Microelectronics Reliability 119 (2021) 114077

Contents lists available at ScienceDirect

Microelectronics Reliability
journal homepage: www.elsevier.com/locate/microrel

Co-simulation of MATLAB and ANSYS for ultrasonic wire bonding

process optimization
Reinhard Schemmel *, Viktor Krieger, Tobias Hemsel, Walter Sextro
Chair of Dynamics and Mechatronics, Paderborn University, Warburger Str. 100, 33098 Paderborn, Germany

A R T I C L E I N F O A B S T R A C T

Keywords: Ultrasonic wire bonding is a solid-state joining process, used in the electronics industry to form electrical con­
Ultrasonic heavy wire bonding nections, e.g. to connect electrical terminals within semiconductor modules. Many process parameters affect the
Co-simulation bond strength, such like the bond normal force, ultrasonic power, wire material and bonding frequency. Today,
ANSYS
process design, development, and optimization is most likely based on the knowledge of process engineers and is
MATLAB
Process optimization
mainly performed by experimental testing. In this contribution, a newly developed simulation tool is presented,
Friction coefficient to reduce time and costs and efficiently determine optimized process parameter. Based on a co-simulation of
Copper-copper MATLAB and ANSYS, the different physical phenomena of the wire bonding process are considered using finite
Aluminium-copper element simulation for the complex plastic deformation of the wire and reduced order models for the transient
dynamics of the transducer, wire, substrate and bond formation. The model parameters such as the coefficients of
friction between bond tool and wire and between wire and substrate were determined for aluminium and copper
wire in experiments with a test rig specially developed for the requirements of heavy wire bonding. To reduce
simulation time, for the finite element simulation a restart analysis and high performance computing is utilized.
Detailed analysis of the bond formation showed, that the normal pressure distribution in the contact between
wire and substrate has high impact on bond formation and distribution of welded areas in the contact area.

1. Introduction contribution focuses specifically on ultrasonic heavy wire bonding.

Approximately 93% of all semiconductor packages are manufactured
Ultrasonic wire bonding is a well-established process used in the using ball-wedge bonding [1]; nearly all dynamic random access
electronics industry to form electrical connections within the important memory (DRAM) chips and most ICs are assembled by ball-wedge
growing market of power modules. The wire bonding process in bonding. In addition, fine wire bonding in general is used for low-
particular is divided into three major processes: thermo compression current connections in devices like lead-frame packages, antenna de­
bonding (TC) with only heat energy (300 ◦ C to 500 ◦ C), thermosonic signs for CMOS wafers where the antenna is designed by the loop of the
bonding (TS) with combined heat (100 ◦ C 150 ◦ C) and additional ul­ wire bond, or disk drives [4,5]. Although most of the annually produced
trasonic energy, and ultrasonic bonding (US) with only ultrasonic en­ wire bonds are made by ball-wedge bonding processes, heavy wire
ergy at room temperature used for wire bonding. A distinction is made wedge-wedge bonding is part of an important growing market of power
between fine wire (d from 12 μm to 75 μm) and heavy wire (d from 50 modules to connect the upper side of chips to substrates and connect
μm to 600 μm) for the different wire diameters d and between the two substrates to terminals. From an application point of view, more than
wire bond forms namely ball and wedge wire [1,2]. In Table 1, the three half of the market is in industrial motor drives, followed by traction
wire bond processes are assigned to the different wire forms and the wire power supplies. Strongly growing market segments are power modules
materials typically used. In fine wire bonding, TC, TS, and US bonding for renewable energies (wind turbines and solar modules) and auto­
are used, depending on the application. In heavy wire bonding on the motive applications [6]. Ultrasonic heavy wire bonding has some ad­
other hand, only US bonding is used in production, although thermo­ vantages over other technologies [7]: the bonding machine with its
sonic heavy wire bonding experiments have been reported lately by three positioning axes can handle high tolerances in height, position and
Hunstig et al. [3] who used a laser for additional heat generation. This orientation of the bond, there is no heat affected zone, and US heavy

* Corresponding author.
E-mail address: [email protected] (R. Schemmel).

https://doi.org/10.1016/j.microrel.2021.114077
Received 13 November 2020; Received in revised form 28 January 2021; Accepted 23 February 2021
Available online 10 March 2021
0026-2714/© 2021 Elsevier Ltd. All rights reserved.
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

Table 1 to diffusion of wire and substrate materials.

The three different process technologies of wire bonding [1,2]. Main process parameters which can be set at the bonding machine
Wire bond form/diameter Processes/typical materials are the bond normal forces FTD and FBN, the ultrasonic voltage uT (in
some cases ultrasonic current iT), the vibration frequency fT, and the
Ball-Wedge (12.5μm < d< 75μm) TC or TS/Au
Wedge-Wedge (d< 75μm) TS or US/Au, Cu, AlSi1 bond duration tB. Additionally, material properties of chosen wire and
Wedge-Wedge (75μm <d< 600μm) US/Al, Cu substrate material like yield strength, surface roughness, and plastic
deformation behavior under influence of ultrasound are highly related
to bond quality, which is determined by destructive testing (shear or pull
wire bonding is a very flexible, and at the same time very cost effective tests) and optical characterization (shear code, shape of the wire and
technology. optical images of cross sections), [10]. Besides the process parameters
Especially for power modules of wind turbines high requirements on and material properties, the bond tool design and the clamping of the
operating lifetime, long term availability, and high reliability must be substrate play an important role.
fulfilled. In automotive applications a trend of steadily increasing the For such a complex process with many parameters affecting the wire
transmittable electrical power and reducing the size of the electronic bond quality it is essential to identify key mechanisms for process
parts for lightweight can be seen [8]. As a result, new challenges are the design, development, and optimization steps described in Fig. 2. Today,
rising demands on the electrical connection with larger junction tem­ process design and development is most likely based on the knowledge
peratures and higher mechanical stresses in the bond connections. of process engineers from previous experimental investigations of
Electrical parts like LED modules, main inverters, and the battery comparable processes. Process characterization is typically done by
management system in automobiles, are highly stressed by temperature experimental investigation using the methods of Design of Experiment
changes and harsh vibration levels [9]. (DoE) to determine the process window. During bond process mea­
To adapt on these demands, new bond processes are developed by surement data at the bonding machine (typically uT, iT, fT, and zWD) is
either changing the wire and substrate materials or the size of the wire. stored which can be taken into account for interpretation of results from
In process development, the wire material and diameter are adapted to the DoE. If needed, process characterization is supported by non-
the specifications in terms of maximum current, switching frequency, destructive optical analysis and additional measurements of the vibra­
and junction temperature of the application. When new heavy wire bond tion of bond tool and substrate, and of bond normal and tangential force
processes are developed, process parameters of the ultrasonic excitation, or by high speed camera videos during the bond process [11,12]. Since
construction of the modules and many more have to be adapted to new many parameters have to be investigated with suitable sample size,
materials and dimensions of the wire and substrate. process characterization is time and cost consuming. Some process pa­
In heavy wire bonding, the wire is clamped with a bond tool by a rameters like the bonding frequency fT typically are not changed,
normal force and ultrasonic vibration is transmitted through the wire because a new bond head including the ultrasonic transducer is needed
into the interface of wire and substrate, see Fig. 1. The process of bond to change the frequency, leading to high additional costs. For process
formation is highly complex and is typically divided into different optimization, results of bond quality, machine data, and the additional
phases: In the pre-deformation phase, a static load FTD is applied to the measurements are analysed to determine optimal process parameters in
bond tool and an initial contact between wire and substrate is created. In terms of high bond strength, low variance, short process time, and
the following phases, ultrasonic vibration xBT is applied to the wire; additional goals like low mechanical stress induced by vibration to
within the cleaning-phase contaminations like oxide layers are removed reduce the risk of substrate damage.
from the contact zone and the surface roughness is reduced by frictional To reduce time and costs but efficiently determine optimized process
processes, thus the effective contact area is increased. In the deformation parameters, process design, characterization, and optimization should
phase, large plastic deformation of the wire occurs because of the so- be enhanced by simulation. By using different simulation tools like
called ultrasonic softening effect and the contact area between wire thermal and mechanical finite element simulation and multi-scale
and substrate further increases. In the interdiffusion phase the intense models of the bond process which have been presented by several re­
contact combined with normal and shear stresses in the interface leads searchers [13–19] in the past, the effects of process parameters on bond

Fig. 1. Ultrasonic transducer system with the tool tip and the interface bond Fig. 2. Development process for ultrasonic wire bonding, enhanced
phases of the faying surfaces of wire and substrate. by simulation.

