ARTIFICIAL INTELLIGENCE PERGAMON Engineering Applications of Artificial Intelligence Il(1998)619-62 Contributed Paper In-situ control of chemical vapor deposition for fiber coating J G. Jones ,* P.D. jero P.H. garrett Materials and Manufacturing Directorate, Air Force Research Laboratory, Wright-Patterson AFB, OH 45433, US.A Electrical and Computer Engineering and Computer Science Department, University of Cincinnati, Cincinnati, OH 45221, USA Received 1 December 1997; accepted I June 1998 Chemical vapor deposition(CvD)is a widely used process for the application of thin films. In this case, CVd is being used to pply a thin-film interface coating (LaAl1O18) to single-crystal, monofilament sapphire fibers for the production of cerami matrix composites. Traditionally, CVD fiber coating has been hampered by problems such as poor coating uniformity, thickness, and chemistry control. To attempt to rectify these problems, a fully-instrumented CVd processing apparatus has been constructed. In addition to more standard sensors such as thermocouples, pressure transducers, and flow meters, sensors capable of monitoring the coating and gas phase properties of interest have been used. Specifically, a quadrupole mass spectrometer and Raman spectrometer are being used to monitor both the gas phase reactants and the coating quality. A fuzzy-logic controller has been developed to control the O, flow rate, based on in situ temperature measurements. Mass spectrometer data and orresponding improved temperature control for a deposition are presented. Raman spectra which were taken off-line are also presented. C 1998 Published by Elsevier Science Ltd. All rights reserved Keywords: Fuzzy logic; In-situ sensing: Process control; Raman spectroscopy; CVD: Composites; Coatings: Thin films 1. Introduction CVd processes have been undertaken( Besmann et al 1994: Strife and Kee, 1996). Various sensors, models, Most ceramic matrix composites (CMCs) require a and control algorithms have been used, some of which fiber interface coating in order to exhibit the desirable may be relevant to fiber coatings. properties that make them potentially useful high-tem This work seeks to advance the state of the art in perature structural materials (Kerans et al., 1989; CVD fiber coating. The goal is to be able to measure Marshall and Evans, 1985). Although interface coat- coating properties(e.g. chemistry, crystallinity, thick ings are widely recognized as a key to these properties, ness, topography)in real time during the coating depo- fiber-coating technology has received relatively little sition. The steps in this process are (1)automate the attention, hindering successful application of CMCs. CVD fiber coating process via computer control/log sistant interfaces, the inability to deposit coatings of ging of all measurable parameters;(2)implement in situ sensors so as to measure coating or gas phase controlled chemistry, thickness, and morphology is a significant problem(Kerans) properties in real time; and (3) implement closed-loop Chemical vapor deposition(CVD) is a well-known process control based on the sensor data, to produc process, used commercially to deposit coatings and coatings with the desired chemistry, crystallinity, thick bulk materials for a variety of industrial applications ness, and topography A voluminous CVD literature exists (Sedgewick and The coating system chosen for study was a lantha Lydtin, 1979). Control of CVD processes can be very num hexaluminate (LaAl1O18)coating on a sapphire complex, and significant efforts in modeling, in situ fiber. Lanthanum hexaluminate is one member of the sensing, and intelligent process control of various B-alumina/magnetoplumbite family of compounds that has been identified as potentially valuable as interface coatings in CMCs (Cinibulk, 1994: Morgan and Marshall, 1993: Sambasivan, 1996). The coating is a 0952-1976/98/S- see front matter 1998 Published by Elsevier Science Ltd. All rights reserved. PI:S0952-1976(98)00027-X
Contributed Paper In-situ control of chemical vapor deposition for ®ber coating J.G. Jones a, *, P.D. Jeroa , P.H. Garrett b a Materials and Manufacturing Directorate, Air Force Research Laboratory, Wright-Patterson AFB, OH 45433, USA b Electrical and Computer Engineering and Computer Science Department, University of Cincinnati, Cincinnati, OH 45221, USA Received 1 December 1997; accepted 1 June 1998 Abstract Chemical vapor deposition (CVD) is a widely used process for the application of thin ®lms. In this case, CVD is being used to apply a thin-®lm interface coating (LaAl11O18) to single-crystal, mono®lament sapphire ®bers for the production of ceramic matrix composites. Traditionally, CVD ®ber coating has been hampered by problems such as poor coating uniformity, thickness, and chemistry control. To attempt to rectify these problems, a fully-instrumented CVD processing apparatus has been constructed. In addition to more standard sensors such as thermocouples, pressure transducers, and ¯ow meters, sensors capable of monitoring the coating and gas phase properties of interest have been used. Speci®cally, a quadrupole mass spectrometer and Raman spectrometer are being used to monitor both the gas phase reactants and the coating quality. A fuzzy-logic controller has been developed to control the O2 ¯ow rate, based on in situ temperature measurements. Mass spectrometer data and corresponding improved