Review

Particle-in-Cell/Test-Particle Simulations of
Technological Plasmas: Sputtering Transport
in Capacitive Radio Frequency Discharges

Jan Trieschmann,* Frederik Schmidt, Thomas Mussenbrock

The paper provides a tutorial to the conceptual layout of a self-consistently coupled Particle-In-
Cell/Test-Particle model for the kinetic simulation of sputtering transport in capacitively
coupled plasmas at low gas pressures. It explains when a kinetic approach is needed and which
numerical concepts allow for capturing the often V (t) V (t) rf B
x-axis
observed nonequilibrium behavior of the charged and
0 L
neutral particles. At the example of a generic sputtering C B

discharge, both the fundamentals of the applied Monte

Carlo methods as well as their conceptual details Charged particles Interaction Neutral particles
are elaborated on. Finally, two assumptions that are
often exploited in the context of sputtering transport Non-equilibrium
Surface
Gas background
simulations, namely on the energy distribution of
impinging ions as well as on the test particle approach, Equilibrium
Gas phase
Energetic neutrals

are validated for the proposed example discharge.

1. Introduction into the electrons. Due to their relatively low mass

compared to ions and their corresponding small inertia,
Within many technological process chains, low tempera- electrons respond to an external excitation almost instan-
ture plasmas are indispensably utilized. One of the most taneously. In contrast, ions typically react on a much slower
important feature to be found in many of these plasmas timescale and, in fact, often only on time averaged
(often referred to as technological plasmas) is the property quantities. These aspects are of major concern regarding
of nonequilibrium, in terms of (i) the individual plasma the operational principle of capacitively coupled plasmas
species relative to one another and/or (ii) a deviation from (CCPs) driven by high-frequency electric fields. However,
thermal equilibrium and thus from a Maxwell-Boltzmann this shall not be elaborated in this work. Great reference
energy distribution. Typically, the individual characteristic is provided elsewhere.[1,2]
kinetic energies of the constituents of the plasma of concern While the understanding of plasma physics based on
(electrons, ions, radicals, etc.) are diversely spread from meV experiments is essential (as all physical phenomena are
for neutral background particles over eV for electrons and incorporated by nature), they often provide only indirect
energetic neutrals (e.g., sputtered particles) up to keV in and/or limited insight due to limited access to the physical
case of ions which have undergone acceleration by the quantities of interest. Moreover, in particular in the context
electric field, which is present in the sheath at all material of industrial manufacturing the prevention of any kind of
boundaries of the plasma. Of importance for these perturbation (e.g., due to applied diagnostic apparatuses) is
discharges is that electromagnetic energy coupled into a clear premise. Simulation approaches, however, offer
the plasma by external sources is predominantly coupled powerful add-ons or even alternatives. At the expense of
limited physical complexity, the most intrinsic physical
J. Trieschmann, F. Schmidt, Prof. T. Mussenbrock processes are actually available for analysis. As such,
Department of Electrical Engineering and Information Science, theoretical models can be utilized for both prediction and
Ruhr University Bochum, Bochum 44780, Germany optimization of relevant plasma processes. It is imperative
E-mail: [email protected] to acknowledge that various conceptually different

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim wileyonlinelibrary.com DOI: 10.1002/ppap.201600140 1
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
 J. Trieschmann, F. Schmidt, T. Mussenbrock

simulation methods are principally applicable in order to excitation, and various dissociation processes in the case
describe technological plasmas. of molecular gases. The electron energy relaxation length
The purpose of this paper is twofold: (i) As a tutorial paper is approximately equal to the effective mean free path
it provides a conceptual layout of a self-consistently for all inelastic processes. The mean free path for each
coupled Particle-In-Cell/Test-Particle model which treats inelastic process depends of course on the details of the
all species of the plasma in a kinetic framework and (ii) as a cross section for the process as well as on the electron
research paper it reports on and discusses new results of the energy distribution function.
physics of sputtering transport in radio-frequency (rf) For neutrals the energy relaxation length is in fact the
driven capacitive discharges at low gas pressures. The paper mean free path. If the mean free path l is larger than the
is organized as follows: Firstly, it is explained when a kinetic typical dimension of the discharge L, the characteristic
approach is needed and which numerical concepts actually Knudsen number Kn ¼ l=L is larger than unity. Under these
allow for a description of the nonequilibrium behavior of circumstances, a kinetic approach for the heavy particles is
the particles. Secondly, using the example of a sputtering needed, even when the particles’ distribution functions are
discharge the fundamentals of Monte Carlo methods as close to thermal.[4] In Figure 1 various modeling approaches
well as the conceptual details in the context of the specified and the respective Knudsen numbers are depicted. (In his
sputtering deposition scenario are elaborated on. Calcula- fundamental work on the direct simulation Monte Carlo
tion results of the given exemplary case are presented and method, Bird provides a very illustrative schematic of the
final conclusions are drawn. physical regimes and conceptual models appropriate[5]).
For the topic of this paper, the transitional flow regime
and the free molecular flow regime and, therefore, the right
2. Simulation Concepts part of Figure 1 are of importance. Precisely, when the
degree of collisional interactions is small but not negligible.
Not all simulation methods are suitable for describing the In this respect, collisions can be assumed to be solely binary,
various kinds of plasmas and all physical phenomena on or interactions with the walls. In the transitional, regime
each length and time scale of interest. Before formulating a particle interactions are predominantly direct binary
model a careful analysis of the time and length scales is encounters (i.e., two particles directly interact). In contrast,
necessary. Based on a scale analysis, one is able to find for a free molecular flow the trace particles merely interact
arguments for and against the choice of a particular model. with a background and can, as such, be treated as physically
For instance, the energy relaxation length is crucial. It independent particles throughout the simulation. Despite
characterizes the length scale of Maxwellization.[3] In low very low gas pressures, in the situations encountered for
temperature, plasmas collisions between charged par- low temperature plasmas, not only neutral but also charged
ticles are relatively rare so that collisions between species are involved. Hence, the constituents can be
charged particles and neutral particles dominate the grouped into: (i) charged species (electrons and ions) and
dynamics. Ions and neutrals are of comparable mass so (ii) neutral species (thermal and fast neutrals, excited
that elastic ion-neutral collisions involve significant species, or sputtered neutral particles). Both groups can be
energy transfer. The energy relaxation length for ions described by means of Lagrangian computer models
is therefore approximately equal to the mean free path for following the Monte Carlo approach.
ion-neutral collisions. The situation is different for
electrons. Due to the large mass difference between
electrons and neutrals, only a small portion of energy is Transitional Molecular
Inviscid Viscous
(Knudsen) (free)
transferred by elastic collisions. That means that a large
number of collisions is needed for Maxwellization of Collisionless
Boltzmann equation Boltzmann
electrons. The energy relaxation length of electrons which equation
only undergo elastic collisions with neutrals is approxi-

1=2 Euler Navier-Stokes
mately le
lel;e mg =me , where lel;e is the mean free
equations equations
path for elastic electron-neutral collisions, while mg and
Computational
me are the mass of a neutral particle and the mass of an Monte Carlo Methods
Fluid Dynamics
electron, respectively. When inelastic collisions of elec-
0 0.01 0.1 1 10 100 ∞
trons and neutrals are significant, the energy relaxation
length will decrease. For large average electron energies Local Knudsen number Kn
the contribution of inelastic collisions may dominate.
Figure 1. Illustration of the respective physical regimes and
Inelastic processes include ionization, electron attach- the corresponding theoretical approaches in the form of a
ment, and electronic excitation in the case of atomic mathematical description as well as computer simulations. The
gases and additionally rotational excitation, vibrational figure is adopted from.[5]

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

2 ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim DOI: 10.1002/ppap.201600140
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
Particle-in-Cell/Test-Particle Simulations of Technological. . .

