`
`Gabriel G. Barna
`Semiconductor Process and Device Center
`Texas Instruments, Dallas, TX 75265
`
`2.
`
`significant time delay relative to the rate of proc-
`essing of wafers.
`In retrospect, it is clear that all of these con-
`straints have to be removed when pursuing the
`ever-increasing demands placed on manufacturing
`operations due to the well-known problems asso-
`ciated with the continuing decrease in feature size.
`Specific requirements are that:
`processing anomalies be determined by ex-
`1.
`amining a much wider domain of parameters
`processing anomalies be detected in shorter
`timeframes; within-wafer or at least wafer-to-
`wafer
`wafer state parameters be measured, or esti-
`mated, frequently
`processing emphasis be focused on decreasing
`the variance of the wafer state parameters in-
`stead of controlling the variance of the set-
`points
`APC is the current paradigm that attempts to
`solve these four specific problems. Under this
`general heading, the FDC component addresses
`the first two requirements, MBPC addresses the
`last two.
`
`3.
`
`4.
`
`Abstract - This paper presents an abridged history
`of Advanced Process Control (APC), including
`both Fault Detection and Classification (FDC) and
`Model Based Process Control (MBPC), both
`within TI and in the semiconductor industry.
`While TI was an early leader in univariate fault
`detection in processing tools, other manufacturers
`have by now implemented such methodologies.
`For MBPC, the MMST program gave TI a lead,
`but others are now following that path. For TI and
`the semiconductor industry as a whole, the current
`thrust is to develop and implement multivariate
`APC methods into the manufacturing operations.
`This paper describes the complexity of the execu-
`tion of these tasks, and lists some of the available
`tools that are requisite for implementing these
`plans.
`
`APC Background
`
`The precursor to APC (Advanced Process Con-
`trol) in semiconductor manufacturing operations,
`has historically been the SPC (Statistical Process
`Control) activities that have become prevalent in
`the 60’s. A fundamental operating principle be-
`hind SPC is that the process parameters - the
`hardware settings - be held invariant over long pe-
`riods of time. SPC then tracks certain unique, in-
`dividual metrics of this process - typically some
`wafer state parameter - and declares the process to
`be out-of-control when the established control
`limits are exceeded with a specified statistical sig-
`nificance. While this approach has some estab-
`lished benefits, it suffers from a) its myopic view
`of the processing domain - looking at one, or only
`a few - parameters and b) its delayed recognition
`of a problem situation - looking at metrics that
`may only be generated once in a while or with a
`
`History of APC at TI
`
`Elements of Advanced Process Control, includ-
`ing both Fault Detection and Classification and
`Model Based Process Control have been utilized
`within various TI Wafer Fabs for about a decade.
`FDC evolved from the early EMS (Endpoint
`Monitoring System) that was an embedded fault
`detection system built into the PAC 150PC series
`of single wafer plasma etchers [l]. This system
`analyzed the endpoint trace of every wafer in rela-
`tion to the endpoint trace of a good “reference”
`wafer, after the completion of the etch process.
`
`0-7803-3371-3/96/$5.00 01996 IEEE
`
`364
`
`1996 IEEE/SEMI Advanced Semiconductor Manufacturing Conference
`
`Authorized licensed use limited to: Christopher Gallo. Downloaded on June 23,2021 at 03:36:23 UTC from IEEE Xplore. Restrictions apply.
`
`Applied Materials, Inc. Ex. 1018
`Applied v. Ocean, IPR Patent No. 6,836,691
`Page 1 of 6
`
`
`
`~
`
`When an anomalous endpoint signal shape was
`detected, the system alerted the operator or shut
`down the etcher, based on a user-defined action
`relative to the severity of the anomaly. This meth-
`odology had a clearly apparent benefit as it auto-
`matically identified anomalous process conditions
`for specific wafers as soon as the wafer finished
`processing; these wafers could be examined indi-
`vidually and disposed of as appropriate. When the
`anomaly was significant, it automatically termi-
`nated processing, hence saving the remaining wa-
`fers from being misprocessed.