temperature control for a deposition are presented. Raman spectra which were taken o-line are also presented. # 1998 Published by Elsevier Science Ltd. All rights reserved. Keywords: Fuzzy logic; In-situ sensing; Process control; Raman spectroscopy; CVD; Composites; Coatings; Thin ®lms 1. Introduction Most ceramic matrix composites (CMCs) require a ®ber interface coating in order to exhibit the desirable properties that make them potentially useful high-temperature structural materials (Kerans et al., 1989; Marshall and Evans, 1985). Although interface coatings are widely recognized as a key to these properties, ®ber-coating technology has received relatively little attention, hindering successful application of CMCs. Additionally, in today's search for new oxidation-resistant interfaces, the inability to deposit coatings of controlled chemistry, thickness, and morphology is a signi®cant problem (Kerans). Chemical vapor deposition (CVD) is a well-known process, used commercially to deposit coatings and bulk materials for a variety of industrial applications. A voluminous CVD literature exists (Sedgewick and Lydtin, 1979). Control of CVD processes can be very complex, and signi®cant eorts in modeling, in situ sensing, and intelligent process control of various CVD processes have been undertaken (Besmann et al., 1994; Strife and Kee, 1996). Various sensors, models, and control algorithms have been used, some of which may be relevant to ®ber coatings. This work seeks to advance the state of the art in CVD ®ber coating. The goal is to be able to measure coating properties (e.g. chemistry, crystallinity, thickness, topography) in real time during the coating deposition. The steps in this process are (1) automate the CVD ®ber coating process via computer control/logging of all measurable parameters; (2) implement in situ sensors so as to measure coating or gas phase properties in real time; and (3) implement closed-loop process control based on the sensor data, to produce coatings with the desired chemistry, crystallinity, thickness, and topography. The coating system chosen for study was a lanthanum hexaluminate (LaAl11O18) coating on a sapphire ®ber. Lanthanum hexaluminate is one member of the b-alumina/magnetoplumbite family of compounds that has been identi®ed as potentially valuable as interface coatings in CMCs (Cinibulk, 1994; Morgan and Marshall, 1993; Sambasivan, 1996). The coating is a Engineering Applications of Arti®cial Intelligence 11 (1998) 619±626 0952-1976/98/$ - see front matter # 1998 Published by Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 2 - 1 9 7 6 ( 9 8 ) 0 0 0 2 7 -X PERGAMON * Corresponding author
J. Jones et al./ Engineering Applications of Artificial Intelligence 11(1998)619-626 complex(multi-cation) oxide in order to test the sensor exhaust gas to the mass spectrometer is also heated. A and control capabilities. single-zone furnace capable of operation to 1200%C, is in-line ahead of the reactor to desize and clean the fiber. A three-zone furnace capable of operation to 2. CVD system 500C. in-line after the reactor. is used for heat treat ment of the fiber coating A hot-wall CVD system capable of coating fiber in a Gas flows are handled by seven mass-flow control continuous manner has been implemented, shown in lers, including five for Ar and two for O2. Of the Ar Fig. 1. The CVD system uses liquid precursor delivery mass-flow controllers, one maintains a constant inert for the precise control of coating stoichiometry, and environment for the HPLC precursor pump, the sec- operates at near atmospheric pressure. Use of inert gas ond controls the carrier gas flow through the vapori- seals obviates the need for a sealed system with large zer, and the third controls the Ar process gas flow spool boxes and their requisite complications. Separate (used to dilute the o2 process gas). The remaining two furnaces(not shown in Fig. 1)are used for desizing Ar controllers feed the gas seals at the system inlet and r prior to deposition, and for heat treatment of outlet. The size of the seal orifices and the Ar flow are the coating after deposition. A supply spool and adjusted such that the seals are slightly pressurized. Ar motorized take-up spool (also not shown in Fig. 1) from the seals is forced both into the reactor and out complete the system of the system, ensuring that the ambient atmosphere is The lanthanum hexaluminate precursors being excluded from the system. The O2 mass-flow control- employed are lanthanum methoxyethoxide(La(mee)3) lers regulate the flow of O2 to the in-line desize furnace and aluminum sec-butoxide (Al(O-S-Bu)3) in excess and the O2 process gas entering the reactor methoxyethanol (mee). The Al(O-s-Bu)3 is viscous The reactor pressure is maintained slightly (l torr (like honey) as received, and is diluted with mee. The typically) below atmospheric pressure through the use total(La+ Al)solution concentration is adjusted to be of a venturi pump system Pressure control is obtained 0.25 M, as recommended by the liquid delivery sys- by regulating the amount of compressed air allowed tem (LDS) manufacturer. Precursor mixtures with into the pump, which in turn controls the amount of La: Al ratios of 1: 4 and 1: 18 were made. The HPlc effluent which is pulled through the pump. a differen- inlet valves can be controlled so as to allow vaporize tial pressure transducer with a 5 V range for a 10 torr tion