It is important to acknowledge that, in particular, for the consequently dominates the collisional sputtered particle
former a solution to the coupled system of Boltzmann’s transport of concern.
equation and Maxwell’s equations is sought after. For the To drive the discharge an ideal voltage source Vrf(t)
present case of weakly ionized plasmas, the proposed is applied at the left electrode (connected via a blocking
approach is well suited. However, the list of methods capacitor with CB ¼ 1 mF). A radio-frequency voltage
mentioned is not comprising. Alternative approaches not waveform V rf ðtÞ ¼ V 0 ½cosðvt þ DfÞ þ cosð2vtÞ is applied
detailed may be applicable as well, for example, the Barnes- with the angular frequency v ¼ 2pf , the driving frequency
Hut tree method.[6,7] These may have additional benefits f ¼ 13.56 MHz, and a voltage amplitude V0 ¼ 300 V  a
and be suitable also for other physical regimes (e.g., for fundamental frequency and its harmonic. Utilizing the
the description of dense plasmas). electrical asymmetry effect (EAE), this geometrically
symmetric discharge can be made electrically asymmetric
with a choice of the relative phase between the two
3. Simulation Scenario harmonics of Df ¼ 0.[10,11] The impact of which will become
more clear in the course of this paper. The right electrode is
Using the example of a capacitively coupled sputtering held at ground potential.
discharge driven by a radio-frequency power supply, all
relevant components required for a consistent theoretical
3.2. Operational Regime
description shall be briefly discussed. This scenario is
particularly chosen as it requests a number of peculiar For what follows a principle understanding of the internal
aspects to be taken into account, which at the same time can kinetics of the discharge is important:[1] (i) Within the CCP
be instructively decomposed individually. The required environment, the plasma is operated in a continuous
background in plasma physics is intentionally kept to a rf mode. A periodic steady-state is strictly expected.
fundamental level of knowledge to not distract from the The discharge is stationary but not static: During an rf-
methodical aspect. cycle electrons are accelerated back and forth within the
discharge gap. Due to their large inertia ions are almost
unaffected by the transient field, but are influenced by the
3.1. Discharge Setup
time averaged electric field. (ii) With a plasma density of
As depicted in Figure 2, the discharge consists of a one ne
4  109 cm3 , the degree of ionization is below 104
dimensional arrangement of two opposing plane, ideal and, consequently, the background gas density can be
electrodes of infinite surface area. These electrodes are safely assumed to be unaffected by the plasma (i.e., gas
separated by a gap of L ¼ 7.5 cm. The left electrode is made of rarefaction due to ionization can be ignored[12]). (iii) An
aluminum for the purpose of sputtering. It is referred to as immanent consequence is a comparable density of
the target. The right electrode is made of stainless steel and sputtered aluminum: The sputtered particle flux is
is referred to as the substrate. In the course of this paper, governed by the flux and energy distribution of ions
the discharge will be viewed under periodic steady-state impinging the surfaces and the corresponding sputtering
conditions. It is therefore important to acknowledge that yield (roughly Y
1). The simplest assumption is the one
during operation the stainless steel surface of the substrate of mono-energetic ions, which strike the surface perpen-
electrode is in fact irrelevant, as its surface properties are dicularly. The flux of impinging argon ions and the flux
governed by a layer of previously deposited aluminum of sputtered aluminum are, therefore, of comparable order
covering the surface. For the purpose of sputtering, the of magnitude. Resulting, the influence of the sputtered
discharge is operated with argon gas at a constant pressure particles on the background gas density in the form of a
p ¼ 0.5 Pa, assuming a background gas temperature of sputtering wind can be neglected as well.[13] The sputtered
T ¼ 650 K.[8,9] A gas density of nAr
5:57  1013 cm3 aluminum is merely a trace within the abundant argon gas.
From a modeling point of view, the charged and neutral
particle contributions can be understood as building
blocks. For these, the above mentioned aspects do have
Vrf (t) VB(t) imperative implications  in terms of the individual
x-axis modules as well as their interaction. The individual
0 L contributions can be interfaced solely through a reduced
CB
set of interactions at the surface and within the gas phase.
This is schematically depicted in Figure 3, where the full
Figure 2. One dimensional discharge setup consisting of two
electrodes (driven: left, grounded: right) connected to an rf complexity of a generic plasma model is reduced to only one
voltage supply V rf via a blocking capacitor CB. The red corridor relevant link for the given discharge. (i) Ions impinging the
illustrates geometric invariance into lateral dimensions. surface may sputter surface atoms of the bulk material,

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.plasma-polymers.org 3
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
 J. Trieschmann, F. Schmidt, T. Mussenbrock

surface-normal direction. This aspect concerns mostly the

Charged particles Interaction Neutral particles
input to the test particle model utilized for the sputtered
particle transport simulation.[9] (B) Secondly, whenever test
Surface
Non-equilibrium Gas background particle simulations are carried out, the fundamental
assumption is that the test particle species are only a trace
Equilibrium Energetic neutrals within an abundant background. In consequence, the test
Gas phase particles are influenced by the abundant species, but not
vice versa. The link is realized unidirectional and the
Figure 3. Decomposition of a generic plasma discharge into only
feedback of the trace particles onto the background is
the relevant components and links for the depicted situation. ignored. Both of which hypotheses are in general highly
The full coupling between the charged particle and the neutral relevant in the context of Monte Carlo sputtering plasma
particle dynamics is reduced to a unidirectional link via the simulations. The benefit of a coupled Particle-In-Cell/Test-
sputtering surface interaction (red bold arrow). Particle model lies within the fact that both assumptions
may (at least up to the fundamental arguments) be
which are the primary source of energetic neutral particles. validated. This topic is to be discussed in the following
In turn, neutral particles (particularly excited species) sections, where related results are presented.
may generate secondary electrons when approaching the
surface. This, however, adds only a small contribution to
the interactions which charged particles experience at the 4. Kinetic Particle Simulation of Sputtering
surfaces anyhow. Consequently, the bidirectional link Plasmas
reduces to a unidirectional input from the charged particle
to the neutral particle model. (ii) In the gas phase, from the In terms of a theoretical model, the previous considerations
above mentioned considerations, a-priori any dynamical raise the question of how to constitute the individual
effects within the background gas such as gas rarefaction modules. At the low pressures of interest, the evaluation of
can be neglected. Moreover, the sputtered particle concen- the dynamics of low temperature plasmas commonly
tration is so low as to render the amount of metal ions requires a kinetic description of the many-particle sys-
formed via electron-impact ionization insignificant (de- tem.[14] Closed analytic formulations of the underlying
spite a lower ionization energy threshold). The correspond- complicated multiphysics/multiscale problems are not
ing bidirectional link fully resolves in favor of the possible and even numerical solutions present an expen-
straightforward assumption of a constant, stationary argon sive computational task. In this paper, answers are sought
gas background for both charged and neutral sputtered by means of Monte Carlo computer simulations, which
particles. (Resonant charge-exchange collisions of argon have proven to be readily applicable for tracing the system
ions with the given argon background are taken into evolution over discrete times.[6,15–19]
account nevertheless.) The coupling between the individual Depending on the problem and physics of interest, a
modules diminishes greatly as feedback onto the charged simple 1-dimensional model might be a sufficient approxi-
particles may be ignored completely. mation. If 2 or 3 dimensions are to be taken into account,
In terms of modeling, two separate modules may be the simulation time increases. Within a d dimensional
implemented, one dealing with a dynamical description of simulation space the number of computational operations
the charged species and the other with the sputtered to be performed scales to the power d. Usually, multi-
neutrals. The principle interfacing stems from the ion flux dimensional simulations are highly parallelized to cope
incident onto the adjacent surfaces, thus serving as input with the computational demands. Within this paragraph,
for the neutral particle module. For the sake of generality it the common fundamental aspects of both the charged
should be noted that, by proper interfacing of the full and the neutral particle module are reviewed. Subse-
linkage, a consistent description may be recovered if quently, the decisive specifics of the two are of concern in
required. A self-consistent model would then need to the respective subsections.
account for a transient neutral gas density within the The overall idea of a Lagrangian particle (flow) picture is
plasma module and a closely coupled loop were to resemble. as follows:[20–22] A given set of super-particles  each of
which represents a collection of physical particles  is
randomly distributed in phase space. Correspondingly,
3.3. Hypotheses
these super-particles represent samples of the distribution
Within the preceding analysis, two important hypotheses function of the species.[15,23] From the ergodic principle
have been postulated. (A) Firstly, it has been proposed that the former defines the latter and vice versa.[24] Most simply,
ions accelerated by the strong electric field in the sheaths hit the super-particles are initially uniformly distributed
the target nearly mono-energetically and predominantly in in configuration space with position vectors ~ rk and

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

4 ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim DOI: 10.1002/ppap.201600140
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
Particle-in-Cell/Test-Particle Simulations of Technological. . .

