Chapter 3
Particles in Semiconductor
Processing
D. Martin Knotter * and Faisal Wali y
*
y
NXP Semiconductors, Nijmegen, The Netherlands
University of Twente, Nijmegen, The Netherlands
1. Introduction
1.1. Impact of Particles in IC
Manufacturing
1.2. Yield Calculation Models
1.3. Origin of Particles
1.4. Determination of Particle
Removal Efficiency (PRE)
1.5. Methods to Remove
Particles
2. Theory
2.1. Particle Removal
Mechanism
2.2. Model of Particle Removal
2.3. Liquid Particle Counter
(LPC) Measurements
2.4. Metal-Ion Core Particles
3. Particle Removal Study
3.1. Tank Dynamics, Impact
of Particle Counter, and
Particle Composition
3.2. Particle Diffusion Out of
the Boundary Layer
3.3. Particle Detachment
4. Conclusions
Acknowledgements
References
1. INTRODUCTION
Advances in integrated circuits (ICs) have a high impact on society. These
advances result in continuously increasing performance of home personal
computers, higher density flash memory chips, faster wireless communication
in combination with smaller antennas, and all kinds of combinations of the
aforementioned components. The main characteristic of these advances has
been the shrinking dimension of the features of which the ICs are made.
1.1. Impact of Particles in IC Manufacturing
Every two years the feature size of the new generation of microprocessors is
reduced with a O2 factor [1]. Since 2004 the smallest size is in the nanoscale
Developments in Surface Contamination and Cleaning
Copyright Ó 2010 Elsevier Inc. All rights of reproduction in any form reserved.
81
82
Developments in Surface Contamination and Cleaning
range which is defined by the 100-nm limit. Simultaneously, particles that can
cause device damage and that are deposited on the product during its
manufacturing are smaller in size. In early road maps, the critical dimension
of the particles was assumed to one-tenth of the minimum device feature size.
Later, this was relaxed from one-third to currently half of the feature size. Today,
it means detection and removal of particles in the range of 10–20 nm.
The manufacturing of microprocessors consists of hundreds of process
steps, many of which can be sensitive to particle contamination. The process
steps that are sensitive to particle contamination can be grouped into several
categories (see Figure 3.1):
a.
b.
c.
d.
Particles in holes or trenches
In-film particles
Particles as mask (not related to pattern)
Patterning deviations where particles become part of the photo mask.
a
c
e-
d
b
s
g
+ ++
d
e-
s
g
+ ++
d
e-
FIGURE 3.1 Cross-sectional details of an IC with different impact of random particles on device
performance. Color coding: white is the insulator; light gray is the semiconductor; dark gray is the
conductor; black is a particle with unknown properties. (a) Particle in a contact hole before the
hole is filled up with metal results in an ‘‘open’’ circuit. (b) Particle on the gate area (g) prior to
gate definition results in poor transistor performance (s ¼ source, d ¼ drain). (c) Particle on an area
that is implanted with low-energy dopants. (d) Patterning problem where the particle is located in
the photo mask pattern, resulting in a masked etch and a ‘‘short’’ circuit.
Particles in Semiconductor Processing
83
These categories do overlap, but they all result in different kinds of failure in
the final device. The commonality is that all these particles can result in random
yield loss.
1. Particles in holes or trenches. The most prevalent failure mode in
manufacturing is category (a), where a particle obstructs the conductivity
between two metal layers. The reason is the process weakness in the
previous steps, i.e. poor definition of holes and trenches in the dielectric
layer. This plasma etch step is contaminating and the subsequent cleaning
step has a relatively small process window. Plasma etching results in
so-called ‘‘sidewall polymers’’ that are actually polymers deliberately
deposited on the sidewall to attain a high degree of anisotropy in the etch
process (sidewall passivation). Furthermore, the sidewall polymers can
contain residue of the reaction product between the plasma gas and the
metal layer that is exposed on the bottom of the hole. Exposure of the metal
to the plasma gas is poorly controlled, resulting in an unpredictable amount
of metal salts in the sidewall polymers. Compared to the cleans earlier in
the process, the clean between the plasma etch and the deposition of the
next metal layer (to fill the hole) has reduced chemical etch activity as
well as physical power, because it is not allowed to etch the exposed metal
nor the dielectric material which, in the latest technologies, is made of
porous silica-based material.
2. In-film particles. Category (b) process failure has been driving the roadmap
for semiconductors with respect to cleaning and cleanliness performance of
the manufacturing facilities. The critical area here is defined as the area
where the gate is going to be made and it is the number of transistors in
the device multiplied by the gate area in each transistor. Surprisingly, this
has remained constant over the years, because the gate area has been halved
with each new generation while the number of transistors has doubled. This
means that the number of allowable particles to achieve a 99% device yield
remains constant. However, the challenge here is that the particles now are
of a smaller size.
Except for the deposition tool itself, there are no typical process-related
particles depositing on the wafer before and during these process steps. In
the deposition tool the material to be deposited is not only deposited on the
wafer but also on the sidewall of the chamber. Several depositions in a row
without a layer-removal step can cause these multilayers to crack from the
stresses, resulting in airborne particles. These particles will be deposited on
the wafer either in the beginning of the deposition, resulting in in-film particles that are non-removable, or at the end of the deposition.
If the particles are not yet within the film, a relatively robust cleaning
process can be used to remove these particles. It can make use of mechanical forces such as megasonic energy without too much concern for pattern
damage. Also, at this stage under-etching is still allowed.
84
Developments in Surface Contamination and Cleaning
3. Particles as mask (not related to pattern). This process failure is similar to
the previous example, where particle contamination is present on a wafer
without a photoresist pattern, but structures can also be present on the
wafer. The critical processes involved can be implantation, oxidation, or
a film-removal step. The particles mask the underlying substrate and they
will leave a ghost pattern in the final device even after the particle has
already been removed.
4. Patterning deviations where particles become part of the photo mask.
Before or during the definition of the patterned photoresist layer, particles
can deposit and can be positioned such that they are on a location where no
photoresist should have been (failure mode d in Figure 3.1). Particles on
photomasks that are used to define the patterned photoresist layer can result
in a similar failure, but this will be picked up as a systematic yield loss
because the image of the particle will be printed on every product at the
same position.
As in the previous case, the particle acts as a mask, but in this case it will
inhibit the etching process on a patterned wafer. The position of the particle
defect can occur anywhere and the particle imprint can no longer be
removed. In the given example (Figure 3.1), the particle results in an
electrical short because the etched metal was not removed completely and
the imprint is short-circuiting two metal lines that should have been
separated. Other failure scenarios are also possible that can result in an
open circuit, if, for example, a dielectric layer is not etched away and the
deposited metal layer later cannot fill that space.
1.2. Yield Calculation Models
The degree of success in IC manufacturing is measured by yield (Y) that is
defined as the ratio of usable devices in respect to potentially usable devices
before starting its manufacturing [2]. Usable devices are defined as those
that pass several physical and electrical tests during or after completion of
the multistep manufacturing process. Knowledge of yield performance of
a manufacturing facility or process is used to predict the yield of new products
that has a higher degree of integration. The economics of the introduction of
such production processes helps in making the decision to build a new fab,
upgrade the existing fab, and the identifications of problematic process steps.
A large portion of the yield loss is caused by contamination present in the
wafer environment and, in this category, particles are the major contributor. To
be able to reduce the impact of contamination on manufacturing, defects have
to be detected. These defects are initially measured after each process step. The
long-term defect probability and yield prediction are related by statistical
probability distribution models.
There are several yield prediction models in the literature. Most of these
models require the detection of defects on the product wafer during the
Particles in Semiconductor Processing
85
multistep process flow and the confirmation of successful or unsuccessful
fabrication of the affected product. This assumes that a wafer consists of
a number of ICs that are sawed out of the wafer at the end of the process. Each
IC consists of a number (N) of transistors. If a single transistor in an IC is not
working, the whole IC is malfunctioning and is considered scrap. Finally,
defects, especially particles, can be deposited randomly on the wafer, which is
a stochastic process.
1.2.1. Wallmark’s Model
In 1960, the first yield prediction model was proposed by Wallmark [3] and is
described by:
Y ¼ ð1 S=100ÞN
(1)
His model was based on the assumption that the percentage (S) of working
transistors that could be produced by a certain process was known from an
existing manufacturing process or was determined experimentally. As the
number of transistors in an IC was in the range of one to several hundred,
determination of such numbers was reliable. The model is based on probability
calculations: if one out of ten transistors is not working, then nine out of ten are
working. Thus, the probability that a device is made out of two working
transistors is 0.92 ¼ 0.81, but with ten transistors the yield would drop to
0.910 ¼ 0.35. An economist could tell if this yield is sufficient to bring such
a product to the market, or that the yield has to be improved first before going
on to integrate the next generation IC into the device.
As the number of transistors in the newer generation of ICs rapidly increased
and the integration climbed to a higher level, this model was no longer suitable.
The main problems were that the method did not identify the yield-determining
process steps and it was not able to predict the impact of the reduction in
transistor dimensions.
1.2.2. Poisson Model
In 1963, Hofstein and Heiman [4] presented some theoretical characteristics
of the ‘‘insulated-gate field-effect transistor’’ (see Figure 3.2). The transistor
was described as a control electrode (gate) insulated from a thin conducting
channel in the surface of a silicon substrate by an oxide film. They proposed
the critical area to be the area under the gate (AG): if a defect occurs in this
area, the transistor will fail and a defect outside this area has no or minor
impact.
The probability that a defect occurs in the critical area is considered to be
totally random and thus independent of surface structure, differences in local
surface composition, or the presence of other defects. Under these conditions,
the Poisson probability function can be used to calculate the yield of a process
for which the defect density (D) is known (equation (2)). This model has proven
86
Developments in Surface Contamination and Cleaning
gate
L
SiO2
channel
Source
drain
W
Silicon wafer
FIGURE 3.2 A typical layout of a field effect transistor (not to scale). W is the gate width that is
considered the critical dimension of a transistor and L is the gate length. The critical area AG ¼ WL.
to be very accurate in predicting yield for products with a total die area below
0.25 cm2.1
Y ¼ eNAG D
(2)
It is the definition of defects resulting in device failure that has been the topic of
dispute. D is composed of a measured defect density, where a certain size is
considered to be critical, and of the kill ratio of the defect. The kill ratio is the
number of particles that causes a defective product divided by the total number
of particles in the critical area.