of mixtures anywhere between these end points pan is used to precisely measure the difference The hot-walled reactor consists of a 2.54 cm diam between atmospheric pressure and the reactor exhaust eter quartz tube 30 cm in length which can be heated pressure, and an MKs 250B pressure controller regu- to 1200.C. At the front of the reactor a 4-way stainless lates the flow of compressed air through the venturi steel cross connects the reactor to the vaporizer, inlet via a throttle valve. The exhaust pressure can be held seal, and the process gas inlet. The cross, seal, and to within +o l torr of the setpoint process gases are all heated to inhibit condensation of The effluent flows through a burn box heated the vaporized precursor. At the exhaust, a 5-way >500C before exiting into the hood. This ensures heated cross is used to connect the reactor to the out- that any residual hydrocarbons are fully oxidized The let seal, mass spectrometer capillary, reactor exhaust, effluent is diluted with compressed air in the venturi and to provide access for in situ thermocouples. TI pump; thus excess oxygen is available if needed I m long capillary tube which connects the reactor The CVD system was automated using LabVIEW 4.0 from National Instruments. a Power Macintosh Exhaust 8100/100 computer system, an N400 expansion chassis from Second Wave Inc, two acquisition boards from National Instruments, and two four-port serial inter- face cards from Applied Engineering. Building or these fundamental components, the entire CVD system HPLC has been automated, and can be operated entirely from the computer platform. The automated com- Hot Wall Rea ponents include the seven mass-flow controllers and corresponding gate valves, the venturi pressure system, 3 PID temperature controllers, 4 thermocouples, 3 pressure transducers, all LDS system functions, and the fiber take-up spool velocity. In situ sensors also integrated into the computer platform, including a mass spectrometer, laser thickness monitor, Raman Fig. I Hot-wall CVD reactor system. spectrometer, and thermocouples inside the reactor
complex (multi-cation) oxide in order to test the sensor and control capabilities. 2. CVD system A hot-wall CVD system capable of coating ®ber in a continuous manner has been implemented, shown in Fig. 1. The CVD system uses liquid precursor delivery for the precise control of coating stoichiometry, and operates at near atmospheric pressure. Use of inert gas seals obviates the need for a sealed system with large spool boxes and their requisite complications. Separate furnaces (not shown in Fig. 1) are used for desizing the ®ber prior to deposition, and for heat treatment of the coating after deposition. A supply spool and motorized take-up spool (also not shown in Fig. 1) complete the system. The lanthanum hexaluminate precursors being employed are lanthanum methoxyethoxide (La(mee)3) and aluminum sec-butoxide (Al(O-s-Bu)3) in excess methoxyethanol (mee). The Al(O-s-Bu)3 is viscous (like honey) as received, and is diluted with mee. The total (La+ Al) solution concentration is adjusted to be 00.25 M, as recommended by the liquid delivery system (LDS) manufacturer. Precursor mixtures with La:Al ratios of 1:4 and 1:18 were made. The HPLC inlet valves can be controlled so as to allow vaporization of mixtures anywhere between these end points. The hot-walled reactor consists of a 2.54 cm diameter quartz tube 030 cm in length which can be heated to 12008C. At the front of the reactor a 4-way stainless steel cross connects the reactor to the vaporizer, inlet seal, and the process gas inlet. The cross, seal, and process gases are all heated to inhibit condensation of the vaporized precursor. At the exhaust, a 5-way heated cross is used to connect the reactor to the outlet seal, mass spectrometer capillary, reactor exhaust, and to provide access for in situ thermocouples. The 1 m long capillary tube which connects the reactor exhaust gas to the mass spectrometer is also heated. A single-zone furnace capable of operation to 12008C, is in-line ahead of the reactor to desize and clean the ®ber. A three-zone furnace capable of operation to 15008C, in-line after the reactor, is used for heat treatment of the ®ber coating. Gas ¯ows are handled by seven mass-¯ow controllers, including ®ve for Ar and two for O2. Of the Ar mass-¯ow controllers, one maintains a constant inert environment for the HPLC precursor pump, the second controls the carrier gas ¯ow through the vaporizer, and the third controls the Ar process gas ¯ow (used to dilute the O2 process gas). The remaining two Ar controllers feed the gas seals at the system inlet and outlet. The size of the seal ori®ces and the Ar ¯ow are adjusted such that the seals are slightly pressurized. Ar from the seals is forced both into the reactor and out of the system, ensuring that the ambient atmosphere is excluded from the system. The O2 mass-¯ow controllers regulate the ¯ow of O2 to the in-line desize furnace and the O2 process gas entering the reactor. The reactor pressure is maintained slightly (01 torr, typically) below atmospheric pressure through the use of a venturi pump system. Pressure control is obtained by regulating the amount of compressed air allowed into the pump, which in turn controls the amount of euent which is pulled through the pump. A dierential pressure transducer with a 5 V range for a 10 torr span is used to