Maxwellian distributed in velocity space with ‘‘position’’ stability criteria.[30,31] Essentially, the ensemble of super-
vectors ~y k . Clearly, k indicates an individual super-particle particles is evolved in time and space subject to internal and
out of the full set of K super-particles. From the statistical external force fields.
phase space sampling of the initial particle distribution, the To describe the system evolution appropriately, colli-
term Monte Carlo emerged. sions play a crucial role. It is important to understand the
Various methods (e.g., Box-Muller or Marsaglia polar dynamics of the system from a conceptual standpoint.
method) have been proposed for the sampling of a thermal At a given instant of time, that dynamical state of each
distribution from (pseudo-)random uniformly distributed super-particle, once in a while also referred to as simulator,
samples Rf 2 ½0; 1, which can be readily applied for the is fully specified by its position and velocity vector and
latter purpose.[25,26] After initialization, the super-particles accordingly located in phase space. During a free flight of
are traced in configuration space according to (non- duration t, the collision time, the trajectory of every super-
relativistic) Newton’s laws particle may be traced. Within every iteration cycle super-
particles are first traced in phase space, then a collision
@~rk
¼~
yk; ð1Þ encounter is evaluated according to its probability and
@t
interaction dynamics.
Following the assumption of a Markov chain, the
@~
yk occasion of an event solely depends on the present state
¼~
ak: ð2Þ
@t and not the past. The sequence of statistically independent
events (i.e., collision encounters) is then based on an
ak ¼ ~
Here, ~ F k =mk is the acceleration of the super-particle exponential probability density pc ðt Þ ¼ expðt=t c Þ.[23]
due to forces of any kind. For a gravitational force ~
F k ¼ mk~g From the collision probability
the mass dependence forthrightly cancels. Under the
influence of an electromagnetic field the Pc ðt  t c Þ ¼ 1  expðt=t c Þ ð5Þ
  force is given by
the Lorentz force ~ F k ¼ qk ~ yk  ~
E þ~ B . The acceleration
acting on a physical particle scales with its charge Qk and the collision time t ¼ t c lnðRf Þ can be evaluated using Rf , a
inversely with its mass Mk . Physical particles with Qk =Mk sample from a uniform random distribution. t is, conse-
therefore experience the same acceleration as the repre- quently, based on the expectation collision time t c (i.e.,
sentative super-particles with qk =mk , given that both are the expectation value of the probability distribution).
based on the same weight w ¼ qk =Qk . (Here, qk ¼ wQk is the The expectation collision  timet c ¼ t c ðyr Þ is specific to the
charge of the super-particle, mk ¼ wMk its mass). In plasma relative speed yr ¼ ~ y bg  of particle k, which is
yk  ~
simulations (accept for dusty plasmas, e.g.,[27]), the force inherently known, and its background collision partner
taken into account is prevalently reduced to the Lorentz (indicated by ‘‘bg’’). At any rate, it is not yet an ensemble
force. Only charged species experience acceleration, neutral mean. In case of a thermal neutral gas background, the
species are unaffected as gravitational forces are neglected. velocity of the background particles ~ y bg may be again
By nature of the computational operations to be sampled from a thermal distribution. The expectation


performed, a discrete time stepping is required. Within collision time is obtained from t c ¼ 1= nbg s ðyr Þyr , which is
the frame of an explicit evaluation scheme, a so-called leap a function of the background density nbg , the velocity (or
frog algorithm may be used.[20,28] Particle positions are energy) dependent cross-section s ðvr Þ and the relative
evaluated at multiples of the time interval Dt, while speed yr . It is the inverse of the collision frequency nc ¼ 1=t c .
corresponding velocities are evaluated at times shifted by Concerning the ensemble average, this is consistent with
Dt=2. Equations (1) and (2) may thus be decoupled following the expectation value for an individual particle to collide
a second order central finite difference scheme (with l with the ensemble of background particles obtained
indicating the discrete time of evaluation), from the Boltzmann collision operator, given by nc;mean ¼
nbg s ðyr Þyr : The overbar indicates the ensemble average
ðlþ1Þ ð lÞ ðlþ1=2Þ
~
rk ¼~
r k þ Dt~
yk ð3Þ over the velocity distribution of the particles within a
control volume. Under acceleration free motion, the
ðlþ1=2Þ ðl1=2Þ ðlÞ
individual particle may be traced for a randomly selected
~
yk ¼~
yk þ Dt~
ak : ð4Þ time of free flight t, terminated by a collision. The
corresponding distance of free flight is l ¼ yk t.
In the presence of a magnetic field, the right hand side of In the presence of an accelerating force  in particular, if
(4) requires additional considerations, for example follow- it is spatially inhomogeneous  the super-particle tracing
ing a procedure proposed by Boris.[29] Alternative schemes procedure has to account for this peculiarity. By employing
may as well be applied to solve the equations of motions (1) the staggered grid approach and a constant time step Dt,
and (2), for instance implicit methods relaxing certain acceleration may be consistently taken into account. The

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.plasma-polymers.org 5
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
 J. Trieschmann, F. Schmidt, T. Mussenbrock

collision probability is checked every Dt, whereas a collision currents to the nodes of the spatial grid. From the charge
is evaluated only when Rf  Pc ðDt  t c Þ. The upper and current density at the grid points the electric (and
bound for the choice of Dt is typically constrained by the magnetic field) is computed. In order to push the particles
scale of the spatial inhomogeneity L of the force field. according to the acting forces, actually the Lorentz force,
By requiring l  L often Dt  t c is inferred, in which the fields first have to be interpolated back to the individual
case the algorithm becomes somewhat inefficient due to particle positions. They are subsequently pushed in
the therefrom arising low collision probability. Therefore position and velocity space according to Newton’s laws.
the specific implementation of the collisional processes As particles traverse through the spatial domain and reach
for both charged and neutral particles differ from the the walls, their interactions with these walls have to be
algorithm shown. accounted for. Finally, collisions are taken into account by
Due to its far reaching implications, the prevalent means of Monte Carlo collisions (MCC) and the cycle starts
assumption of taking into account only collisions between over. Although, previously described quite generally, in
plasma and, respectively, energetic neutral species with a the following considerations the problem complexity is
thermal neutral gas background may need to be reconsid- reduced by the absence of both externally applied and self-
ered for different situations (e.g., ionized physical vapor consistent magnetic fields. Consequently, the electrostatic
deposition). The individual aspects specific to the plasma approximation of Maxwell’s equations is justified, leading
and neutral species are left for later consideration in to Poisson’s equation, which has to be solved within
subsections 4.1 and 4.2, where also the interactions with the each cycle.
walls are considered. To conclude, the described iterative
procedure is evolved until a steady-state situation is
4.1.1. Charge Assignment
reached. Finally, the particle distributions in position and
velocity space can be obtained by averaging over an The conceptual notion of the PIC method lies within the
adequate period of time. If a periodic steady-state is of treatment and decoupling of charge interaction processes.
concern, time-resolved averaging may be performed Charged particles do not only experience short range
accordingly. It should be clear that, for this purpose, the interactions (i.e., elastic or inelastic collisions), but also long
physical observables to be decided on should be specified range Coulomb forces. In principle, these can be incorpo-
beforehand. rated into the simulation by direct evaluation of the
individual interactions for each pair of charged particles
respectively, ignoring self-interactions. If no methodical
4.1. Particle-In-Cell Scheme
modification is desired,[6,7] this leads to the dilemma of
The particle-in-cell (PIC) method conceptually relies on the unfeasible computational requirements. Within PIC, rather
previously introduced Monte Carlo method.[20,21,32] As than to directly evaluate the individual Coulomb inter-
suggested by its name, a division of the computational actions, by appropriate spatial mapping and interpolation,
(spatial) domain into discrete grid cells is exploited. self-interaction is circumvented relaxing the coupled
Specifically, in the one dimensional setup depicted in solution problem to a feasible task  following the
Figure 2, the discharge gap L may be discretized into N Clouds-in-Cells idea.[28] Direct Coulomb particle-particle
equidistant segments of length Dx. interactions are disregarded, which implies that also the
The evolution of the system is traced in an iterative corresponding detailed spatial information is lost. This
evaluation cycle.[22] A schematic of one cycle of duration Dt translates to a finite spectral resolution in space and time,
is given in Figure 4. In brief, the cycle starts with an which may be problematic.[7,28] To obtain the charge
assignment of the individual super-particle charges and density r at the grid nodes, the charges of the ions and
electrons given at the individual particle positions ~ rk need
to be weighted onto the discrete grid nodes, for instance
Calculate electric Weight fields to Calculate new
potential and positions and the m-th node at xm ¼ mDx. Different weighting methods
particle positions
field velocities are possible.[20,28] One of the simplest energy bounded
interpolation schemes is a linear weighting.[28] In one
Weight particle Calculate wall- dimension, assuming an equidistant mesh, first a triangle
charges to grid interaction
of width 2Dx and centered at the continuous particle
position is considered. The super-particles point charge is
MCC-Process
Remove Create new distributed onto the grid nodes covered by the triangle 
Adjust
velocities particles particles due to its limited width at most two nodes. By ensuring
that the total charge distributed equals the charge of
Figure 4. Schematic illustration of the simulation cycle during the super-particle qk , the grid based charge density is
one time step Dt iterated until simulation convergence. obtained.