Defects are measured and sized with optical techniques. Bare or patterned
wafers are scanned with a laser beam and when the beam illuminates a defect
the light is scattered. The scatter intensity is a measure for the defect size.
However, the scatter intensity not only depends on the defect size but, among
other things, it also depends on the defect composition. Defects measured are
particles, pits and asperities, and pattern deformation. Only the first category,
particles, is of interest in this chapter. Their composition remains unknown.
The size at which particles become a killer is rather arbitrary. As mentioned
before, the size depends on the most critical dimension (i.e. W) of the device
and is defined as a fraction of that size. This fraction is currently 0.5. In fact, it
does not matter what fraction is chosen, because the kill ratio will compensate
for errors in the chosen size.
The kill ratio is determined by actually measuring a certain number of
defects of the same category and it determines the number of devices that failed
due to these defects. It is a calibration factor. It becomes clear that if the size
at which a defect was measured is chosen to be smaller, more defects will
be measured and the kill ratio will decrease accordingly. Besides the size
1
Instead of one specifc critical area, the gate area (NAG), most yield models are more general and
use A as total area in a certain process step. Therefore, for NAG also read A.
87
Particles in Semiconductor Processing
compensation, the kill ratio is also needed because the tools to measure defect
density do not measure the composition of the defect. Some defects are not
killer, e.g. organic particles will disappear in a furnace oxidation. Since in
manufacturing the compositions of defects in most process steps are unknown,
the kill ratio has to be determined for each process separately.
1.2.3. Murphy’s Model
The Poisson model is correct for one process with a known fixed defect density.
However, the defect density is not constant in time nor is the defect density
uniform over the wafer. It is subject to random variation that results in some
clustering of defects. The effect of such clustering is that the chance that an IC
in such area or time frame is affected by two defects will increase. This IC can
only fail once. If these clusters had spread out over the whole wafer or in time
these two defects would have hit two ICs and both would have been scraped.
Since the relation between defect density and yield is not linear, the long-term
yield will be underestimated (see Figure 3.3).
Murphy [5] included the long-term variation in defect density in the longterm yield calculations by using a (normalized) distribution function f(D) that
describes this variation:
ðN
eNAG D f D dD
(3)
Y ¼
0
1
0.8
Yield (fraction)
0.17
0.6
-0.13
0.4
0.2
0
0
10
20
30
40
50
Defect density (AU)
FIGURE 3.3 Relation between defect density and yield according to the Poisson model. If the
average defect density is 10 and it fluctuates randomly between 5 and 15, then the long-term yield
will be underestimated by a maximum of 2%.
88
Developments in Surface Contamination and Cleaning
The function can be determined from the manufacturing history or it can be
assumed to have a certain shape, such as rectangular, bell curve, triangular, box,
or line. Murphy chose the triangular shape and obtained:
2
NAG D
(4)
Y ¼ 1e
NAG D
1.2.4. Negative Binomial Model
A more physical meaning can be given to the variation in defect density.
Incidents are the reason that defects cluster in place or time. Wafers are hit by
splashes, dry-in marks, scratches, or some other localized random events,
causing many defects to be created in a localized area. Also, one of the many
processes or process tools can run out of control at some moment in time. The
key property is that the final defect count is the sum of many processes that can
run out of control, adding a little to the total defect count. Since most processes
and tools run within control, the defect distribution becomes skewed with
a maximum close to zero defects. The statistic that best describes such
nonsymmetric variation is the gamma distribution. Therefore, Stapper [6,7]
expressed the variation in defect concentration by a gamma distribution function and used that function for f(D) of equation (3). After integration, he
derived:
NAG D a
(5)
Y ¼ 1þ
a
The new parameter a is a clustering parameter that is equal to D2/s2, where s2 is
the variance in the defect density. According to Stapper [6], a varies from 0.3 to
5, but is typically between 2 and 3 [8]. More clustering (more incidents) results
in a larger variance, smaller values of a and, thus, a larger predicted yield than
predicted with Poisson’s model with the same average defect density. As an
effect of the gamma distribution function, this model becomes a sort of unifying
model.
If a becomes larger than 10, this models starts to overlap with the Poisson
model (i.e. no clustering).
When a is between 4 and 7, the yield prediction is similar to Murphy’s
model.
With a ¼ 1, this model is equal to Seed’s model [2], which is often used as
an alternative to Poisson’s model.
In 1991, Stapper [9] emphasized that the negative binomial yield model has
found general acceptance in semiconductor manufacturing in Canada,
Europe, and the USA. The Poisson yield model is the model of choice for
comparing data from single process steps and is used in ITRS roadmap
discussions [1].
Particles in Semiconductor Processing
89
1.2.5. Random Defects and Non-Random Deposition Model
In all yield models defects are systematic, identified with in-line defect
inspection tools, and well classified. Furthermore, the defects are assumed to
deposit randomly on any area of the whole wafer. In a recent study [10], defect
density in the wafer environment was related to yield. This means that the
measured particles in ultra-pure water (UPW), which is used for device
manufacturing, relate to the in-line measured defect density on the respective
devices. The defect density (D) due to particles present in UPW can be
described by our predictive model:
D ¼ Np SKR Pd
(6)
Here Np represents the particle concentration (particles cm3 in UPW),
S (L cm2) is the amount of UPW that contacts the wafer during the fabrication
at the critical process steps, KR represents the fraction of killer particles, and Pd
is the probability that particles deposit on to critical areas. The value of Np and S
can be measured. The unknown value KR varies with the particle composition
and size. Pd depends on the process settings of the manufacturing process step
generating contamination, particle composition, and wafer surface composition. It is Pd that is the subject of our further investigations as it can have a large
impact on yield, if the particle would have a preference to deposit on critical
areas. It was shown [11] that particles do deposit on specific areas during the
drying sequence of a cleaning procedure.
1.3. Origin of Particles
The composition of the particles is diverse and it depends on location and time.
Failure analysis of particles using SEM-EDX that resulted in a device failure
shows, in many cases, the presence of Si, but whether it is Si, silica, silicon
nitride, or organosilicon compounds (e.g. dimethylsiloxane) remains unknown
in many cases. Another category of particles detected in failure analysis is
particles related to photoresist residue. Organic material is difficult to discern,
but fluoride residues can be measured and they are very likely related to the
fluoride used in the plasma to etch patterns. These residues are either not
removed after the plasma etch step, or they are redeposited out of the cleaning
solution used to do the post-etch cleanup.
The non-process-related particles originate from the wafer environment,
such as cleanroom air and process liquids (chemicals and water), or from the
wafer edge.
Cleanroom air. Manufacturing of ICs is done in cleanrooms, but the
handling of wafers is done in an even cleaner microenvironment. Process
tools are completely enclosed with their own filter system and the wafers
are stored and transported in closed pods that have standard interfaces to
connect the pod to the tool. This enables the loading and unloading of
90
Developments in Surface Contamination and Cleaning
wafers into the tool without exposing the wafer to cleanroom air. Thus, the
probability that particles come from the cleanroom air is relatively low.
Process liquids. This is an important source of particles. If particles are
present in the liquids they have plenty of opportunity to end up on the wafer
surface (see Figure 3.4) [11]. During the immersion of the wafers into the
liquids, hydrodynamic forces act on particles floating on the liquid whereby
the particles are ‘‘stamped’’ on the wafer surface [12]. When the wafers are
in the submerged state, particles are deposited due to electrostatic forces
and van der Waals forces [13]. The withdrawal of the wafers from the liquid
determines the amount of liquid left to dry on the surface [14]. All contamination in the drying liquid will be left on the surface. The final step, drying,
does not determine the number of particles left on the wafer, but rather the
location on the wafer [11]. Wafers are exposed to liquids during lithography,
cleaning, wet etching, galvanic deposition, spin-on layers, and polishing
steps. These process steps make up more than 50% of the total number of
process steps.
Wafer edge. The edge of the wafer is a known source of contamination. At
the edge contamination is accumulated and generated. Accumulation can
occur because many cleans do not target the cleaning of the wafer edge.
Generation is due to mechanical mishandling of the wafer that is done
with grippers manipulating the wafer, or during transport where wafers
often bump into solid surfaces. Particles are also generated at the edge,
because the deposited layer ends at the edge in an uncontrolled way and
the layers can peel and flake off. If these wafers are immersed, particles
can dislodge and redeposit on the wafer. Usually, the pattern of deposition
is recognizable with inspection tools as it occurs as a smear from the edge.
Besides the aforementioned resist residues the major source of process-related
particles are deposition tools. During deposition, material is deposited on the
wafer and on the wall of the reactor. If the layers on the reactor become too
thick, these layers can crack and flakes will come loose. In some cases, it is the
gas mixture at the start of the deposition process that is of incorrect composition
which causes particles to form in the gas. In all these cases, particle deposition
1
3
4
2
FIGURE 3.4 Opportunities for particles to deposit on a wafer during the bath cleaning process:
(1) immersion; (2) submersion; (3) withdrawal; (4) drying.
91
Particles in Semiconductor Processing
is unanticipated and the risk can be reduced by increased tool cleaning
sequences, or by adding an extra wafer-cleaning step after the deposition step.
Chemical mechanical polishing (CMP) is another process step that adds
particles to the wafer. However, this is anticipated and there are dedicated postCMP cleaning tools and processes available. Many of these processes do not
target the wafer edge properly.
If particles deposit on the wafer, it is assumed to be a random process. Yield
models are using this boundary condition (see Section 1.2). It has been shown
that particles coming out of a liquid do not deposit randomly [11]. If a surface
has mixed areas that are hydrophobic and hydrophilic, particles tend to deposit
on the hydrophilic areas, depending on the nature of the particles. If the surface
has structures, the particles will tend to deposit next to the sidewalls.
Convection flows and lateral capillary forces during the drying step drive these
processes. A problem for technologies with smaller feature sizes is that the
smaller particles that might be critical have a higher tendency to deposit
preferentially.