2
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

quality can be investigated early in the process design phase. Results of program. For the software tools, in this contribution, the software ver­
the experimental process characterization can be combined with simu­ sions ANSYS 18.1 and MATLAB 2018b are used. In the following sec­
lation results for the interpretation of effects of process parameters on tions, the submodels are described.
bond quality and thus to reduce the overall amount of needed experi­
mental parameter combinations and additional measurements; by that 2.1. Finite element model
the amount of experimental effort and thus time and costs can be
reduced significantly. The schematic workflow for generating the finite element model for
In this contribution, a new simulation tool based on a co-simulation the co-simulation and the mechanical model itself are shown in Fig. 4.
with MATLAB and ANSYS for enhanced process design and optimization
in ultrasonic heavy wire bonding is presented. First, the model setup of
the co-simulation is described in Section 2. The model parameters such
as the time- and material-dependent coefficients are experimentally
determined with a test rig in Section 3. The features of the simulation
tool - the interface between MATLAB and ANSYS, the restart option in
ANSYS, and high performance computing results - are presented in
Section 4. In Section 5, results of the simulation of bond formation are
presented.

2. Model setup for the co-simulation

The different physical phenomena of the wire bonding process are

considered in different submodels as shown in Fig. 3. The complex
plastic deformation behavior of the wire is modeled with finite element
method using ANSYS Mechanical. In MATLAB, the ultrasonic excitation
is calculated by Reduced Order Models (ROM) for transient simulation
of transducer, bond tool, substrate substructure vibration and to calcu­
late the bond formation and ultrasonic softening. Additionally, a
graphical user interface allows the user to choose bond parameters,
configuration settings of the simulation and post processing options. The
submodels in MATLAB and ANSYS are coupled by a CORBA-interface
between the two software tools to apply the normal force and material
softening factor of the wire material to the finite element model. As a
result of the finite element simulation, the normal force distribution in
the contact between wire and substrate and the vertical displacement
zWD of the bond tool are returned to the MATLAB simulation. In this co-
simulation, MATLAB acts as the master program and ANSYS as the slave Fig. 4. Finite element model for the co-simulation.

Fig. 3. Program structure of the co-simulation of the bond model.

3
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

The finite element model is built in ANSYS Workbench, with the benefits frequency, for which the model parameters have been identified.
of the graphical user interface and automatic mesh generation. The Alternatively a signal generator can be chosen that sets directly the
geometry is imported from an CAD-model into ANSYS Workbench and values for excitation frequency fT and vibration amplitude xBT that are
modified by merging surfaces and edges for easier contact definitions. defined by the user in the Graphical User Interface (GUI).
Within a static structural analysis, contact and boundary settings, the
mesh types and element sizes for the different geometry regions are set.
The bond tool is modeled as a rigid body, since it is made of tungsten 2.4. Wire model
carbide and compared to aluminium or copper wire material, it is about
ten times more rigid. The wire and substrate are modeled with nonlinear The wire is modeled as a rigid body with the mass mW, which is
plastic deformation behavior; on both, the wire and substrate material, connected to the tool tip via a friction contact that transmits the
the softening factor kUSS is applied from MATLAB, which lowers the tangential force Ft, TW to the wire and on the bottom side, the reaction
stress-strain curve of the material model. The contacts between bond force Ft, WS from the contact between wire and substrate acts on the wire,
tool and wire, and between wire and substrate are modeled with the see Fig. 5. The shear stiffness in direction of the vibration is represented
penalty based Augmented Lagrange contact formulation which is suitable by ct, W(zWD).
for robust convergence behavior and good accuracy for calculating the The differential equation for the wire mass is:
normal force distribution between wire and substrate. mW ẍW = Ft,TW − Ft,WS . (1)
A half model of the bond tool, wire and substrate is used with the x-z
plane as the symmetry plane; fixed support is defined for the surface on The contact between tool and wire is modeled with the coulomb
the bottom side of the substrate. The wire is bend, which corresponds to friction law:
the real shape of the wire before the touch down. In this contribution, ( )
geometries of a heavy wire bonding tool produced by Small Precision Ft,TW = μTW k FBN sign ẋBT − ẋW (2)
Tools (SPT) with the type number OSG7-16-M-3,18-2,733-G, a 400 μm
wire and a substrate with 30 μm thickness and 1.5 mm × 0.6 mm top The bond normal force FBN is magnified by the constant factor k,
surface dimensions are modeled. For meshing, linear hexahedral ele­ which mainly depends on the tool opening angle αBT of the v-groove, see
ments and inflation layer for the wire were used; the mesh size is Fig. 6. Additional information on the transmittable tangential force and
adapted to the regions of interest with higher density in the contact magnification of bond normal force depending on different bond tool
region between wire and substrate. The model shown in Fig. 4 has designs were reported by Althoff et al. in [17]. The bond normal force
62,881 nodes. FBN splits to the two contact points P1 and P2, which are merged to the
When generating the model with ANSYS Workbench, a database file one single point contact in Eq. (2). When assuming frictionless contact as
(here called M_AaaS.dat) is saved automatically in the solver directory shown in Fig. 6, the magnification factor k can be calculated by
and contains all information like the mesh node coordinates, contact k FN,P1 + FN,P2 1
formulations and material properties. The database file is then edited by = = (α ) ∀ αBT ∈ (0, π] (3)
FBN sin
BT
deleting the solution commands and editing the material definitions 2
which are set during the solution of the co-simulation by macros with
The assumption to calculate k in Eq. (3) is, that there is no friction in
inputs from MATLAB, so these commands are not needed in the database
the contact between bond tool and wire. Without this assumption, a
itself. The M_AaaS.dat file is than used to perform the co-simulation with
three-dimensional contact problem is obtained where the forces interact
ANSYS Mechanical APDL by setting the boundary conditions (bond
normally and tangentially to the tool flank. In this case no simple
normal force on the bond tool) and material softening from MATLAB
analytical solution can be found, as all three forces interact with each
using the interface described in Section 4.
other.
The neglect of friction in the contact between bond tool and wire
2.2. Ultrasonic softening model
represents a strong simplification since friction is mandatory to transmit
tangential force FTW. Therefore a finite element simulation with the wire
To calculate the material softening factor kUSS, a material model in
bond model from Fig. 4 was carried out to estimate the magnification
MATLAB is used, which can have several inputs received from trans­
factor k numerically. Three different forms of simulation were carried
ducer, bond formation, and substrate model. From current state of
out: with frictionless contact between tool and wire, with frictional
knowledge, it is known that ultrasonic softening is based on three
contact (μTW equal to 0.4 [17]) and with frictional contact and a hori­
mechanisms: stress superposition, friction reduction, and metallurgical
zontal deflection of xBT equal to15 μm; a deflection of 15 μm was chosen
volume effects, [20]. Different parameters like the excitation frequency,
to ensure, that the transition from stick to slip occurs to receive the
vibration amplitude, and even duration of ultrasonic excitation have an
maximum transmittable tangential force. The simulation was carried out
impact on the material softening. For wire bonding materials, no phys­
with a bond normal force of 15 N and copper for the wire material, see
ical material softening model is known and is currently part of research
activities. In this contribution, a simplified material model using the
force Ft, WS between wire and substrate as an input is used.

2.3. Transducer model/ultrasonic excitation

For the transducer, an equivalent model which has been presented in

[21] is used; the model describes the electromechanical behavior of a
piezoelectric element and the mechanical structure of the transducer
close to its resonance frequency and includes a Phased Locked Loop
(PLL) controller, which controls the excitation frequency fT to the
resonance frequency of the transducer. Inputs of the transducer model
are ultrasonic voltage uT and tangential contact force FT, TW between
bonding tool and wire at the tool tip. The outputs are the excitation
frequency fT of the PLL and the resulting tool tip amplitude xBT. With the Fig. 5. Model of the wire as a rigid body and the two frictional contacts be­
transducer model, the excitation frequency is fixed to the resonance tween bonding tool and wire and between wire and substrate.