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

6 ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim DOI: 10.1002/ppap.201600140
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
Particle-in-Cell/Test-Particle Simulations of Technological. . .

ðlþ1=2Þ
4.1.2. EM Field Solver and External Lumped Circuits using (3). The velocity ~
vk is updated based on (4) with

  ðlÞ 
a k ¼ qk =mk ~
ðlÞ
For the calculation of the forces acting on the super- acceleration ~ E xk , directly obtained from
particles, the electric field, and if necessary also the ðlþ1Þ
the Lorentz force. The new position ~
rk is then calculated.
magnetic field, has to be determined. In addition to the
ðlþ1Þ
internal forces, technological plasmas are almost exclu- In a one dimensional case, position vector ~
rk is replaced
sively driven via externally supplied power. The corre- ðlþ1Þ
by the scalar position xk .
sponding forces need to be included in the simulation
efficiently, as outlined in this paragraph at the example 4.1.5. Wall Interaction
of the electric field.[20,33] The electric potential and
the electric field are defined on the same grid as the Depending on the case of interest, the walls may act as
charge density. To calculate the fields, Poisson’s equation particle sinks, reflect particles or  as in the exemplary
r2 Fð~ r Þ ¼ rð~
r Þ=e0 is discretized with a second order finite setup  constitute a source for sputtered particles. If the
difference scheme and solved numerically given the latter is the case, the energy and angle of the incoming ions
charge density rð~ r Þ. In the one dimensional case, this is are essential, once they reach the wall. These values are
ðFmþ1  2Fm þ Fm1 Þ=Dx2 ¼ rm =e0 , where m indicates taken as input for the computation of the sputtering yield
the respective grid point on an equidistant mesh. From Fð~ r Þ, and distribution. The details of which are covered in
or Fm , respectively, the electric field is easily obtained subsection 4.2. The ion itself is almost exclusively neutral-
via ~ E ð~
r Þ ¼ rFð~ r Þ, in the one dimensional case via ized in the process of a wall interaction and contributes
Em ¼ ðFmþ1  Fm1 Þ=2Dx. The external excitation ap- to the neutral background. Therefore, given a constant
plied to the electrodes specifies the boundary conditions background density, it is simply removed from the
through their surface charges s. This holds if the source is simulation. The same is done with wall interactions for
connected via an external lumped circuit. The surface which the walls act as perfect particle sinks, as for
charge s is used as the interface connecting the plasma aluminum particles in the exemplary setup (whether ions
and the external circuit. Simply speaking, the boundary or not).[34,35] When particle reflection has to be taken into
condition is chosen in a way that the continuity equation is account and the (energy dependent) reflection probability
implicitly satisfied throughout the simulation. For the Prefl is known, a simple approach is to compare Prefl with a
specific details of the implementation, ref.[33] is strongly uniform random number Rf . If Rf  Prefl , the particle is not
recommended. reflected and removed from the simulation. Otherwise, the
reflection is calculated. For the easiest case of a specular
reflection (typically a good approximation for clean metal
4.1.3. Field Interpolation
surfaces),[5] the velocity component perpendicular to the
Once the electric (and if necessary magnetic) field has wall is simply inverted. Another effect which may be
been obtained, the forces acting on the ensemble of super- important is secondary electrons emission. It can be taken
particles have to be taken into account. As the fields and into account similarly on a probabilistic basis. Various
corresponding forces are known on the numerical grid particles (e.g., ions, electrons, excited species, or energetic
only, an intermediate interpolation is required to obtain neutrals) may give rise to this phenomenon. However, for
the forces at the specific super-particle positions ~
rk and the sake of simplicity, it is neglected in the present case.
avoid self-forces. Again, a straightforward approach is
the linear interpolation as previously mentioned in the 4.1.6. Particle Collisions
context of charge assignment.[28] The formalism for a one-
dimensional equidistant mesh is principally the same: An aspect of major importance for kinetic simulation
Taking a triangle of width 2Dx centered at the position of methods like PIC are the particle collisions and the way they
the super-particle, the force from the two adjacent nodes are incorporated in the simulation framework.[22] A
is averaged (inversely weighted by distance to the grid fundamental factor defining the collision dynamics is the
points Dx). energy dependent cross section s ðeÞ, where the interaction
energy e is related to the relative speed as e ¼ 0:5 mr y2r , with
mr ¼ m1 m2 =ðm1 þ m2 Þ being the reduced mass. Within PIC
4.1.4. Particle Pushing
models the cross section is usually utilized in the form of a
The method of how to apply the forces on the particles and look-up table from experimental measurements. Depend-
to calculate their new positions and velocities depends on ing on the kind of collision process of interest (e.g.,
the specific situation. If a magnetic field is involved, the ionization, charge exchange, elastic scattering) and the
particle pusher developed by Boris has proved to be involved species, different data sources for the cross section
efficient.[29] In the absence of a magnetic field, the situation are available in the literature, particularly in the LXcat
simplifies greatly. The new position can be calculated database.[36]

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.plasma-polymers.org 7
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
 J. Trieschmann, F. Schmidt, T. Mussenbrock

For weakly ionized low temperature plasmas, the neutral are typically much faster than ions, which is why they are
gas in the vacuum chamber is most abundant. Hence, only the limiting factor regarding the time step. Different
collisions of the various plasma species with the back- evaluation schemes have been proposed, relaxing the
ground gas are to be taken into account. It is computation- mentioned constraints, in particular relying on an implicit
ally exceptionally expensive to repetitively calculate the evaluation of the PIC cycle.[31,39–42]
collision probability for each possible kind of process The number of simulated super-particles is a different
individually. More cost effective approaches have been matter. Depending on the dimensionality of the simulation
proposed  specifically, the null collision method.[37,38] and the physical environment, the particle weight has to be
From the definition of a cumulative energy dependent defined. Common definitions include the number of
P
collision cross section s sum ðeÞ ¼ Jj¼l s j ðeÞ, summed over simulated particles per Debye length N D and per grid cell
all energies a maximum collision frequency N C . Since a proper grid cell size DV also depends on lD , N C is a
good parameter of choice. A good estimate for many
h pffiffiffiffiffiffiffiffiffiffiffiffiffii applications is to set the particle weight so that N C  20.[32]
ð6Þ
nc;max ¼ nbg s sum ðeÞ 2e=mr Different implications come into play when considering
max
cases with reduced dimensionality (e.g., one dimensional as
is obtained. Subsequently, using equation (5) the energy in this tutorial). Firstly, within the Poisson solver, three
independent total collision probability Pc;tot is calculated dimensional point charges translate to planes of charges
(once during initialization of the simulation). This corre- in a one dimensional evaluation. Invariance in lateral
sponds to an upper probability bound. Throughout the direction is inherent to the simulation, implying the
simulation, irrespective of the actual interaction energy e; peculiar physical behavior. Moreover, even in situations
collisions are taken into account and evaluated with where a one dimensional representation in position space
probability Pc;tot given a fixed Dt for each time step as seems applicable, in velocity space it may not. Symmetries
described previously. Whenever a collision takes place, the in configuration space translate into velocity space. For
type of collision has to be selected as well. For the specific the fully symmetric 1D situation described, it would be
interaction energy of concern, the cumulative cross section sufficient to maintain only two velocity components 
s sum ðeÞ is virtually complemented with a so-called null- parallel and transversal to the configuration axis. However,
collision process s null ðeÞ so that nc;max nc;sum ðeÞ þ nnull ðeÞ. A in the presence of a magnetic field, symmetry is broken
random collision sample is obtained by multiplication with and three velocity components are required. The most
a new random sample Rf , so that 0  nrand  nc;max . Since prominent choice, therefore, is to reduce dimensionality in
every possible collision is included, the value of nrand position space, but fully maintain a three dimensional
selects the l-th type collision process  essentially from velocity space. Thereby, particle positions ~ rk translate to
Pl1 Pl scalar positions xk , while the three dimensional velocities~ yk
j¼l nj ðeÞ  nrand  j¼l nj ðeÞ. Of course, for collisions
involving an energy threshold (e.g., ionization or excitation) are preserved. In particular, throughout collisional pro-
the interaction energy has to be checked additionally. In cesses scattering into different directions in velocity space
case a null collision event is chosen, the evaluation is is accounted for. The model is consequently referred to as
stopped and no collision takes place. ‘‘1D3V.’’

