Computational Materials Science 95 (2014) 288301

Contents lists available at ScienceDirect

Computational Materials Science

journal homepage: www.elsevier.com/locate/commatsci

A practical ow diagram for the solution of complex non-linear

thermo-mechanical numerical models
D.G. Karalis a,, N.G. Tsouvalis b, V.J. Papazoglou b, D.I. Pantelis b
a

Hellenic Navy, Hellenic Naval Academy, Mechanics & Materials Division, Marine Materials Laboratory, Hazjikyriakou Avenue, Piraeus 185 39, Greece
Shipbuilding Technology Laboratory, School of Naval Architecture and Marine Engineering, National Technical University of Athens, 9 Heroon Polytechniou Avenue, Zografou,
Athens 157 73, Greece
b

a r t i c l e

i n f o

Article history:
Received 13 November 2013
Received in revised form 30 May 2014
Accepted 27 July 2014

Keywords:
Thermo-mechanical modelling
Finite element analysis
Convergence ow diagram
Repair welding

a b s t r a c t
In this paper the authors propose a practical ow diagram for the systematic development and solution of
complex thermo-mechanical nite element analysis models. The proposed diagram consists of three different phases and provides a step-by-step guide for the development of the nal thermo-mechanical model,
taking into account convergence issues, mesh density and estimation of time step magnitude. In phase I, a
preliminary thermo-mechanical analysis is carried out in order to get an idea of the model behaviour, the
required resources and the feasibility of the overall analysis. In phase II the nal thermal model is developed
in full, taking into account the mechanical results obtained at the end of phase I, whereas in phase III the
nal mechanical model is generated on the basis of a continuously modied thermal model. The proposed
procedure presented herein in the form of a ow diagram provides the capability for gradual output of the
numerical results (preliminary results, thermal results, mechanical results), while paying attention to the
time-consuming problem of results convergence required for a numerically accurate analysis. The former is
an important issue for large-scale complex simulation projects, whereas the latter provides evidence that
the development of the numerical model has been realized on the basis of the modelling laws. For better
presentation and understanding, the proposed procedure is applied to the study of a nite element analysis
thermo-mechanical model, where increased intricacy generally exists.
2014 Elsevier B.V. All rights reserved.

1. The problem of modelling consecutive phenomena

1.1. Introduction
The thermo-mechanical response of steel or aluminium plates
during welding or plate forming by line heating has been investigated by several researchers during the last decades. Most of the
research is focused on either or both the thermal and the mechanical part of the structural response through a combination of experimental and numerical simulations. The numerical part of the
investigation still attracts high interest due to its extreme intricacy
and the uncertainty in predicting the structural response prior to
the treatment (welding or line heating) itself. An extensive review
has been conducted in [14].
In a fully uncoupled thermo-mechanical nite element model,
the analysis is usually carried out in a staggered approach: the thermal problem is solved rst, followed by the solution of the mechanical problem. The latter mechanical analysis runs on the basis of the
thermal results in order to account for the thermal stress and phase
Corresponding author. Tel.: +30 210 45 81 656, +30 697 37 97 661.
E-mail address: KaralisDimitris@TEEmail.gr (D.G. Karalis).
http://dx.doi.org/10.1016/j.commatsci.2014.07.045
0927-0256/ 2014 Elsevier B.V. All rights reserved.

change effects on the structural response of the structure. This is performed by importing to the mechanical model the nodal temperatures at each time increment and calculating the thermal strain.
From the aforementioned staggered approach it is deduced that
both thermal and mechanical models must normally run with the
same analysis parameters, namely time step magnitude and mesh
density. If, for example, the material undergoes phase transformation accompanied by volume change during a specic short temperature range, a small constant time step and a ne mesh are required
in the areas of transformation for both the thermal and the mechanical analysis. This allows for the accurate monitoring of the transient
stress developed during the transformation temperature range [5].
Thus, for the entire analysis there should be an exact correspondence between the mesh density and the time step magnitude
between the two models. This requirement renders the whole procedure of model development very complex and time-consuming.
1.2. The three major problems: mesh density, time step and
convergence of results
The rst problem arising during the thermo-mechanical modelling is that the thermal and the mechanical models are completely

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

different in nature, as they model different physical phenomena.

Therefore the mesh density selected for the solution of the thermal
problem is, in most cases, inappropriate for the solution of the
mechanical problem.
Secondly, the time step required for the accurate solution of the
mechanical analysis may be too large compared to the time step
required for the accurate solution of the heat ow problem, where,
for example, extreme temperature gradients are encountered. The
latter is also valid in the opposite case as, at high temperatures, the
structure may exhibit extreme material non-linearities.
A third problem pertains to the results convergence criteria.
The development of a numerical model by means of the nite element method is generally terminated when the analysis has
reached (a level of) results convergence. For example, classical
convergence criteria are based on the stabilization of nodal results,
such as temperatures or displacements with regard to mesh density and time step. It is actually not worthy remeshing the model
or reducing the time step if the nodal results do not change values
versus simulated time.
It should be emphasized at this point that in general there are
four types of convergence in nite element analyses:
i. convergence of equilibrium iterations due to non-linearities
(e.g. material, contact or geometrical non-linearities),
ii. convergence in the solutions of the linearized systems of
algebraic equations in case of iterative solvers,
iii. convergence of the results due to mesh renement and
iv. convergence of the results due to time step reduction.
In most commercial nite element software platforms, specic
optimum values and tolerances are already pre-set in order to
control best the convergence of the equilibrium iterations due to
non-linearities and convergence of the equations in case of iterative solvers. In the present study, emphasis is given only to the last
two convergence types, namely time step and mesh renement, as
they are the main user-dependent parameters that strongly inuence the entire simulation and results convergence. The procedure
followed towards the convergence of results governs directly the
overall simulated time, the numerical analysis cost and affects
the accuracy of the results. For example, some of the complex simulations presented in [6,7] have lasted a few days, time that could
have been strongly increased if a few more additional analyses
have been required due to convergence issues. At this point, it
should be mentioned that in most publications dealing with complex thermo-mechanical simulations the convergence criteria have
not been described at all, as the authors provide only the models
setup and the numerical results. Hence, the end reader of the aforementioned papers comes to understand that the authors have
somehow performed a convergence analysis prior to publishing
the results obtained by means of the nite element method. This
convergence analysis is of great interest as it is complicated,
time-consuming and strongly user-dependent.
In sum, a common time step and mesh density are normally
required for both the thermal and the mechanical analysis. These
two common parameters must allow both physical problems to
be modelled satisfactorily, but they must also provide an acceptable level of results convergence for both models.
1.3. The aim of this paper
The authors aim at proposing a practical ow diagram for the
systematic development and solution of complex FEA thermomechanical models. In this ow diagram a progressive development of several thermal and mechanical models will be presented
on the basis of different mesh densities and time steps, aiming at
reaching the convergence of the thermal and mechanical results.