4
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

and for the simulation with frictional contact and horizontal deflection k
is equal to 1.5 for αBT equal to 70◦ .
The magnification decreases for both simulations with friction and
increases for the simulation with friction and horizontal deflection
compared to the simulation with friction but without deflection. The
explanation for this behavior of k lies in the additional forces that occur
with friction: on the one hand parallel to the tool flank (in direction of ̃ y
in Fig. 6) and on the other hand in x-direction the force FTW when the
tool is deflected. Any non-zero force in direction of ̃ y also has a
component in the direction of FBN, so the magnitude of FNP, 1 in Fig. 6
will be reduced in this case. If there is no tool deflection, there is only the
force component in direction of ̃ y, whereas with a deflection, the addi­
tional force FTW in the x-direction is present. The additional force
component FTW decreases again the friction force in direction of ̃ y and
therefore increases FN, P1, because the absolute value of the frictional
force vector is limited by the value of μTW FN, P1. This is why k increases
with bond tool deflection compared to no deflection in Fig. 7a). Since the
factor k is a function of the tangentially applied force FTW and in order to
reduce the complexity of the model, the average value 1.35 between the
upper and lower limit when considering friction is assumed for k for the
following investigations; otherwise k would be time-variant and would
Fig. 6. Reaction forces between bond tool and wire for a frictionless contact oscillate with FTW causing higher numerical effort and convergence
between bond tool and wire. problems in the solution process of the model.
Since the maximum transmissible tangential force is an important
parameter of the bond tool design, the normalized maximum trans­
Table 3 for the material parameters. The results are shown in Fig. 7a): for
missible force FTW received from the simulation has been plotted in
the simulation with frictionless contact, the results from finite element
Fig. 7b); it can be seen that the maximum occurs for the standard angle
simulation and the analytical solution of Eq. (3) correlate well; for αBT
of αBT equal to 70◦ , so that this opening angle can be regarded as optimal
equal to 70◦ , k is equal to 1.74. For the simulation with frictional contact
in this respect. In fact, αBT also has an influence on other process vari­
but without any horizontal deflection the magnification k is equal to 1.2
ables such as the vertical deformation behavior measured as vertical
position zWD: the smaller αBT is selected, the faster the tool penetrates
into the wire, so that early tool/substrate contacts can occur. Therefore,
the choice of αBT is also a compromise between maximum transferable
tangential force and acceptable wire deformation behavior and it is a
complex relationship how the choice of αBT corresponds to the bond
quality [22].
The frictional contact between bond tool and wire is modeled using
the Karnopp friction model extended by the switch model presented by
Leine et al. in [23]. When modelling dry friction contacts, especially the
transition from stick to slip state is crucial for numerical efficiency
because of the discontinuity of the signum function. The Karnopp switch
model treats the system as three different sets of ordinary differential
equations (ODE): one for the slip phase, a second for the stick phase and
a third for the transition from stick to slip (algorithm 1). The constant
η << ẋBT − ẋW is defined for numerical efficiency to determine the slip
mode: the system is considered to be in the slip mode if the relative
velocity is larger than η. If equilibrium between the reaction force Ft, WS
and the breakaway friction force μTW k FBN is reached, the system is
considered to be in transition from stick to slip. During the sticking
phase, the acceleration of the wire mass is equal to the acceleration of
the bond tool tip, as both, tool tip and wire, are moving equally.
Algorithm 1. Set of three differential equations for the wire model,
using the Karnopp switching model.
l
if ∣ẋBT − ẋW ∣ > η then
2
Sliding:
⎡ ⎤
[ ] ẋW
3 ẋW ⎢ ( ( )) ⎥
=⎢⎣ 1 − F
⎥
ẍW + μ k F sign ẋ − ẋ ⎦
mD
t,WS TW BN BT W

4
else if ∣Ft, WS ∣ > μTW k FBN then
5
Transition from stick to slip:
⎡ ⎤
[ ] ẋW
6 ẋW ⎢ ⎥
Fig. 7. Results of finite element simulation for a) the factor k for different ẍW
=⎢ 1 ( ( )) ⎥
⎣ − Ft,WS + μTW k FBN sign Ft,WS ⎦
contact conditions between bond tool and wire and b) the maximum trans­ mD
mittable and normalized force FTW.

5
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

7
else simulation was performed for aluminium (Al) and copper (Cu) wire
8
Stick: materials with the material properties shown in Table 2.
[ ] [ ]
The results for the wire shear stiffness are shown in Fig. 8b). For both
9 ẋW ẋ
ẍW
= W
ẍBT wire materials, the shear stiffness increases progressively as the bonding
10
tool is lowered. When comparing the results at 190 μm to 10 μm, the
end shear stiffness of the wire at the end is about 16 times higher compared
to the beginning).
2.4.1. Wire shear stiffness The results from the FE-simulation were approximated by a poly­
The shear stiffness ct, W(zWD) of the wire - which is summarized nomial fit of second order:
within the spring in the contact between wire and substrate - depends on
ct,W (zWD ) = p1 z2WD + p2 zWD + p3 . (4)
the vertical position zWD of the bonding tool. The shear stiffness ct, W is
not directly used in the equation system of the wire model in algorithm The polynomial coefficients are listed in Table 3. Within the co-
1, but in the discretized friction model in Section 2.6 to calculate the simulation, the polynomial function is implemented in MATLAB to
tangential force Ft, WS between wire and substrate which acts on the calculate the current shear stiffness of the wire depending on the vertical
bottom side of the wire. The experimental identification of the tangen­ position zWD which is received from the ANSYS simulation.
tial stiffness of the wire is very challenging, as in experiments the stiff­
ness of the test rig impacts the measurement. For this reason the FE 2.5. Substrate model
model from Section 2.1 has been used to determine the shear stiffness of
the wire with 400 μm diameter for aluminium and copper material. For calculating the system dynamics, the substrate is modeled as a
Therefore the substrate was modeled as a rigid body, thus only the wire flexible structure. For this, state space models based on the mode su­
had flexible body behavior. In the simulation the vertical displacement perposition method are used; the modal matrix and eigenfrequencies are
zWD of the bonding tool was increased from 10 μm 190 μm in 20 μm steps derived from numerical modal analysis using a finite element model of
in relation to the initial position of the undeformed wire; for 190 μm, the wire and substrate substructure; modes for the state space model are
tool/substrate contact occurred, thus this was the maximum vertical observed by Model Order Reduction (MOR) within MATLAB using
deflection of interest. For each height the bonding tool was deflected Hankel Singular Values. Details of the MOR used in this contribution are
once sinusoidally with an amplitude of 3 μm and as a result, the described in [19].
tangential reaction force Ft, WS between wire and substrate was recorded.
To determine the tangential stiffness, the slope of the positive flank of
2.6. Bond formation model
the hysteresis was evaluated using linear regression, see Fig. 8a). The

To calculate the distributed bond strength, a discretized friction

model shown in Fig. 9 and the differences equation system in Eq. (5) are
used. A uniform mesh grid with coordinates of center points of partial
contact areas is created within MATLAB with input parameters nx, and
ny the number of partial areas in x- and y- direction respectively and lx,
and ly the outer dimensions of the grid; the input parameters define the
resolution Δx and Δy and total number nPA of the partial areas. To
calculate the partial contact normal forces FN, i, the FE-model results are
interpolated to the mesh grid of the friction model using the scatter­
edInterpolant function within MATLAB and linear interpolation method.
Between the updating time steps of the FE-model, the interpolated
normal force distribution is normalized to the current bond normal force
FBN(tSim) to satisfy static equilibrium condition. For each single point
contact, the tangential stiffness is calculated from the overall shear
stiffness ct, W(zWD) (Section 2.4.1) by dividing this stiffness with the
number of elements in contact.
The bond formation is modeled using the discretized friction model
and it is assumed that the whole interface area between wire and sub­
strate can be divided into three parts (Fig. 10):
• Oxide layer area: initial contaminations like oxides prevent the
bond formation. To remove contamination frictional work Wf is needed
to clean the surface.
• Activated area: the clean parts of the interface are activated for
bond formation. The activation state τ = Sactivated
S is increased by cleaning
and is mandatory for bond formation.
• Bonded area: frictional work in activated areas leads to bond
formation and bonding state γ = Sbonded
S is increased.
The differential equation system published in [24] has been extended
to the following differences equation system by integration over one

Table 2
Material parameter of 400 μm aluminium and copper wire.
Fig. 8. a) Hysteresis of the tangential force of the copper wire for two different Material Young’s modulus/GPa Yield strength/GPa
heights (70 μm and 130 μm) with the slope of cT, W at the left flank. b)
Al 70 41.85
Tangential stiffness for the copper and aluminium wire from the FE-simulation
Cu 100 96.40
and curve of the polynomial best fit.