`MBPC was first developed in SFAB for epi
`deposition [2]. The model was simply:
`deposition thickness = rate * time
`and this model was adjusted based on intermittent
`measurements of the actual thickness. Using the
`adjusted model, the deposition time was recalcu-
`lated to keep the deposition thickness at the target
`value. This method improved Cpk, as expected,
`over running in an open-loop configuration using a
`set deposition time. In addition, it provided other
`significant benefits by reducing the number of
`qualification runs, as well as the time and the
`number of pilots required to requalify a process
`after an R&M operation.
`From a historical perspective, these methods
`were readily accepted and integrated into produc-
`tion because:
`both techniques were conceptually simple,
`1.
`hence easy to implement and disseminate
`there was a clearly defined benefit for both
`techniques
`the original implementation of these tech-
`niques was a reasonably “low-cost” effort from
`both a hardware and software points of view
`Current Status of APC
`
`2.
`
`3.
`
`processing tool that generated some signature (e.g.
`furnace temperature profile, current profile of the
`power supply during resist spin operation, etc.). So
`the methodology was provided in a stand-alone
`application, available to connect to various tools.
`This embodiment, ECR (Electronic Chart Re-
`corder), has been fanned out and is currently op-
`erational in most of TI Wafer Fabs. A slightly dif-
`ferent embodiment (Cruiser) was generated locally
`in one Fab where it is connected to a very wide
`base of processing equipment. In both cases, the
`analysis is univariate, where faults are detected
`based on data from a single sensor.
`MBPC has taken two distinct paths within TI.
`SFAB has continued with the upgrade of their
`deposition control activities, with “home-grown”
`software, and have shown significant operational
`benefits using univariate MBPC. Meanwhile,
`during the MMST program [3], TI developed the
`basis of what later has become known as
`ProcessWORKS [4], a system that contains all the
`elements necessary for automated multivariate
`MBPC. This embodiment is currently being de-
`veloped in SPDC and SFAB, in anticipation of TI-
`wide deployment.
`
`Status at Other Semiconductor Manufac-
`turers
`
`Most semiconductor manufacturers are currently
`using sensor signals in a univariate fashion for
`applications such as endpointing for oxide CMP,
`tool-state information from RF sensors, etc. Table
`1 summarizes these activities, and basically shows
`that everyone participates in univariate FDC and
`some manufacturers are heading to multivariate
`FDC. One major remaining question is whether to
`do these analyses at the tool or factory level.
`
`Status at Texas Instruments
`
`APC Directions
`
`The FDC activity started by the EMS system has
`diffused throughout all the Wafer Fabs across TI.
`The driving force was the realization that the
`methodology was readily applicable to any other
`
`It is clear from interactions at SEMATECH and
`other national level meetings that APC is being
`accepted and pursued by all major semiconductor
`manufacturers. One interesting point is the differ-
`
`Authorized licensed use limited to: Christopher Gallo. Downloaded on June 23,2021 at 03:36:23 UTC from IEEE Xplore. Restrictions apply.
`
`365
`
`1996 IEEElSEMl Advanced Semiconductor Manufacturing Conference
`
`Applied Materials, Inc. Ex. 1018
`Applied v. Ocean, IPR Patent No. 6,836,691
`Page 2 of 6
`
`
`
`Company
`AMD
`HP
`IBM
`Intel
`Motorola
`TI
`
`APC Activities
`Wafer Sleuth; AMD/HoneywelVNIST APC activity
`originated Wafer Sleuth
`univariate FDC with Process Guard
`univariate FDC for several years
`multivariate FDC with Modelware
`univariate FDC, univariate and multivariate MBPC, Wafer Sleuth,
`PCARTSPC for multivariate FDC
`Table 1. Summary of APC activities at the major semiconductor manufacturers
`ent emphasis being put on the various aspects of of years. With the evolution of process sensors
`that provide a variety of signals from the process
`this problem by the different players. For some,
`the emphasis is on Fault Detection, to prevent
`in real time, the trend is clearly towards multivari-
`further wafer misprocessing. For others, the pre- ate FDC. With the multiplicity of sensor signals,
`diction of wafer state properties is more signifi-
`the chance of detecting faults is significantly in-
`cant. Such a diversity in the visions and expecta- creased. However, this multivariate FDC places
`tions for FDC is also complicated by the variety of
`the problem in a significantly different realm, due
`the nomenclature used by different organizations.