precisely measure the dierence between atmospheric pressure and the reactor exhaust pressure, and an MKS 250B pressure controller regulates the ¯ow of compressed air through the venturi via a throttle valve. The exhaust pressure can be held to within 20.1 torr of the setpoint. The euent ¯ows through a burn box heated to >5008C before exiting into the hood. This ensures that any residual hydrocarbons are fully oxidized. The euent is diluted with compressed air in the venturi pump; thus excess oxygen is available if needed. The CVD system was automated using LabVIEW1 4.0 from National Instruments, a Power Macintosh 8100/100 computer system, an N400 expansion chassis from Second Wave Inc., two acquisition boards from National Instruments, and two four-port serial interface cards from Applied Engineering. Building on these fundamental components, the entire CVD system has been automated, and can be operated entirely from the computer platform. The automated components include the seven mass-¯ow controllers and corresponding gate valves, the venturi pressure system, 13 PID temperature controllers, 4 thermocouples, 3 pressure transducers, all LDS system functions, and the ®ber take-up spool velocity. In situ sensors were also integrated into the computer platform, including a mass spectrometer, laser thickness monitor, Raman Fig. 1. Hot-wall CVD reactor system. spectrometer, and thermocouples inside the reactor. 620 J. Jones et al. / Engineering Applications of Arti®cial Intelligence 11 (1998) 619±626
J. Jones et al./ Engineering Applications of Artificial Intelligence 11(1998)619-626 Associative H-Defuzzificati CVD Process 3. In situ sensors molecular structure. These lines are shifted in fre- quency by an amount, Vy, from that of the exciting In situ measurements of the gas stream are made radiation, vI. This frequency shift is expressed with using a mass spectrometer at the exhaust of the reac- the dimensions of wavenumbers, or inverse centi- tor, and thermocouples inside the reactor tube. (Jones, meters, cm, and is independent of the excitation 1997)In situ measurements of the fiber are accom- plished using a Raman spectrometer. Using the infor Preliminary in-line and off-line experiments to mation recorded from the gas stream and fiber characterize the coating material using Raman spec provides for the opportunity of process discovery. By troscopy have been performed. This analysis has beer discovering discrepancies between what is expected, performed using an EIC fiber-optic Echelle Raman based on the theoretical models, and what actually spectrograph with one wavenumber resolution and f1. 8 occurs,improvements in the fidelity of the models and collection optics. This instrument has a useable range the understanding of the process can be made. Efforts of 250 to 3000 cm-I on the Stokes side of the 532 nm to date have relied primarily on data from the mass spectrometer and thermocouples excitation wavelength. Using this instrument, a col- lected spectrum which has an acceptable signal-to- The exhaust gas from the Cvd reactor is analyzed noise ratio (SNR) typically requires a 60s collection using a Balzers Qms 420 mass spectrometer. It utilizes a non-discriminating gas inlet system which includes a time. Software from Galactic Industries, GRAMS/ 386, is used to process and analyze the collected spec I m long resistively heated 0 150 mm ID capillary tube which connects the reactor exhaust to a flange on the mass spectrometer vacuum column. This configuration A fuzzy-logic(Klir and Yuan, 1995) based control- sults in a two-stage reduction of the pressure from shown in Fig. 2, which adjusts the oxygen flow atmospheric to 10-4 torr, with the partial pressure rate to regulate the in situ temperature, was developed composition of the gas mixture remaining unchanged using Lab vIEw (Jones, 1997). The fuzzy controller The actual analysis is performed at high vacuum, typi- uses two inputs, error and error rate, to generate a cally 10 torr, using a quadrupole mass filter. The change in the output flow rate of oxygen comma as entering the mass spectrometer is ionized by an ion Both the error and error- rate inputs were filtered using source, while the mass filter selects the appropriate a 3rd-order Butterworth low-pass filter with a normal mass ratio, which will be counted by the ion detector. ized cutoff of on=0. 2. Based on the fuzzy variables Integration of the ion-detector signal provides shown in Table 1, triangular membership functions measure of ions detected. based on the ion current were used for the fuzzification of both the error and and hence a relative intensity of each AMU. Spectrum error rate, shown in Fig 3. Defuzzification was accom- analyses can be recorded at intervals as small as N5s plished through the use of a 3 by 5 fuzzy associative Three type-K thermocouples are located at selected memory(FAM)rule-base, defining the output change locations within the reactor tube. Another type-K ther- mocouple resides outside the tube and is used for fur- nace control. Each thermocouple is connected to a Table I nonlinearized signal-conditioning module with a 10 V Fuzzy variables for fuzzy associative memory range for -100 to 1350.C. Each signal is acquired using 12 bit A/D conversion and linearized using a Error Error rate Change in O, flow rate polynomial PL Positive large P Positive PF Positive full Raman spectra arise from the inelastic scattering of PS Positive small PH Positive half his igh-intensity laser radiation incident on a Raman ZE Zero NS Negative small active material. Vibrational modes of a molecule give NL Negative large H Negative half tive full rise to characteristic Raman lines, suggestive of the