4.1.7. Limitations and Constraints 4.2. Test Particle Method

It is important to keep in mind the conceptual assumptions Neutral particles within the plasma may be treated
and limitations. As mentioned previously, the predominant separately from the charged particle contribution. They
assumption within PIC lies in the replacement of discrete are, however, coupled via the input from the ion flux and
multi-particle Coulomb interactions with Coulombic parti- energy distribution at the surfaces. In the given scenario,
cle-field interactions. ‘‘Microscopic’’ details are smoothened a kinetic transport description is required due to the
and spatial as well as temporal resolution is lost.[28] comparably large Knudsen number. One should keep in
Consequently, PIC may not be applicable if these details mind that the sputtered particles have an inherently non-
are essential.[7] In addition, for the explicit evaluation thermal velocity distribution. A kinetic description is
scheme described within this manuscript, the size of one therefore advisable. If only a rough description of the
grid cell DV is subject to limitations. In order to be energy particle fluxes to the walls is desired, alternative methods
bound, the numerical cell dimensions are limited to may be applicable.[43] The neutral particle model is
DV 1=3  lD (i.e., much smaller than the Debye length concerned with solely the collisional transport of energetic
lD ¼ ðe0 kB T e =ne e2 Þ1=2 ).[20] To ensure stability, the choice sputtered particles. Self-interaction (i.e., metal with metal)
of the time step is constrained to Dt  v1 p , with the is neglected, as arguable by the condition of a constant
1=2
electron plasma frequency vp ¼ ðne e2 =e0 me Þ . Electrons background gas and only a trace of aluminum therein. The

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

8 ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim DOI: 10.1002/ppap.201600140
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
Particle-in-Cell/Test-Particle Simulations of Technological. . .

individual aluminum interactions with the argon back- nature of the collision. For practical purposes, often the total
R1
ground are assumed to be of binary nature. Due to the low collision cross section sðyr Þ s T ðyr Þ ¼ 2p 0 bðyr Þdb is
pressure three-body collisions can be safely neglected. For used for evaluation of the collision probability. Two
this situation, the test particle method (TPM) is most premises are: Firstly, a finite distance of interaction is
appropriate. The heavy particle dynamics can be efficiently required so that the upper bound of the integral is bmax and
‘‘decoupled’’ from the fast electron and plasma dynamics. s T ðyr Þ ¼ pb2max ðyr Þ irrespective of scattering details. And
Principally, TPM can be also applied not only to the secondly, the details of the interaction depend on a specific
transport of neutral particles but also to the general case of scattering law (i.e., even a large total cross section may
ion transport taking into account the interaction of the lead to only small angles of deflection). In order to be
particles with the background atoms as well as with a applicable in theoretical models, the cross section has
known electric field (e.g., provided from PIC simulations).[44] to be numerically known (may it be in the form of an
This shall not be detailed in this work. As a major analytic formula or numerical look-up tables). Analytic
prerequisite, it is important to notice that any type of expressions have the advantage of well known scattering
feedback onto the plasma is inherently omitted then. dependencies. A drawback is that barely all the details of a
Regarding the methodology, various variants are docu- scattering interaction are captured. The collision process is
mented in the literature, which have been widely applied best viewed in the co-moving frame of the collision.[5,57]
over decades for the simulation of transport phenomena The most prominent choice is the variable hard
and gas kinetics.[45–53] A peculiar variant is the test multi sphere (VHS) model[58] with an isotropic distribution of
particle method (TMPM).[9,54,55] The general concept of the scattering angle xðbÞ ¼ 2arccosðb=bmax Þ on a unit
TMPM again lies within the separation of the physical sphere  a function of the impact parameter.[5] Due to
processes under consideration. That is, book keeping and its isotropic angular distribution, the VHS model is
numerical handling is one rather technical aspect, the easily transferred into the laboratory frame (i.e., it is
physical considerations are a different one. isotropic in either coordinate system) thus requiring no
coordinate transformation. An energy (and thus velocity)
4.2.1. Collisional Transport dependence is commonly implied following a power law
for the corresponding viscosity mðT Þ ¼ mðT ref ÞðT=T ref Þ v ;
The previous considerations can be applied to the case of specified for combinations of reference temperature
energetic neutral particle transport subject to binary T ref and viscosity index v (e.g., fitted to experimental
collisions within a vacuum environment. For the general- data[5,9,18,59]). From these considerations, the maximum
ized inhomogeneous situation, a constant time step Dt is impact parameter can be defined as[5,18]
suggested. Based on the previous considerations and, in
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
particular, the ensemble averaged expectation value for the 
1 2kB T ref v0:5
collision frequency the transport can be described. On the bmax ðyr Þ ¼ dref : ð8Þ
G ð2:5  vÞ mr y2r
one hand, taking equation (5) and linearizing with t Dt
and t c 1=nc;mean yields In an alike collision event of two particles of species a
this can be directly evaluated using mr ¼ ma =2 and the
Pc ¼ nbg s ðyr Þyr Dt: ð7Þ reference diameter for this species
On the other hand, given a Lagrangian collection of super- sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
particles, the ensemble average is equivalently evaluated 15 ma kB T ref =p
dref ¼ : ð9Þ
by exercising the calculation of the collision probabilities 8ð2:5  vÞð3:5  vÞmðT ref Þ
for all particles individually. In principle, the most
For an unlike collision expression (8) can be employed
imperative constraint is that Dt be small enough, so

using and a linear interpolation dref ¼ dref;a þ dref;b =2
Pk;bg  1. This concept in general works quite well.[54,55]

and v ¼ va þ vb =2 evaluating (9) for species a and b,
Problematically, the above mentioned constraint renders
respectively.
the evaluation quite inefficient. For improved efficiency
More complicated choices for the collision model are
the no-time counter scheme can be an alternative. This
available,[51,53] an instructive example being the M1
shall not be the focus of this manuscript, but is detailed
collision model  a modified VHS model.[54,55,59] These
elsewhere.[5,9,18,56]
models typically share an anisotropic scattering nature,
which does have direct implications on particle transport
4.2.2. Binary Interactions
in particular for the directional flow of for example
For evaluation of the individual binary interactions, the sputtered particles. Note that this affects mainly the spatial
differential collision cross section s d ðyr ; xÞ is important. Its distribution. The amount of particles is specified by flux
velocity and angular dependence differs depending on the continuity. For the M1 model this is xðbÞ ¼ pð1  b=bmax Þ,

