Surface & Coatings Technology 241 (2014) 45–49

Contents lists available at ScienceDirect

Surface & Coatings Technology

j ourna l homepage: www.e lsev ie r .com/ locate /sur fcoat

Numerical optimization of baffles for sputtering optical precision filters☆

Andreas Pflug ⁎, Michael Siemers, Thomas Melzig, Daniel Rademacher, Tobias Zickenrott, Michael VergöhlFraunhofer Institute for Surface Engineering and Thin Films IST, Bienroder Weg 54e, 38108 Braunschweig, Germany

☆ This manuscript is based on work presented at the SAnnual Technical Conference in Providence, Rhode Island⁎ Corresponding author.

E-mail address: andreas.pflug@ist.fraunhofer.de (A. Pfl

0257-8972/$ – see front matter © 2013 Elsevier B.V. All rihttp://dx.doi.org/10.1016/j.surfcoat.2013.11.008

a b s t r a c t

a r t i c l e i n f o

Available online 15 November 2013

Keywords:Process modelingOptical precision coatingsDirect Simulation Monte CarloMagnetron sputtering

The optimization of the coating uniformity of precision optical filters generally is a critical and time consumingprocedure. The present paper demonstrates this optimization procedure on anewoptical precision sputter coater“Enhanced Optical Sputtering System (EOSS)” at Fraunhofer IST. The coater concept is based on dual cylindricalsputtering sources and a rotating turn-table as sample-holder. For compensating non-uniformity introduced bythe particle flux profile and the radially dependent track speed on the turntable, baffle elements have to be de-signed and inserted beneath the substrates.For that purpose thedistribution of the particleflow from the cylindricalmagnetron aswell as the resulting thick-ness profile for different shaper designs is simulated using Direct SimulationMonte Carlo (DSMC) transport sim-ulation. For comparison, experimentally obtainedfilm thickness profiles are evaluatedby spectrophotometry andellipsometry. The simulationmodel is used for optimization of the baffle geometry aswell as investigation on therole of long term drifts caused by target erosion and mechanical tolerances.

© 2013 Elsevier B.V. All rights reserved.

1. Introduction

Advanced optical precision applications such as spectroscopy, lifesciences or fsec lasers require coatings with high quality and precision.Due to the high number of layers, optical filters are quite sensitive tolayer optical thickness errors; thus many products require maintainingfilm thickness deviations below ±0.5% or even better. With the trendof increasing substrate sizes the fabrication of such products becomesmore and more a challenge.

Recently, the so-called “Enhanced optical sputtering system”

(EOSS) — a novel coater concept for deposition of dielectric layers withhigh long-term stability and low absorption and particle contaminationlevels — has been introduced [1]. This coater consists of two compart-mentswith dual rotatable sputter cathodes in a sputter-up configuration.A third compartment is equipped with an ICP source for plasma oxida-tion. Up to 10 substrates with a diameter of up to 200 mm can bemounted on a turntable above the sputter sourceswith amaximum rota-tion speed of 250 rpm.

Due to the radial dependency of the track speed of the turntable, thedeposition rate on the substrates would decrease inversely with radialposition. A bafflemountedbetween substrate and sputter sourcewith ap-propriate geometry serves as uniformity mask for compensating this ef-fect. Since the flux of sputtered particles is not evenly distributed on thesubstrate plane, the proper design of the baffle usually requires numerousdeposition experiments and iterations for fine-tuning. As an alternative,this work demonstrates a method of numerically optimizing the baffle.

ociety of Vacuum Coaters 56th, April 20-25, 2013.

ug).

ghts reserved.

2. Simulation setup

The numerical optimization is accomplished by computing the fluxof gas and sputtered metal within the sputter recipient. The descriptionof particle transport under rarefied gas flow conditions is possible viathe “Direct Simulation Monte Carlo” (DSMC) method [2], whereof weuse an in-house developed, massive parallel implementation [3,4].

The geometric model of the sputtering recipient as shown in Fig. 1comprises dual rotatable targets, gas inlets mounted at left and rightsides, shutters, the turntable, and the baffle elements mounted just2 mm beneath the turntable. It is assumed that the planar substratesare aligned in-line with the lower plane of the turntable.