3. In situ sensors In situ measurements of the gas stream are made using a mass spectrometer at the exhaust of the reactor, and thermocouples inside the reactor tube. (Jones, 1997) In situ measurements of the ®ber are accomplished using a Raman spectrometer. Using the information recorded from the gas stream and ®ber provides for the opportunity of process discovery. By discovering discrepancies between what is expected, based on the theoretical models, and what actually occurs, improvements in the ®delity of the models and the understanding of the process can be made. Eorts to date have relied primarily on data from the mass spectrometer and thermocouples. The exhaust gas from the CVD reactor is analyzed using a Balzers QMS 420 mass spectrometer. It utilizes a non-discriminating gas inlet system which includes a 1 m long resistively heated 0.150 mm ID capillary tube which connects the reactor exhaust to a ¯ange on the mass spectrometer vacuum column. This con®guration results in a two-stage reduction of the pressure from atmospheric to 10ÿ4 torr, with the partial pressure composition of the gas mixture remaining unchanged. The actual analysis is performed at high vacuum, typically 10ÿ6 torr, using a quadrupole mass ®lter. The gas entering the mass spectrometer is ionized by an ion source, while the mass ®lter selects the appropriate mass ratio, which will be counted by the ion detector. Integration of the ion-detector signal provides a measure of ions detected, based on the ion current, and hence a relative intensity of each AMU. Spectrum analyses can be recorded at intervals as small as 05 s. Three type-K thermocouples are located at selected locations within the reactor tube. Another type-K thermocouple resides outside the tube and is used for furnace control. Each thermocouple is connected to a nonlinearized signal-conditioning module with a 10 V range for ÿ100 to 13508C. Each signal is acquired using 12 bit A/D conversion and linearized using a polynomial. Raman spectra arise from the inelastic scattering of high-intensity laser radiation incident on a Raman active material. Vibrational modes of a molecule give rise to characteristic Raman lines, suggestive of the molecular structure. These lines are shifted in frequency by an amount, nn, from that of the exciting radiation, nI. This frequency shift is expressed with the dimensions of wavenumbers, or inverse centimeters, cmÿ1 , and is independent of the excitation wavelength. Preliminary in-line and o-line experiments to characterize the coating material using Raman spectroscopy have been performed. This analysis has been performed using an EIC ®ber-optic Echelle Raman spectrograph with one wavenumber resolution and f1.8 collection optics. This instrument has a useable range of 250 to 3000 cmÿ1 on the Stokes side of the 532 nm excitation wavelength. Using this instrument, a collected spectrum which has an acceptable signal-tonoise ratio (SNR) typically requires a 60 s collection time. Software from Galactic Industries, GRAMS/ 3861 , is used to process and analyze the collected spectra. A fuzzy-logic (Klir and Yuan, 1995) based controller, shown in Fig. 2, which adjusts the oxygen ¯ow rate to regulate the in situ temperature, was developed using LabVIEW1 (Jones, 1997). The fuzzy controller uses two inputs, error and error rate, to generate a change in the output ¯ow rate of oxygen command. Both the error and error-rate inputs were ®ltered using a 3rd-order Butterworth low-pass ®lter with a normalized cuto of on= 0.2. Based on the fuzzy variables shown in Table 1, triangular membership functions were used for the fuzzi®cation of both the error and error rate, shown in Fig. 3. Defuzzi®cation was accomplished through the use of a 3 by 5 fuzzy associative memory (FAM) rule-base, de®ning the output change Fig. 2. Fuzzy controller block diagram. Table 1 Fuzzy variables for fuzzy associative memory Error Error rate Change in O2 ¯ow rate PL Positive large P Positive PF Positive full PS Positive small Z Zero PH Positive half ZE Zero N Negative ZO Zero NS Negative small NH Negative half NL Negative large NF Negative full J. Jones et al. / Engineering Applications of Arti®cial Intelligence 11 (1998) 619±626 621
J. Jones et al./ Engineering Applications of Artificial Intelligence 11(1998)619-626 NL NS ZE PS nip -6 0 3 6 Error('c -0.30 0.3 Error Rate( C/min) Fuzzy Input Membership Functions. 0.33-0.1500.150.33 Fuzzy Output Membership Functions. Fig 3. Fuzzy membership functions of oxygen in terms of five different fuzzy variables, as Y correspond to the degree of membership of the indicated in Table 2(Kosko, 1992) FAM input fuzzy sets(the antecedents) and Zi corre- which correspond to each FAM rule are combined in output is calculated ae. zy sets(the consequents), the The degree of membership of the input fuzzy sets spond to the output an AND fashion to produce the degree of clipping of each specified output fuzzy set. In fuzzy logic an AND ∑zmin(x,y) corresponds to taking the minimum value of the degree of membership of each corresponding fuzzy RUles membership function. The output membership fund min(Xi, Yi tions were chosen to be triangular, see Fig. 3 The correlation-minimum inference method with centroid defuzzification was used to generate the out ut command(Kosko, 1992). The control 4. Process-control example Fig. 4, shows the output command change over the range of inputs for error and error rate. Using this Figures 5 and 6 show the temperature and mass method, the resulting output value is the center of petra data, respectively, collected over the course of gravity of the clipped output fuzzy sets. Letting Xi and one deposition using the fuzzy control algorithm. The Mass Spectrometer Control Surface 0 Error Rate(C/Minute Fig 4. Fuzzy control surface