289

Fig. 1. The welded bracket used as an example to present the ow diagram.

It ought to be mentioned here that a ow diagram for the solution

of such staggered thermo-mechanical models is missing from the
international literature and that the whole process is a real labyrinth for both experienced and inexperienced users dealing with
thermo-mechanical modelling. Please note that the aim of the
authors is to discuss the proposed ow diagram and present the
steps followed for creating the nal thermo-mechanical model
with regard to mesh density and selection of time step and not
to provide the mathematically-based analysis for its development.
The latter has already been discussed in the literature [835]. The
implementation of the proposed ow diagram requires a commercial thermal and mechanical or multi-physics FEA software package for which code verication has been already performed.

2. A typical example to explain the ow diagram

2.1. The physical model
In order to discuss the proposed ow diagram, a thermomechanical simulation will be employed. The latter concerns the
weld treatment of a welded bracket under load, by means of tungsten inert gas (TIG) welding. The whole conguration of the simulation is presented in Fig. 1.
The bracket shown in Fig. 1 is made of typical carbon structural
steel (containing 0.45% w/w carbon) and consists of a bent ange
and a triangular reinforcing web welded on the ange. The welds
AC and AB exist along both sides of the web. Treatment is performed along the AB weld on the side towards the +z semi-axis
(the one that is visible in Fig. 1) using a TIG torch without ller
metal and aims at treating the existing weld close to the melting
temperature. Such treatments are applied to repair in-situ cracked
or defected welds (repair welding). The welded bracket is xed at
its smaller side (see red1 triangles in Fig. 1 that refer to the xations). After the material has cooled to ambient temperature, uniform pressure is applied on the other side of the bracket (see red
arrows in Fig. 1) tending to buckle the triangular reinforcing web.
The latter pressure simulates the operational load present on the
bracket after the completion of the treatment. The treated length
lAB is equal to 128 mm, whereas the ange and web thicknesses
are equal to 25 mm and 12.5 mm respectively. The power of the
welding torch was set equal to Q = 3770 W whereas the speed was
set equal to v = 6 mm/s. This simulation is quite complex involving
the existence of extreme non-linearities as temperatures are raised
to the steel melting point. It represents a difcult-to-solve numerical
analysis, as thermal, mechanical and thermally-induced mechanical
1
For interpretation of color in Fig. 1, the reader is referred to the web version of
this article.

290

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

phenomena coexist and strongly affect the solution and convergence

procedure.
2.2. The numerical model
From the description of the aforementioned weld treatment, it
is deduced that the numerical analysis consists of two parts: thermal and mechanical. Normally, during the rst thermal part, the
transient temperature distribution for the whole bracket is calculated, whereas in the second mechanical part, the total transient
displacements and stresses are calculated, including any residual
stresses. Note that, during the second part of the analysis, apart
from the thermal stresses derived from the thermal treatment,
additional stresses are generated due to the externally applied
pressure. It is concluded from the above that the mesh density
and the time step of both models should be able to model all transient phenomena related to the weld treatment (temperature distribution, thermal stresses, residual stresses and distortion), as
well as the general mechanical response like local stresses raised
at the geometrical discontinuities of the structure and induced
by the externally applied forces and the weld treatment itself [1].
Typical questions during model development pertain to (a) the
ow diagram proposed in the following to reach the common mesh
density and time step that offer adequate convergence of the
results for both models, and (b) the overall time and computing
resources required to complete the analysis. These questions
become more critical and difcult to be answered as the modelled
physical structure becomes more complicated and bigger in size
[6,36,37]. A large-scale structure implies that each trial run of
the thermo-mechanical analysis aimed at reaching an acceptable
level of convergence will last at least for a considerable amount
of time.
For the thermomechanical simulation of the aforementioned
treatment, a three dimensional nite element model was set up
using ALGOR nite element code [22]. A staggered approach
was employed by solving at rst for temperatures and then for
displacements and stresses (uncoupled formulation). First-order
eight-nodded solid heat transfer elements were used for the
thermal part and rst-order eight-nodded solid thermoplastic elements (instead of second order elements with midside nodes [2,3])
were used for the mechanical part in order to account for the worst
scenario with respect to available element types. Mesh compatibility was retained between the two analyses. The heat source was
modelled by employing a moving Gaussian distribution. The kinematics, the constitutive formulations, the modelling, as well as the
boundary conditions were applied as per [6,7]. The temperature
eld was considered unaffected by the structural response. The
steel was modelled as isotropic, having yield stress equal to
380 MPa [6,7,38,39] and temperature dependent properties including plasticity and strain hardening. Cooling was implemented by
means of conduction, convection and radiation. At the beginning
of the simulation, the bracket is considered free from welding
residual stresses. The stress free reference temperature of the
material was set at ambient temperature (25 C).
3. The philosophy of the spatial and temporal renement
aiming at results convergence
Prior to presenting the ow diagram some logical questions
arise: how is the mesh renement and time step modication performed, in order to achieve satisfactory convergence of the results?
What is the philosophy behind this temporal and spatial renement? When convergence can be considered as satisfactory?
The progressive time step reduction and the gradual mesh
renement play an important role affecting the accuracy of the