6
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

Table 3
Polynomial coefficients of the fit to the shear stiffness of the aluminium and
FT,i,max = τS Δx Δy. (7)
copper wire. The bond formation between wire and substrate is modeled using a
p1/N/m/μm p2/N/m/μm p3/N/m variable coefficient of friction μ(γ)= based on the equation reported in
Al 247.37 361.12 ⋅ 102 843.83 ⋅ 103 [14]:
Cu 254.66 436.04 ⋅ 102 126.55 ⋅ 104
μ(γ) = (1 − γ)μOx + μMet γ. (8)
When modelling the bond between wire and substrate using the
coefficient of friction μ(γ) in Eq. (8), the maximum tangential force FT,
max depends on the applied normal force FN following the law FT, max =
μ(γ) FN. When using a discretized friction model with normal force
distribution, even a high coefficient of friction for bonded areas can lead
to partial slippage because of low local normal forces. For this reason,
the friction coefficient for bonded area - where no sliding is allowed - is
modeled in Eq. (9) depending on the normal force FN, i in the ith area
element to satisfy the condition FT, i(γ = 1) = τS Δx Δy. This is
mandatory to achieve constant boundary condition for bonded areas
within contact area between wire and substrate, independently from
normal force distribution FN(x, y). The coefficient of friction has an
initial value of μi, ox and with rising state γi, the coefficient of friction μi
increases to the maximum value calculated from shear strength:
( ) τs Δx Δy
Fig. 9. Discretized friction model with partial area elements, each modeled μi γi , τi , FN,i = (1 − γi − τi )μOx + τi μMet + γi . (9)
using Jenkin element.
Fn,i

3. Experimental identification of coefficients of friction

The measurement of the friction coefficients μTW between the bond

tool and wire and μWS between wire and substrate is important for the
parametrization of the friction models within the bond process model.
The aim of the measurements is to determine the coefficient of friction
for the different material pairings for 400 μm aluminium and copper
wire. For this purpose a novel test rig based on a universal testing ma­
chine - extended by a dynamical excitation using a shaker - has been
Fig. 10. Interface between wire and substrate subdivided into areas of oxide developed, see Fig. 11.
layers, activated areas and bonded areas. The crosshead of the testing machine is guided on spindles and can
be moved in vertical direction with a speed from 1 mm/min 500 mm/
min. The traverse below the crosshead is linearly guided in vertical di­
period length TP, n of the ultrasonic excitation:
rection on a steel plate, which is mounted behind the testing machine.
Δγi,n ΔWf,i,n The traverse is supported by a rod on a pressure spring above the
=β τi,n−
crosshead. Due to its own weight, the traverse achieves the equilibrium
1
Δx Δy
(5)
Δτi,n ΔWf,i,n [ ] position on the spring, so that from the touchdown between wire and
=α
Δx Δy
1 − γi,n− 1 − τi,n− 1 − Δγi,n . substrate onwards there is a linear increase in force according to the
spring characteristic; this results in a more robust and stable force
Both states τ and γ are dimensionless and the changes Δτi, n and Δγi, n control for the touchdown and bond normal force compared to a rigid
of states of the ith area element within one period of the nth oscillation coupling between traverse and crosshead. The maximum normal force is
cycle with period length TP, n is described by the right hand sides. For the determined by the dead weight of the traverse and the components
activation state τi, n only the amount of frictional work in the oxide layer mounted on it. The traverse is supported by a linear guide in vertical
area is taken into account. At the same time, when activated areas have direction, which is mounted on a steel plate behind the testing machine,
been bonded, τi, n decreases by the amount of Δγi, n. For the bonding which supports the traverse in direction of the dynamic excitation to
state γi, n, when no activated areas are left, γi, n reaches it’s maximum reduce machine vibrations. On the bottom side of the traverse, a
value of one. The coefficients α and β are model parameters, which are piezoelectric three component force sensor (Kistler 9347C) is mounted
identified from measurements. to measure the bond normal force. Additionally, the tangential force
The single friction contact of i-th partial area is modeled by Coulomb during low frequency wire bonding experiments can be measured. On
friction model coupled with a spring cj, i, called “Jenkin element” in the bottom side of the force sensor, a substrate holder is mounted, which
literature, using evolution differential equation presented in [18] with fixes the DCB using a pressure hose. When pressure is applied to the
u̇ = ẋW − ẋS : tube, it expands inside the substrate holder housing, fixing the DCB. The
( ⃒
) )⃒ FT,i ⃒m
⃒ ) substrate holder has a L-shape to make the bond tool tip visible from the
ḞT,i 1( (
= cJ,i u̇ 1 − 1 + sign u̇ FT,i ⃒⃒ ⃒ (6) front and right hand side for laser measurements and high speed camera
2 μi FN,i ⃒
recordings.
In Eq. (6), the value μi FN, i defines the maximum tangential force Ft, i Contrary to typical bonding machines, the orientation of substrate
that can be transmitted by the Jenkin element. Depending on the state γi, and bonding tool is upside down: the DCB is fixed above the bonding
the value of coefficient of friction μi changes; in case of fully bonded tool with the tool-tip pointing upwards. The bonding tool is fixed in a
partial area element, the maximum tangential force is calculated from holding adapter with two screws in a drilling. The bonding tool from
shear yield stress τY of the material using the relationship τY = 0, 5…0, Kulicke & Soffa (K&S) with the type number 153-16A-UVU31 has been
6 Rm where Rm is the tensile strength: shortened to increase the bending stiffness and next to the bonding tool a

7
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

Fig. 11. CAD image of the test rig for low frequency wire bonding experiments.

wire guide is mounted which was newly designed for the test rig; both,
Table 4
the substrate holder and the wire guide were produced using a 3d printer
Settings for the experiments for the contacts Tool/Wire (T/W) and Wire/Sub­
using polylactide (PLA). The adapter including the bonding tool is fixed
strate (W/S).
on a linear guide with flexure hinges, which are used to achieve high
fatigue strength under dynamic excitation with small vibration ampli­ Contact Material Freq./Hz FBN/N Oscillations

tudes. The linear guide is mounted on a steel plate, that is mounted on T/W Al 5.0 3.0 60
the load frame of the testing machine. Additionally, this plate is con­ T/W Cu 5.0 5.0 40
W/S Al 5.0 6.0 100
nected by two braces to the vertical steel plate behind the testing ma­
W/S Cu 5.0 4.5 100
chine, to further increase the stiffness of the testing machine in
horizontal direction to reduce machine vibration under dynamic exci­
tation. The adapter on the linear guide is connected by a rod to the the test rig on the measurements (especially on the force measurement)
flange of the shaker to excite the linearly guided bond tool. A capacitive and achieve quasi-static conditions. It cannot be denied that at higher
distance sensor is mounted underneath the traverse to measure the frequencies in the range of the bonding frequency, dynamic effects such
distance between sensor and lower steel plate of the traverse; this dis­ as an influence of the relative velocity in the contacts affect the coeffi­
tance can be considered equivalent to the vertical wire deformation zWD cient of friction. On the other hand, the measurement of forces in the
measured at the bonding machine. ultrasonic range is still part of research [25–29] and there are no cali­
To control the movement of the bonding tool and the bond normal brated force sensors for these high frequencies, so that the quasi-static
force by setting the vertical movement of the traverse, a control board friction coefficient is determined in the first step.
was built, which allows for setting different test setups. To examine both contacts independently, once the wire was fixed to
the substrate so that only the bond tool could slide on the wire and the
other time the wire was clamped in the v-groove of the bond tool only so
3.1. Experimental setup and measurements that the wire could slide on the substrate. To fix the wire on the sub­
strate, a bond loop was ultrasonically bonded and formed as low as
For measuring the coefficients of friction, the bond normal force FBN, possible; the bond tool in the test rig was than positioned in the middle
the excitation frequency, and the number of oscillation cycles can be set between source and destination bond, which fixed the wire on the
by the user with the control board; the settings for the different mea­ substrate.
surements are listed in Table 4. In general, the coefficients of friction in The amplifier of the shaker was excited with a constant voltage level
both contacts between bonding tool and wire and between wire and during the measurement, which has been adjusted before the experi­
substrate vary during the measurements due to frictional processes like ments to achieve a certain amplitude. Since the voltage level of the
fretting and cleaning processes. The number of oscillation cycles were shaker was constant, the vibration amplitude of the bond tool generally
chosen such that a stationary value for the coefficient during the mea­ decreased during the measurements, because the tangential force
surement was achieved. The excitation frequency of the bonding tool increased and with that, the mechanical resistance for the bond tool, see
was set to 5 Hz to avoid dynamical impacts of resonance frequencies of

