Kein Folientitel

1
DIAGNOSTICS IN INDUSTRIAL REACTORS
WHY DIAGNOSTICS (sensors) ? Structure of a
semiconductor production plant Financial
aspect Trend and consequence for plasma
diagnostics PLASMA DIAGNOSTIC IN
PRODUCTION Role of a sensor in a
system Different levels of warning NEEDS AND
TRENDS IN THE EQUIPMENT INDUSTRY Today, problems
of scale up Longer term, total feedback
control Self bias and its meaning Examples of
todays stupidity Loss of the viewing
angle Total automatic control Relevance of
measurement System self teaching Variations
reactor to reactor A new problem
electrostatics SUMMARY AND CONCLUSIONS
2
Possible architecture of a plant
plasmas
3
Example of a fraction of process flow
4
Financial aspect
Investment from 0.5 to 1 billion US !
5
Financial aspect
6
TRENDS IN THE ELECTRONIC INDUSTRY
Investment is GIGANTIC Process is very
COMPLEX Main keys to success DELAY (very
short) YIELD (very high)
Consequences on equipment choice NO RISK -
proven technique preferred - long delay for
introducing innovation (probation phase) -
simple and stupid far better than
sophistication CONSISTANCY MORE IMPORTANT THAN
PERFORMANCE - production scale up based on
copy and paste - process and equipment should
work in robust zones - equipment long MTBF
and short MTTR MTBF Mean Time Between
Failure, MTTR Mean Time To Repair
7
PLASMA DIAGNOSTIC IN PRODUCTION
Main consequences for diagnostic sensors
attached to the plasma production
tools RELIABILITY - the sensor must be a
source of improved reliability - the sensor
should not interfere with the process - sensor
failure should not jeopardise the benefit
related to the presence of the sensor. - simple
and stupid is always a winner - smart and
complex sent back to development PRODUCTION
WORTHY - ideally the sensor should give an
early warning that the process is drifting
away from its optimum before the product is
out of its specifications. - however false
alarms are strictly forbidden. - the
understanding of physics is irrelevant, what
counts is that the sensor sensitivity and its
correlation with process quality.
8
PLASMA DIAGNOSTIC IN PRODUCTION
WARNING This paper is (on purpose) provocative
to the scientific community. Hence what is
stressed is what scientist are not already
familiar with. PLEASE REMEMBER It needs
sometimes far more genius to implement something
stupid and simple instead of a sophisticated
and complex measurement. Speaking of reliability
or yield, it generally takes as much work (and
creativity) to go from 50 to 98 than to go from
98 to 99 . Unfortunately the business margins
are in the last points !!
9
THE 3 ROLES AND LOCATIONS OF A SENSOR
PLASMA
SENSOR
To next process
1) REMOTE POST PROCESS ANALYSIS Not a real
plasma sensor, analyses the result Touching the
substrate is usually forbidden Optical technique
(reflectometry, ellipsometry) Still very
effective if fast and reliable Allow s an early
stop of defective substrates Can detect process
drift if sensitive enough
Gives only one value (or one set of value) per
run.
10
THE 3 ROLES AND LOCATIONS OF A SENSOR
PLASMA
SENSOR
2) REAL TIME CONTROL OF A PARAMETER Measures
one process parameter. Needs not to perturb the
process (watch out for window or probe
perturbation) Possibly analyses the substrate
(reflectometry) Possibly analyses the plasma
(probe, RF voltage, etc) Possibly analyses the
gas phase (pressure, QMS, etc..) Possibly
analyses a combination (OES, self bias, etc) Can
detect process drift if sensitive enough Trade
off needed between number of data and memory size
Gives a full time sequence of measurement per run
11
THE 3 ROLES AND LOCATIONS OF A SENSOR
Input parameter
(End point detection is part of this case)
Feed back controller
PLASMA
SENSOR
3) FEED BACK CONTROLLING SENSOR Similar as case
2) but also The sensor is used to stabilise
the process At least one process input parameter
is automatically varied to keep constant the
measurement of the sensor. Some examples Match
box tuning (Reflected RF / 2 setting of match box
capacitors) Pressure control (pressure / throttle
valve)
The offset of the sensor (departure from set
point) can be recorded as process data
12
WHAT TO DO WITH THE DATA ?