entire simulation. Taking into account the variety of different elements and analysis types that exist nowadays in most commercial
nite element platforms, the development of an efcient set of
equations between the reduction of the time step and the gradual
mesh renement that leads to results convergence is a very
difcult and triggering task. In the current proposed diagram, this
spatial and temporal renement is based on the repetitive execution of the thermal and the post-mechanical model. This execution
provides feedback pertaining to the appropriateness of the spatial
and temporal renement that was applied. The latter methodology
has the advantage of applicability in most thermomechanical
simulations except of casting simulation where it is not directly
applicable due to material ow.
The gradual reduction of the initial time step that is applied by
the analyst is strongly affected by all the temperature and time
dependent phenomena that take place during the entire simulation. It is well known, that in a typical transient non-linear thermomechanical analysis, temperature and time dependent magnitudes
exist. Temperature dependent magnitudes can refer, for example,
to the material properties, coefcient of heat convection and convection heat, radiation; whereas time dependent magnitudes can
refer to the moving heat source, heat convection, operational loads,
pressures, existence of gaps, etc.
As a basis for the discussion of the next paragraphs, Fig. 2
depicts typical examples of the temperature dependent heat
capacity, the thermal conductivity, the convection lm coefcient,
the yield stress and the thermal dilatation of a typical mild steel
that undergoes several microstructural transformations depending
on the peak austenitization temperature (Tpeak) [6,7]. In the same
gure, the time dependent moving heat source of a welding arc
is also presented [6]. Furthermore, in Fig. 2f, the temperature
depended axial stress response of an axially xed steel specimen
that undergoes phase change transformation is shown [5,6].
In nite element simulations, the temperatures in the thermal
analysis and the displacements in the mechanical analysis are calculated for every node of the model and are exported at every time
step. Therefore the applied gradual reduction of the initial time
step value should nally:
i. Provide small temperature differences at every node of the
model between all successive analysis steps, so that the temperature dependent phenomena are accurately modelled.
For example, a very small time step can result in very small
nodal temperature differences between all successive analysis steps. It is up to the researcher to decide, whether the
latter temperature difference can accurately model the
non-linear material properties at the areas of solid state
transformations (see Fig. 2a, c and d) or whether it is enough
to accurately account for the convection heat losses (see
Fig. 2b). On the other hand, a relatively bigger time step
can provide larger nodal temperature differences between
successive analysis steps and thus hiding or articially
minimizing the effects of phase change on the transient
mechanical response of the structure. Taking into account
that the steel phase transformation temperature range is
approximately DTtr = 300 C, a practical guideline is to select
the maximum allowable time step that provides temperature differences of maximum 30 C, or 10% of DTtr. This small
temperature difference will later provide the basis for an
accurate mechanical analysis where small thermally
induced stress differences are also required. Therefore the
selected maximum allowable time step should additionally
keep the thermally-induced stress differences between successive analysis steps in a stress analysis smaller than a
small percentage of the material yield stress. In current
paper the value of 5% of the material yield stress is

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

291

Fig. 2. An example of temperature and time dependent magnitudes in a typical thermomechanical analysis, reprinted from [6,7]. (a) Material heat capacity and thermal
conductivity, (b) surface convection lm coefcient, (c) material yield stress, (d) material thermal dilatation, (e) three dimensional moving heat source, (f) axial stress
response of an axially xed steel specimen that undergoes phase change transformation.

suggested. Similar guidelines can also be applied for the

strain. In conclusion, the maximum allowable temperature
difference required for the realistic modelling of the temperature dependent phenomena denes the maximum allowable time step value to be used. The latter will be obtained
after the repetitive execution and post-processing of the
thermal and mechanical models in all three phases of the
ow diagram that will be proposed later.
ii. Allow the time dependent magnitudes to be adequately
taken into account during the entire analysis. For example,
if the actual velocity of the moving source is high (see
Fig. 2e), a relative small time step is required in order to

accurately capture the steep shape of the source along

length a1. Furthermore, if instant cooling of the weld metal
is applied during welding (e.g. underwater welding), only a
very small time step value will be able to capture the instant
change of the heat transfer coefcient, and thus correctly
calculate the transient heat transfer phenomena. Taking into
account that in most conventional welding simulations (a)
the size of the moving heat source is equal to several millimetres along the three axes, (b) the torch speed is equal to
several millimetres per second and (c) no forced convection
exists, a simple guideline is to set the initial value of time
step not larger than 1 s. Normally, the latter value is later

292

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

expected to be strongly reduced to a small percentage of its

initial value in order to satisfy the convergence criteria.
Alternatively, a second practical guideline for the selection
of the initial value of time step pertains to divide the total
duration of the steepest part of the curve of the time dependent magnitude into minimum 3 different equal time steps.
In conclusion, the steepest part of the curve of the time
dependent magnitudes denes the maximum allowable
time step value to be applied. The latter will be deduced
after the iterative execution and post-processing of the thermal and mechanical results.
From the discussion above it is concluded that the nal time
step value that provides satisfactory convergence of the results
should be the minimum of the two maximum time steps derived
from items (i) and (ii) above.
As far as the gradual renement of the initial mesh density is
concerned, similar rules and observations that were described previously are valid. More specically, the applied gradual renement
of the initial mesh density should aim at providing enough elements at the areas of interest in order to accurately model the
material that is affected by:
i. The abrupt change of the temperature dependent magnitudes. For example, the narrow heat affected zone that is
generated between the weld pool and the base metal is
strongly affected by the temperature (and phase) dependent
material properties. A few elements along this zone would
not sufce to accurately obtain the highly transient phenomena that occur in this area. A practical guideline pertains to
employ an initial mesh density of minimum 3 elements
along the heat affected zone. In conclusion, the mesh where
an abrupt change of the temperature dependent magnitudes
takes place should be highly rened.
ii. The abrupt change of the time dependent magnitudes. For
example, in order to accurately apply the power of the heat
source along length a1 (see Fig. 2e), many elements are
required to be present along this length. It is obvious that
only one or two elements along this area would not sufce.
A practical guideline pertains to employ an initial mesh density of 3 elements per the shortest length among a1, b or c
(see Fig. 2e) for the whole area where the arc distribution
is applied. In conclusion, the mesh where an abrupt change
of the applied time dependent magnitudes takes place
should be highly rened.
iii. The stress concentrations generated by both residual or
applied operational loads. The former is of great importance
in case of welding residual stress analysis and requires a
rened mesh along the three axes especially at the vicinity
of the weld metal and the heat affected zone where residual
stresses present strong variability. The latter stress concentrations can be derived from a static stress analysis.
Similarly to the time step comments, the above discussion
shows that the nal mesh density should at least satisfy items
(i), (ii) and (iii) above, in order to provide satisfactory and accurate
modelling.
The third issue pertains to the criteria of the results convergence acceptance. This dilemma that is set at every sub-step of
the analysis (more specically at every rhombus of the ow diagram that will be proposed later) strictly depends on the
researcher and the way he deals with the scope of the analysis,
the areas of interest and the magnitudes being monitored. As
stated previously, it is actually not worthy remeshing the model
or reducing the time step if the nodal displacements (or temperatures) do not change values versus simulated time in a dis-

