. (1988). Theoretical stress-strain model for confined concrete.
Journal of Structural Engineering 114:8, 1804-1823.
Neuenhofer, A. and Filippou, F.C. (1997). Evaluation of nonlinear frame finite-element models. Journal of
Structural Engineering 123:7, 958-966.
OpenSees (2010). Open System for Earthquake Engineering Simulation (online). Available from
http://opensees.berkeley.edu
Paulay, T. and Priestley, M.J.N. (1992). Seismic Design of Reinforced Concrete and Masonry Buildings, John
Wiley & Sons, New York.
Priestely, M.J.N., Calvi, G.M. and Kowalsky, M.J. (2007). Displacement-based Seismic Design of Structures,
IUSS Press, Pavia, Italy.
Scott, M.H. and Fenves, G.L. (2006). Plastic hinge integration methods for force-based beam-column elements.
Journal of Structural Engineering 132:2, 244-252.
Scott, M.H. and Hamutuoglu, O.M. (2008). Numerically consistent regularization of force-based frame
elements. International Journal for Numerical Methods in Engineering 76:10, 1612-1631.
Tanaka, H. and Park, R. (1990). Effect of Lateral Confining Reinforcement on the Ductile Behaviour of
Reinforced Concrete Columns, Report 90-2, Department of Civil Engineering, University of Canterbury,
New Zealand.