13
LEVELS OF WARNING FROM A SENSOR
A process controlling sensor should have several
levels of warning depending of the departure from
the set point
SENSOR OUTPUT
IMMEDIATE STOP
Defective system Stop at the end of the
run Product in process are not qualified
Product still OK Warning message System check-up t
o be planned as soon as possible
OK
Central value
14
EXAMPLE MFC DRIFT
Mass Flow Controllers, when handling highly
reactive gases and vapours, are among the most
sensitive subsystems in plasma processor. Usually
they do not break, they drift. An erroneous gas
mixture can bring thin films out of
specifications (leaky insulator, wrong taper).
15
IMPROVING PROCESS SAFETY double checking
16
NEED AND TRENDS IN THE EQUIPMENT INDUSTRY
Note the point of view expressed here is
mostly related to the Display industry, however
some of the long term trends are also valid for
the semiconductor industry.
SHORT TERM Cost issue - Added value per unit
surface, sensor cost Basic questions -
Interpretation of a sensor drift Substrate size
issue - Optical measurement and viewing
angle - Self bias (meaning of the
measurement) Reactor to reactor - difference
between 2 reactors LONG TERM Feed back
control - Status of today - Some possible
improvements - Long term future Sensor
research - Relation to process results -
Multi-step processes - Data compression
17
SENSORS Cost issue
18
SENSORS Cost issue
The Display Industry will not grow unless it
learns how to produce at low cost, same goes for
the solar cell industry. The Semiconductor
Industry is gradually becoming a low margin
business (pressure from far east
countries). Equipment suppliers are being
transmitted the cost squeezing pressure system
cost of ownership shall drop down COO COST
PER UNIT PROCESS IN PRODUCTION ( Yearly
amortisation / Yearly throughput) gasses
electricity water other fluids man
power maintenance floor space rental
yield decay due to this system The investment
cost of a system is not related to the weight or
the steel work, but is rather related to the
complexity of the system (electronic and
electric, interlocks, software, troubleshooting,
etc). Sensors can make a definitive impact on
system complexity. Stupid simple more than ever
the best choice !
19
RELATIONS PROCESS-SENSOR
The ideal sensor gives a signal which varies with
the maximum sensitivity with process quality. The
relation can be experimentally established by
optimised orthogonal planning (work intensive).
In the attached figure, the process parameter
space is only 2-d. It is usually much higher
dimension. Example of process parameters -
RF power, match box setting - Pressure, gas
flow and composition - Temperature Attention
quality can be also a multiple parameter
concept. There is no need to understand the
detail of the physics in the relation between
sensor signal and process quality. It is more
important not to neglect hidden parameters
(purity, substrate history, etc.).
the process quality is shown in the colour scale
20
ISSUES RELATED TO SUBSTRATE SIZE Trend in the
Display Industry
21
ISSUES RELATED TO SUBSTRATE SIZE self bias
sensitivity and interpretation (1)
DC Voltage measurement
RF filter
RF PLASMA
Matching network
Self bias good example of stupid simple -
Easy to implement - External to the process
zone - Varies with some process
parameters Self bias is universally used in RF
plasmas Still used for probing display
processing What is the meaning of self bias ?
RF GENERATOR
22
ISSUES RELATED TO SUBSTRATE SIZE self bias
sensitivity and interpretation (2)
VRF
VRF
Vbias
SHEATH (electrode)
C ? Selectrode
SHEATH (electrode)
SHEATH (g.)
Vplasma
Vplasma
SHEATH (ground)
SHEATH (ground)
C ? Sground
Equivalent circuit in blue RF component in cyan
DC component
SHEATH
Voltage
Vplasma
average
Ions
Vplasma
electrons
23
ISSUES RELATED TO SUBSTRATE SIZE self bias
sensitivity and interpretation (3)
Plasma sheath vacuum capacitance equivalent
thickness ? with ? ? V?
Sheath rectification implies that Ve ? ½ Ve
(peak to peak RF voltage) Vg ? ½ Vg
Constant RF current ?? Ce Ve ? Cg Vg hence Se
Ve / V?e Sg Vg / V?g
24
ISSUES RELATED TO SUBSTRATE SIZE self bias
sensitivity and interpretation (4)
Vbias Vg - Ve ½ (Ve - Vg) ½ VRF (1-Rn) /
(1Rn) with R Se/Sg (relative electrode
surface ratio)
WAFER
For R? 1 , Vbias is a function of both VRF and R
the surface ratio. Before is was just ½ VRF
Vbias becomes a less simple tool
25
ISSUES RELATED TO SUBSTRATE SIZE self bias
sensitivity and interpretation (5)
Equivalent circuit
RF
Plasma
Dust grain below the substrate
The series capacitance of the substrate changes
the sheath capacitance of the ground, hence it
modifies the equivalent surface ratio. If a dust
grain is below the substrate, the substrate
capacitance is modified, a change is seen in
Vbias. If a Vbias drift is observed, shall we
change the RF generator for calibration or open
the reactor for cleaning?