placement (or temperature) analysis. The same observation is

also valid for the stresses or heat uxes or other magnitudes
of interest. It must be mentioned here, that the convergence
study must be realized on the basis of the predicted results: a
literature research [33] has shown that authors have often used
reaction forces or displacements to perform their convergence
tests and subsequently made predictions for other results such
stresses or strains. Furthermore the tolerance of the results convergence criteria may strongly differ among researchers; a 15%
difference between the temperatures of two successive thermal
analyses results may seem enough for the termination of the
repetitive execution of the models in a general heat treatment
simulation; whereas it may be dealt as a considerable difference
in the case of a material phase-change response analysis. Furthermore, the criteria of the results convergence depend also
on the area of interest. In a residual stress analysis for example,
the researcher is mainly focused on the weld metal and the heat
affected zone; thus the residual stress results on these zones
must converge satisfactorily when the researcher decides to terminate his investigation (or the repetitive loop of the ow diagram that will follow). On the other hand, in a residual
displacement analysis, the area of interest is usually the far
end of the plates that present the maximum distortion. From
the discussion above it is derived that (a) the aim of the thermomechanical simulation, (b) the response of the area of interest of
the model and (c) the magnitudes being monitored, should be
the governing parameters that will allow the termination or
the continuation of the convergence study. Here the authors
would like to emphasize on the fact that an accurate thermalstress analysis requires a very accurate thermal analysis. It is
thus very important for the analyst to have the thermal analysis
accurately performed and the thermal results converged to an
acceptable level. Finally, in addition to what was discussed previously, a good practice in a thermal analysis where extreme
thermal gradients exist is to monitor the minimum temperature
calculated by the software at every time step. For example, in a
typical welding simulation without preheating, the minimum
calculated transient temperature of the model should always
be at least the initial material temperature. The execution of
the analysis with an inappropriate time step in combination
with a rough mesh can provide smaller or negative transient
minimum temperatures compared to the initial temperature of
the material.
To summarize the discussion that deals with the convergence
criteria, in current paper the analysis will be terminated when
the relative difference of the magnitudes being monitored (temperatures, stresses and displacements) between two successive
analyses is less than 10%. The latter value provides an acceptable
level of convergence in engineering terms taking into account that
welding simulations are very complex non-linear thermomechanical problems. Of course a smaller value can also be applied, which
represents a more strict criterion but has a negative impact on
computer resources and cost.
4. The proposed ow diagram
The proposed ow diagram consists of three different phases
(phases IIII), each one containing several sub-steps and execution
loops. For the readers convenience, each phase and sub-step are
discussed separately and are supported by gures of the relevant
nite element model. In the proposed ow diagram a progressive
development of several thermal and mechanical models will be
presented on the basis of different mesh densities and time steps,
aiming at reaching the convergence of the thermal and mechanical
results. In current FEA simulation the main magnitudes being monitored are:

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

i. The transient temperature distribution at the mid-length of

the treated zone, which provides evidence that the desired
maximum temperature has been reached.
ii. The temperature distribution during cooling, in order to conrm that the residual stresses refer to the totally cooled condition of the material.
iii. The transient von Mises stress at the mid-length of the treated zone, which provides an indication of the plastically
deformed zones.
iv. The residual von Mises stress after the material has cooled to
ambient temperature and prior to the application of the
external load, in order to assess the resistance of the structure against brittle fracture and susceptibility to environmental (season) cracking.
v. The thermally-induced residual displacement after the
material has cooled to ambient temperature and prior to
the application of the external load.
vi. The nal von Mises stress, which provides the operational
stress.
vii. The nal displacement of the structure, which provides its
shape during operation.
4.1. Phase I: preliminary thermal and mechanical analysis
4.1.1. The linear loop
Phase I pertains to the preliminary investigation carried out
prior to the main analyses (phases II and III) aiming at obtaining
a general idea of the model behaviour, checking the feasibility of
the analysis and getting feedback regarding the time required to
complete the analysis and computer resources needed. The preliminary analysis is considered to be very important, as it is worth
knowing whether the analysis lasts a day or a week or whether
the non-linearities encountered will allow the completion or
cancellation of the simulation. Phase I contains two sub-steps,
the preliminary linear and non-linear thermomechanical analyses.
The ow diagram of phase I is shown in Fig. 3.
The preliminary linear transient thermal analysis, followed by a
linear transient (or static) mechanical analysis including external
forces and thermal stresses at several time steps, can provide some
answers to all of the questions raised, namely the location of areas
of greatest interest, mesh density and time step magnitude. At this
linear sub-step it is important to mention that both hand or analytical solutions and user experience can help determine the areas of
greatest importance or interest and can thus contribute to the
selection and estimation of a proper initial mesh density for both
models. As a general guide, a ner mesh is required in thermomechanical modelling at areas close to any heat input, heat sinks,
geometrical discontinuities and in areas of material, boundary and
geometrical non-linearities. Thus, the initial design of the mesh
density must take into account at least the aforementioned factors.
Initial time step selection is also discussed in [25,26]; a relative
small value has been suggested in previous sections in order to
capture the linear transient heat transfer phenomena to a satisfying level. In the case of welding, for example, the speed of the
welding torch and the size of the arc distribution selected for modelling the heat input [40] can help determine an initial value of the
time step. Notice that, in most thermo-mechanical simulations
(excluding repair welding of dynamically excited structures),
modal effects are not taken into account, which facilitates initial
time step estimation [13,22,26]. It has to be emphasized here that
at this step it is not important to accurately and precisely estimate
the aforementioned initial values. On the basis of the ow diagram
shown in Fig. 3, the repetitive execution of the linear models will
nally lead to the time step value and mesh density that provide
satisfying convergence of the preliminary results. Normally a gradual decrease of the initial time step and a mesh renement at the