of oxygen in terms of ®ve dierent fuzzy variables, as indicated in Table 2 (Kosko, 1992). The degree of membership of the input fuzzy sets which correspond to each FAM rule are combined in an AND fashion to produce the degree of clipping of each speci®ed output fuzzy set. In fuzzy logic an AND corresponds to taking the minimum value of the degree of membership of each corresponding fuzzy membership function. The output membership functions were chosen to be triangular, see Fig. 3. The correlation-minimum inference method with centroid defuzzi®cation was used to generate the output command (Kosko, 1992). The control surface, Fig. 4, shows the output command change over the range of inputs for error and error rate. Using this method, the resulting output value is the center of gravity of the clipped output fuzzy sets. Letting Xi and Yi correspond to the degree of membership of the FAM input fuzzy sets (the antecedents) and Zi correspond to the output fuzzy sets (the consequents), the output is calculated as: O NRules P i1 Zi min Xi; yi NRules P i1 min Xi; Yi : 1 4. Process-control example Figures 5 and 6 show the temperature and mass spectra data, respectively, collected over the course of one deposition using the fuzzy control algorithm. The Fig. 3. Fuzzy membership functions. Fig. 4. Fuzzy control surface. 622 J. Jones et al. / Engineering Applications of Arti®cial Intelligence 11 (1998) 619±626
J. Jones et al./ Engineering Applications of Artificial Intelligence 11(1998)619-626 Table 2 Fuzzy associative memory bank PH fuzzy controller was exercised to cause the in situ tem- the hydrocarbon-rich atmosphere in the reactor results perature to be regulated at 570.C by automatically in significant heat generation through oxidation reac- adjusting the o, flow into the reactor. Adding O to tions (combustion). In thi operational 115 Insitu Temperature 85 oxyger Flow ate Reactor wal Temperature: Time(minutes) Fig. 5. Reactor temperature resulting from deposition with fuzzy controller. Run Number 24. Mass Spectra Regulating Temp. During Deposition. Time(minutes) Fig. 6. Mass spectra resulting from deposition with fuzzy controller
fuzzy controller was exercised to cause the in situ temperature to be regulated at 5708C by automatically adjusting the O2 ¯ow into the reactor. Adding O2 to the hydrocarbon-rich atmosphere in the reactor results in signi®cant heat generation through oxidation reactions (combustion). In this particular operational Table 2 Fuzzy associative memory bank Error rate P PF PH NH NH NF Z PF PH ZO NH NF N PF PH PH NH NF NL NS ZE PS PL Error Fig. 6. Mass spectra resulting from deposition with fuzzy controller. Fig. 5. Reactor temperature resulting from deposition with fuzzy controller. J. Jones et al. / Engineering Applications of Arti®cial Intelligence 11 (1998) 619±626 623
J. Jones et al./ Engineering Applications of Artificial Intelligence 11(1998)619-626 Reactor Temperature and sP from Run 27 Insitu Temperature Reactor Wall Temperature Setpoin Time(minutes) Fig. 7. Reactor temperature during uncontrolled deposition. regime (e.g. reactor temperature, precursor and gas ecies remain essentially constant, with con flow settings, etc. )the in situ temperature is regulated temperature(15-47 min) using O, to within +1. 25.C over 30 min, (15-47 An uncontrolled deposition, seen in Fig. 7, shows min, Fig. 5). The initial response time (7- 15 min that the in situ temperature varied substantially with Fig. 5)of 8 min. could be improved by increasing the (3- 41 min. due to the ebb and flow of number of error and output fuzzy sets, and thus exothermic oxidation reactions(combustion)occurring in the reactor(the reactor atmosphere is essentially a increasing the maximum step size above 0.4 SCCM mixture of alcohol vapors and oxygen in excess Ar) update, shown in Fig 4. Even with the change in reac This despite the fact that the reactor wall temperature tor wall setpoint (540 to 545C)at 20 min, the in situ (the process variable)remained approximately at the emperature was maintained. Fig. 6 shows that the gas setpoint. This deposition was performed under the Run Number 27. Mass Spectra Regulating Temp. During Deposition 8 Time(minut Fig 8. Mass spectra of exhaust during uncontrolled deposition