`to new issues that now have to be addressed, such
`However, all these tasks are part of the APC “big as:
`sensor data acquisition, data transfer and
`picture”, which is shown in Figure 1. This figure 0
`summarizes the APC paradigm, as perceived in
`communications to a computational algorithm
`the SEMATECH J-88-E project [5,6]. The three
`have to be seamlessly automated
`major blocks represent the hardware, Process and
`data pre-treatment algorithms are required
`wafer state parameters. Assuming that direct wa-
`0 complex computations (PCA, p u , time-series
`fer-state sensors are not being used (as those are
`analysis ...) have to be performed to accommo-
`typically unavailable in OEM equipment, at this
`date for the correlated, redundant data sets
`time), all APC activities can be represented by the
`from different sensors
`models f, g and h. The use and requirements of
`these algorithms have to be made robust
`these models will be elucidated in the following
`against the natural drift in the sensors and the
`sections.
`intermittent step-changes in the system at
`times when the machine is cleaned
`these computations have to be automated, so
`
`Fault Detection
`
`Q.d
`
`RF Pmnr
`TCP pewv
`cmporltion
`1
`
`\
`
`t u t”
`
`3
`Virtual Sensor Model
`
`’ r
`, w,
`
`&2rp2L
`
`Lh. m
`RMKIlon
`
`Etch Rab
`
`defines a multitude of data acquisition and
`analysis options and requirements
`It is worth emphasizing that the first task of FDC
`is fault detection, i.e. the determination that during
`the processing of a particular wafer, the sensor
`signatures indicate a “non-normal” state. This re-
`quires that a model h be generated that determines
`process state anomalies from the multitude of ma-
`chine state and process state sensor data. This is a
`bigger problem than might first be envisioned,
`
`2
`Figure 1. Model representation of APC compo-
`nents
`
`Authorized licensed use limited to: Christopher Gallo. Downloaded on June 23,2021 at 03:36:23 UTC from IEEE Xplore. Restrictions apply.
`
`366
`
`1996 IEEElSEMl Advanced Semiconductor Manufacturing Conference
`
`Applied Materials, Inc. Ex. 1018
`Applied v. Ocean, IPR Patent No. 6,836,691
`Page 3 of 6
`
`
`
`since the “normal” state is not a stationary point,
`but a slowly changing trajectory through time. The
`clearest example of this is the noticeable degrada-
`tion of the endpoint signal obtained from the ma-
`chine state data in any plasma etcher. This signal
`decreases continuously, as the window transmit-
`tance degrades due to the gradual accumulation of
`a residue on this window. So measured by this pa-
`rameter, the normal processing state metric de-
`creases continuously, while there is of course no
`“fault”, or even a drift, in the system. The FDC
`methods have to distinguish between this type of
`sensor drift, and the other source of drift; that due
`to the changing “state” of the processing chamber.
`This is typically attributed to things such as: wear
`of the electrodes, buildup of residue on the walls
`of the process chamber, etc.
`Given these constantly changing sensor signals
`and chamber state during the normal operating cy-
`cles between successive cleans of the equipment, it
`is clear that models can not be generated based on
`the concept that specific settings on the equipment
`generate a specific set of values for all the sensor
`signals. Models have to comprehend the concept
`of moving means and the correlation between
`multiple sensor signals (a covariance matrix).
`So an absolute requirement of these FDC meth-
`ods is that they be robust against the normal drifts
`in the system for extended time periods. The pri-
`mary reason for this, of course, is that the use of
`any such methodology in a manufacturing opera-
`tion requires
`that
`there be minimal model
`“upkeep” as well as a minimal number of false
`alarms. The “extended time period” is somewhat
`undefined, and will be different for different tools.