All inner chamber surfaces relevant for gas and metal transport arerepresented by triangularly meshed surfaces. The geometric modeland the mesh are constructed using the open source mesh generatorGMSH [5].

For the sputter sources, the race tracks on the sputtering targets areapproximated by four elongated rectangles (see Fig. 2). This is possiblesince with a target length of 550 mm and a substrate diameter of200 mm, the end blocks have no significant contribution to the film de-posited on the substrate. The target to substrate distance depends onthe actual thickness of the target material and is in the range of 80 mm.

The angular andenergy distribution of the sputtered particles is non-thermal and obeys a so-called Thompson energy distribution with

f E; θð Þ ¼ cosn θð Þ EE þ Ubð Þ3 ð1Þ

where E is the energy, θ the polar emission angle, n is an index of thecosine angular distribution, and Ub is the binding energy of thematerialat the target surface.

Rotatable targets

Ar inlet Ar inlet

Turntable

Baffle

Turbo molecularpump

ShutterShutter

Fig. 1. Scheme of the geometry of a sputter compartmentwithin the EOSS optical precisioncoater.

46 A. Pflug et al. / Surface & Coatings Technology 241 (2014) 45–49

These data are strongly material dependent and can be found in lit-erature e. g. in Ref. [6] or by applyingMonte Carlo simulations of the col-lision cascade in the targets by using simulation packages such as SRIM[7]. The right graph of Fig. 2 shows the energy distribution of sputteredNb in comparison with a Maxwellian distribution with the same maxi-mum in energy.

The angular positions of the race tracks have been previously deter-mined by magnetic field computation: The maximum plasma densityoccurs at positions, where the normal component of the magneticfield with respect to the target surface becomes zero.

With a backgroundAr gas pressure in the range of 300 mPa, themeanfree path is in the order of 1–2 cm. Hence, to maintain proper collisionstatistics in the DSMC simulation, a cell dimension of about 8 mm is suf-ficiently small. In the actual DSMC simulation case, the geometric modelis embedded into a large cuboid volume sized 1060 × 1226 × 635 mm3,the number of cells in X, Y and Z direction are 128, 150, and 80. It shall benoted that the triangular surface mesh elements visible in Figs. 1 and 2are only relevant in the particle trajectory computation. Unless the geo-metric shape becomes distorted by very large elements, their size hasno impact on the simulation.

For parallelization, the whole volume is sub-divided into 80 cuboidsegments; a load balance algorithm measures the approximate CPUload of each segment and performs a proper assignment of volume

Approximatedrace tracks

Fig. 2. Locations of sputter particle sources on rotatable targets (left) and energy

segments to CPU cores. By this way, a good parallelization efficiency inthe range of N70% is obtained for up to 40 CPU cores.

The time step used in the DSMC iterations is 1 μs implying a meantravelling path of a thermalized Ar atom of about 0.4 mm per iteration;for sputtered Nb with a kinetic energy of 10 eV, the travelling distancewould be 5 mm. Thus, high energetic sputtered species are a limitingfactor for the maximum possible iteration time step.

The statistical weighting factors of Ar and Nb are 3 × 1012 and 1010,respectively; i.e. one Nb simulation particle represents 1010 real Nbatoms. A homogeneous Ar pressure distribution of 320 mPawas initial-ly applied to the whole simulation volume. After about one real timesecond corresponding to 106 DSMC iterations the gaseous and metalflux profiles are in equilibrium. For the actual computation we use aLinux cluster based on four computing nodes whereof each is equippedwith four 12-core AMD “Magny-Cours 6172” CPUs. One DSMC simula-tion run typically uses 30–40 cores and overall computation timesbetween 18 and 24 hours.

For better statistics of the resulting profiles, the simulation data aretime-averaged over a further time interval of 0.2 s. Cut plane views ofthe resulting Ar and Nb fluxes through the reactor are shown in left andcenter picture of Fig. 3. With respect to baffle optimization, the most rel-evant result from the simulation is the cumulated Nb deposition profileonto the non-shaded area of the turntable, as shown in the right pictureof Fig. 3. Besides the substrate plane, also parasitic deposition onto thebaffles and onto other inner chamber walls can be visualized from thesimulation data. By integration over all deposited surfaces, the fractionof material which is effectively used for deposition can be evaluated.