regime (e.g. reactor temperature, precursor and gas ¯ow settings, etc.) the in situ temperature is regulated using O2 to within 21.258C over 30 min., (015±047 - min., Fig. 5). The initial response time (07±015 min., Fig. 5) of 8 min. could be improved by increasing the number of error and output fuzzy sets, and thus increasing the maximum step size above 0.4 SCCM/ update, shown in Fig. 4. Even with the change in reactor wall setpoint (540 to 5458C) at 20 min., the in situ temperature was maintained. Fig. 6 shows that the gas phase species remain essentially constant, with constant temperature (015±047 min.). An uncontrolled deposition, seen in Fig. 7, shows that the in situ temperature varied substantially with time (03±041 min.) due to the ebb and ¯ow of exothermic oxidation reactions (combustion) occurring in the reactor (the reactor atmosphere is essentially a mixture of alcohol vapors and oxygen in excess Ar). This despite the fact that the reactor wall temperature (the process variable) remained approximately at the setpoint. This deposition was performed under the Fig. 7. Reactor temperature during uncontrolled deposition. Fig. 8. Mass spectra of exhaust during uncontrolled deposition. 624 J. Jones et al. / Engineering Applications of Arti®cial Intelligence 11 (1998) 619±626
J. Jones et al./ Engineering Applications of Artificial Intelligence 11(1998)619-626 4000 3000 1000 Wavenumber(cm) Fig 9. Raman spectra of coated Al]O3 monofilament. same conditions as that utilizing the fuzzy logic con- ify the presence of LaAl1O18, while the peak at oxygen 83 cm m laa 5. and was set to 300 SCCM. Since deposition rate is extre-(from polycrystalline Al2O3)can be used to verify mely sensitive to temperature, it is desirable to keep proper stoichiometry and complete reaction. (The pre the temperature precisely controlled. Fig 8 shows that sence of residual LaAlO3 and/or Al2O3 indicates that there was also a drift in the gas species additional heat treatment is necessary, or that the coat- ng stoichiometry is incorrect.) 5. Raman spectroscopy 6. Summary The Raman spectra of an uncoated monofilament sapphire fiber, not shown, shows three peaks indicative An MOCVD system capable of continuous fibe of ax axis growth: 377, 418, and 645 cm-1 coating has been assembled, in which all of the system (eziorowsl 996). Fig. 9 shows the Raman spectra ra parameters are logged by and controlled by computer of an early coated monofilament. The peaks observed: The system is designed for the coating of continuous 377. 418. 428. 450. 578. 645. and 750 cm-I, corre- fibers for CMCs. The system employs liquid precursor spond to those previously reported for polycrystalline delivery for precise precursor stoichiometry control Al,,(Jeziorowski, 1996). The broad peak at inert gas seals for near-atmospheric operation, a ma 293 cm is an instrument artifact. The implication is spectrometer, thermocouples, and a Raman spec that the La2O3 was not incorporated into the coating, trometer for in-situ measurements. Process-control ex or was volatilized during heat treatment. Additional work needs to be performed to correlate Raman spec- tra of fiber coatings having different stoichiometries. thicknesses. and heat treatments. a polycrystalline bulk sample of LaAl O18(phase pure by XRD) was also analyzed, and is shown in Fig. 10. The spectrum shows a number of peaks, as well as a significant amount of fluorescence. Several of the peaks overlap those of polycrystalline Al2O3, and so cannot be used unambiguously to verify the pre- F zool sence of LaAl1O18 (in a fiber coating). The peak at 483 cm overlaps that of LaAlO3, a compound that forms termediary to LaAlnO18. The peaks at 300400500 7008009001000 3,714,and1015 respect to these materials) to LaAl1O18. The peaks at 553.714 and 1015 cm-l can therefore be used to ver Fig. 10. Raman spectra of lanthanum hexaluminate
same conditions as that utilizing the fuzzy logic controller, with the exception that the oxygen ¯ow rate was set to 300 SCCM. Since deposition rate is extremely sensitive to temperature, it is desirable to keep the temperature precisely controlled. Fig. 8 shows that there was also a drift in the gas species. 5. Raman spectroscopy The Raman spectra of an uncoated mono®lament sapphire ®ber, not shown, shows three peaks indicative of ax axis growth: 377, 418, and 645 cmÿ1 (Jeziorowski, 1996). Fig. 9 shows the Raman spectra of an early coated mono®lament. The peaks observed: 377, 418, 428, 450, 578, 645, and 750 cmÿ1 , correspond to those previously reported for polycrystalline Al2O3 (Jeziorowski, 1996). The broad peak at 293 cmÿ1 is an instrument artifact. The implication is that the La2O3 was not incorporated into the coating, or was volatilized during heat treatment. Additional work needs to be performed to correlate Raman spectra of ®ber coatings having dierent stoichiometries, thicknesses, and heat treatments. A polycrystalline bulk sample of LaAl11O18 (phase pure by XRD) was also analyzed, and is shown in Fig. 10. The spectrum shows a number of peaks, as well as a signi®cant amount of ¯uorescence. Several of the peaks overlap those of polycrystalline Al2O3, and so cannot be used unambiguously to verify the presence of LaAl11O18 (in a ®ber coating). The peak at 483 cmÿ1 overlaps that of LaAlO3, a compound that forms as an intermediary to LaAl11O18. The peaks at 553, 714, and 1015 cmÿ1 appear to be unique (with respect to these materials) to LaAl11O18. The peaks at 553, 714, and 1015 cmÿ1 can therefore be used to verify the presence of LaAl11O18, while the peak at 483 cmÿ1 (from LaAlO3) and 450, 645, and 750 cmÿ1 (from polycrystalline Al2O3) can be used to verify proper stoichiometry and complete reaction. (The presence of residual LaAlO3 and/or Al2O3 indicates that additional heat treatment is necessary, or that the coating stoichiometry is incorrect.) 6. Summary An MOCVD system capable of continuous ®ber coating has been assembled, in which all of the system parameters are logged by and controlled by computer. The system is designed for the coating of continuous ®bers for CMCs. The system employs liquid precursor delivery for precise precursor stoichiometry control, inert gas seals for near-atmospheric operation, a mass spectrometer, thermocouples, and a Raman spectrometer for in-situ measurements. Process-control exFig. 9. Raman spectra of coated Al2O3 mono®lament. Fig. 10. Raman spectra of lanthanum hexaluminate. J. Jones et al. / Engineering Applications of Arti®cial Intelligence 11 (1998) 619±626 625