8
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

3.2. Results

In Fig. 14 the results for the friction coefficient between bonding tool
and wire are shown; for the calculation of the friction coefficient, the
magnification factor k equal to 1.35 from Section 2.4 was considered.
For the copper wire in Fig. 14a), the mean value of μTW starts at approx.
0.47 and increases to 0.66 at the end of the measurement. The friction
coefficient for aluminium wire starts at a higher level at approx. 1.12
and decreases over the duration of the measurement to a level of approx.
0.74, see Fig. 14b). For the copper wire, after 5 s (25 oscillation cycles),
the friction coefficient reaches a value of 0.62 which is approx. 95% of
the end value. For the aluminium wire, after 10 s (50 oscillation cycles),
a value of 0.79 is reached, which corresponds to 105% in relation to the
Fig. 12. Measurement of the tangential force and vibration amplitude of the
tool tip. final value. Comparing these results with a typical ultrasonic excitation
with a frequency of 60 kHz for heavy wire bonding, the two durations
correspond to 0.4 ms and 0.8 ms respectively; typical process times for
Fig. 12. A closed loop control was not used, because the reduction of the
heavy wire bonding are between 100 ms to 500 ms. This means that even
vibration amplitude does not affect the measurement of the coefficient
if the friction coefficient between bonding tool and wire changes at the
of friction as long as the sliding mode occurs. Additionally, an open loop
beginning and reaches a stationary final value only after a few oscilla­
control is more robust and easier to use, because no outliers of the vi­
tion cycles, this value can be assumed to be constant in good approxi­
bration occur due to instabilities of a closed loop controller and no
mation. Therefore, the final values of the measurements are used for the
control parameters need to be identified.
wire model parameter μTW.
The friction coefficient is determined by evaluating the plateau of the
In Fig. 15 the results of the measurements of the friction coefficient
hysteresis; exemplary results are shown in Fig. 13 for the contact be­
between wire and substrate are shown; the results are plotted over the
tween bond tool and copper wire. In Fig. 13a), a detail view on the
dissipated frictional energy Wf per contact area A; the contact areas have
measured time series of the tangential force Ft, TW, the normal force FBN,
been measured after the experiments using digital microscopy. During
and the vibration xBT of the bond tool tip are shown. The stick slip
the first measurements, the copper wire in particular hardly deformed at
behavior can be clearly seen in the shape of the time series of the
all due to the low normal forces and only a very small line-contact area
displacement and tangential force; the displacement abruptly increases
was obtained, which varied in size and form from experiment to
when sliding occurs (tangential force reaches its maximum). The hys­
experiment. In order to achieve consistent test conditions, the wire was
teresis in Fig. 13b) relates clearly to dry friction related to the “Jenkin
therefore loaded with a static pre-load of 50 N for the copper wire and
Element” explained in Eq. (6); the tangential stiffness leads to steadily
20 N for the aluminium wire before vibration excitation, resulting in a
increasing tangential force in the sticking phase until the transition from
uniform elliptical contact area between wire and substrate; all experi­
stick to slip occurs.
ments were performed on DCB substrate. The friction coefficient for the
copper wire in Fig. 15a) initially starts at μOX equal to 0.21 and increases
up to 0.45. For the aluminium wire in Fig. 15b), the friction coefficient
has an initial value of μOX equal to 0.23 and a final value of 0.56.

3.3. Discussion

The two copper oxides, Cu2O and CuO are both very soft, [30]. The
aluminium oxide Al2O3 on the other hand is hard and brittle and has
proved to crack under the bonding load, [31]. Rabinowicz found in [32],
that when oxide hardness is no greater than three times the metal
hardness, the soft oxide layer act as solid lubricants; for the copper wire,
this ratio is less than three and for the aluminium wire, it is greater than
three. This is one possible explanation, why the friction coefficient in­
creases for copper and decreases for aluminium wire as the oxide layers
are worn out from the contact during the experiments; either way, the
end values of the measurements are used for the parametrization of the
model parameters, as they are achieved within fairly short time after 25
(copper) and 50 (aluminium) oscillation cycles.
For the contact between wire and substrate, the coefficient of friction
starts at approx. 0.2 for both wire materials. Tabor stated in [33], that
for metals, the friction coefficients can be calculated from the ratio be­
tween shear strength and yield pressure of the underlying softer mate­
rial; for ideal plastic materials, this ratio is about 0.2 which is in good
correlation to the measurements in Fig. 15. With increasing frictional
work Wf due to sliding of the wire on the substrate, the friction coeffi­
cient increases. Microscopy observation showed that the contact areas
were not fully cleaned and activated after the measurements. The con­
tact areas were classified as 50% cleaned and activated for the copper
wire and 75% for the aluminium wire. In case of the copper wire, the
Fig. 13. a) Time series of the tangential force, tool tip displacement and bond contact area was partially bright and shiny indicating fretting processes
normal force and b) hysteresis of the ratio between Ft, TW and FBN for evaluation and for the aluminium wire, a thin aluminium layer was left on the DCB
of the friction coefficient (dashed line). substrate within the cleaned and activated areas, thus these areas were

9
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

Fig. 14. Measurements of the friction coefficient between bonding tool and wire for a) the copper wire and b) the aluminium wire. The grey lines relate to the single
measurements, the black solid line to the mean time series.

Fig. 15. Measurements of the friction coefficient between wire and substrate on DCB substrate for a) the copper wire and b) the aluminium wire. The grey lines relate
to the single measurements, the black solid line to the mean time series of the single measurements.

segmented from the rest of the contact area to classify the areas using the section are summarized in Table 5.
methods described in [34]. Liu found in [35], that the coefficient of
friction for fully cleaned surface between copper and copper in air is 4. Program structure of the co-simulation
approx. 1. Assuming this value for the fully cleaned contact area and
normalizing the final value in Fig. 15a) to 1, the results correlate well to Within MATLAB, object oriented programming was used; Fig. 17
the classification of about 50% cleaned contact area. The coefficient of
friction μMet for fully cleaned contact area is 0.9 for copper and 0.75 for
aluminium wire, taking into account that about 50% respectively 75% of Table 5
the contact area were cleaned and activated. Model parameters, identified from measurements.
In Fig. 16, the results from Fig. 15 have been normalized to the Material μTW/− μOX/− μMet/− α/mm2/mJ
percentage value of cleaned and activated area which can be referred to
Al 0.74 – – –
the cleaning state γ in the bond formation model in Eq. (5). Within the Cu 0.66 – – –
bond formation model, the model parameter α is from interest, which Al – 0.23 0.75 0.025
has been identified in Fig. 16. All model parameters identified in this Cu – 0.21 0.90 0.121

a) b)
1 1
/-

0.5 0.5
WS

0 0
2 4 6 10 20 30 40
2
Wf / A / mJ /mm Wf / A / mJ /mm2

Fig. 16. Cleaning state τ for a) the copper wire and b) the aluminium wire. The slope of the dashed line relates to the cleaning coefficient α.

10
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

Fig. 17. Object structure of the bond model within MATLAB.

gives an overview and description of the objects. A graphical user Main.