293

critical areas should be applied. Notice that, as stated previously,

this sub-step aims at obtaining a general idea of the model behaviour; thus a few repetitive executions should sufce.
In our example, the linear sub step has started by employing an
initial time step of Dt = 1 s and a coarse uniform mesh containing
NE = 4208 elements. From the executions of the linear thermal
analyses it was derived that the maximum thermal gradients
appear close to the weld line (line AB, see Fig. 1), whereas the
maximum stress gradients derived from the linear static stress
analysis, including both external forces and thermal stresses at
specic time steps, appear along the weld and at the triangular
reinforcing web. After the termination of the linear sub-step, the
nal locally rened mesh contained NE = 6079 elements and the
nal time step had been gradually reduced to Dt = 0.1 s. The loop
of the linear step was terminated when the relative difference of
the maximum thermally-induced residual displacement (prior to
the application of the external pressure) between two successive
analyses was less than 20%. The same criterion was also applied
for the von Mises stress and the maximum temperature developed
at mid-length of the treated zone. The latter value of 20% is larger
compared to the desirable value of 10% that was described in 3
but is considered as sufcient for the preliminary stage of the analysis. At this point, the maximum temperature difference of every
node between all successive time steps was less than 85 C which
refers to about 30% of DTtr (which is larger compared to the desirable value of 10%). The nal linear analysis results are depicted in
Fig. 4.
4.1.2. The non-linear loop
Following the ow diagram depicted in Fig. 3, the initial execution of the non-linear analyses should be carried out on the basis of
the time step value and mesh density derived from the previous
linear sub-step. Thus, the thermal model is executed followed by
the execution of the non-linear mechanical one. For this reason,
the calculated temperatures obtained from the thermal analysis
are imported into the mechanical model; the external forces are
then applied, after the weld treated bracket has cooled to room
temperature. This staggered execution is performed a few times
(see also Fig. 3), each time with a modied time step and mesh
density. Normally a gradual decrease of the initial time step and
an increase of mesh density at the critical areas should be applied.
Notice that, at this point, the model modications are performed
on the basis of both the thermal and the mechanical results, as
the common mesh density and time step must take into account
both heat ux and stress gradients. At the end of this loop the
model is expected to be able to model the non-linear transient
phenomena, to some extent at least, as it contains a crude mesh
that is locally rened to a small degree close to the areas of high
stress and heat ux gradients. It should be emphasized at this point
once again that high accuracy of temperatures and displacements
is not required at this phase (accuracy issues of the calculated
results will be discussed in the following phases) and, therefore,
a relative small number of loops is suggested. At the end of phase
I, apart from the preliminary level of the results convergence
obtained, the user can reach some basic conclusions regarding
the demands of the overall numerical simulation, computer
resources, time requirements and cost. It is also important to mention that, at this stage of the ow diagram, it is up to the analyst to
decide whether to undertake the simulated project or not.
In our example, the nal thermal and mechanical model at the
end of phase I contained NE = 7587 elements, whereas the time
step of the analysis was equal to Dt = 0.08 s. The loop of the nonlinear step was terminated when the relative difference of the
maximum thermally-induced residual displacement between two
successive analyses was less than 15%. The same criterion was also
applied for the von Mises stress and the maximum temperature

294

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

Fig. 3. Flow diagram of phase I.

developed at mid-length of the treated zone. Notice that at this

point, the relative difference approximates the target value of
10% that is required to terminate the convergence study but is still

not acceptable in engineering terms. At the end of the non-linear

sub-step the maximum temperature difference of every node
between all successive time steps was equal to 75 C which refers

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

295

Fig. 4. The results at the end of the linear loop (NE = 6079 elements, Dt = 0.1 s). (a) Transient temperature at the mid-length of the treated zone (Tmax = 864 C, Tmin = 25 C),
(b) temperature distribution during cooling (Tmax = 33 C, Tmin = 25 C), (c) von Mises static stress at the mid-length of the treated zone (rmax = 1063 MPa, rmin = 0 MPa), (d)
von Mises static stress due to the applied operational pressure only (rmax = 205 MPa, rmin = 0.5 MPa), (e) nal displacement magnitudes due to the applied operational
pressure only (dmin = 0 mm, dmax = 0.515 mm, scale factor 10).

to about 25% of DTtr. The maximum von Mises stress difference of

every node along the treated zone between all successive time
steps was equal to 27 MPa which refers to 7% of the yield stress.
The results of the non-linear analysis are illustrated in Fig. 5. From
the results presented previously it is deduced that further executions are required until the termination of the project as the convergence of the results being monitored is larger than 10% and
the temperature and the stress differences between all successive
time steps are larger than 10% of DTtr and 5% of yield stress
respectively.
The basic differences between the model shown in Fig. 4 and
the updated one shown in Fig. 5 are concentrated around the areas
of high gradients, like those at both sides of the weld AB, at the
heat-affected zone and around points B and C (see also Fig. 1).
More specically, the model shown in Fig. 5 contains a ner mesh
along both sides of weld AB at its heat affected zone, along the
free edge of the triangular web and close to points B and C, where
higher thermal and stress gradients develop.