3. Optimization method

If the baffle is located close to the substrate, it can be assumed thatno significant metal–gas interaction occurs between baffle and sub-strate. In this case, it is a valid approach to simply cut the metal fluxtowards the substrate according to the baffle geometry.

The benefit of this approach is that we need only one time-consuming DSMC computation without baffles for obtaining thewhole metal flux profile at the substrate plane and can subsequentlycompute the radial deposition profile within a simple interpolationand cutting algorithm based on the baffle geometry. The latter steponly takes a few milliseconds on a single CPU; this enables to performa systematic optimization of the baffle geometry for uniform depositionprofile according to the overall scheme shown in Fig. 4. For that purpose,the inner boundary of the baffle elements is assembled from multiplecircular arcs, whereof the junction coordinates and radii can be freely

0 50 1000

100

200

300

400

500

600 SRIM histogram Fit with Ub=7.39 eV Maxwellian distribution with <E>=7.39 eV

Inte

nsity

[a.u

.]

Energy [eV]

distribution of sputtered particles in the case of Nb as target material (right).

Fig. 3. Cut plane views of simulated fluxes of Ar (left) and Nb (center) as well as Nb deposition profile on walls (right). All plots are time-averaged during between 1.6 and 1.8 s.

47A. Pflug et al. / Surface & Coatings Technology 241 (2014) 45–49

parameterized. This geometric representation is compatible with a CNCcutting tool used for fabrication of the baffle element afterwards. Theinset in Fig. 4 shows both, the shape of the initial and numerically opti-mized baffle in geometric relation to the sputtering targets.

With the given geometric boundary of the baffle, themetal flux pro-file is integrated along the track lines of the turntable motion; themetalflux profile is thereby interpolated at the integration points and the cut-off behavior of the baffle is taken into account. The radial depositionprofile on the substrate is obtained as a result.

In order to optimize the deposition profile, a simplex fitting algo-rithm [8] is applied which uses the radial deposition flux and a goalfunction as input and modifies the coordinates and radii of the circlearcs of the baffle. The goal function is a constant line; whereof the abso-lute valuemay be varied to a certain extend corresponding to a variationin baffle transmission coefficient.

This method can be also applied on other sputtering materials suchas Ta or Si. For each newmaterial at least one 3D DSMC simulation is re-quired,where theparametersUb and n fromEq. (1) are adjusted accord-ing to the material's specific energy and angular emission profile. Afterthe resultingmetal flux profile at the substrate plane has been obtained,the remaining optimization procedure is the same as described above.

4. Optimization results

The results of the DSMC simulation runs and the baffle optimizationare summarized in Fig. 5: In the case of no baffle, the radial depositionprofile depends reciprocally on the radial position. An experimental baf-fle optimization procedure involves several iterations, where film

Fig. 4. Procedure for extracting the optimized baffle geometry from DSMC calculatedmetal fluxgeometric position. Additional shielding plates near the baffles are omitted for clarity.

thicknessmeasurements and appropriate modification of the baffle ele-ment are required.

The black line and symbols in Fig. 5 show the situation for such baffleafter two experimental optimization iterations; here the remaining filmthickness deviations are still about ±2.5%.

In contrast, with numerically optimized baffles, the film thicknessdeviations could be further reduced below ±0.35%; the correspondingoptimized profile shown in Fig. 5 originates from ellipsometricmeasurements.

5. Tolerances and long term drifts

Besides optimization, the simulation model allows for analyzing theimpact of mechanical tolerances and long term drifts on the coating re-sult. With respect to long term drifts, the erosion of the rotating targetcould influence the deposition profile. For investigating that, we com-pared two DSMC simulation runs with fixed baffle geometry but withdifferent target diameters of 152 mm and 138 mm, respectively. With132 mm as diameter of the backing tubes, this corresponds to targetmaterial thicknesses of 10 mm and 3 mm. A magnetic field computa-tion at the cylindrical targets surface shows that with increasing targeterosion, the angular positions of both race tracks slightly shift towardsinside (see Fig. 6, left graph). This leads to a more focused emissionprofile.

The quotient of both deposition profiles, as shown in the right graphof Fig. 6, indicates that this introduces a linear shift in the radial deposi-tion profile.