1
interface is used to set the input parameters for the simulation, such like function obj = simulation(obj)
2
the solver settings (number of cores for the FE-simulation, solver type, while tSim < tB do
3
etc.), the bond parameters for a single bond simulation or for parameter obj.Bondmodel.SOC([0,TP],uT(tSim),FBN(tSim));
4
sweeps the array with parameter combinations. The settings are stored if ΔFBN > FBN, crit or ΔSSoft > SSoft, crit then
5
in the Cache.m file, which is loaded by the object Main when starting the SSoft =
6
simulation. The method simulation of the object Main runs the co- obj.Bondmodell.USS.refresh(xW(tSim),fT);
7
simulation within MATLAB; the pseudo code of the method is shown obj.Bondmodel.MAaaS.Solve_Next_LS(FBN,SSoft);
8
in algorithm 2. While the simulation time tSim is smaller than the bond else
9
duration tB, one single vibration cycle with period length TP after the No update of the finite element model necessary.
10
other is simulated. Therefore, the method SOC of the object Bondmodel is end
11
used with the input parameters simulation duration TP, ultrasonic fT = obj.Bondmodel.PLL.update(uT,iT);
12
transducer voltage uT, and bond normal force FBN to solve the ODE- obj.Bondmodel.BondFormation.update();
13
system. The FE-simulation is updated, when - compared to the previ­ end
ous updated load step - either the bond normal force FBN or the material
softening factor SSoft changed more than the criteria FBN, crit and SSoft, crit,
which are defined in the Cache.m file. When the FE-model needs to be 4.1. MATLAB-ANSYS as a server (MAaaS) interface
updated, first the object USS is refreshed to calculate the current soft­
ening factor SSoft for the wire material. Afterwards the next Load Step For the co-simulation, a Common Object Request Broker Architec­
(LS) is calculated by the method Solve_Next_LS of the object MAaaS with ture (CORBA)-interface between MATLAB and ANSYS as a server
the input parameters bond normal force FBN and softening factor SSoft. (MAaaS) was programmed in MATLAB; the AaaS feature is provided by
Before the next vibration cycle is simulated, the object BondFormation ANSYS as a MATLAB Toolbox [36] that allows local or remote client
and the object PLL are updated to calculate the growth of bond forma­ applications to access and interact with a running session of ANSYS
tion and the excitation frequency fT for the transducer based on the re­ Mechanical. The AaaS interface exists alongside the graphical user
sults of the simulated vibration cycle. interface (e.g. ANSYS Workbench or ANSYS Mechanical) and the batch
Algorithm 2. Pseudocode of the method simulation of object (obj) mode interface. It is different from batch mode in that APDL commands,
initialization, and intermediate and final result reporting can be issued

11
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

to the running session at any time, as long as the ANSYS solver has
reached a stable point, e.g. after successfully ending an equilibrium
iteration. This allows for solution steering without exiting from the
running simulation [37]. The interface was programmed in the object
MAaaS including an error handling to catch messages from the running
ANSYS session, a license manager to limit the maximum of used licenses
of ANSYS and MATLAB when using High Performance Computing
running multiple simulations in parallel, and several methods to run
ANSYS Parametric Design Language commands from MATLAB during
the simulation.

4.2. Multi frame restart

The results of the FE-Model for the set FTD value are needed as the
starting point for the co-simulation and the simulation steps before FTD
are not relevant for the wire bond simulation itself. To reduce simulation
time, the Multi Frame Restart option in ANSYS Mechanical [38] is used,
to calculate the load steps up to a maximum value of FTD, which can then
be reused for each simulation unless the fundamental simulation settings
of the FE-model like the number of cores, or solver settings are changed.
The load steps up to FTD, max are simulated once before the wire bond
simulation and the restart files for each load step are stored for the future
simulations. Using the restart option, regardless the following load steps
after touchdown, reduction of simulation duration for a wire bond
simulation using a specific restart point is always the same. Using this
restart feature, about 50% of the overall simulation duration of the Co-
simulation can be saved; especially the first load steps are time
consuming, because the contact status between wire and substrate
changes quickly with many elements coming into contact, so the time
increment for solving the load steps in the beginning is relatively small
for convergence reasons and most of the simulation time can be saved
here.

4.3. Program sequence

The program sequence for starting a AaaS session, performing a

restart for the finite element model and running the co-simulation is
shown in Fig. 18. After starting a bond model simulation, first the files
for the restart analysis are checked, if they are matching the simulation
settings; if the database of the FE-model or the number of cores have
been changed or the restart files are missing, the restart files have to be
generated before the simulation and are stored for the current and future
simulations. When the files are available for the restart analysis, the
AaaS session is started by executing a command line to run the ANSYS
Mechanical executable in the AaaS mode with specified options
including the path of the workspace for the FE-simulation, number of
cores, main memory, parallel processing method (shared memory,
distributed memory), and the paths of the input- and output-files of the
simulation. After starting the session, a keyfile including the information Fig. 18. Program sequence plan of the co-simulation to run one simulation of
for MATLAB to connect to the AaaS session is automatically created. For the bond model.
this, the Java class ORB is initialized which allows for CORBA objects
communicating with each other by connecting ‘objects making requests’ with the new input parameters; for communication between MATLAB
(clients, here MATLAB) with ‘objects servicing requests’ (servers, here and the AaaS-session, whether the macros finished successfully or not,
ANSYS Mechanical) [39]. After the MATLAB client is connected to the semaphores are used.
running AaaS session, the restart files are loaded into the workspace of
the FE-simulation to perform a restart starting from the set value of FTD.
In the final step of preparing the wire bond simulation, the object of the 4.4. High performance computing
class Main is generated and the initial values for the values for the bond
parameters and solver settings are stored in the object. Afterward, the The program structure allows for duplicating the simulation to be
wire bond simulation starts using the method described by the pseudo solved in parallel with different input parameters, by copying all model
code in algorithm 2 updating the running FE-simulation using the files to a temporary directory and running the simulations indepen­
CORBA-interface and the features of the AaaS-session to simulate the dently from each other there. A typical simulation run of the wire bond
next Load Step with updated material softening SSoft and bond normal model using a personal computer with Intel i7-4790 CPU with four cores
force FBN. For the steps of solving the next Load Step and loading the and 16 RAM can take up to approx. 8 h simulation duration for large
contact results from the solver files of the FE-simulation, macros have finite element models with more than 100.000 nodes, thus parameter
been programmed to execute the APDL commands by the AaaS session sweeps with many simulation runs are very time consuming. To reduce

12
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

computing time, paralleled High Performance Computing (HPC) is uti­

lized using the computer cluster “OCuLUS” of the Paderborn Center for
Parallel Computing (PC2) at Paderborn University [40]. On OCuLUS, the
computing resources 552 “standard nodes” with 16 cores and 64 GB
RAM each can be accessed.
On a PC (Fujitsu Esprimo E900) with Intel i7-4790 CPU with four
cores and 16 GB RAM running the simulation on a Solid State Drive
(SSD) the simulation of the wire bond model (FE-mesh with 119,677
nodes) takes about 6.95 h. Using eight cores on OCuLUS, simulation
duration can be decreased by 45% to 3.63 h; with this configuration an
Array Job - multiple single simulations (“Jobs”) running in parallel -
with 16 Jobs takes approx. 9.5 h to finish all Jobs; the simulation
duration is longer compared to the Single Job, because preparation of
the simulation, waiting for licenses, and how the Workload Manager
distributes the Jobs to the resources is different compared to the Single
Job. This means that 16 Jobs running on the OCuLUS take only 2.55 h
longer than one single simulation on the available PC with four cores
running on the SSD showing the benefits of HPC especially for parameter
sweeps; array jobs with even larger amount of single jobs are possible on
the OCuLUS without increasing the overall simulation duration of the
array job. For further detailed information on the workflow for running
the simulations on the computer cluster it is referred to the contribution
of Schemmel et al. in [41].