1.4. Determination of Particle Removal Efficiency (PRE)
In order to evaluate a cleaning process, a method is required to determine the
particle removal efficiency of the process. This has to be done with particlecontaminated wafers. In early manufacturing, these wafers were prepared
by putting the wafers through a process or tool that was known to contaminate
the wafer. Subsequently, the number of particles was measured (pre-count),
the wafers were cleaned, and the number of particles was re-measured (postcount). The PRE was calculated from:
PRE% ¼ 100%
ðPre-count Post-countÞ
Pre-count
(7)
This was a very pragmatic approach that could indeed give some indication
of the performance of two processes under evaluation at the same time. The
weakness of this approach is that the particles are of unknown origin and the
composition can change on a day-to-day basis. Also, the amount of deposited
particles could vary from wafer to wafer. Thus, for scientific purposes, this
method is unsuitable.
A second drawback of the method is that it does not account for particles
added by the cleaning process itself. This means that if the pre-count is
relatively small, the determined PRE will be offset by these added particles. An
improvement is to use a series of particle-contaminated wafers with variable
particle concentrations. If the pre-count and the post-countpre-count differences are plotted on an x–y plot, the slope of the linear regression represents the
PRE and the intercept represents the amount of particles added by the clean
itself. An example of such an experiment is given in Figure 3.5.
92
Developments in Surface Contamination and Cleaning
Pre-count–post-count
1100
900
y = 0.93x - 24
700
500
300
100
-100
0
250
500
750
1000
Pre-count
FIGURE 3.5
Example of a particle removal test: PRE ¼ 93% and added particles by the clean is 24.
The methods to contaminate the wafers have been improved by contaminating wafers with particles of known composition. In manufacturing,
suspensions are generated by breaking silicon wafers in/above water (to
generate silicon particle suspension) or by diluting slurries that are used for
the polishing process (silica particles). These suspensions are applied on to
wafers by spinning, or the wafers are immersed in these suspensions. For
scientific studies, silicon nitride particles tend to be preferred, because these
particles are believed to be more difficult to remove than silica or silicon.
Problems related to this particle-application method will be discussed in
Section 3.3.1.
It is not common practice in the industry to monitor particle removal efficiency with particle-contaminated wafers on a regular basis. Instead, practical
clean wafers are processed in cleaning tools and the amount of added particles
(D[x]) by process and tool is monitored. It is wrongly assumed that the small
amount of particles already present on the wafers is not removed. It was
demonstrated that from such a monitor program the particle removal efficiency
could be determined [15], because the small amount of particles is partially
removed. By describing the cleaning process as an equilibrium reaction
between particle attachment and detachment the following equation can be
derived:
(8)
D x ¼ Ad 1 ekr t
Here the parameter Ad describes the amount of particles that are added by the
process at infinite process time minus the initial particle count. kr is the removal
rate constant of the particle detachment process. This formula makes more
sense with the following definition of particle removal efficiency:
PRE% ¼ 100% 1 ekr t and PRE ¼ 1 ekr t
(9)
Particles in Semiconductor Processing
93
The performance of the cleaning process could be calculated by taking Ad and
kr to be normally distributed [15].
1.5. Methods to Remove Particles
Particle contamination accounts for 90% of the contaminants and is responsible
for 80% of the defects [16]. The easiest way to avoid the impact of particles is
to prevent the particles from depositing on the wafer. If this is not possible, the
particles have to be removed. In the semiconductor industry there are many
methods to remove the particles. They are based on chemical principles,
physical forces, or a combination of these. Since the structures made on the
surfaces are so small and fragile, there is a delicate balance between particle
removal and structure damage.
The main chemical principle is undercutting of the particle by etching the
substrate. Additionally, surfactants can be added to either aid the detachment or
prevent the redeposition of particles. Physical methods that aid the chemical
processes are ultrasonics or megasonics [17], brush cleaning [18], or centrifugal forces [19]. More or less pure physical methods are bombardment of the
contaminated surface with high-speed droplets [20,21], or with ice particles
[22,23], or with laser-assisted cleaning [24,25]. Many other methods are
under investigation, but have not found serious application in semiconductor
manufacturing.
Solutions that are used to remove particles are aqueous based. Solutions
such as HF etch silicon oxide but not silicon, while such solutions as ammonia–
hydrogen peroxide mixture etch silicon and silicon oxide at more or less the
same rate. The problem with the latter is that it can etch silicon in an uncontrolled way, resulting in increased silicon surface roughness [26]. A disadvantage of the HF solution is that the electrostatic double-layer thickness is
minimal and repulsive forces that prevent redeposition of particles are weak.
Furthermore, using megasonic cleaning in combination with HF can cause
pitting of the substrate [27].
The first systematically developed silicon wafer cleaning process is the RCA
clean introduced in 1965 by Kern and Puotinen [28]. It is called the standard
clean (SC) and is based on a two-step process with intermediate water rinses.
The first step is aqueous ammonium hydroxide–hydrogen peroxide–water
mixture (APM), also called standard clean 1 (SC1), and the second step is
hydrochloric acid–hydrogen peroxide mixture (HPM), or standard clean 2
(SC2). It is the APM step that targets the removal of particles and the HPM step
that adds particles.
APM has been the major workhorse for particle removal in the semiconductor industry and has been upgraded throughout the years. Both APM
and HPM were originally used at 75–80 C which resulted in excessive
hydrogen peroxide decomposition. For this reason, nowadays APM is used at
94
Developments in Surface Contamination and Cleaning
lower temperatures and hydrogen peroxide is left out of the HPM. Also, the
concentration has been a target of improvement. APM was originally used in
a mixing ratio of 1:1:5–8 (NH4OH (25%)–H2O2 (30%)–H2O) and this has gone
down to as low as 1:1:500 [29]. The addition of ultrasonic energy or megasonic
energy allowed higher dilution ratios and lower operating temperatures [30].
Megasonic and ultrasonic cleaning are both acoustic cleaning methods.
Megasonic cleaning uses a higher frequency (typically 800 kHz up to 2 MHz)
than ultrasonic cleaning (<100 kHz). The principal mechanism of acoustic
cleaning is pulsating gas bubbles moving across a particle-contaminated
surface and thereby initiating the liftoff of particles [31,32]. Furthermore, these
gas bubbles together with the sound waves cause micro-streaming which is
a relatively non-harmful mechanism to enhance particle diffusion. However,
these same gas bubbles can also collapse completely, creating high-energy hot
spots, so-called cavitation [33]. Cavitation causes structure damage. Since in
ultrasonic cleaning cavitation bubbles are larger than in megasonic cleaning,
the damaged area is larger.
Brush cleaning can be applied in two modes: contact mode and non-contact
mode. In the contact mode, the brush directly touches the surface and applies
a force on the particles present on the surface of the substrate. In the noncontact brush-cleaning mode, the spinning brush generates a fluid velocity that
removes the particles by shear forces. Contact brush cleaning is mainly used on
flat surfaces, for example, after the chemical mechanical polishing process,
since full contact of a brush will damage the device structure.
The mechanism of removal by high-speed droplets or high-speed ice
particles is slightly different. High-speed ice particles (usually made up of solid
CO2 or argon, formed by a sudden rapid expansion of the respective gases)
require direct collision with the particle on the surface, whereby impulse is
transferred and the particle is knocked away, or explosive evaporation detaches
the particle from the surface. In the case of droplets, this mechanism is also
possible, but it is also likely that the advancing front of an impinging droplet
close to the particle generates enough shear forces to detach the particles. Highspeed droplets are created by aerosol formation with a high pressure of N2 in
a dedicated spray nozzle. The most important parameter to obtain high particle
removal efficiencies and low substrate damage is the control over the droplet or
ice-crystal size distribution [34].
A laser can heat a localized spot very rapidly, which can cause an explosive
expansion of the material. It induces either an explosive evaporation of a
condensed liquid film between the particle and the substrate, or it generates
a shock wave by heating the liquid just above the particle. Commercial tools are
available for so-called laser cleaning that use the explosive evaporation
mechanism. Some believe that the liquid film is not required and the particle is
ejected from the wafer by the rapid expansion of the substrate [35]. Laser shock
wave cleaning is in development [25].
Particles in Semiconductor Processing
95
2. THEORY
2.1. Particle Removal Mechanism
In order to remove particle contamination from wafer surfaces, the removal
forces that act on the particles should be greater than the attractive adhesion
forces. Adhesion forces are extensively described in many textbooks
[36,37]. However, the most important adhesion forces are van der Waals
forces and electrostatic forces [38]. The strength of the van der Waals forces
depends on the particle material as well as the particle size. Electrostatic
forces depend, in addition to material properties, on the pH of the solution
and can be attractive or repulsive. It is a longer range force than the van der
Waals force.
The major challenge of particle contamination in a semiconductor
manufacturing environment is that particles are made up of a collection of
different materials (see Section 1.3). Furthermore, some particles are deposited from liquid and others from air; some particles adhere chemically, others
adhere purely physically; and some particles dissolve, while others are
insoluble in the cleaning liquids. Therefore, the adhesion strength varies and
the required force to remove them can be excessive for the majority of the
particles. It is the detachment process that gets the most attention of
researchers.
As described in the previous paragraph, both chemical and physical forces
are used to remove particles. There are many ways to apply a force on
a particle, but the most important one is by shear forces. Shear forces are
forces that act in the plane of the substrate, while the particle moves
perpendicular to this plane. Due to the shear forces, the particle can break the
strong chemical bonds and can start to roll or slide. The required liftoff forces
are introduced if a sliding or rolling particle contacts a surface asperity or
structure [39].
A mechanism that is underestimated by the industry is due to surface
tension when a particle on a surface passes through a liquid–gas interface.
This is the case during the immersion of wafers into a liquid (Section 3.2),
cleaning with high-velocity droplets, and also megasonic cleaning where
gas bubbles move over the wafer surface. A very instructive description is
given by Leenaars [40]. In his patent application, a laser creates a gas
bubble and this gas bubble is moved over a contaminated surface (see
Figure 3.6). The forces generated by the advancing and receding contact
angles can result in enough liftoff force to overcome the adhesion force
(FA).
The maximum lift force (Fl,max) that acts on the particle is given in
equation (10), where R is the particle radius, g is the surface tension of the
liquid, q is the wetting angle of the particle, and a ¼ the contact angle of the
liquid on the substrate.