Due to the more focused emission character for small target diame-ters, the effective deposition rate increases where the baffles are most

profile. The inset shows the shape of the initial and numerically optimized baffle and their

Fig. 5. Measured and modeled radial deposition profile for the case of (i) no baffle (greyline and circles), (ii) experimentally optimized baffle (black line and circles), and (iii) nu-merically optimized baffle (grey squares, measurement only).

500 550 600 650 700

98

99

100

101

102

103

Rel

ativ

e de

spos

ition

rat

e [%

]

Radial position on substrate [mm]

Simulated rate profiles with optimized baffle Magnetron tilting 5° Magnetron tilting 10° (default position) Magnetron tilting 15°

Fig. 7. Impact of magnet bar tilting on simulated deposition profile.

48 A. Pflug et al. / Surface & Coatings Technology 241 (2014) 45–49

closed, i.e. at the inside position of the targets. This effect is comparablysmall but not negligible for high-precision optical coatings.

It shall be noted that the noise in the right graph of Fig. 6 stems fromthe limited number of particles in the simulation. With the given simu-lation parameter settings, each point in the radial distribution resultsfrom approx. 4 × 106 simulation particles. By assuming Poisson statis-tics, this leads to a relative noise of ±0.05% for one radial distributionand ±0.1% for the shown quotient of two distributions.

An examplewith respect tomechanical tolerances is the angular ori-entation of themagnet barswithin the targets. In the original configura-tion, both magnet bars are tilted 10° towards the center in order toincrease the deposition flux in substrate area. Fig. 7 shows a simulateddeposition profile with optimized baffle geometry in comparison witha second simulation run, where the magnet bars are tilted by only 5°.It turns out that the deposition profile is highly sensitive on the exacttilting angle. As a result, a significant linear gradient in the radial filmthickness profile occurs as can be seen in Fig. 7. On the other hand, if a

Fig. 6. Effect of the target material consumption from 10 mm down to 3 mm on the a

linear film thickness gradient is observed on experimentally obtainedsamples, the simulationmodel indicates that this could be compensatedby adjusting the magnet bar tilting.

6. Conclusion

The DSMCmethod is capable of describing gas and particle fluxes ina PVD reactor with high accuracy. Due to massive parallel algorithms, itis possible to apply this method on recipient geometries with industrialrelevant size and complexity.

In combination with an optimization procedure it is possible toapply DSMC simulation for refinement of the recipient geometry to-wards improved film thickness uniformity; in our example we reacheda uniformity level of ±0.35 % on a substrate sized 200 mm in diameter.Furthermore it is possible to investigate the impact of long term driftsand mechanical tolerances on the deposition result within the simula-tion model, which is quite important for the practical design of noveltypes of PVD coating chambers.

ngular position of the racetracks (left) and on relative deposition profile (right).

49A. Pflug et al. / Surface & Coatings Technology 241 (2014) 45–49

Acknowledgement

The development of the DSMC/PIC-MC simulation tool was partlysupported by the VolkswagenStiftungHannoverwithin the joint project“Cosmos” (I83/234), which is greatly acknowledged by the authors.

References

[1] D. Rademacher, G. Bräuer, M. Vergöhl, B. Fritz, T. Zickenrott, Proc. SPIE 8168 (2011)81680O.

[2] G.A. Bird, Oxf. Eng. Sci. Ser. 42, 1994, 1–458.[3] A. Pflug, M. Siemers, B. Szyszka, H. Gross, Proc. 52nd Annual Technical Conference of

the Society of Vacuum Coaters, 2009, pp. 364–369.[4] C. Schwanke, A. Pflug, M. Siemers, B. Szyszka, Lect. Notes Comput. Sci 7133 (2012)

213–219.[5] C. Geuzaine, J.-F. Remacle, Int. J. Numer. Methods Eng. 79 (11) (2009)

1309–1331.[6] Y. Yamamura, T. Takiguchi, M. Ishida, Radiat. Eff. Defect Solids 118 (3) (1991)

237–261.[7] J.F. Ziegler, Software and web site, http://www.srim.org.[8] J.A. Nelder, R. Mead, Comput. J. 7 (4) (1965) 308–313.