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.plasma-polymers.org 9
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
 J. Trieschmann, F. Schmidt, T. Mussenbrock

anisotropically distributed over a unit sphere. Regarding sputtering event, a particle is sputtered given that Rf  Y.
the energy scaling, equations (8) and (9) can be readily On the occasion that the sputter yield is larger than one,
employed. The reference data available in the literature for the check can be performed N times with Rf  Y=N and
VHS canpbe usedffi for M1 with the reference diameter
ffiffiffiffiffiffiffiffiffiffiffi N ¼ ceilðY Þ. For each successful check, a new particle is
dref;M1 ¼ ð4=3Þdref;VHS .[59] Not of concern in this work, but inserted into the volume with an initial position equal to
relevant when analytic models are favored instead of look- the location of ion impact.
up tables, are theoretical approaches to collisional inter- In order to fully specify a sputtered particle to be inserted,
actions of charged particles. An illustrative approach is the its velocity and angle of emission are to be known. These
one originally published by Langevin in which the charge may be conveniently sampled from a given particle
induced interaction is consistently taken into account.[60,61] distribution. Notably, however, what is actually specified
A detailed study of this model and its implementation has is the velocity distribution f ð~y Þ of the particle flux, not the
been proposed by Nanbu et al.[62,63] velocity distribution of the particle density (within a given
In practical terms, either collision model requires for the volume). Both of which are indeed related by a scaling with
evaluation of a random collision event a choice of the the normal component of the velocity ~ n ~y . In the case of
impact parameter. This is to be sampled from a random sputtering, it is important to acknowledge that a sputtered
distribution. A probability density uniformly distributed particle flux behaves vastly different than what is to
over a cross section with the substitution B ¼ b2 and dB ¼ be expected from a thermal influx. The former being
2bdb implies significantly more energetic with energies on the order of
eV, in contrast to thermal particles with typical energies of
dBdc 2bdbdc 25 meV. Sputtered particles are commonly assumed to
ps ds ¼ ¼ : ð10Þ
sT sT follow a Sigmund-Thompson distribution.[66,67] Several
variants are published in the literature.[9,55,68] The details
The probability density in terms of the impact parameter
depend on the specific situation. For a more accurate
b and the azimuthal angle c is therefore linear in b
description, the output from transport of ions in matter
(consistent with the Boltzmann collision integral). The
(TRIM) or even better molecular dynamics (MD) simulations
corresponding impact parameter distribution is sampled by
pffiffiffiffiffiffi may be readily employed.[65,69,70]
the method of inversion from b0 ¼ bmax Rf :[5] To fully
The individual particle properties may be obtained
specify the scattering direction also a random polar angle
from an acceptance-rejection sampling of a given energy
c0 ¼ 2pRf is chosen. In order to calculate the post-collision
distribution, paired with an emission angle obtained
velocity of the scattered particles in the laboratory frame in
from sampling of a correspondingly given angular distri-
the anisotropic situation, also a consecutive coordinate
bution.[5] The most common choice is the assumption
transform is required.[5,18,57]
of a cosine distribution f ðuÞ cosðuÞ and therefore

pffiffiffiffiffi
u ¼ arccos Rf .[47,71,72]
4.2.3. Wall Interactions
When a neutral particle impinges a wall, several path-
An aspect not detailed in subsection 4.1 concerns the ways are possible. It may be specularly reflected, be
implications of the plasma-wall interaction for the neutral thermally re-emitted, or stick to the wall. While the former
particle model  more precisely, its manifestation in terms two choices are common for inert gas flows, the latter is
of a sputtered particle flux. It is the imperative component most relevant for the scenario analyzed in this paper.[5,34] In
regarding the coupling from the PIC to the TPM model. this case, the walls can be envisioned as perfect particle
The sputtered particle flux (i.e., the number of particles sinks and any particle approaching is removed from the
emanated from the surface per unit area and unit time) is simulation. The absorbed flux, however, is highly relevant
most easily realized for super-particles of same weight for the purpose of a later analysis.
within PIC as well as TPM. In detail, rather than to a priori
specifying the total number flux of sputtered particles, for
4.2.4. Limitations and Constraints
each ion hitting the wall the interaction can be evaluated on
a single event basis. That is, for every sputtering event the Throughout a TPM simulation, a few constraints have to be
probability can be calculated from a given sputtering yield accounted for. The foremost important parameter concern-
Y ðE0 ; u0 Þ, which is a function of the projectile’s energy E0 and ing collisional transport is the discrete time step Dt. It is
its angle of incidence u0 . For a vast amount of projectile- required to be chosen small enough to properly account for
target combinations fitted parameters are available from all collisions encountered. That is, given at hand equation (7)
Behrisch and Eckstein. These parameters can be used to governed by the expectation collision time t c , a correct
evaluate the respective analytic expressions.[64,65] The physical representation is ensured only, when Dt  t c is
sputtering yield can be interpreted from a probabilistic satisfied. Once again, only binary collisions are taken into
perspective: If an ion fulfills the initial criteria to perform a account, limiting the model’s range of applicability.

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

10 ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim DOI: 10.1002/ppap.201600140
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
Particle-in-Cell/Test-Particle Simulations of Technological. . .

In the inhomogeneous situation Dt, moreover, has to be

chosen small enough to properly resolve any local
gradients. This is straightforwardly linked to the distance
a particle travels during one iteration. When a mesh with
cell size DV is employed, at all times this distance shall not
exceed DV 1=3 and therewith any given gradient length
scale.[73]
In terms of the diagnostics, it is important to distinguish Figure 6. Ion energy distribution functions of ions impinging
the various TPM simulation types. In mesh based TPM the target (left; solid black) and substrate (right; dashed red)
simulations, the number densities may be forthrightly electrode, respectively.
obtained from a cell based average of the particles residing
within. In contrast, within meshless simulations with respective voltage drop in front of the individual electrodes.
adaptive time steps Dt t c , the particle density is obtained To a certain degree, the IEDs are spread around the peak
taking into account the specific ‘‘flight duration.’’ A time value, which can be argued from the fact that also ions do
weighted evaluation is obligatory.[47] not truly experience the average electric field and potential,
but to some extent the time varying. (Electrons respond to
the transient electric field in any case.) As such, ions are
5. Simulation Results slightly modulated as well, as there are times when the
potential drop in front of the electrodes is smaller or greater
Equipped with the details specified in the preceding than the averaged value.[74–76] In addition, a low energy
sections a brief review of some exemplary results is an plateau with periodic humps is vaguely observed for both
instructive task. Figures 5–9 show simulation results for the energy distributions. These are known to be due to elastic
exemplary case specified in section 3. PIC is used for the resonant charge exchange collisions.[77]
plasma simulation, whereas a consistently coupled TPM is The quantitative aspects of the plasma dynamics are
used for the sputtered neutral species. The average electric largely determined by processes taking place in the gas
potential is plotted in Figure 5. As is to be seen therein, phase. Of particular interest are the mean plasma quanti-
a DC self-bias voltage of
225 V establishes due to the ties: The mean electron and ion densities, their mean fluxes,
electrically asymmetric excitation of the plasma. More and velocities. Each of which is depicted in Figure 8. Several
importantly, in front of both electrodes a substantial points are important: (i) technological radio-frequency
voltage drop is found, approximately 350 V (left), 125 V plasmas are predominantly driven by selective external
(right). Consequently, ions which undergo acceleration in energy coupling into the electron component. Correspond-
this sheath potential more or less permanently impinge ingly, their density and energy distribution is of great
on the electrodes with substantial energies and may interest. Quasineutrality typically establishes within the
possibly sputter the electrode surface. This is important plasma volume and so the ion and electron density roughly
for a twofold reason: Firstly, the main achievement is the equalize in the bulk of the plasma. As a function of distance,
sputtering of the target surface (left). Secondly, also re- the average densities are presented in Figure 8a. A
sputtering of the substrate surface (right; initially stainless characteristic charge depletion in the boundary sheath
steel, but in fact aluminum coated) is to be taken into can be clearly identified. (ii) The particle fluxes to the walls 
account. For both electrodes, the corresponding ion energy one of the most influential parameters governing the
distributions (IEDs) of argon ions impacting on the sputtering process  follow accordingly. In periodic steady
electrodes are shown in Figure 6. The IEDs are clearly
centered around mean energies corresponding to the

Figure 5. Spatial profile of the average electric potential Figure 7. Contour plot presenting the energy and angular
obtained from the coupled PIC-TPM simulation for the distribution of ions impinging the target. Colorscale: Blue
specified parameters. represents zero, red maximum.