26
ISSUES RELATED TO SUBSTRATE SIZE self bias
sensitivity and interpretation (6)
Vbias as a plasma monitoring tool is so popular
that people are still using it in the RIE etching
of insulating substrate. Question what is the
meaning of such a measurement when all the
electrode is protected from direct exposure to
the plasma to avoid sputtering of the metal?
Sacrificial quartz liners
RIE ETCH PLASMA
to Vbias measurement
FILTER
RF
The measured signal is due to the faint plasma
which penetrates in the electrode/ground gap. One
may wander about effects such as geometry,
thermal expansion, surface oxidation,
electronegative gases, etc.. Again it is clear
that self bias gives an information with value
(it is even used as an end point), but the open
question is how to interpret signal level
variations, drift, etc..
27
ISSUES RELATED TO SUBSTRATE SIZE Viewing angle
The plasma gap cannot be varied the process
which was qualified with smaller substrate must
be preserved
2 severe problems - incidence angle is very
small, reflectometry or ellipsometry are not
at their best. Technology to be
revisited. -the optical aperture is very
small, non coherent sources are difficult to
use. This optical etendue problem is also an
issue for local analysis of the plasma
spontaneous emission.
During process, part of the analysed light
emission is reflected on walls. The reflection
coefficient is modulated by the film thickness
variations. The signal is found to vary while the
plasma is stable. This oscillation would have
some value if the reflection angle was well
defined, but the solid angle would need to be
very small and the signal would be very weak.
28
REACTOR TO REACTOR VARIATIONS
KAI Parallel processing PECVD production system
(20 identical reactors in parallel). Ideally they
should provide identical results. In reality they
give an excellent /- 3 reactor to reactor
thickness variation for SiN deposition. Note that
this level of variation is of the order of the
the uncertainty of the sensors attached to each
reactor. However after very careful analysis and
cross-correlation we have identified, for this
specific SiN process, what is responsible of the
3 box to box average thickness variation -
about 1 is due to gas flow variation (well
related to the accuracy of our gas flow
divider) - about 1 is related to the RF
generator calibration accuracy (made on a 50?
resistive load) -about 1 is related to the
reactor capacitance fluctuation (match box losses)
29
REACTOR TO REACTOR VARIATIONS
RF power distribution
PLASMA
OHMIC LOSSES IN MATCH BOX FEED THROUGH
50 ?
30
FEED BACK CONTROL IN PROCESS SYSTEMS
Classical feed-back sub-system units found today
in most standard process equipment
31
FEED BACK present status
These independent feedback loops are actually
interacting via the plasma. In some cases, this
interaction can result into long relaxation time
, even instability
Example Pressure regulation versus match box
Reflected RF
Process exhaust flow
(for one given match box setting)
0
0
Effective RF power
Pressure
Effective RF power
Exhaust gas flow
Process pressure
Plasma impedance
Reflected RF power
Match box adjustment
Both pressure control and match box setting are
coupled via the plasma response. If the gain of
the feedback loops is too large and for some
parameter coupling configuration, the ensemble
can ring and offer very poor control. All this
can be stabilised by proper setting of the PID.
32
FEED BACK present status
33
LONG TERM Multiple feedback
Modern fighter planes, Most advanced robots are
driven by multiple feedback from a simple central
unit
The central unit computer must be a fast real
time unit. Such units are today available on the
market (originating mostly from the military
market)
34
LONG TERM Multiple feedback
The computer establishes the process result as a
combination of sensor measurement, then detect
the differential from optimum, finally it
calculates the variations of all control
parameter settings in order to bring the process
as close as possible to its optimum process point.
35
LONG TERM Multiple feedback
Necessary conditions for multiple feedback
implementation - Drop the individual sensor
feedback concept. - Relation process quality /
sensor response must be known - Sensors must be
absolutely reliable
One finds here again the main issues related to
sensor/diagnostics in industrial environment
Such a system can define its best response by
self-learning The response of all process
component is analysed for a step like
perturbation of all input parameters, including
the controlled parameters. The system response is
locally linearized and the response calculation
is similar to a matrix inversion logic.