5. Simulation results

For the simulation, following settings of the wire bond model have
Fig. 19. Results of the distribution of bond formation: bonding state γ steadily
been chosen:
increases with the bond duration. Outer dimensions of the subplots: 1.15 mm ×
0.36 mm.
• tR = 15 ms
• tB = 300 ms
standard deviation of zWD are shown. At the beginning of the simulation,
• xBT = 3 μm (sine excitation by signal generator with 58 kHz excita­
the FE-model is updated several times in short intervals, because the
tion frequency)
bond strength increases quickly, thus the update criterion described in
• FTD = 10.4 N
Section 4 is fulfilled more often than in the later stages of the bond
• FBN = 26 N
process, when the bond formation reaches the saturation. In the begin­
○ Wire material: Copper (dia. 400 μm)
ning stage of the bond process, the wire deformation is mainly driven by
○ Finite element model: 62881 nodes
the increasing bond normal force, which is ramped up from FTD= 10.4 N
to FBN= 26 N. Afterward, FBN and xBT are constant, but due to the
In Fig. 19 the distributed bond formation in the contact between wire
increasing oscillating shear forces acting on the wire because of the
and substrate is shown at eight different time steps of the bond process.
increasing bond strength, the material softening factor SSoft increases
In the beginning stages of bond formation, it can be seen that the bond
and so does the vertical displacement zWD.
growth rate in the peripheral parts of the elliptical interface area is
In Fig. 22(a), exemplary the states γ and τ of one selected partial area
higher compared to the central area. With increasing bond duration, the
are shown. As described in Section 2.6, partial areas have to be activated
bonding state γ steadily increases in the partial areas and the contact
first before bond formation can take place and both, activation and bond
gets more homogeneously bonded over the bond duration. The distri­
formation are driven by the friction work within the partial area. First,
bution of bond formation can be explained by the inhomogeneous dis­
the activation state γ increases in the beginning of the bond process
tribution of the bond normal forces FN(x, y) in the contact shown in
while γ only slowly increases. The slope of the curve of γ rises with
Fig. 20. Especially at the beginning of the bond process, the highest
increasing activated areas which can be bonded and flattens with
normal forces occur at the periphery of the contact area and when
decreasing activated areas which are bonded. On the one hand, the
sliding between wire and substrate begins, the highest amount of fric­
decreasing activation state τ negatively affects the differences equation
tional work occurs in these parts of the interface; in the symmetry plane,
of γ (less activated areas can be bonded), and on the other hand, with
the minimum values of the normal forces occur; in these parts of the
increasing bonding state γ, the coefficient of friction increases following
contact area, the bond strength γ in Fig. 19 is the lowest. With increasing
Eq. (9), leading to sticking and less frictional work in the partial area. In
bond duration, the normal force distribution gets more homogeneous
Fig. 22(b), the calculated shear force FS is compared to the end value
and the maximum of the normal forces occurs in the middle part of the
measured by shear testing after bonding. To calculate the shear force,
contact area underneath the bonding tool; mean reason for this is the
the bonded area from the simulation is multiplied with the shear
increase of the material softening factor SSoft with rising bond strength
strength (165 MPa) of the used copper wire. The shape of the curve of
and the changing geometry of the wire because of the plastic
the shear force value over the bond duration is in good agreement with
deformation.
measurements of shear force values over the bond duration, reported in
In Fig. 21, the vertical displacement zWD of the bonding tool tip from
[24] for 300 μm copper wire.
simulation is shown and compared to measurements at the bonding
machine. The bonding experiments were carried out with the same
6. Conclusion
bonding parameters set at the bonding machine as for the simulation;
the vibration amplitude has been measured at the tool tip and the mean
The presented program structure of the co-simulation including the
amplitude over the bond duration was approx. 3 μm; in the shaded error
restart option together with high performance computing allows for
bar in Fig. 21, the mean value of 30 bonds and the boundaries of the

13
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

Fig. 20. Distribution of normal forces in the contact between wire and substrate received from the FE-model at selected updating steps of the FE-model. Outer
dimensions of the subplots: 1.15 mm × 0.36 mm.

Fig. 21. Comparison of vertical displacement zWD of the bonding tool between
simulation and measurement at the bonding machine. The load steps, when the
FE-model is updated, are marked by the red dots.

complex simulations. Parameter sweeps with many parameter combi­

nations can be simulated in short time, using the resources of a computer
cluster. Using submodels for the different physical phenomena in ul­
trasonic wire bonding is suitable for the process design process.
Changing geometries of the bonding tool, wire, and substrate, using
different materials for the wire and substrate, and varying the excitation Fig. 22. (a) States τ and γ of one selected partial area and (b) calculated shear
parameters like the vibration amplitude and excitation frequency is easy force FS and measurement of the shear force value at 300 ms (bottom).