96
Developments in Surface Contamination and Cleaning
Fl
FH
gas
gas
liquid
Fl
FH
liquid
FA
FA
a: particle moves into liquid
b: particle moves out of liquid
FIGURE 3.6 Particle removal mechanism with a gas bubble (only the gas–liquid–solid interface
is depicted). When a particle goes through a gas–liquid interface it will feel two major forces. FA is
adhesion force between particle and substrate and FH is force caused by the surface tension
difference at the liquid–air interface [41].
q
Fl;max ¼ 2pRg sin
cos a
2
2
(10)
From this equation three conclusions can be drawn about the cleaning system:
1. Best cleaning is done with liquids with high surface tension, e.g. water, and
without surface tension-reducing agents (surfactants) added.
2. Substrate surface should be wetted by water and thus should be as hydrophilic as possible. Surface with contact angle larger than 90 will draw
the particles to the surface.
3. It is much easier to remove hydrophobic particles.
Leenaars [40] noted that the speed of the gas bubble should not exceed
10 cm s1, because the interface of the liquid cannot adapt faster and the
cleaning becomes less effective. On the same principle, Kittle [41] used a very
large amount of gas bubbles as foam to remove particles.
2.2. Model of Particle Removal
The removal of particles is considered to be a three-step process (see
Figure 3.7). Physical and chemical parameters, such as temperature, viscosity,
pH, and the presence of surfactants, can impact each step differently. The first
step (ka) is the detachment of the particle from the surface, which brings the
particle into the boundary layer. The second step (kb) is the diffusion of the
particle through the boundary layer into the bulk of the process liquids. If this
does not occur before the wafer is dried, the particle will redeposit on the wafer
surface. Finally, the particle has to be carried away from the wafer environment
(kc) by filtration or draining of the contaminated process liquid [42].
97
Particles in Semiconductor Processing
Boundary layer
ka
Bulk liquid
kb
kc
Silicon
wafer
Filter
Measure concentration
FIGURE 3.7 The particle removal process is described by three consecutive process steps: ka,
the detachment; kb, the diffusion; kc, the removal out of the system.
Physical–chemical processes control the detachment process. The rate at
which particles are detached from the surface (ka) depends on a number of
factors:
Forces acting on the particle. The higher the force, the faster the particles
will start to move and detach.
Etch rate. The higher the etch rate of the substrate surface, the faster the
undercutting process and the subsequent particle detachment.
Redeposition. The lower the redeposition rate, the better the average
particle removal efficiency. Making use of repulsive electrostatic forces
between the substrate surface and the particles by selecting the proper pH
can prevent redeposition. Addition of surfactants can also reduce
redeposition.
Once the particles are detached they have to diffuse through the boundary layer
into the bulk liquid with a linear diffusion rate constant kb. The thickness of the
boundary layer determines, to a large extent, the residence time of the particles.
Thinner boundary layers will reduce the migration distance, which can be
achieved by applying higher liquid flows, for example by using higher flow
rates, applying megasonic energy, or using centrifugal forces. The rate of
diffusion is also determined by the size of the particles, as well as viscosity and
temperature of the liquid.
The last process step is the actual removal of particles from the process
(kc). For example, in a tank process where wafers are fully immersed in
a process liquid, the process liquid is agitated and pumped around in a recirculation loop. It leaves the top of the tank and re-enters the bottom of the tank
after passing a pump and a particle filter. If mixing in the tank is thorough,
then the particle concentration will decrease exponentially over time.
The time constant of this decay is determined by the flow rate and the particle
filter efficiency.
98
Developments in Surface Contamination and Cleaning
This full process can be monitored by measuring the particle concentration
in the bulk of the liquid with a liquid particle counter (LPC). The LPC uses light
scattering for particle detection. To avoid any possible quantification and sizing
issues, a monodisperse silica suspension is used for this purpose [43]. It is
possible to measure the effect of the above-mentioned physical and chemical
parameters on the rate constants ka, kb, and kc.
2.3. Liquid Particle Counter (LPC) Measurements
The total removal process of particles can be considered as a chemical reaction
process, where the particle concentration on the surface is expressed as [A], in
the boundary layer as [B], and in the bulk liquid as [I]. First-order reaction
kinetics are assumed and the respective reaction rate constants, k, can be
determined from dedicated experiments.
kc. If only the bulk of the liquid is filled with particles, some of the particles
will diffuse into the boundary layer, but most of them are removed by the
filtration action. The removal can be expressed by the differential in equation (11) using the continuously stirred tank reactor (CSTR) model and
neglecting the particles that diffuse into the boundary layer:
d½I
¼ kc ½I
dt
(11)
where kc is the tank turnover frequency, which is determined by the flow
rate divided by the tank volume. In this experimental setup, kc for a tank
filled with ultrapure water is 0.0098 s1. In subsequent experiments, kc
can vary since the flow rate depends on the viscosity of the cleaning
liquid, but it can always be determined independently from the other rate
constants.
kb. If the boundary layer is filled with a number of particles [B]0, some particles will precipitate on the surface (and become the particle concentration
A), but most of them will diffuse out of the boundary layer, where they are
subsequently removed. The concentration as a function of time that we are
able to measure, [I], can be expressed by equation (12), which makes use of
linear diffusion rates:
kb kb t
ekc t
e
(12)
½I ¼ ½B0
kc kb
In our experiments, we found that we needed two types of corrections to fit
the experimental data with the theoretical model: the fraction of particles
removed upon immersion and non-ideal CSTR behavior.
During the immersion process of the surface into the cleaning liquid,
a fraction (f) of the particles is immediately removed by the shear forces
99
Particles in Semiconductor Processing
of the liquid, causing intermixing of a part of the boundary layer and the
bulk liquid. Unfortunately, in this experimental setup f is not a constant,
as it depends on the speed and the angle at which the wafers are immersed
into the liquid. This was done manually.
A tank filled with 25 wafers does not behave as an ideal CSTR. ki is
a time constant that is a measure of the rate of the system to come from
a non-ideal CSTR state to the ideal state.
Equation (13) describes the corrected particle concentration in the bulk
of the liquid as measured with a liquid particle counter:
kb k t
kc t
kc t
b
þ 1f
e
e
1 eki t
(13)
½I ¼ ½B0 f e
kc kb
It appears that many parameters are necessary to fit some experimental
data, but some of the parameters (ki ¼ 0.0166 s1, kc ¼ 0.0098 s1) are
determined only one time and are kept constant for all experiments.
In Figure 3.8, an example is given with kb ¼ 0.00503 s1, f ¼ 0.432, and
[B]0 ¼ 18,135 particles mL1. The need to use a non-ideality constant
appears obvious from these results.
5000
Exptl. data
non-ideal CSTR
particles (n/mL)
4000
ideal CSTR
3000
2000
1000
0
0
200
400
600
800
1000
t (s)
FIGURE 3.8 Example of experimental data fitted (least squares) by a model with ideal CSTR
behavior and one that is corrected for non-idealities; n is the number of particles.
ka. Similar to the previous derivation, the concentration of particles in the
bulk could be modeled by equation (14) if the particles on the surface would
immediately start to detach as a kind of first-order reaction rate:
ka kb
eka t ekb t þ ekc t eðka kb þkc Þt (14)
½I ¼ ½A0
ðkb ka Þðkc kb Þ
Also, this can be corrected for the non-ideal CSTR model and for the
fraction of particles this is immediately sheared off during the immersion of
100
Developments in Surface Contamination and Cleaning
the wafers into the cleaning liquid. The complete model that describes the
process in Figure 3.7 is given by:
ka kb
½I ¼ ½A0 f ekc t þ 1 f
eka t ekb t þ ekc t
ðkb ka Þðkc kb Þ
ðka kb þkc Þt
(15)
e
1 eki t
The results we have obtained so far for the removal of particles from
a silicon surface cannot be fitted with equation (15). The reason is that the
removal process (going from A to B) is not correctly modeled. In
Figure 3.9, data from a typical removal experiment are depicted and the
particles are removed in two steps. The best fit is obtained by a model that
assumes that a fraction (f) is removed by shear forces and a second fraction
(1 f) is removed according to a stochastic process. A Gaussian distribution with an average particle cleaning time and a standard deviation (s)
describes the latter process, but the distribution becomes skewed by the
process mentioned earlier.
The correction for the skewness and determination of the correct
average cleaning time is done with equation (13), but in this case [B]0 is
replaced by the Gaussian error function that was approached by summation over time. The fitting parameters are given in the caption of
Figure 3.9. It should be noted that ki and kc are identical (these were kept
fixed) to what was previously measured, kb is slightly smaller (perhaps due
to the slightly longer average diffusion path), and f is larger.
Exptl. data
best fit
Particles (n/mL)
400
300
200
100
0
0
1000
2000
3000
4000
5000
t (s)
FIGURE 3.9 Example of experimental data fitted (least squares) by a model with non-ideal
CSTR behavior whereby particles are detached after an average cleaning time of 2280 s with s ¼
228 s. Other parameters: kb ¼ 0.0048 s1; kc ¼ 0.0098 s1; ki ¼ 0.0166 s1; f ¼ 0.93; background
offset ¼ 17 n/mL1; [A]0 ¼ 1290 n/mL1; n ¼ number of particles.
Particles in Semiconductor Processing
101
2.4. Metal Ion Core Particles
Removal of nanosized particle contamination from the surface of the substrate
is an essential requirement for IC manufacturing with critical dimensions
smaller than 100 nm. However, for an effective particle removal study, proper
detection of the particles is important. The conventional and previously
described method to detect the particles is based on light scattering. The
problem with this method is that it has size limitations (>100 nm) and it cannot
distinguish between the signals of nanoparticles and background noise or
surfactant micelles. To circumvent this problem, particles for use with an
alternative detection method were synthesized.
In 1968, Stöber et al. [44] developed a method known as the Stöber synthesis
capable of producing nanosized monodisperse silica particles using ammonia,
ethanol, and tetraethoxysilane (TEOS). Van Blaaderen et al. [45] described the
so-called seed technique to synthesize particles with a luminescent core. First,
particles were formed together with the luminescent material and, second, the
particles were grown further by adding clean TEOS to the suspension. The net
effect is a silica particle with a pure silica shell and a slightly contaminated
core. UV detection tools are not sensitive enough to measure the four to five
orders of magnitude of concentration changes required for the particle removal
study.
The seeded technique was used to make metal-ion core particles. Instead of
the luminescent material, a diluted scandium nitrate solution was added.