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.plasma-polymers.org 11
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
 J. Trieschmann, F. Schmidt, T. Mussenbrock

respect, another vastly influential parameter is the mean

velocity of the incident ions, displayed in Figure 8c. It is
inherently linked to the IED formerly addressed. A decisive
difference is that the mean velocity is an example of a
volumetric quantity, while the IED corresponds is a fluxual
property. Of relevance for the sputtering process, however,
is solely the latter in particular at the plasma-wall interface.

5.1. Sputter Deposition

When the peculiar focus is the understanding of

the sputter deposition, the described model approach
requires a distinctive coupling of the plasma related
properties into the neutral particle model. The flux of
sputtered particles originating from the surfaces mani-
fests in a spatial profile of the aluminum density and
mean velocity consistently obtained within the coupled
PIC-TPM simulation. This is depicted in Figure 9. The flux
of particles is not of primary interest as particles are
assumed to be solely generated at or lost to the walls.
The metal flux is continuous throughout the discharge
Figure 8. a) Temporally averaged spatial distribution of the volume. (This is different for the plasma, as in this case
electron (dashed red) and argon ion (solid black) density. b) electron-ion pairs are generated in the bulk and lost to
Corresponding average electron and ion flux. c) Respective the walls. Nearly no ionization takes place in the
average mean velocities calculated for both electrons and ions.
boundary sheaths, which manifests in a straight line in
Figure 8 indicating flux continuity.) The distinct profiles
state operation, on average the volumetric generation and of the sputtered species’ density and velocity stem
loss terms balance. Under the assumption of a Bohm from the interaction of particle fluxes originating from
flux leaving the plasma bulk and entering the bounding the target (left to right) as well as re-sputtered from
plasma sheath, the fluxes of electrons and ions are on the substrate (right to left). Both electrodes act as particle
average equal. This is also observed from Figure 8b. In sources and sinks. As such, minima in the density are
terms of quantitative considerations, this ion flux (i.e., the observed near the boundaries. Of particular importance in
number of ions per unit area and unit time) impinging the this respect is the collisional transport through the
target has strong implications for the flux of sputtered discharge gap. Particles suffering a collision are isotrop-
metal originating from the very given surface. (iii) In this ized (i.e., may be forward or backward scattered). As the
sputtered particle flux directed left to right is larger
than right to left, the mean density and velocity profiles
become asymmetric. Due to the backscattered component
and the requirement for flux continuity, the velocities are
observed to increase adjacent to the surfaces. This can be
understood from the details of the particle distributions
within the discharge gap. Details can be found for
example in refs.[9,47] and shall not be reconsidered at this
point.

5.2. Hypotheses Validation

Finally, due is a discussion of the previously in section 3.3
introduced assumptions often used in the context of
sputter transport simulations: Assumption (A) can be
readily verified from the ion energy distribution. As
Figure 9. Spatial profiles of sputtered aluminum within the displayed in Figure 6, ions do not hit the surfaces strictly
discharge gap: (a) Average density and (b) the corresponding mono-energetically. However, considering the width-to-
average mean velocity. peak ratio DE=Emax
0:1 at both the target and the

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

12 ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim DOI: 10.1002/ppap.201600140
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
Particle-in-Cell/Test-Particle Simulations of Technological. . .

substrate, the spread in energy is negligible. It results from are verified, providing valid justification for more
collisions as well as the radio-frequency modulation. In simplified models relying solely on the Test-Particle
conclusion, mono-energetic impingement is a good ap- approach.[9]
proximation. Regarding the angular distribution, Figure 7
suggests an angular dispersion of less than 18 off surface
normal, as easily verified. Acknowledgments: This work is supported by the German
Assumption (B) can only indirectly be validated. That is, Research Foundation (DFG) in the frame of Collaborative Research
Centre SFB/TRR 87. The authors gratefully thank D. Kru € ger, T.
the proposed model does not permit to justify whether
Gergs, S. Gallian, and R. P. Brinkmann from the Institute of
the influence of the plasma and the energetic neutral Theoretical Electrical Engineering, Ruhr University Bochum and S.
particles is irrelevant. In contrast, what can be verified is Ries, N. Bibinov, and P. Awakowicz from the Institute for Electrical
the contents of the respective species as quantified by Engineering and Plasma Technology, Ruhr University Bochum for
their partial densities in comparison with the argon fruitful discussions.
background gas. On the one hand, the degree of ionization
is nArþ =nAr 9104 and rendering the corresponding
rarefaction insignificant. On the other hand, the neutral Received: July 13, 2016; Revised: August 22, 2016; Accepted:
aluminum to neutral argon fraction is even lower with August 23, 2016; DOI: 10.1002/ppap.201600140
nAl =nAr 95  105 and gas rarefaction due to sputtering
Keywords: cold plasma; sputtering; Monte Carlo; modeling;
wind effects is immaterial as well. Taking into account transport
ionization of the metal as well, the density of singly
ionized aluminum nAlþ (not depicted, but incorporated
into the simulation) is roughly a factor 1800 smaller
[1] M. A. Lieberman, A. J. Lichtenberg, ‘‘Principles of Plasma
than the argon ion density nArþ and thus absolutely Discharges and Materials Processing,’’ 2nd edition, Wiley,
obsolete.[78] Hoboken, USA 2005.
In conclusion, both assumptions previously postulated [2] P. Chabert, N. Braithwaite, ‘‘Physics of Radio-Frequency
and exploited throughout this paper are validated by Plasmas,’’ Cambridge University Press, Cambridge, UK 2011.
[3] U. Kortshagen, L. D. Tsendin, Electron Kinetics and Applica-
corresponding observations in the obtained simulation
tions of Glow Discharges. vol. 367 ‘‘NATO ASI Series B:
results. The reduced model (Figure 3) is unconditionally Physics,’’ Kluwer Academic Publishers, New York, USA 2002.
justified. [4] K. Bobzin, R. P. Brinkmann, T. Mussenbrock, N. Bagcivan, R. H.
Brugnara, M. Scha €fer, J. Trieschmann, Surf. Coat. Technol.
2013, 237, 176.
6. Conclusion [5] G. A. Bird, ‘‘Molecular Gas Dynamics and the Direct Simulation
of Gas Flows,’’ Clarendon Press; Oxford University Press,
Oxford; New York 1994.
The aim of this manuscript is to provide an audit of [6] J. Barnes, P. Hut, Nature 1986, 324, 446.
simulation approaches to low pressure plasma discharges [7] A. J. Christlieb, R. Krasny, J. P. Verboncoeur, IEEE T. Plasma. Sci.
coupled to a dynamic neutral particle component. Starting 2004, 32, 384.
[8] S. Bienholz, N. Bibinov, P. Awakowicz, J. Phys. D 2013, 46,
from a critical statement of the scenario provided in section
084010.
3, a brief introduction into Lagrangian based Monte Carlo [9] J. Trieschmann, T. Mussenbrock, J. Appl. Phys. 2015, 118,
modeling is provided. In section 4, at the example of a radio- 033302.
frequency driven capacitive sputtering discharge, the [10] B. G. Heil, U. Czarnetzki, R. P. Brinkmann, T. Mussenbrock, J.
peculiarities and difficulties involved with a representative Phys. D 2008, 41, 165202.
[11] U. Czarnetzki, B. G. Heil, J. Schulze, Z. Donko, T. Mussenbrock,
description of these complicated many-particle systems are
R. P. Brinkmann, J. Phys. 2009, 162, 012010.
worked out. Specifically, the scenario is assessed initially [12] M. A. Raadu, I. Axna €s, J. T. Gudmundsson, C. Huo, N. Brenning,
from the plasma side by means of the Particle-In-Cell (PIC) Plasma Sources Sci. Technol. 2011, 20, 065007.
method. A similar task is exercised for the conceptual [13] D. W. Hoffman, J. Vac. Sci. Technol. A 1985, 3, 561.
design and singular features of a neutral Test-Particle [14] I. D. Kaganovich, V. I. Kolobov, L. D. Tsendin, Appl. Phys. Lett.
model. Emphasis is put onto the concept of the coupling of 1996, 69, 3818.
[15] N. Metropolis, S. Ulam, J. Am. Stat. Assoc. 1949, 44, 335.
both respective models. [16] F. J. Alexander, A. L. Garcia, Comput. Phys. 1997, 11, 588.
Finally, in section 5 results for the described exemplary [17] H. Gould, J. Tobochnik, W. Christian, ‘‘An Introduction to
discharge are presented with the focus on merely the Computer Simulation Methods: Applications to Physical
physical and conceptual plausibility. For the given setup, Systems,’’ 2nd edition, vol. 3rd Pearson Addison Wesley,
a number of characteristic features of the plasma as well San Francisco (USA) 2007.
[18] A. Kersch, W. J. Morokoff, ‘‘Transport Simulation in Microelec-
as the neutral particles are addressed, which may provide tronics.’’ Birkha€user, Basel 2011.
comparison and/or guidance for the well-disposed reader. [19] A. L. Garcia, ‘‘Numerical Methods for Physics,’’ 2nd edition,
In this context, the two main postulates initially proposed Alejandro L. Garcia, San Jose (USA) 2015.