14
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

to be done. Based on the co-simulation, overall process optimization is [11] A. Unger, Modellbasierte Mehrzieloptimierung zur Herstellung von Ultraschall-
Drahtbondverbindungen in Leistungshalbleitermodulen, Ph.D. thesis, Paderborn
possible.
University, 2017.
Further on, a modified model for the bond formation and the coef­ [12] Y. Long, Investigations on the mechanisms of ultrasonic wire bonding, Ph.D. thesis,
ficient of friction have been presented. Based on two states τ and γ, the 2019, https://doi.org/10.15488/4829.
cleaning and the bond formation itself are modeled separately; an [13] M. Mayer, J. Schwizer, Ultrasonic bonding: Understanding how process parameters
determine the strength of au-al bonds, Proc. International Symposium on
extended friction law has been presented to model bonded areas inde­ Microelectronics IMAPS (2002) 626–631.
pendently from the applied normal force using the coefficient of friction. [14] H. Gaul, M. Schneider-Ramelow, H. Reichl, Analytic model verification of the
The friction coefficients between bond tool and wire and between interfacial friction power in al us w/w bonding on au pads, IEEE Transactions on
Components and Packaging Technologies 33 (3) (2010) 607–613, https://doi.org/
wire and substrate were identified experimentally for the aluminium 10.1109/TCAPT.2010.2049847.
wire and copper wire on DCB substrate. For the contact between wire [15] Y. Ding, J.-K. Kim, P. Tong, Effects of bonding force on contact pressure and
and substrate, where the bond formation takes place, significant dif­ frictional energy in wire bonding, Microelectron. Reliab. 46 (7) (2006) 1101–1112,
https://doi.org/10.1016/j.microrel.2005.09.010.
ferences between copper and aluminium wire were observed, as for the [16] M. Sbeiti, Thermomechanische beschreibung der ausbildung einer
aluminium wire five times more frictional energy compared to the intermetallischen phase beim ultraschall-wedge/wedge-drahtbonden im rahmen
copper wire is needed to achieve a certain cleaning and activation of the der theorie der materiellen kräfte, Ph.D. thesis, 2013.
[17] S. Althoff, T. Meyer, A. Unger, W. Sextro, F. Eacock, Shape-dependent
contact area. transmittable tangential force of wire bond tools, in: IEEE 66th Electronic
The simulation results allow for detailed analysis of the bond for­ Components and Technology Conference, 2016, pp. 2103–2110, https://doi.org/
mation in the early stages of process development by analysing the bond 10.1109/ECTC.2016.234.
[18] T. Meyer, A. Unger, S. Althoff, W. Sextro, M. Brökelmann, M. Hunstig, K. Guth,
strength distribution, analysing the dynamical behavior of the wire and
Modeling and simulation of the ultrasonic wire bonding process, in: 2015 17th
substrate, and calculating shear force values. Electronics Packaging Technology Conference, 2015, https://doi.org/10.1109/
Future work will further concentrate on the parametrization of the EPTC.2015.7412377.
model parameters - especially the ultrasonic softening model - for the [19] R. Schemmel, T. Hemsel, W. Sextro, Numerical and Experimental Investigations in
Ultrasonic Heavy Wire Bonding, 6th European Conference on Computational
copper and aluminium wire on DCB substrate. In future, the wire bond Mechanics (ECCM 6), 2018.
model will be used for optimization of the bond parameters and iden­ [20] J. Hu, T. Shimizu, M. Yang, Investigation on Ultrasonic Volume Effects: Stress
tification of process windows. In addition, the usability of the model for Superposition, Acoustic Softening and Dynamic Impact, Ultrasonics
Sonochemistry, 2018.
other wire materials and also for other wire bonding technologies such [21] A. Unger, W. Sextro, T. Meyer, P. Eichwald, S. Althoff, F. Eacock, M. Brökelmann,
as thin wire bonding will be investigated. M. Hunstig, K. Guth, Modeling of the stick-slip effect in heavy copper wire bonding
to determine and reduce tool wear, in: 2015 17th Electronics Packaging
Technology Conference, 2015, https://doi.org/10.1109/EPTC.2015.7412375.
CRediT authorship contribution statement [22] T. Xu, T. Walker, B. Poncelet, J. Fu, C. Luechinger, Consumable and process
improvement for large copper wire bonding, International Symposium on
Reinhard Schemmel: Conceptualization, Methodology, Software, Microelectronics 2016 (2016) 000445–000449, https://doi.org/10.4071/isom-
2016-THA34.
Validation, Investigation. [23] R. Leine, D. Campen, A. Kraker, L. Steen, Stick-slip vibrations induced by alternate
Viktor Krieger: Software. friction models, Nonlinear Dynamics 16 (1998) 41–54, https://doi.org/10.1023/A:
Tobias Hemsel: Writing- Review & Editing. 1008289604683.
[24] R. Schemmel, S. Althoff, W. Sextro, A. Unger, M. Hunstig, M. Broekelmann, Effects
Walter Sextro: Supervision.
of different working frequencies on the joint formation in copper wire bonding, in:
CIPS 2018; 10th International Conference on Integrated Power Electronics
Declaration of competing interest Systems, 2018, pp. 1–6. URL, http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&ar
number=8403137isnumber=8402818.
[25] R. Kumme, G. Lauer, M. Peters, A. Sawla, Development of Methods for Dynamic
The authors declare that they have no known competing financial Force Calibration, Part 1: Dynamic Calibration of Force Transducers Based on the
interests or personal relationships that could have appeared to influence Determination of Inertia Forces with Interferometrically Calibrated Acceleration
Transducers, 1990.
the work reported in this paper. [26] G. Lauer, Development of methods for dynamic force calibration: final report, in:
Absolute Calibration of Piezoelectric Force Transducers by Laser Interferometry,
References Commission of the European Communities, 1990.
[27] J. Schwizer, M. Mayer, D. Bolliger, O. Paul, H. Baltes, Thermosonic ball bonding:
friction model based on integrated microsensor measurements, in: Twenty Fourth
[1] Z. Lai, J. Liu, The nordic electronics packaging guideline, chapter a: wire bonding,
IEEE/CPMT International Electronics Manufacturing Technology Symposium (Cat.
Tech. Rep, 2000. URL, http://extra.ivf.se/ngl/A-WireBonding/ChapterA.htm.
No.99CH36330), 1999, pp. 108–114, https://doi.org/10.1109/
[2] R. Schemmel, T. Hemsel, C. Dymel, M. Hunstig, M. Brökelmann, W. Sextro, Using
IEMT.1999.804803.
complex multi-dimensional vibration trajectories in ultrasonic bonding and
[28] E. Korkmaz, B.A. Gozen, B. Bediz, O.B. Ozdoganlar, Accurate measurement of
welding, Sensors and Actuators A: Physical 295 (2019) 653–662, https://doi.org/
micromachining forces through dynamic compensation of dynamometers,
10.1016/j.sna.2019.04.025. URL, http://www.sciencedirect.
Precision Engineering 49 (2017) 365–376, https://doi.org/10.1016/j.
com/science/article/pii/S0924424718321733.
precisioneng.2017.03.006. URL, http://www.sciencedirect.com/science/article/
[3] M. Hunstig, A. Unger, M. Brökelmann, H.J. Hesse, Process advantages of
pii/S0141635916301994.
thermosonic wedge-wedge bonding using dosed tool heating, International
[29] M. Michalski, U. Leicht, A. Heath, M. Merklein, Dynamic correction of oscillatory
Symposium on Microelectronics 2019 (1) (2019) 000519–000523, arXiv:
forces during ultrasonic-assisted metal forming, Production Engineering 11 (08
https://doi.org/10.4071/2380-4505-2019.1.000519.
2017), https://doi.org/10.1007/s11740-017-0762-3.
[4] P. Chauhan, Z.W. Zhong, M. Pecht, Copper wire bonding concerns and best
[30] D. Gross, S. Haag, M. Reinold, M. Schneider-Ramelow, K.-D. Lang, Heavy copper
practices, J. Electron. Mater. 42 (8) (2013) 2415–2434, https://doi.org/10.1007/
wire-bonding on silicon chips with aluminum-passivated cu bond-pads,
s11664-013-2576-1.
Microelectronic Engineering 156 (2016) 41–45, mAM (Materials for Advanced
[5] D. Kim, R. Willmot, D. Peroulis, A high-efficiency low-cost wire-bond loop antenna
Metallization) 2015, https://doi.org/10.1016/j.mee.2015.12.017. URL, http://
for cmos wafers, in: 2009 IEEE Antennas and Propagation Society International
www.sciencedirect.com/science/article/pii/S0167931715301180.
Symposium, 2009, pp. 1–4.
[31] Y. Long, F. Dencker, A. Isaak, J. Hermsdorf, M. Wurz, J. Twiefel, Self-cleaning
[6] T. Stockmeier, From packaging to “un”-packaging - trends in power semiconductor
mechanisms in ultrasonic bonding of al wire, J. Mater. Process. Technol. 258
modules, in: 2008 20th International Symposium on Power Semiconductor Devices
(2018) 58–66.
and IC’s, 2008, pp. 12–19, https://doi.org/10.1109/ISPSD.2008.4538886.
[32] E. Rabinowicz, Friction and wear of materials, in: Wiley Series on the Science &
[7] M. Hunstig, W. Schaermann, M. Broekelmann, S. Holtkaemper, D. Siepe, H.
Technology, Wiley, 1965. URL, https://books.google.de/books?id=kuV
J. Hesse, Smart ultrasonic welding in power electronics packaging, in: CIPS 2020;
SAAAAMAAJ.
11th International Conference on Integrated Power Electronics Systems, 2020,
[33] D. Tabor, Junction growth in metallic friction: the role of combined stresses and
pp. 1–6.
surface contamination, Proceedings of the Royal Society of London A:
[8] R. John, O. Vermesan, R. Bayerer, High temperature power electronics igbt
Mathematical, Physical and Engineering Sciences 251 (1266) (1959) 378–393,
modules for electrical and hybrid vehicles, IMAPS, High Temperature Electronics
https://doi.org/10.1098/rspa.1959.0114.
Network (HiTEN) 1 (2009) 199–204.
[34] A. Unger, R. Schemmel, T. Meyer, F. Eacock, P. Eichwald, S. Althoff, W. Sextro,
[9] A. Karch, Neue Lötlegierung für erweiterte Einsatztemperaturen, 2018.
M. Brökelmann, M. Hunstig, K. Guth, Validated simulation of the ultrasonic wire
[10] DVS - German Welding Society, Technical Bulletin DVS 2811 - Test Procedures for
Wire Bonded Joints, Tech. Rep, 2017.

15
R. Schemmel et al. Microelectronics Reliability 119 (2021) 114077

bonding process, in: Ultrasonics Symposium (IUS), 2016 IEEE International, IEEE [39] Oracle [link]. URL, https://docs.oracle.com/javase/7/docs/api/org/omg/CORBA
CPMT Symposium Japan, 2016, pp. 251–254. /ORB.html, 2020.
[35] T. Liu, Sliding friction of copper, Wear 7 (2) (1964) 163–174, https://doi.org/ [40] Paderborn Center for Parallel Computing, Technical documentation wiki of oculus,
10.1016/0043-1648(64)90051-1. URL, http://www.sciencedirect.com/science/ URL, https://wikis.uni-paderborn.de/pc2doc/OCuLUS, 2020.
article/pii/0043164864900511. [41] R. Schemmel, V. Krieger, T. Hemsel, W. Sextro, Co-simulation of MATLAB and
[36] I. ANSYS, ANSYS as a Server Example: MATLAB Setup, 2019. ANSYS for ultrasonic wire bonding process optimization, in: 2020 21st
[37] I. ANSYS, ANSYS Mechanical APDL as a Server User’s Guide, 2018. International Conference on Thermal, Mechanical and Multi-Physics Simulation
[38] I. ANSYS, ANSYS Mechanical APDL Basic Analysis Guide, 2018. and Experiments in Microelectronics and Microsystems (EuroSimE), 2020,
pp. 1–11, https://doi.org/10.1109/EuroSimE48426.2020.9152679.

16