Scandium was chosen because it is rarely used in the microelectronic environment (low background intervention) and this metal ion attaches very
strongly to silica. The size of the particles is controlled by varying the reactant
quantities [46]. Quantification of the particle concentration can be done by
measuring the metal ion concentration, either in the solution or on the wafer
surface with inductively coupled plasma–mass spectroscopy (ICP-MS). It was
found that the scandium does not leach out of the particles even at pH 1 [47].
3. PARTICLE REMOVAL STUDY
3.1. Tank Dynamics, Impact of Particle Counter, and Particle
Composition
In the previous section, a differential equation was derived to measure the tank
dynamics in a CSTR (equation (11)). Solving this equation results in a function
where the particle decay is exponential with time:
I ¼ ½I0 ekc t
(16)
To establish the parameters that describe this process, a tank was filled with
particles and homogenized before the recirculating liquid was filtered. As the
goal was to use HF-based cleaning solution, particles with a low etch rate in
102
Developments in Surface Contamination and Cleaning
HF were used for this study. They were composed either of a mixture of
different size polystyrene latex spheres (PLS) or Si3N4 particles with a wide
size distribution. It was found that the apparent removal rate of particles
depends on the particle composition and the particle size.
Figure 3.10 summarizes the results of particle removal experiments using
PLS and Si3N4 particles. The removal rate is expressed in half-life time, t½, that
relates to kc as t½ ¼ ln 2/kc. In this case, the theoretical half-life time, which is
based on the measured tank volume and flow rate, would have been 49.6 s. In
the data supplied by the liquid particle counter, this value is only measured for
the large Si3N4 particles. After correction of the data for optical coincidence, all
sizes of Si3N4 particles are removed at the rate expected on the basis of tank
dynamics, while PLS is filtered out much faster than expected. This does not
change much after a second correction for optical coincidence resulting in size
promotion (two particles coincide, resulting in a measurement of one particle of
the next larger category). As the filter used had a capture efficiency >99% for
particles with sizes larger than 0.05 mm, this effect could not have been caused
by the difference in filter efficiency.
An explanation for the faster than theoretical removal of PLS out of the
recirculating liquid and the filter tank is that there is an extra transport mechanism for the particles towards the surface of the liquid. Since the tank overflows and the liquid leaving the tank comes from the surface, a higher than
average concentration of particle-contaminated water leaves the tank. A
particle with an apparent weight smaller than the total pull of the meniscus
60
0.1−0.12 µm
0.12−0.15 µm
0.15−0.25 µm
55
t½ (s)
50
45
40
35
PLS raw
Si3N4 raw
PLS cor
Si3N4 cor
FIGURE 3.10 Half-life time of the filtration process of particles out of a recirculated tank. Raw
data represent data as given by the measurement tool; ‘‘cor’’ is the same data after correction for
optical particle coincidence in the measurement tool. The solid black line is the theoretically
expected t½ with 100% filter efficiency (49.6 s) [44].
Particles in Semiconductor Processing
103
around it will float, or it will have a high tendency to be located at gas–liquid
interfaces [48]. Since PLS is organic in nature, it will have low surface tension
and its preference for location at the gas–liquid interface will be much higher
than that of the hydrophilic Si3N4 particles. Furthermore, small gas bubbles
that are present due to agitation of the liquid will be the extra transport carrier
for the particles to become submerged, as particles will attach to the gas–liquid
interface of the bubble. This property makes PLS unsuitable as model particles
in particle removal studies in liquids. Additionally, if these particles would have
deposited on the silicon surface, the lift forces induced by the advancing
contact angle will significantly aid the detachment of PLS from the surface
(equation (10)). Indeed, PLS deposited on silicon have been found to be
removed in large part upon the immersion in water, while particles that are
present on product surfaces are not removed in such a step.
The disadvantage of Si3N4 is that it is commercially not available as
a monodisperse suspension. Silica has similar surface properties as Si3N4 and
is available as monodisperse suspensions [49]. However, besides the fact that
silica is etched in HF solutions, experiments with the removal of silica ran into
a completely different problem. As the refractive index of silica is so much
closer to water than Si3N4, the scatter intensity becomes less efficient. Therefore, particles with a nominal mean diameter of 0.16 mm are invisible to
a detector with a detection limit lower than 0.1 mm latex sphere equivalent
(LSE), while 0.33 mm particles will appear at 0.15 mm and smaller, which is
close to the detection limit. Still, silica will be the particle of choice for further
particle removal studies, because they can be made with a marker that allows
alternative detection methods.
The preference for Si3N4 for particle removal studies is based on the more
challenging conditions for its removal compared to that of silica [50]. This
difference can be caused by the methods of detection, which are all based on
light scattering. The scatter cross-section of Si3N4 is much larger than that of
silica for a particle of the same size. In water the difference is almost three
times (see Figure 3.11) and in a vacuum it is 1.3 times. Consequently, a study
using a liquid particle counter in which 100 nm (LSE) Si3N4 and silica particles
are being removed is actually removing 70 nm Si3N4 and 200 nm silica
particles. It is easier to remove larger particles and therefore Si3N4 seems more
difficult to remove [51].
It is important in these kind of experiments to check the presence of dead
volume, which is areas in the tank where liquid is only refreshed by diffusion. If
a dead volume becomes filled with a significant amount of particles, it will
deliver particles to the bulk slower than the removal of particles out of the bath.
Consequently, the determined ka value will be larger than it actually is. It was
found that a full batch of 25 wafers placed in a specific way into the tank
behaves as a semi-closed box. The flow through the wafers was limited. If the
particle level in the tank was reduced to nearly zero and the wafers were taken
out of the solution, a peak in particle concentration could be observed.
104
Developments in Surface Contamination and Cleaning
0.3
PLS
0.25
Si3N4
LSE (µm)
0.2
0.15
SiO2
0.1
0.05
0
0
0.05
0.1
0.15
0.2
0.25
0.3
Particle diameter (µm)
FIGURE 3.11 The scatter cross-section of particles in water (nliq ¼ 1.34) using the Rayleigh
approximation expressed as latex spheres equivalent (LSE) [44].
Reducing the number of wafers to avoid flow restrictions between wafers was
not an option, because a large contaminated surface area is required to enable
a particle concentration 1000 times above the background level.
3.2. Particle Diffusion Out of the Boundary Layer
The boundary layer is a mathematical concept and describes the thickness of
the liquid layer on top of a solid surface where the liquid is not flowing, while
the bulk of the liquid is flowing. However, the velocity of the flowing liquid
does not decrease abruptly from maximum velocity to zero on approach to
a solid surface. Therefore, the boundary layer is defined as the distance in the
liquid from the solid surface where the velocity is smaller than 99% of the
velocity at infinite distance. In the boundary layer approximation, the boundary
layer thickness is assumed to be constant over the whole wafer surface, which
is only true when the liquid velocity is constant over the whole surface. Since
the boundary layer is a liquid layer with almost zero velocity, within the
boundary layer the only mass transport mechanism will be diffusion. In the
laminar flow regime, the boundary layer thickness depends reciprocally on
the square root of the water velocity at infinite distance and is around 1 cm in
most semiconductor applications.
A second layer that is of interest is the carry-over layer. This is the layer that is
left behind on the surface after taking the surface out of a liquid (Figure 3.12).
The carry-over layer is much thinner than the boundary layer and, in most cases,
105
Particles in Semiconductor Processing
v (cm/s)
Carry-over layer ~ 0.0007 v 2/3
~ 20 µm
FIGURE 3.12
of the liquid.
Carry-over layer is the liquid that remains on the surface after removal of the bulk
is of the order of 20 mm. All the contaminants that are left in the carry-over
layer will remain on the wafer surface when the liquid is evaporated. It is
assumed that the concentration of contaminants in the boundary layer is
uniform and, therefore, the same as in the carry-over layer. Therefore, by
using an LPC that enables the measurement of the cleaning rate of the
boundary layer, it also measures the cleaning rates of the carry-over layer.
Besides the distance a particle has to travel, its velocity also determines the
residence time in the boundary layer. Particles are placed in motion due to
random collisions with liquid molecules, which results in so-called Brownian
motion. The average kinetic energy of the particles, {1/2}mv2 with v ¼ average
velocity and m ¼ particle mass, is discrete values of {1/2}kT (here k is the
Boltzmann constant). The resulting diffusion rate of particles is described by
Einstein’s law of diffusion. This law, combined with the Stokes friction factor,
results in:
D ¼
kT
3phd
(17)
This equation shows that the diffusion coefficient (D) depends linearly on
temperature (T) and relates reciprocally to the viscosity (h) and the particle
diameter (d); the latter can also be used to convert to the mass of the particle.
To determine the effect of all variables on diffusion rates of particles out of
the boundary layer, the boundary layer is filled with particles by making use of
the carry-over layer. Immersing a batch of dry silicon wafers in a slightly
alkaline suspension of silica particles, and subsequently taking them out, will
result in a carry-over layer on the wafers filled with particles. Alkalinity
stabilizes the suspension and prevents the particles from depositing on the
wafer surface. During reimmersion of these wafers in a clean process tank,
a fraction (f) of the carry-over layer with particles will be rinsed off immediately, but a significant fraction will remain in the boundary layer from which
particles slowly diffuse into the bulk of the liquid. Measuring the particle
concentration in the bulk results in a concentration profile ([I]) described by
equation (15). An example of such a measurement was previously given in
Figure 3.8.
106
Developments in Surface Contamination and Cleaning
3.2.1. Effect of pH on kb
The effect of pH on the linear diffusion rate of particles leaving the boundary
layer was measured for 0.33 and 0.43 mm diameter silica particles [41]. The pH
of the cleaning solution was adjusted with ammonia and nitric acid and the
diffusion rate constants were calculated from the average of three to five
particle concentration measurements. The method described in Section 2.3 has
been used to fit the data and establish the one-dimensional diffusion rate
constants. Figure 3.13 summarizes the results.
The smaller silica particles indeed diffuse faster than the larger silica
particles. Surprisingly, there was a strong effect of the pH on the linear diffusion rate. For both particle sizes, the diffusion rate is highest at a pH of 1.8,
which is also known to be the point of zero charge for silica surfaces. This
leads to the conclusion that the thickness of the electrostatic double layer
(EDL) around the particles affects the effective radius of the particle and
effectively slows down the diffusion rate. In this case, the EDL thickness would
be thinnest at the point of zero charge unless the EDL is compressed by a high
salt concentration.