4.2. Phase II: nalizing the thermal model and obtaining the thermal
results
Phase II pertains to the main analysis of the thermal problem. It
aims at calculating the transient and residual temperature distribution of the structure under investigation. The ow diagram of
phase II is shown in Fig. 6.
At the beginning of this phase, the thermal model is executed
using the mesh density and the time step derived from phase I.
Depending on the convergence criteria, the models mesh is progressively rened until the desired level of accuracy in the areas
of interest is attained [33]. For example, the gradual increase of
mesh density as we approach the weld line AB will allow better
estimation of the size of the heat affected zones where phase
transformations occur during the overall treatment. Note that
the mesh of the thermal model must also be rened but to
a lesser degree at areas of high stress gradients of the mechanical model, as observed at the end of phase I. Notwithstanding

296

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

Fig. 5. The results at the end of phase I (NE = 7587 elements, Dt = 0.08 s). (a) Transient temperature at the mid-length of the treated zone (Tmax = 954 oC, Tmin = 25 C), (b)
temperature distribution during cooling (Tmax = 34 C, Tmin = 25 C), (c) transient von Mises stress at the mid-length of the treated zone (rmax = 360 MPa, rmin = 3 MPa), (d)
von Mises residual stress (rmax = 393 MPa, rmin = 0.5 MPa), (e) nal von Mises stress (rmax = 382 MPa, rmin = 0.6 MPa), (f) nal displacement magnitudes (dmin = 0 mm,
dmax = 0.522 mm, scale factor 10).

that the latter local mesh renement does not necessarily contribute to the accuracy of the thermal analysis results, it will contribute to the faster solution and convergence of the mechanical
model carried out in phase III. As long as thermal results convergence has been attained for specic mesh density, the thermal
model is re-executed by reducing the time step in order to conrm that changes of the time step do not signicantly affect the
temperature results. This procedure may require a few more
loops to complete in order to provide a better estimation of
the maximum temperatures reached in the heat affected zones
and to calculate the cooling rates in the transformation areas,
necessary for the post thermal-stress analysis (phase III). At this
stage of phase II, further remeshing or time step reduction are
not expected to strongly affect the results, thus the nal thermal
results, such as maximum temperatures and cooling rates, can
be obtained. Notice that mesh renement and step time reduction can be performed simultaneously for the case of analysts

that are very experienced with thermomechanical modelling. Here

it has to be emphasized that the thermal results include the temperature ranges and time steps of any microstructural transformations realized in specic areas of the model. These microstructural
changes can play an important role in the mechanical response of
the structure analysed in phase III; thus this stage requires maximum attention [15,39]. It should be mentioned at this point
that normally a distinct or some deviation (if any) is expected
between numerical and experimental thermal results due to the
unknowns involved in the analysis (e.g. heat input, arc efciency,
material properties, heat loss); thus adaptation or calibration of
the thermal model may be necessary, see Refs. [4,41]. The adaptation procedure may require a few more re-executions of the thermal model with modied input data. These adapted thermal
numerical results, later used for post-mechanical analysis (phase
III), are expected to more accurately address the problem of the
transient and residual response of the structure.

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

297

Fig. 6. Flow diagram of phase II.

On the basis of the aforementioned discussion, the mesh of our

example required severe renement along both sides of weld AB
and its heat-affected zone in order to capture the high temperature
gradients and cooling rates (AB line, see Fig. 1). Local remeshing
has also been performed along the webs free edge and areas B
and C, where high stress gradients appeared during the end of
phase I. The latter remeshing is carried out in order to prepare
the mesh density for the mechanical analysis that will follow in
phase III. The mesh density, consisting of NE = 20,080 elements at
the end of phase II and the nal non-linear thermal analysis results
obtained using Dt = 0.037 s are presented in Fig. 7. The nal loop of
phase II was terminated when the relative difference of the maximum temperature at mid-length of the treated zone between two
successive analyses was less than 10%. The latter value satises the
criterion that was described in 3 for the termination of the
thermal part. At the end of phase II, the maximum temperature
difference of every model node between all successive time steps

was equal to 30 C (or 10% of DTtr) which is also within the desirable range. From the results depicted in Fig. 7 it is also deduced
that the maximum temperature calculated by the software
(Tmax = 1390 C) was not above the melting point of the material
which is equal to 1537 C [42]. Thus no melted zone was created.
If the actual treatment had resulted in the formation of a melted
zone then the thermal model developed during this phase should
have been further adapted with respect to the experimental or
actual results (if any), as indicated in Fig. 6 in order to compensate
for the unknown parameters that are involved in the numerical
analysis. The duration of the analysis of the nal converged
thermal model (single run) was 3.5 CPU-hours2 whereas its size
was equal to 3 GB.

2
CPU Intel Core i3 M350 @ 2.27 GHz (dual core, dual thread), RAM 4 GB, HDD,
Windows 64 bit OS.

298

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

Fig. 7. The nal mesh density of the thermal model as obtained at the end of phase II (NE = 20,080 elements, Dt = 0.037 s). (a) Transient temperature at the mid-length of the
treated zone (Tmax = 1390 C, Tmin = 25 C), (b) temperature distribution during cooling (Tmax = 39 C, Tmin = 25 C).

4.3. Phase III: nalizing the mechanical model by modifying the

thermal model
Phase III aims at nalizing the mechanical model for the
mechanical analysis. Contrary to the procedure described in phase
II for the nalization of the thermal part, the modications per-

formed on the mechanical model during this phase require the

thermal model to be modied as well. Thus the user should modify
both the thermal and the mechanical models which is a time-consuming and computer-demanding procedure. The latter strongly
affects the overall analysis cost. The ow diagram of phase III is
depicted in Fig. 8.

Fig. 8. Flow diagram of phase III.