`- 5000 wafers processed over a period of - one
`As a guideline, models have to be valid for at least
`
`month, including changes incurred at the periodic
`chamber-clean operations.
`Once a fault is detected, the next task is fault
`clusszjkutiun. This task is performed with models
`f’ and g (which have to be generated for a given
`system). These specify the possible causes of a
`fault (e.g. pressure was too high) and determine
`the effect of the fault on the wafer state (e.g.
`
`etched linewidth is too narrow), respectively.
`Model g is typically called a “Virtual Sensor”, as it
`provides wafer state data from the available sensor
`signals in the absence of actual wafer state meas-
`urements. The previously described issues with
`robustness to sensor and chamber state drifts apply
`to these models as well.
`Components of multivariate FDC are being en-
`abled by a number of software vendors [7-141 with
`capabilities for analyzing machine and sensor data.
`These algorithms still have to be integrated into
`the data acquisition scheme used in a particular
`installation, and this is a significantly complex
`task in itself. But at least the data analysis is fa-
`cilitated by the availability of such commercial
`software.
`
`Model Base Process Control
`For MBPC, there is a similar trend away from
`the control of a single wafer state parameter to-
`wards multivariate, sensor and model-based proc-
`ess control. This evolution requires:
`sensors for wafer state measurements, if possi-
`ble (e.g. Full Wafer Interferometry [15] )
`“virtual sensors” for parameters that can not be
`directly measured
`sensor data acquisition, data transfer and
`communications to a computational algorithm
`have to be seamlessly automated
`models and control algorithms to be used for
`control
`a controller that performs these necessary
`computations
`a feedback loop to the machine, that allows
`newly calculated recipes to be downloaded to
`the machine and executed for the next wafer
`(run-to-run control)
`Figure 2 shows an example of a system archi-
`tecture, generated during the MMST program, that
`can provide these capabilities. This controller, be
`it imbedded in the OEM tool or added in a piggy-
`back fashion, has to have capability to perform the
`multiple major tasks -defined by the boxes in
`Figure 2 - with a user-friendly GUI, otherwise the
`
`367
`
`1996 IEEElSEMl Advanced Semiconductor Manufacturing Conference
`
`Authorized licensed use limited to: Christopher Gallo. Downloaded on June 23,2021 at 03:36:23 UTC from IEEE Xplore. Restrictions apply.
`
`Applied Materials, Inc. Ex. 1018
`Applied v. Ocean, IPR Patent No. 6,836,691
`Page 4 of 6
`
`
`
`It is clear, without going into more detail on
`each component, that APC is a complex system
`requiring interaction from a multiplicity of techni-
`cal disciplines and between the tool vendor, the
`sensor and controller manufacturers and the ulti-
`mate Wafer Fab user. But this task is also bene-
`fiting for the development of commercial software
`for executing the components of MBPC [3,16].
`In summary, multivariate APC is now a wide-
`spread goal that is within the near-term plans for
`the semiconductor industry. However, in contrast
`to the early univariate APC activities at TI, multi-
`variate APC can be conceptually rather complex
`and can be implemented only after a rather signifi-
`cant investment in hardware, software and new
`operating methods. The complexity of this APC
`methodology has been an impediment to its dis-
`semination and widespread use. But with all the
`ongoing activities in semiconductor companies,
`and the interactive programs with OEM, sensor
`and controller suppliers through SEMATECH, it is
`clear that multivariate APC will be implemented
`in mass production within the next 3-5 years.
`References
`
`[ 11 G. G. Barna and C. Ratliff, “Process and Appa-
`ratus for Detecting Aberrations in Production
`Process Operations,” U.S. Patent 4,847,792 issued
`Jul. 11, 1989
`[2] R. L. Wise, “Advanced Process Control for
`CVD of Epitaxial Silicon.” Proceedings of SEMI
`Automation
`Conference,
`SEMI-
`CON/Southwest’86, Ocbber 14- 16,
`Dallas, Texas, pp. 19-30.