3.2.2. Effect of Salt Concentration on kb
The hypothesis of the impact of the compressed EDL can be tested by adding
extra salt to the aqueous solution while keeping the pH fixed and subsequently
determining kb. The results of such an experiment are summarized in Figure 3.14.
The pH of the solution is fixed with ammonia at 9.5. By adding extra NH4NO3 the
salt concentration increases. Particles diffuse faster out of the boundary layer
with high salt concentration than with low salt concentration. This confirms that
20
330 nm
430 nm
kb (ms-1)
15
10
5
0
1
1.8
3
6-7
10
11.5
12
pH
FIGURE 3.13 Linear diffusion rate constant for 0.33 and 0.43 mm silica particles in aqueous
solution at different pH values [43].
107
Particles in Semiconductor Processing
18
15
kb (ms-1)
12
9
6
330 nm
3
430 nm
0
0
10
20
30
40
50
[NH4+] (mmol/L)
FIGURE 3.14 Linear diffusion rate constant of 0.33 and 0.43 mm diameter silica particles in
aqueous solution at pH 9.5 and various salt concentrations (NH4NO3). The dashed line is the
maximum measured diffusion rate at pH 1.8.
the thickness of the EDL has an effect on the diffusion of particles out of the
boundary layer. However, the maximum diffusion rate at pH 9.5 with increased
salt concentration is still lower than the diffusion rate at a pH of 1.8. Several
explanations can be proposed for this phenomenon.
3.2.3. Effect of Temperature and Viscosity on kb
According to equation (17), increasing temperature will accelerate the diffusion
process and thus the removal of the particles out of the boundary layer. This is
due to the higher velocity of the particles, but also due to the reduced viscosity
of the liquid resulting in thinning of the boundary layer. Consequently, the
distance through which the particle has to migrate is shorter. The measured
diffusion rate in this experimental setup is a combined effect. To study the effect
separately, a constant viscosity at different temperatures is obtained by the
addition of polyethylene glycol (PEG) to the cleaning solution. The side-effect
of adding PEG, although concentrations are kept low (0.005–0.04 mmol L1), is
that the dielectric constant of the liquid medium changes as well, which in turn
influences the EDL. The results for 0.43 mm particles are summarized in
Figure 3.15.
It was no surprise that the diffusion rate increased with increasing temperature and decreased as the viscosity increased by the addition of PEG. However,
looking at the data points marked with an asterisk in Figure 3.15, which have an
identical viscosity at different temperatures, the diffusion rate does not increase
with increasing temperature, instead it decreases by 24%. This means that the
PEG, which is also a non-ionic surfactant, chemically or physically interacts
with the particles, which results in an increase in the effective radius/mass of
the particles and thereby decelerates the diffusion process.
108
Developments in Surface Contamination and Cleaning
18
15
kb (ms-1)
12
PEG (g/L)
0.00
9
1.62
4.63
6
13.89
3
0
*
20
*
*
25
30
*
35
T (ºC)
FIGURE 3.15 Linear diffusion rate constant of 0.43 mm silica particles as functions of
temperature and PEG 8000 concentration. Bars marked with an asterisk have identical viscosity at
a fixed pH of 3.5 [43].
3.2.4. Effect of Surfactant on kb
Surfactants are considered to aid the particle removal process, while in the
previous experiments it was shown that the surfactant properties of PEG reduce
the diffusion of the particles out of the boundary layer. Therefore, smaller
surfactant molecules were used to study the specific effect of surfactants on kb.
The selected surfactants were a cationic surfactant, benzalkonium chloride
(BAC; 0.02 mmol L1), and a non-ionic surfactant, polyoxyethylene(12) tridecyl ether (0.006 mmol L1), which is similar to PEG but with a much lower
molecular weight.
The same experiments as with PEG were repeated, but now in the presence of
a surfactant (BAC) whose concentration was kept well below the critical micelle
concentration (CMC). At pH 3.5 the diffusion rate is considerably reduced by the
addition of a surfactant (compare Figure 3.16 with Figure 3.15). The effect of
PEG in the solution containing BAC has become insignificant and the only
remaining effect is the impact on viscosity. In solutions with constant viscosity,
the diffusion rate now tends to increase slightly with increasing temperature.
In the solution without any PEG, the effect of surfactant is greater than if
PEG was present. To visualize this effect, the change in diffusion rate constant
has been plotted as a ratio of kb with surfactant to kb without surfactant
(Figure 3.17). The effect of temperature was removed by taking the average
value of the four temperatures.
At pH 3.5, both the cationic and non-ionic surfactants decrease the diffusion
rate in solutions without PEG, while at pH 10 the addition of the surfactants
seems to have no effect. Increasing the PEG concentration, the effect of
109
Particles in Semiconductor Processing
18
15
kb (ms-1)
12
PEG (g/L)
0.00
9
1.62
4.63
13.89
6
3
*
0
*
20
*
25
30
*
35
T (ºC)
FIGURE 3.16
added [43].
The results of the same experiments as in Figure 3.15, but with 7 mmol L1 BAC
additional surfactant at pH 3.5 is canceled out, whereas at pH 10 the diffusion
rate initially is accelerated with increasing PEG concentration, but this effect
subsequently also fades.
All these observations are explained by a mechanism that impacts either the
effective particle radius, i.e. the particle diameter and some part of the electrostatic double layer, or the effective particle mass, i.e. the particle mass plus
the material adsorbed on it.
No surfactant and no PEG. The pH impacts the effective radius of the
particle (see Figure 3.13). At pH 3.5, the zeta potential of the silica particle
is smaller than at pH 10 and, therefore, the electrostatic double layer at
pH 10 (R5 in Figure 3.18) is thicker than R1.
With surfactant and no PEG. If only a cationic surfactant is adsorbed on the
surface, the surface charge is shielded, resulting in a thinner electrostatic
double layer compared to no adsorption (R1 and R5 versus R2 and R6
respectively). Also, the effective mass of the particle slightly increases.
The net effect on the diffusion rate is that at pH 3.5 it is significantly
reduced, and at pH 10 there is a balance and no effect is measured
(Figure 3.17 at zero PEG concentration).
With surfactant and with PEG. Surfactants and PEG have different impacts.
These molecules are in a competition for adsorption sites on the silica particle.
Replacing a surfactant molecule with PEG implies less charge shielding and
thus an increased effective radius, but the total mass of the particle will
increase. For this reason, the diffusion rate constant increases when the PEG
concentration is not too large (between 0 and 10 g L1 in Figure 3.17).
110
Developments in Surface Contamination and Cleaning
1.2
cationic
kb + surf / kb no surf
non-ionic
1.0
0.8
0.6
pH 3.5
0.4
0
5
10
15
PEG (g/L)
1.6
kb + surf / kb no surf
cationic
non-ionic
1.4
1.2
1.0
pH 10
0.8
0
5
10
15
PEG (g/L)
FIGURE 3.17 Ratio of kb with surfactant and kb without surfactant at different PEG concentrations (average overall temperatures). Left figure is at pH 3.5. Right figure is at pH 10 [43].
No surfactant and with PEG. PEG adsorbed on the particle surface will have
only a minor effect on the effective radius, but it will increase the effective
mass. Besides the increased viscosity, the higher mass also reduces the diffusion rate (Figure 3.15). In the presence of the surfactant, PEG competes for
surface sites. If the PEG concentration is high enough, the surfactant molecules
are expelled from the surface by PEG. Therefore, the effect of the combination
of surfactant and PEG at 14 g L1 is essentially negligible (Figure 3.17).
One of the contributing actions of megasonics, spinning, or droplet bombardment is that it reduces the thickness of the boundary layer. A thinner boundary
layer means that the particles are removed away faster from the wafer surface.
This results in increased particle removal efficiency.
+
-
-
-
-
+
-+
R2
+
+
+
+
pH
=
10
-
+
R5
+
--
+
-
- +
-
+
+
+
+
+
+
-
m6
+
-
+
+
R7
+
+
-
-
-
-
-
-
-
-
+
-
+
-
+
m7
R8
+
+
-
-
+ = cationic surfactant
+
+
+
m8
= PEG
Impact of pH, surfactant, and PEG on the thickness of the electrostatic double layer and the effective mass of the particle. Sizes are relative and
111
FIGURE 3.18
notional.
-
R6
-
m5
+
+
+
+
-
+
m4
+
+
-
+
+
+
+
-
-
m3
R4
-
+
+
m2
m1
-
-
-
+
R3
-
-
-
-
+
+
+ PEG, no surfactant
+
+
+
+
-
-
+
R1
-
+
+ PEG, + surfactant
+
+
pH
=
3.5
no PEG, + surfactant
Particles in Semiconductor Processing
no PEG, no surfactant
112
Developments in Surface Contamination and Cleaning
3.3. Particle Detachment
Once all rate constants have been determined, the first step of the removal
process can be studied, i.e. the detachment of particles from the surface by
chemical processes. To this end, monodisperse silica particles (average size
0.33 mm) were deposited from an aqueous suspension on to wafers. Using
a diluted APM solution at 35 C as a slow-etching cleaning solution [52], the
freshly deposited particles came from the dried wafer surface in two waves
(Figure 3.9). In the first wave, the loosely bonded particles come off upon the
immersion of the wafers in the cleaning solution. This is probably due to
hydrodynamic forces (either surface tension or drag forces) acting on the
particles at the gas–solid–liquid interface. The second wave started much later.
This peak is almost Gaussian in shape, but is slightly skewed because of the
earlier-determined diffusion (kb) and filtration (kc) processes (Figure 3.19). The
time scales of the detachment process and the subsequent diffusion and
filtration are so much different that the integration of the respective time
function (equation (15)) is not required. This can change if the detachment
process is accelerated by higher etch rates or more physical power.
3.3.1. Impact of Deposition Condition on ka
The measurement system to determine the particle removal rate is strongly
impacted by the storage conditions of the particle-contaminated wafers prior
to the cleaning process [53]. The effect on the process is observed in the
height of the first peak and the average removal time of the second peak. If the
Hydrodynamic Forces
4000
delays caused by
kb(diffusion), kc
(filtration)
3000
height
Number of particles (/mL)
5000
2000
Etching of SiO2
1000
0
0
1000
2000
3000
time (s)
average removal time
FIGURE 3.19 Two key parameters, height of first peak and average removal time, used to
described the particle aging process.
