You are on page 1of 4

Ion Implant Data Log Analysis for Process Control

and Fault Detection


Michael J. Rendon, David C. Sing, Marcy Beard, Michael Hartig, John C. Arnold
Motorola – Technology and Manufacturing
Dan Noble Center
Austin, TX 78721 US

Abstract— Data mining techniques have been introduced to the statistics of a user selectable recipe or set of recipes. However,
semiconductor industry in recent years. In many applications the data export capability of the on-board SPC function is often
real time process data from plasma etch or deposition tools is limited to transferring a single variable at a time to a 3.5-inch
collected and powerful statistical methods are used to detect floppy disk, and the system cannot be easily accessed remotely.
trends and correlations among variables; the results of which can
be used for process control and fault detection. The introduction
of such methods to ion implant process control has been slow,
In this paper we report on progress in developing a network
partly due to the large number (from hundreds to thousands) of
based implant data log (IDL) data analysis system which can
process recipes which can be run on a given tool, and the relative
lack of real time process signals which are available during a be used to access data from multiple tools and multiple recipes.
typical ion implant process. However, valuable run summary The system that is currently under development can be used to
data is available on virtually every ion implant tool in the form of generate individual control charts from any of over 100 process
implant data logs (IDL’s). The introduction of modern computer variables for a user selectable process recipe or from all
networking technology makes it possible to automatically collect implants in the database. Multi-variable models are being
data from IDL’s on multiple implant tools and store them in a developed to compare relationships among process variables.
database. Once the database is formed a multitude of analysis These models calculate predicted values of process variables,
tools can be applied for process control, fault detection, and and the differences between the model and the actual variables
ultimately scrap reduction and yield enhancement. Traditional are used as indicators of process drift, hardware malfunctions,
data mining techniques such as principle component analysis can recipe integrity, and in some cases mis-processing.
be applied to seek out correlations among the many different
signals available. However, the precise nature of the ion implant
process allows some variables to be analyzed in a more
deterministic manner. For example, a precise and predictable II. GLOBAL PROCESS MODELS
functional relationship is expected to be present between analyzer
magnet current and the implant species charge, mass, and
energy. A model can be generated to predict the magnet current A common problem an implant process engineer is faced
and the value in the IDL can be compared against it. The with is that a given implant tool or toolset utilizes a very large
deviations between model predictions and actual implant number of process recipes. Even within a tool type, for
parameters can be used as the input to a response system which, example high current implanters, doses can range over four
given the nature of the signal and the size and statistical
orders of magnitude (10 to 10 ions/cm ), energies can range
12 16 2

significance of the deviation, would generate the appropriate


response, if any. The level of the response could vary from an
over nearly two orders of magnitude (3 to 160 keV), and ion
email notification to an engineer (for example, the detection of masses can vary over one order of magnitude (11 AMU for
recipe with a bad source setup history which could be improved Boron to 121 AMU for Antimony). Control charts for a single
to reduce autotune times) to the automatic shutdown of the tool process variable for a single recipe can be useful for detecting
(if the wrong species was determined to have been implanted). tool faults and process drifts, but faced with hundreds of
process recipes and over 100 variables for each recipe to
Keywords-Ion Implantation; Automatic Process Control; Mass review, a process engineer can easily be overwhelmed by the
Analysis; Charge Control) crush of data, especially since each tool often produces over
100 IDL files each day. Multi-variable process models which
I. INTRODUCTION can be applied to all the IDL files regardless of the specific
process recipe, or at least a large subset of available process
In the past decade tool manufacturers have improved upon recipe, provide the basis for a more powerful process control
the quality and quantity of summary statistics files available on system. Multi-variable process models for analyzer magnet
ion implant tools. This trend has been facilitated by the current, secondary electron flood gun current, and source arc
increase in processing and storage capability of the PC and current are presented in the following section.
UNIX based workstations that have been coupled to the tool’s
automation systems. Several implant tool suppliers offer SPC
packages that provide the capability to review summary

- 331 - 0-7803-7155-0/02/$10.00 © 2002 IEEE


III. EXAMPLES OF GLOBAL PROCESS VARIABLE MODELS
Figure 1 shows the analyzer magnet current vs stiffness plot
for four high current tools installed at Motorola’s MOS13
A. Analyzer Magnet Universal Curve facility. Data is plotted for every implant from June 1 to July
31, 2002 for three tools, and from November 2001 to January
2002 for a fourth. Over 20,000 implants are represented in this
The heart of virtually all ion implant tools is the mass plot. The November to January dataset includes a case where
analysis magnet. The fundamental equation which governs the an AMU calibration error caused a scrap incident. The bad
operation of the magnet is the well known relationship between implants are clearly separated above the bulk of the data, and
the charge Q, mass M, and energy E of an ion and its radius of are marked with an ‘X’. Figure 2 shows the control chart for
curvature R as it moves in magnetic field of strength B: the November to January data, where the relative error between
the actual magnet current and a fourth order polynomial fit to
the magnet curve data is plotted. The bad implants are again
R*B = sqrt ( 2*M*E / Q ) (1)
2
easily identified as flyers well separated from the bulk of the
data.