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

Analysis starts by executing the mechanical model on the basis

of the mesh density and time step derived at the end of phase II for
the thermal model. Note that, at this point, the mesh density as calculated from the thermal model can capture to some degree the
stress gradients of the mechanical analysis. As long as the mechanical model does not meet the convergence criteria, it is further
modied and executed on the basis of a more detailed mesh in
the areas of interest. Again, as suggested in phase II, a relative wide
range of different mesh densities must be tested [33]. For this purpose, the thermal model is necessarily further rened and executed
as well, focusing also on the areas of high stress gradients derived
from the mechanical analysis. Again, attention of mesh renement
is paid to the areas of the isolated boundary conditions, the nodes
where the heat input is delivered, the geometrical discontinuities
of the model, the areas of microstructural transformations and
the areas where non-linearities are observed. Please note that the
modication and re-execution of the thermal model should not
provide better accuracy of thermal results in engineering terms,
as the nal temperature results have already been obtained at
the end of phase II; thus, the thermal model is further modied

299

only in order to provide the appropriate basis for the execution

of the mechanical analysis. It ought to be mentioned here that
the current loop offers a good chance for the researcher to conrm
that the thermal model has met the convergence criteria required
for the termination of phase II.
As long as the mechanical results have converged to the desired
level (which implies for example that maximum displacements
stabilization has been observed with respect to mesh renement
in a displacement analysis), both thermal and mechanical models
can be further executed with their time step modied in order to
check the sensitivity of the mechanical results on time step. The
latter loop process may require a few successive executions until
the nal time step values are stabilized. Simultaneous mesh renement and time step reduction can also be applied for the case of
expert analysts. At the end of this phase, mechanical results like
transient displacements and stresses and residual stresses can be
obtained. Once again, note that at the end of this phase the thermal
model has been further modied but the nal thermal results have
already been obtained at the end of phase II. The thermal model
obtained at the end of phase II should provide almost the same

Fig. 9. The nal mesh density of the mechanical model as obtained at the end of phase III (NE = 42,773 elements, Dt = 0.025 s). (a) Transient temperature at the mid-length of
the treated zone (Tmax = 1516 C, Tmin = 25 C), (b) temperature distribution during cooling (Tmax = 39 C, Tmin = 25 C), (c) transient von Mises stress at the mid-length of the
treated zone (rmax = 385 MPa, rmin = 0 MPa), (d) von Mises residual stress (rmax = 423 MPa, rmin = 0.2 MPa), (e) nal von Mises stress (rmax = 407 MPa, rmin = 0.4 MPa), (f)
nal displacement magnitudes (dmin = 0 mm, dmax = 0.55 mm, scale factor 10).

300

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

accuracy in engineering terms as the one derived at the end of

phase III, but due to the lesser degree of freedom it contains, it is
quicker to execute and can thus be used for further investigation.
In our example, strong stress gradients that appeared during
the mechanical analysis along the free edge of the web, on both
sides of welds AB and AC and at the geometrical discontinuities (areas B, C) have led to further local mesh renement of
the model. The model at the end of phase III containing
NE = 42,773 elements and the nal non-linear analysis results
obtained using Dt = 0.025 s are shown in Fig. 9. The nal loop
of phase III was terminated when the relative difference of the
maximum thermally-induced residual displacement between
two successive analyses was less than 8%. The same criterion
was also applied for the von Mises stress and the maximum
temperature developed at mid-length of the treated zone. At
the end of phase III, the maximum temperature difference of
every model node of the thermal analysis between all successive
time steps was reduced to 13 C (or 4.3% of DTtr) whereas the
maximum von Mises stress difference of every node along the
treated zone between all successive time steps of the treatment
was equal to 12 MPa (or 3% of the yield stress). From the above
it is derived that all criteria described in 3 were satised; thus
termination of the overall analysis was well applied. Here it has
to be mentioned, that the converged model presented in Fig. 9 is
able to capture the residual stresses developed along the treated
weld (see Fig. 9d). These stresses are also visible in the operational condition after the external pressure has been applied
(see Fig. 9e). Furthermore, as it is derived from the thermal
results, the difference of the calculated maximum temperatures
at mid-length of the treated zone between the thermal analyses
of phases II and III (1390 C and 1516 C respectively) was less
than 10%, which on the basis of the criteria described in 3 conrms the termination of phase II. Of course, in case of more
strict criteria (for example 5% of temperature differences) phase
II should have been further executed. The duration of the analysis of the nal converged mechanical model (single run) was 333
CPU-hours (Footnote 2), whereas it size was equal to 98 GB.

iii. Maximum von Mises stress difference of every node along

the treated zone between all successive time steps equal to
or smaller than 5% of the yield stress of the material.
It should be mentioned at this point that the overall simulation
described in Section 4 could have been performed directly in a single phase by employing a very dense mesh throughout the whole
model and a very small time step for the entire simulated time.
This cursory methodology, although being direct and simple,
does not provide evidence of results convergence, is not feasible
for large scale models and requires the largest computer resources
with respect to memory capacity and running CPU-hours. On the
other hand, it should be also stated that the proposed ow diagram
is certainly not unique and converged results may have been
obtained by carrying out a sensitivity analysis probably in a different manner instead of using the aforementioned diagram as a
systematic solution guide. It is again emphasized that there is very
limited literature available on how to perform a systematic model
development for thermomechanical simulations.
Concluding, in this paper a practical ow diagram for the systematic model development and solution of complex non-linear
thermo-mechanical nite element analysis models is presented.
The proposed diagram consists of three phases. In phase I, a preliminary thermo-mechanical analysis is carried out in order to get
an idea of the model behaviour, cost and feasibility of the overall
analysis. During this phase, both thermal and mechanical models
are progressively modied until a preliminary level of results convergence is met. In phase II the nal thermal model is developed in
full, taking also into account the mechanical results encountered at
the end of phase I, whereas in phase III the nal mechanical model
is generated on the basis of a continuously modied thermal
model. The proposed procedure, which has been presented in the
form of a ow diagram, allows for the gradual output of the numerical results (preliminary results, thermal results, mechanical
results), assuring at the same time that these results are the outcome of converged analyses. The gradual output of the numerical
results is an important issue for large-scale simulation projects;
whereas converged analyses is evidence that the development of
the numerical models has been run on the basis of modelling laws.

5. Discussion and conclusions

Acknowledgements
In conclusion, on the basis of the proposed diagram the analyst
has managed to determine the exact model density and time step
(that were unknown at the end of phase I, see Fig. 4) in order to
accurately capture the thermo-mechanical response of this specic
treatment by performing the minimum number of computer runs
(end of phases II and III, see Figs. 7 and 9 respectively) while being
able to provide a preliminary answer with respect to the analysis
feasibility at the end of phase I. Furthermore, in the example presented in this study, focus was given only on the seven items
described at the beginning of 4. Focusing on different magnitudes
and at different time steps may have required more or less number
of numerical executions compared to these applied here (phases I
III). In any case, the criteria of convergence acceptance strictly
depend on the researcher. Summarizing, in current paper the
authors have adopted the following criteria for the termination
of the proposed ow diagram:
i. Relative difference of the maximum thermally-induced
residual displacement between two successive analyses less
than 10%. The same criterion was also applied for the von
Mises stress and the maximum temperature developed at
mid-length of the treated zone.
ii. Maximum temperature difference of every model node
along the treated zone between all successive time steps
equal to or smaller than 10% of DTtr.