113
Height 1st peak (1000 particles/mL)
Particles in Semiconductor Processing
6
40%
70%
100%
5
4
3
2
1
0
0.1
1
10
100
1000
Storage time (h)
FIGURE 3.20 Intensity of first peak as a function of storage time of particle-contaminated
wafers (RH ¼ 40%, 70%, and 100%).
Average removal time (s)
particle-contaminated wafers were stored for longer times, the height of the
first peak decreases and the average removal time of the second peak changes.
Under normal cleanroom conditions (relative humidity (RH) between 40% and
70%), the first peak disappears within 10 hours of storage (Figure 3.20). If RH
during storage was set at 100%, the intensity of the first peak remained constant
for the first 24 hours, after which the peak slowly decreased.
The dependence of the average removal time on the storage conditions is
depicted in Figure 3.21. The time to maximum of the peak increased at RH
40–70% from 2000 to 3000 seconds within the first 6 hours. After 6 hours the
average removal time remained constant at around 3000 seconds. This means
that more etching is required to remove all particles. During the first 24 hours of
storage at RH of 100% (Figure 3.22), the removal time remained constant at
around 1700 seconds; thereafter it increased to around 2300 seconds from
36 hours’’ storage onwards. Although in the experiments under dry conditions
4000
3000
2000
1000
0
0
2
4
6
8
Storage time (hours)
FIGURE 3.21 Average removal time as a function of storage time of particle-contaminated
wafers (RH ¼ 40%, 70%). The dots are 70% RH and the triangles are 40% RH measurement points.
114
Developments in Surface Contamination and Cleaning
Average removal time (s)
3000
2000
1000
0
1
10
100
1000
Storage time (hours)
FIGURE 3.22 Average removal time as a function of storage time of particle-contaminated
wafers (RH ¼ 100%).
(RH < 6%) the humidity conditions decreased during the first few hours, the
average removal time was high and remained high (between 3000 and
4000 seconds) right from the start.
With the assumption that longer etching means stronger particle–substrate
interaction, the experiments seem to indicate that removal of particles becomes
more difficult upon storage. After the application and dry-in, particles are
attached to the surface with certain strength, which will not be equal for all
particles. On the basis of Figure 3.9, this strength is considered to be normally
distributed over the particles (Figure 3.23). If external forces are applied during
the immersion, the more weakly bonded particles will yield and detach from the
surface, resulting in the first particle wave. During the subsequent etching
process, the remainder of the particles will be removed, giving rise to the
second particle wave. The effect of aging/storage is that the average adhesion
Number of particles
aging
Immersion
forces
2nd peak
particles
1st peak
particles
Particle adhesion force
FIGURE 3.23
Adhesion force distribution of particles attached to a surface.
115
Particles in Semiconductor Processing
strength increases and that the immersion forces are not sufficient to remove
particles. As a result, the first peak will vanish and the second peak that is fed by
the slightly weaker bonded particles could show up initially a little earlier in the
etch process, but eventually it will shift to longer etch times for complete
removal.
The effect of moisture on aging can be understood by the capillary forces
causing a thin condensed water layer to exist between the particle and the wafer
surface (Figure 3.24). Dry-in of such a condensation layer results in a shorter
particle–substrate distance up to an extent that the two interfaces are in the
range of the van der Waals force. Subsequently, capillary forces acting on the
particles cause the particles to deform, whereby a larger contact area is created
and thus stronger adhesion is established by van der Waals forces. In the case of
100% humidity, the condensed layer is too thick to cause capillary forces to act
on and deform the particles. However, in the first step, evaporation/condensation is likely to occur at 100% RH and, therefore, the first particle wave also
disappears under these conditions (Figure 3.24).
These results can be explained by assuming the first particle wave is due
to particles floating on a thin water layer between the particle and the
substrate. In the second wave, particles are detached from the surface, which
are in the range of van der Waals forces. The increase in the adhesion force
of the particles in the second wave is due to the particle deformation induced
by capillary forces. Additionally, as explained later, the increase in adhesion
can be a result of dissolved and subsequently precipitated silicates, which
create actual chemical bonds between the particle and the substrate during
the aging process.
We have shown that the storage conditions during the time the particles are
applied and removed have a great impact on the adhesion strength of particles
on a silicon surface. This implies that particle removal studies using this kind of
model system will have a problem with reproducibility and repeatability. This
will depend on the laboratory, or even on the researcher, as to what the absolute
outcome will be of a particle removal study. However, general trends will
remain evident if the contamination part of the cleaning experiment remains
exactly the same.
evaporation /
condensation
van der Waals forces
capillary forces
aging
FIGURE 3.24 Evaporation of the condensation layer will result in stronger particle adhesion and
eventually lead to a deformed particle having a larger contact area.
116
Average removal time (s)
Developments in Surface Contamination and Cleaning
2000
1500
1000
500
0
1
3
5
7
9
11
13
pH
FIGURE 3.25 Effect of pH of the particle suspension used to contaminate the wafers on the
removal process.
3.3.2. Impact of Deposition Conditions
Particles that are deposited from a dry aerosol are relatively easy to remove.
This is disadvantageous for particle removal studies because the method does
not discriminate between a good and a better clean [54]. To make this more
challenging, these particle-contaminated wafers can be immersed in a liquid,
dried, and subsequently used for a cleaning experiment. The particles become
more difficult to remove, which means that the deposition conditions impact
upon the adhesion of the particles.
In the previous experiments, silica particles were deposited from a neutral or
slightly alkaline solution. If the deposition is done from a particle suspension at
pH 2, the particles are very easy to remove. All the particles were removed or
detached by immersion of the wafers into the liquid, i.e. there was no second
peak. Increasing the pH of the particle suspension, the second peak appeared
again and the average removal time increased with increasing pH (Figure 3.25).
The first peak from the experiment with particles deposited at pH 2 was broader
than expected for a process controlled only by ka. This indicates that a fraction
of the particles ends up in the boundary layer upon immersion, which causes kb
to become dominant in the removal rate.
An explanation for the effect of pH on the adhesion strength of the deposited
particles is that particles deposited at higher pH could be ‘‘glued’’ on the
surface [55]. By dissolution and redeposition of silica at the particle–substrate
interface (Figure 3.26), the particles become chemically bound to the surface.
SiO2 + OH-
FIGURE 3.26
SiO3H-
Particles are glued to the surface.
Particles in Semiconductor Processing
117
FIGURE 3.27
SEM picture (left) of a surface formally covered with 1.5 mm sized silica
particles leaving residues that have also been measured with AFM (right). (see colour plate section
at end for coloured version)
Courtesy of Frank Holsteyns [57].
This would not happen at low pH, because the solubility of SiO2 is much lower.
Holsteyns [55] has obtained SEM images of the residue of such ‘‘glue’’ after
the particles were removed (Figure 3.27).
4. CONCLUSIONS
Particle research in the semiconductor industry is very pragmatic, but also very
challenging. A state-of-the-art complementary metal oxide semiconductor
(CMOS) manufacturer is interested in the removal of particles with no or
minimal substrate loss, while in mature manufacturing the goal is to have ‘‘zero
defects’’, i.e. less than one out of a million products that are allowed to fail. Due
to this competitiveness, advances in fundamental research on particle removal
are commercially implemented within a couple of years.
Research is focused on mechanical and/or chemical enhancement of particle
removal, but with minimal substrate etching. Addition of surfactants to
cleaning solutions to aid particle removal might inhibit efficient particle
removal if surface tension forces are required to lift the particle, or if particle
diffusion away from the surface is a rate-limiting step.
Prevention of particle deposition is not yet a major research topic. Particles
in the wafer environment causing random yield loss are not part of current yield
models. It is our intention to do so in future.
ACKNOWLEDGEMENTS
The authors would like to acknowledge all the students who have worked on
this particle project throughout the years: Yolaine Dumesnil, Sander Wolters,
Roy te Brake, Michiel Enkelaar, Melvin Kasanrokijat, Romuald Roucou, Remi
Peyrin, Florian le Goupil, Federic Michel, Adrien Maurel, Michel van Straten,
Wybe Roodhuizen, and Clement Sieutat.
118
Developments in Surface Contamination and Cleaning
REFERENCES
[1] The International Technology Roadmap for Semiconductors, 2008 edition, International
Sematech, San Jose, CA; http://www.itrs.net/, 2009.
[2] T. Kim, W. Kuo, Modeling manufacturing yield and reliability, IEEE Trans. Semicond.
Manuf. 12 (1999) 485.
[3] J.T. Wallmark, Design considerations for integrated electronic devices, Proc. IRE 48 (1960) 293.
[4] S.R. Hofstein, F.P. Heiman, The insulated-gate field-effect transistor, Proc. IEEE 51 (1963) 1190.
[5] B.T. Murphy, Cost-size optima of monolithic integrated circuits, Proc. IEEE 52 (1964) 1537.
[6] C.H. Stapper, Defect Density Distribution for LSI Yield Calculations, IEEE Trans. Electron
Dev. 655 (1973) 20.
[7] W. Kuo, T. Kim, An Overview of Manufacturing Yield and Reliability Modeling for
Semiconductor Products, Proc. IEEE 1329 (1999) 87.
[8] R.S. Hemmert, Poisson Process and Integrated Circuit Yield Prediction, Solid-State Electr.
511 (1981) 24.
[9] C.H. Stapper, Integrated Circuit Yield Statistics, IEEE Trans. Semicond. Manuf. 294 (1991) 4.
[10] F. Wali, D.M. Knotter, F.G. Kuper, Impact of Particles in Ultra Pure Water on the Random
Yield Loss in IC Production, Microelectronic Eng. 140 (2009) 86.
[11] F. Wali, D.M. Knotter, T. Bearda, P.W. Mertens, Local Distribution of Particles Deposited on
the Structured Surfaces, Solid State Phenom. 65 (2009) 145–146.
[12] L. Mouche, F. Tardif, Mechanisms of Particle Removal from Silicon Wafer Surface in Wet
Chemical Cleaning Process, J. Electrochem. Soc. 1684 (1995) 141.