This equation can be re-written in terms of the analyzer B. Secondary Electron Flood (SEF) Primary Current
magnet current I and the extraction voltage V as:

The high current implant tools employ an SEF charge


I = K * sqrt ( M*V / Q) (2) control system, the operating parameters are set by the recipe
values. The standard recipe setting of the SEF primary current
is ten times the ion beam current. Occasionally, when a beam
where K is a constant which contains the geometry factors current in a recipe is changed, or a recipe is copied from
(radius R) and other fixed factors. The quantity sqrt ( M*V / Q) another recipe, the SEF current will be incorrectly set. A global
is referred to as the magnetic stiffness of the beam. In reality, check of all implants is implemented by calculating the ratio of
the relationship between magnet current and magnetic stiffness the beam current to plasma flood from the IDL files, This ratio
becomes non-linear at high magnetic fields due to saturation of should be near 10 for all implants. Figure 3 shows the ratio of
the magnet core. However, a plot of analyzer magnet current Primary SEF current to Beam current for one tool (#417) for
versus magnetic stiffness will form a smooth curve for implants two months of implants. This control chart can then be used to
of all species, charge states, beam currents, and extraction flag IDL files for a follow up to check to see if the recipe called
voltages for a properly calibrated analyzer magnet. A similar for the correct SEF primary current or if the SEF system was
model can also be constructed for other magnets which are operating normally.
used to steer the beam, for example, the final energy magnet of
many high energy implanters.

Figure 1. Analyzer Magnet Universal Curve (data from 4 tools Figure 2. Magnet Curve Control Chart (data from 1 tool). ‘X’
overlayed).’X’ marks fliers due to AMU calibration error. marks fliers due to AMU calibration error.

- 332 - 0-7803-7155-0/02/$10.00 © 2002 IEEE


calculated by linear interpolation the manufacturer’s table of
maximum spec beam currents as a function of energy. Figure 4
shows the predicted arc current vs the actual arc current for a
single tool (#417). A good correlation is observed, with a
correlation coefficient of greater than 0.9. A control chart of the
relative error between the actual arc current and the predicted
arc current is shown in Figure 5. The data shows this tool
generally sets up within 20% of the predicted value, and the
flyers can then be analyzed to identify recipes with excessive
arc current that may require retuning.

IV. A PROTOTYPE IDL ANALYSIS SYSTEM

As was mentioned earlier, it is increasingly common for


implanters and many other process tools to have dozens or
hundreds of different recipes active at any given time. Even
when the number of recipes is smaller, other factors such as
Figure 3. Ratio of primary flood gun current to beam current (data from wafer pattern density, film thickness, etc. can cause a given
1 tool)
tool to have many distinct sets of operating conditions. This
proliferation of states presents a new challenge for traditional
C. Species Specific Arc Current Models fault detection methods, which typically demand that the tool
owner either explicitly set limits for each parameter or
Arc current is the single best indication of good source and manually classify runs as “good” or “bad” for the purposes of
extraction performance. A poorly built source or poorly aligned training a statistical model. As the number of “normal”
extraction assembly often results in excessive arc current operating states increases, the burden associated with
required to produce a given beam. Poor source or extraction establishing and maintaining limits becomes unmanageable.
performance can be identified by comparing the actual arc To mitigate this effect, a different approach was employed
current against the arc current predicted by a model. A simple in this work. Rather than attempting to automatically
model was constructed for As+ arc current (Iarc) using a power disposition each run as “good” or “bad,” the analytical system
law model: was designed to flag anomalous runs for further engineering
Iarc = A * (Inorm)
B
(3) attention. Although this approach theoretically leaves more
opportunity for wafer scrap, particularly if the cycle time for
engineering response is long, the much lower cost in up front
In (3) normalized beam current (Inorm) was calculated using the effort makes it more attractive than the conventional methods.
average beam current from the IDL file divided by the An analysis program was written using a commercially
maximum beam current. The maximum beam current is available software package to demonstrate the approach. A new

Figure 4. As+ arc current model results (data from 1 tool). Figure 5. As+ arc current control chart (data from 1 tool).

- 333 - 0-7803-7155-0/02/$10.00 © 2002 IEEE


type of chart was created, with limits automatically added as The initial version of the system is interactive and requires
soon as any data is entered. The limits are based on the mean user participation. Design of an automated system is underway,
and standard deviation of the data up to that point, and they with a goal of automatically collecting the data, processing it
change with each new point entered. Initially all points are through the analysis routines, and reporting on all flyer points
included in the limits calculations. After a user-configurable [1]. A graphical user interface will be included for viewing
“ignore time” any point outside the limits will be considered an charts and updating user input parameters such as how many
anomalous run (a “flyer”) and will not be included in points are needed before making a model. Default values for
subsequent calculations. In addition, a cap is placed on the these parameters will allow the system to run without user
number of points used in the limits calculation so the dataset setup, but the engineer may wish to modify the parameters for a
rolls as more points are added to the chart.. particular variable to allow for known trends or different levels
of variation. System tweaks will be required to limit reporting
of false flyers while still capturing real problems, but it is
Models have been created within the program for the expected that the time spent doing so will be much less than
analyzer magnet, plasma flood gun, and arc current curves. setting up user-defined limits in the first place.
After a certain number of points are entered, a model is
automatically generated and residuals are calculated against it.
The residual values are plotted and limits are set up as REFERENCES
described above in order to flag points that land a significant
distance from the model curve. Figures 1 through 5 are [1] M. Rendon, D. C. Sing, “Local Area Network LAN Address
examples of the models and control charts created using this Manufacturing and Development Implant Tool Issues,” International Ion
algorithm. Implant Technology Conference 2002, unpublished.

- 334 - 0-7803-7155-0/02/$10.00 © 2002 IEEE

You might also like