The authors acknowledge that the necessity of a ow diagram

for the systematic development and solution of complex non-linear thermomechanical models arose during the work within the
EU funded project Shipbuilding Low Cost, Versatile and Safe
Welding by YAG Laser Applications SHIPYAG (contract number
G3RD-CT-2000-00251). The authors gratefully acknowledge that
part of the work presented in this paper was funded by the aforementioned European program.
References
[1]
[2]
[3]
[4]
[5]
[6]

[7]
[8]
[9]
[10]
[11]
[12]
[13]

D. Radaj, Heat Effects of Welding, Springer-Verlag, Berlin, Heidelberg, 1992.

L.-E. Lindgren, J. Therm. Stresses 24 (2001) 141192.
L.-E. Lindgren, J. Therm. Stresses 24 (2001) 195231.
L.-E. Lindgren, Comput. Methods Appl. Mech. Eng. 195 (2006) 67106736.
K. Satoh, Trans. Jpn. Weld. Soc. 3 (1972) 125134.
L. Andersen, Residual Stresses and Deformations in Steel Structures, PhD
Thesis, Department of Naval Architecture and Offshore Engineering, Technical
University of Denmark, 2000.
H.B. Clausen, Plate Forming by Line Heating, PhD Thesis, Department of Naval
Architecture and Offshore Engineering, Technical University of Denmark, 2000.
P. Tong, T.H.H. Pian, Int. J. Solids Struct. 3 (1967) 865879.
P. Tong, T.H.H. Pian, Int. J. Solids Struct. 5 (1969) 463472.
R. Scott, Numer. Math.-Numer. Math. 21 (1973) 317327.
W. Lord, J.H. Hwang, Comput. Electr. Eng. 1 (1974) 513520.
G.E. Ramey, N. Krishnamurthy, Comput. Struct. 4 (1974) 11851206.
K.J. Bathe, E.L. Wilson, Numerical Methods in Finite Element Analysis,
Prentice-Hall Inc., New Jersey, Canada, 1976.

D.G. Karalis et al. / Computational Materials Science 95 (2014) 288301

[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]

[30]
[31]

G. Petruska, Comput. Math. Appl. 4 (1978) 6771.

M. Berkovic, Comput. Struct. 10 (1979) 195202.
S.E. Laux, R.J. Lomax, Solid-State Electron. 24 (1981) 485493.
G.A. Mohr, I.C. Medland, Eng. Fract. Mech. 17 (1983) 481491.
R.J. Melosh, Finite Elem. Anal. Des. 7 (1990) 115121.
J.H. Keyak, H.B. Skinner, J. Biomed. Eng. 14 (1992) 483489.
L.W. Marks, T.N. Gardner, J. Biomed. Eng. 15 (1993) 474476.
R.J. Melosh, Finite Elem. Anal. Des. 13 (1993) 105113.
C. Spyrakos, Finite Element Modelling in Engineering Practice, Algor
Publishing Division, Pittsburgh, PA, 1994.
R.S. Bains, J. Sugimoto, Eng. Anal. Boundary Elem. 14 (1994) 267275.
L. Demkowicz, Comput. Math. Appl. 27 (1994) 6984.
Y.Y. Zhu, S. Cescotto, Comput. Struct. 53 (1994) 275304.
R.D. Cook, Finite Element Modelling for Stress Analysis, John Wiley & Sons, Inc.,
1995.
P. Morin, R.H. Nochetto, K.G. Siebert, SIAM Rev. 44 (2002) 631658.
J.-F. Hiller, K.-J. Bathe, Comput. Struct. 81 (2003) 639654.
Q. Gu, J.P. Conte, Convergence studies on nonlinear nite element response
sensitivity, in: Proceedings of the Ninth International Conference on
Applications of Statistics and Probability in Civil Engineering (ICASP9), San
Francisco, USA, July 69, 2003.
ABAQUS Version 6.5, Analysis Users Manual, ABAQUS Inc., 2004.
W. Huang, Int. J. Numer. Anal. Modell. 2 (2005) 5774.

301

[32] C. Carstensen, Appl. Numer. Math. 59 (2009) 21192130.

[33] H. Schmidt, T. Alber, T. Wehner, R. Blakytny, H.-J. Wilke, J. Biomech. 42 (2009)
19261934.
[34] C. Veyhl, I.V. Belova, G.E. Murch, A. chsner, T. Fiedler, Finite Elem. Anal. Des.
46 (2010) 371378.
[35] K.-J. Bathe, P.-S. Lee, Comput. Struct. 89 (2011) 285301.
[36] D. Berglund, H. Alberg, H. Runnemalm, Finite Elem. Anal. Des. 39 (2003) 865
881.
[37] C.D. Elcoate, R.J. Dennis, P.J. Bouchard, M.C. Smith, Int. J. Press. Vessels Pip. 82
(2005) 244257.
[38] H.J.M. Geijselaers, Numerical Simulation of Stresses due to Solid State
Transformations. The Simulation of Laser Hardening, Thesis University of
Twente, Enschede, Hellendoorn, The Netherlands, 2003.
[39] D.G. Karalis, D.I. Pantelis, V.J. Papazoglou, Theoret. Appl. Fract. Mech. 54 (2010)
3743.
[40] J. Goldak, M. Asadi, R.G. Alena, Comput. Mater. Sci. 48 (2010) 390401.
[41] D.G. Karalis, V.J. Papazoglou, D.I. Pantelis, The accurate prediction of the
thermal response of welded structures based on the nite element method:
myth or reality?, in: Proceedings of the First International Conference of
Engineering against Fracture, Springer Science + Business Media, BV, 2009, pp.
513529.
[42] G.L. Huyett, Engineering Handbook, G.L Huyett, 2004.