[13] W. Fyen, K. Xu, R. Vos, G. Vereecke, P.W. Mertens, M. Heyns, Particle Deposition from
a Carry-Over Layer During Immersion Rinsing, in: K.L. Mittal (Ed.), Particles on Surfaces 8:
Detection, Adhesion and Removal, VSP, Utrecht, The Netherlands, 2003, pp. 77–128.
[14] W. Fyen, P.W. Mertens, Study of the Carry-Over Layer and its Importance for Wet Processing
of Semiconductor Substrates, J. Electrochem. Soc. 443 (2006) 153.
[15] D.M. Knotter, Cleaning for Sub 0.1 mm Technology: a Particular Challenge, in: C. Claeys,
F. Gonzalez, R. Singh, J. Murota, P. Fazan (Eds.), Proc. ULSI Process Integration III,
PV2003-6, The Electrochemical Society, Pennington, NJ, 2003, p. 349.
[16] T. Hattori, Detection and Analysis of Particles in Production Lines, in: T. Hattori (Ed.), Ultra
Clean Surface Processing of Silicon Wafers, Springer-Verlag, New York, 1998, pp. 243–258.
[17] A. Mayer, S. Schwartzman, Post-CMP Cleaning Using Acoustic Streaming, J. Electronic
Mater. 855 (1979) 6.
[18] K. Xu, Nano-Sized Particles: Quantification and Removal by Brush Scrubber Cleaning, PhD
Thesis, K.U. Leuven, Leuven, Belgium (2004).
[19] J. Visser, Particle Adhesion and Removal: A Review, Particulate Sci. Technol. 169 (1995) 13.
[20] I. Kanno, N. Yokoi, K. Sato, Wafer Cleaning by Water and Gas Mixture with High Velocity, in:
J. Ruzyllo, R. Novak (Eds.), Proc. Cleaning Technology in Semiconductor Device
Manufacturing V, The Electrochemical Society, Pennington, NJ, 1998, p. 54. PV 97–35.
[21] H. Hirano, K. Sato, T. Osaka, H. Kuniyasu, T. Hattori, Damage-Free Ultradiluted HF/
Nitrogen Jet Spray Cleaning for Particle Removal with Minimal Silicon and Oxide Loss,
Electrochem. Solid-State Lett. 62 (2006) 9.
[22] N. Narayanswami, J. Heitzinger, J. Patrin, Development and Optimization of a Cryogenic
Aerosol-Based Wafer Cleaning System, in: K.L. Mittal (Ed.), Particles on Surfaces 5&6:
Detection, Adhesion and Removal, VSP, Utrecht, The Netherlands, 1999, pp. 251–266.
[23] C. Toscano, G. Ahmadi, Particle Removal Mechanism in Cryogenic Surface Cleaning,
J. Adhesion 175 (2003) 79.
[24] G. Vereecke, E. Röhr, M.M. Heyns, Laser-Assisted Removal of Particles on Silicon Wafers,
J. Appl. Phys. 3837 (1999) 85.
[25] I. Varghese, C. Cetinkaya, Non-Contact Removal of 60 nm Latex Particles from Silicon
Wafers with Laser Induced Plasma, J. Adhesion Sci. Technol. 795 (2004) 18.
Particles in Semiconductor Processing
119
[26] D.M. Knotter, S. de Gendt, P.W. Mertens, M.M. Heyns, Silicon Surface Roughening Mechanisms in Ammonia Hydrogen Peroxide Mixtures, J. Electrochem. Soc. 736 (2000) 147.
[27] G. Briend, P. Besson, T. Salvetat, S. Petitdidier, Impact of Megasonic Activation with
Different Chemistries on Silicon Surfaces in Single Wafer Tool, Solid State Phenom. 15
(2009) 145–146.
[28] W. Kern, D.A. Puotinen, Cleaning Solutions Based on Hydrogen Peroxide for Use in Silicon
Semiconductor Technology, RCA Review 187 (1970) 31.
[29] M. Meuris, S. Verhaverbeke, P.W. Mertens, H.F. Schmidt, A.L.P. Rotondaro, M.M. Heyns,
A. Philipossian, A New Cleaning Concept for Particle and Metal Removal on Si Surfaces, in:
J. Ruzyllo, R.E. Novak (Eds.), Proc. Cleaning Technology in Semiconductor Device
Manufacturing III, The Electrochemical Society, Pennington, NJ, 1994, pp. 15–25. PV 94–07.
[30] A. Mayer, S. Schwartzman, Megasonic Cleaning: A New Cleaning and Drying System for
Use in Semiconductor Processing, J. Electronic Mater. 855 (1979) 8.
[31] G. Vereecke, E. Parton, F. Holsteyns, K. Xu, R. Vos, P.W. Mertens, M. Schmidt, T. Bauer,
Evaluation of Megasonic Cleaning for Sub-90-nm Technologies, Solid State Phenom. 146
(2004) 14.
[32] A. Otto, T. Nowak, R. Mettin, F. Holsteyns, A. Lippert, Characterization of a Cavitation
Bubble Structure at 230 kHz: Bubble Population, Sonoluminescence and Cleaning Potential,
Solid State Phenom. 11 (2009) 145–146.
[33] A. Phillipp, W. Lauterborn, Cavitation Erosion by Single Laser-Produced Bubbles, J. Fluid.
Mech. 75 (1998) 361.
[34] K. Xu, S. Pichler, K. Wostyn, G. Cado, C. Springer, G. Gale, M. Dalmer, P.W. Mertens,
T. Bearda, E. Gaulhofer, D. Chen, Removal of Nano-Particles by Aerosol Spray: Effect of
Droplet Size and Velocity on Cleaning Performance, Solid State Phenom. 31 (2009) 145–146.
[35] A.C. Engelsberg, Removal of Surface Contaminants by Irradiation from a High-Energy
Source, US Patent (1991) 5,024,968.
[36] K.L. Mittal (Ed.), Particles on Surfaces 8: Detection, Adhesion and Removal, VSP/Brill,
Leiden, The Netherlands, 2003.
[37] K.L. Mittal (Ed.), Particles on Surfaces 9: Detection, Adhesion and Removal, VSP/Brill,
Leiden, The Netherlands, 2006.
[38] J. Visser, in: E. Matijevic (Ed.), Surface and Colloid Science, Adhesion of Colloidal Particles,
Vol. 8, John Wiley, New York, 1976, pp. 3–84.
[39] K. Bakhtari, A.O. Guldiken, A.A. Busnaina, J. Park, Experimental and Analytical Study of
Submicrometer Particle Removal from Deep Trenches, J. Electrochem. Soc. 153 (2006) 9.
[40] A.F.M. Leenaars, Methods of Removing Undesired Particles from a Surface of a Substrate,
US Patent (1988) 4,781,764.
[41] P.A. Kittle, Surface Treatment of Semiconductor Substrates, US Patent (2000) 6, 090, 217.
[42] D.M. Knotter, R. te Brake, M. Enkelaar, In Situ Particle Removal Studies Using an Optical
Particle Counter, Solid State Phenom. 189 (2008) 134.
[43] D.M. Knotter, S.A.M. Wolters, M.A. Kasanroijat, Particle Concentration Measurements in
Process Liquids Using Light-Scattering Techniques, Particulate Sci. Technol. 435 (2007) 25.
[44] W. Stöber, A. Fink, E. Bohn, Controlled Growth of Monodisperse Silica Spheres in the
Micron Size Range, J. Colloid Interface Sci. 62 (1968) 26.
[45] A. Van Blaaderen, J. Van Geest, A. Vrij, Monodisperse Colloidal Silica Spheres from
Tetraoxysilanes: Particle Formation and Growth Mechanism, J. Colloid Interface Sci. 2 (1992) 152.
[46] F. Wali, D.M. Knotter, J.J. Kelly, F.G. Kuper, Deposition and Detection of Particles During
Integrated Circuit Manufacturing, in: Proc. 9th SAFE Workshop Semiconductor Advances
for Future Electronics and Sensors, Technology Foundation STW, The Netherlands, 2006,
pp. 483–487.
[47] F. Wali, D.M. Knotter, J.J. Kelly, Preparation of Mono-Disperse Silica Particles with MetalIon Tracer, paper presented at 11th International Symposium on Particles on Substrate:
Detection, Adhesion and Removal, Orono, ME (2008).
120
Developments in Surface Contamination and Cleaning
[48] D.J. Shaw, Introduction to Colloid and Surface Chemistry, third ed. Butterworth, London,
1980, Chapter 6.
[49] M. Tokuse, R. Oyama, H. Nakagawa, I. Nakabayashi, Characterization of the Oxidized Si3N4
Whisker Surface Layer Using XPS and TOF-SIMS, Japan Soc. Anal. Sci. 281 (2001) 17.
[50] R. Vos, I. Cornelissen, M. Meuris, P. Mertens, M. Heyns, Optimisation of dHF-based
Cleaning Recipes: To Remove or Not to Remove a Particle? in: T. Hattori, R.E. Novak,
J. Ruzyllo (Eds.), Proc. Cleaning Technology in Semiconductor Device Manufacturing VI
The Electrochemical Society, Pennington, NJ, 1999, pp. 461–468. PV 99-36.
[51] A.A. Busnaina, J. Taylor, I. Kashkoush, Measurements of Adhesion and Removal Forces of
Submicron Particles on Silicon Surfaces, J. Adhes. Sci. Technol. 441 (1993) 7.
[52] D.M. Knotter, R. Roucou, R. Peyrin, Reduced Particle Removal Efficiency upon Wafer
Storage, Solid State Phenom. 61 (2009) 145–146.
[53] G. Vereecke, F. Holsteyns, S. Arnauts, S. Beckx, P. Jaenen, K. Kenis, M. Lismont, M. Lux,
R. Vos, J. Snow, P.W. Mertens, Using Megasonics for Particle and Residue Removal in Single
Wafer Cleaning, Solid State Phenom. 141 (2005) 103.
[54] N. Narayanswami, P.A. Ruether, G. Thomes, J.F. Wygand, N. Lee, K.K. Christenson,
J.W. Butterbaugh, Method for Evaluation and Optimization of Particle Removal Processes,
Electrochem. Soc. Proc. 469 (1999) 99.
[55] F. Holsteyns, Removal of Nanoparticles Contaminants from Semiconductor Substrates by
Megasonic Cleaning, PhD Thesis, K.U. Leuven, Leuven, Belgium (2008).