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim www.plasma-polymers.org 13
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!
 J. Trieschmann, F. Schmidt, T. Mussenbrock

[20] C. K. Birdsall, A. B. Langdon, ‘‘Plasma Physics via Computer [47] G. M. Turner, I. S. Falconer, B. W. James, D. R. McKenzie, J. Appl.
Simulations,’’ McGraw-Hill, New York, USA 1985. Phys. 1989, 65, 3671.
[21] R. W. Hockney, J. W. Eastwood, ‘‘Computer Simulation Using [48] A. M. Myers, J. R. Doyle, D. N. Ruzic, J. Appl. Phys. 1992, 72,
Particles,’’ Taylor & Francis Group, New York, USA 1988. 3064.
[22] C. Birdsall, IEEE T. Plasma Sci. 1991, 19, 65. [49] G. M. Turner, I. S. Falconer, B. W. James, D. R. McKenzie, J. Vac.
[23] W. Feller, ‘‘An Introduction to Probability Theory and Its Sci. Technol. A 1992, 10, 455.
Applications,’’ vol. 2. 2nd edition, John Wiley & Sons, New [50] F. Clenet, P. Briaud, G. Lemperiere, G. Turban, Thin Solid Films
York, USA 1971. 1999, 346, 290.
[24] C. Cercignani, The Boltzmann Equation and Its Applications. [51] K. van Aeken, S. Mahieu, D. Depla, J. Phys. D 2008, 41, 205307.
vol. 67 ‘‘Applied Mathematical Sciences,’’ 2nd edition, Springer [52] D. Depla, W. Leroy, Thin Solid Films 2012, 520, 6337.
New York, New York, NY 1988. [53] D. Lundin, C. Vitelaru, L. d. Poucques, N. Brenning, T. Minea, J.
[25] G. E. P. Box, M. E. Muller, Ann. Math. Stat. 1958, 29, 610. Phys. D 2013, 46, 175201.
[26] G. Marsaglia, T. A. Bray, SIAM Rev. 1964, 6, 260. [54] A. Kersch, W. Morokoff, C. Werner, J. Appl. Phys. 1994, 75,
[27] P. K. Shukla, A. A. Mamun, ‘‘Introduction to Dusty Plasma 2278.
Physics.’’ CRC Press, Boca Raton, USA 2001. [55] V. V. Serikov, K. Nanbu, J. Vac. Sci. Technol. A 1996, 14, 3108.
[28] C. K. Birdsall, D. Fuss, J. Comput. Phys. 1969, 3, 494. [56] G. A. Bird, Prog. Astronaut. Aeronaut. 1989, 118, 211.
[29] J. Boris, ‘‘Relativistic Plasma Simulation-Optimization of a [57] W. G. Vincenti, C. H. Kruger, ‘‘Introduction to Physical Gas
Hybrid Code,’’ in Proceedings of 4th Conference on Numerical Dynamics,’’ Krieger, Malabar, FL, USA 1967.
Simulation of Plasmas, Naval Research Laboratory, Wash- [58] G. A. Bird, Prog. Astronaut. Aeronaut. 1981, 74, 239.
ington DC 1970, p. 3. [59] W. J. Morokoff, A. Kersch, Comput. Math. Appl. 1998, 35, 155.
[30] R. E. Wirz, Discharge Plasma Processes of Ring-Cusp Ion [60] P. Langevin, Annales de chimie et de physique 1905, 5, 245.
Thrusters. PhD thesis, California Institute of Technology, [61] H. Hasse, Phil. Mag. 1926, 1, 139.
2005. [62] K. Nanbu, Y. Kitatani, J. Phys. D 1995, 28, 324.
[31] A. B. Langdon, B. I. Cohen, A. Friedman, J. Comput. Phys. 1983, [63] K. Nanbu, IEEE T. Plasma Sci. 2000, 28, 971.
51, 107. [64] R. Behrisch, W. Eckstein, Sputtering by Particle Bombard-
[32] M. M. Turner, A. Derzsi, Z. Donko , D. Eremin, S. J. Kelly, T. ment. vol. 110 ‘‘Topics in Applied Physics,’’ Springer Berlin
Lafleur, T. Mussenbrock, Phys Plasmas (1994-present), 2013, Heidelberg, Berlin, Heidelberg 2007.
20, 013507. [65] W. Eckstein, Vacuum 2008, 82, 930.
[33] J. P. Verboncoeur, M. V. Alves, V. Vahedi, C. K. Birdsall, J. [66] M. W. Thompson, Philos. Mag. 1968, 18, 377.
Comput. Phys. 1993, 104, 321. [67] P. Sigmund, Phys. Rev. 1969, 184, 383.
[34] U. Hansen, P. Vogl, V. Fiorentini, Phys. Rev. B 1999, 59, R7856. [68] M. Stepanova, S. K. Dew, J. Vac. Sci. Technol. A 2805, 19, 2001.
[35] H. Wu, A. Anders, J. Phys. D 2010, 43, 065206. [69] W. Eckstein, Computer Simulation of Ion-Solid Interactions.
[36] http://www.lxcat.laplace.univ-tlse.fr, 2016. vol. 10 ‘‘Springer Series in Materials Science,’’ Springer, Berlin
[37] H. R. Skullerud, J. Phys. D 1567, 1, 1968. 1991.
[38] S. L. Lin, J. N. Bardsley, Comput. Phys. Commun. 1978, 15, 161. [70] H. M. Urbassek, Nucl. Instrum. Methods Phys. Res. Section B
[39] B. I. Cohen, A. B. Langdon, A. Friedman, J. Comput. Phys. 1982, 1997, 122, 427.
46, 15. [71] M. Knudsen, Annalen der Physik 1916, 353, 1113.
[40] A. Friedman, J. Comput. Phys. 1990, 90, 292. [72] J. Greenwood, Vacuum 2002, 67, 217.
[41] V. Vahedi, G. DiPeso, C. K. Birdsall, M. A. Lieberman, T. D. [73] Z.-X. Sun, Z. Tang, Y.-L. He, W.-Q. Tao, Comput. Fluids 2011, 50,
Rognlien, Plasma Sources Sci. Technol. 1993, 2, 261. 1.
[42] G. Chen, L. Chaco n, J. Comput. Phys. 2013, 247, 79. [74] P. Benoit-Cattin, L.-C. Bernard, J. Appl. Phys. 1968, 39, 5723.
[43] B. Rother, Surf. Coat. Technol. 1994, 64, 155. [75] A. Metze, D. W. Ernie, H. J. Oskam, J. Appl. Phys. 1989, 65, 993.
[44] T. M. Minea, C. Costin, A. Revel, D. Lundin, L. Caillault, Surf. [76] T. Mussenbrock, Contrib. Plasma Phys. 2012, 52, 571.
Coat. Technol. 2014, 255, 52. [77] C. Wild, P. Koidl, J. Appl. Phys. 1991, 69, 2909.
[45] R. E. Somekh, J. Vac. Sci. Technol. A 1984, 2, 1285. [78] L. L. Shimon, E. I. Nepiipov, I. P. Zapesochnyi, Sov. Phys. Tech.
[46] T. Motohiro, Y. Taga, Thin Solid Films 1984, 112, 161. Phys. 1975, 20, 434.

Plasma Process Polym 2016, DOI: 10.1002/ppap.201600140

14 ß 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim DOI: 10.1002/ppap.201600140
R

Early View Publication; these are NOT the final page numbers, use DOI for citation !!