`[3] TI Technical Journal, 9(5), Sept-Oct, 1992.
`[4] ProcessWORKS is a memeber of the WORKS
`family of products being commercialized by Texas
`Instruments, Dallas, TX 75243
`[5] G. G. Barna, “Deliverable #1: “Lam 9600 Sen-
`sor Loading for Sensor Based Fault Detection and
`Classification (FDC) and Advanced Process Con-
`trol (APC),” Sematech Technical Transfer Docu-
`ment #: 94082503A-ENGY Sept. 1994.
`
`. M Y
`
`FEEDFORWARD
`
`-------
`
`WAQHos18 cu#l
`MAINTENANCE
`Figure 2. Fundamental components of APC
`
`I
`complexity of the controller will deter its use. The
`first major issue is data acquisition and reduction.
`If process state sensors are monitored for APC, the
`next hurdle is the numerous dimensions to deal
`with, such as: a large number of parallel signals
`(significantly increased if looking at spectral data)
`to be analyzed for within-wafer, wafer-to-wafer,
`within-lot and lot-to-lot variability. The controller
`also has to have univariate and multivariate SPC
`capabilities. Another fundamental requirement for
`APC is that input/output models be available for
`the modeling of the sensor readings to the wafer
`properties or hardware states. Polynomial model
`generation has become routinely available and
`utilized for process characterization, but other
`techniques such as Neural Network modeling are
`also being investigated. In some cases (e.g. spec-
`troscopic measurements of plasma properties), a
`key issue becomes the temporal stability of these
`models, given
`that measurements are made
`through optical windows that continually degrade
`in transmittance as the reactor ages between
`cleaning cycles. The controller also has to have
`optimization capabilities so the available control
`algorithms can act effectively on the process mod-
`els. Finally, the controller has to be able to com-
`municate back to the processing tool, typically
`through a SECSII interface, and the tool has to be
`ready to accept recipe changes generated by this
`controller.
`
`368
`
`1996 IEEElSEMl Advanced Semiconductor Manufacturing Conference
`
`Authorized licensed use limited to: Christopher Gallo. Downloaded on June 23,2021 at 03:36:23 UTC from IEEE Xplore. Restrictions apply.
`
`Applied Materials, Inc. Ex. 1018
`Applied v. Ocean, IPR Patent No. 6,836,691
`Page 5 of 6
`
`
`
`[6] G. Barna, “Deliverable Wt: “Characterization
`of the Hardware, Sensors, Process and Analysis
`Techniques for Sensor Based Fault Detection and
`Classification (FDC) and Advanced Process Con-
`trol (APC),” Sematech Technical Transfer Docu-
`ment #: 95012668A-ENG, Feb. 1995.
`[7] BBN Domain Corporation, Cambridge, MA
`02 140
`[8] Perception Technologies, Albany, CA 94706
`[9] Triant Technologies, Nanaimo, BC, Canada
`V9S 1G5
`
`[ 101 Umetrics, Winchester, MA 01890
`[ 1 11 Brookside Software, San Mateo, CA 94402
`[ 121 Brooks Automation, Richmond, BC, Canada
`V7A 4V4
`[13] Pattern Associates Inc., Evanston, IL 60201
`[14] Real Time Performance, Sunnyvale, CA
`94086
`[15] Low Entropy Systems, Boston, MA 02135
`[ 161 Honeywell, Minneapolis, MN, 554 18
`
`369
`
`1996 IEEElSEMl Advanced Semiconductor Manufacturing Conference
`
`Authorized licensed use limited to: Christopher Gallo. Downloaded on June 23,2021 at 03:36:23 UTC from IEEE Xplore. Restrictions apply.
`
`Applied Materials, Inc. Ex. 1018
`Applied v. Ocean, IPR Patent No. 6,836,691
`Page 6 of 6
`
`