626 J. Jones et al./ Engineering Applications of Artificial Intelligence 11(1998)619-626 periments employing a fuzzy logic algorithm have been Strife, J.R Kee, R- 1996. Manufacturing science of silicon nitride used to regulate the in situ temperature to within chemical vapor deposition report WL-TR-96-4091. Wright +1. 25.C, and thereby to minimize drift of the gas aboratory, Air Force Materiel Command, WPAFB, OH phase products Authors'Biographies Patrick H. Garrett, PhD, PE, is a professor at the University of Cincinnati and Visiting Scientist at the Materials Directorate rences Wright-Patterson Air Force Base. He has 30 years of industrial and academic experience in instrumentation and control systems for man- Besmann, T M, Lowden, R H, Stinton, D P, McLaughlin, J.C. ufacturing processes. His research includes the development of com- Sheldon, BW.. Starr. T L, Smith, AW. 1994. Pre prehensive new process performance accountability measures. for chemical vapor deposition. report WL-TR-94 Laboratory, Air Force Materiel Command, WP John G. Jones received a Bachelor Science in electrical Cinibulk, M K., 1994. Magnetoplumbite compounds as a fiber coating Engineering From Rose-Hulman Institute of Technology in 1991 in oxide/oxide composites. Cer. Engng. Sci. Proc. 15(5). 721-730 He received a PhD from the Department of Electrical and Computer 996. Fourier transform Raman spectroscopy and Engineering and Computer Science in 1997 from the University of laser flash measurements of alumina single crystals. In: S.A. Asher, P. Stein(Ed ) In: Proceedings of the xvth International has been in the development of a fiber chemical vapor deposition Conference On Raman Spectroscopy, Pittsburgh, PA. Wiley, New process in conjunction with other researchers at the Materials York,pp.827-827 Directorate located at Wright-Patterson AFB, OH. Dr Jones is cur- Jones,JG, 1997. Intelligent process control of fiber chemical vapor rently a member of the process design group at the Materials deposition. Ph. D. dissertation, University of Cincinna Directorate. Dr Jones has been a member of the Institute of erans, R.J., private communication. Electrical and Electronic Engineering for over 10 years, and is cur. Kerans, R., Hay, R.S., Pagano, N.J., Parthasarathy, TA 1989. The role of the fiber-matrix interface in ceramic composites. Bull. A Cer. Commerce and the Riverside Junior Chamber of Commerce Soc.68(2),429442. Klir, G.J., Yuan, B, 1995. Fuzzy Sets And Fuzzy Logic Theory and Paul Jero was born in Washington, Illinois on March 9, 1961.He Applications. Prentice-Hall. Englewood Cliffs. attended the University of llinois and earned a Bachelor of Science Kosko, B. 1992. Fuzzy associative memories. Neural Networks And in Ceramic Engineering(highest honors)in 1983. Continuing at the Fuzzy Systems A Dynamical Systems Approach To Intelligence. Prentice-Hall. Englewood Cliffs, Pp. 314-32Machine University of Illinois, he earned a Master of Science in 1985 and a PhD in 1988, both in Ceramic Engineering. Dr. Jero joined the Marshall, D B, Evans, A.G., 1985. Failure mechanisms in ceramic Materials Directorate, Wright-Patterson AFB. OH, in April 1988 fiber/ceramic matrix composites. J. Am. Cer. Soc. 68(5), 225-231 and became a member of the ceramics group. His work has focused Morgan. P.E. D, Marshall. D.B., 1993. Functional interfaces for characterization of fiber/matrix interfaces in ceramic composites. oxide/oxide composites. Mat Sci. Engng. A162, 15-25 and more recently on development of oxidation resistant interface Sambasivan, S, Moris, J.A., Petusky, W.T., 1996. Rb B-alumina as an coatings for CMCs. His work on micro-tests for measurement of nterface coating in oxide CMCs Cer. Engng. Sci. Proc. 17(4), 250- interfacial properties in ceramic composites has led to establishment f a world class facility within the laboratory, as well as significant Sedgewick, T.O., Lydtin, H. ed, 1979. Proceedin extension of scientific knowledge in the area of interfacial pheno International Conference on Chemical Vapor Deposition, Los ena. Dr. Jero is a member of the American Ceramic Society, and Angeles, CA. The Electrochemical Society, Princeton, NJ. serves as a technical reviewer for the society
periments employing a fuzzy logic algorithm have been used to regulate the in situ temperature to within 21.258C, and thereby to minimize drift of the gas phase products. References Besmann, T.M., Lowden, R.H., Stinton, D.P., McLaughlin, J.C., Sheldon, B.W., Starr, T.L., Smith, A.W., 1994. Processing science for chemical vapor deposition. report WL-TR-94-4044, Wright Laboratory, Air Force Materiel Command, WPAFB, OH.. Cinibulk, M.K., 1994. Magnetoplumbite compounds as a ®ber coating in oxide/oxide composites. Cer. Engng. Sci. Proc. 15 (5), 721±730. Jeziorowski, H.P., 1996. Fourier transform Raman spectroscopy and laser ¯ash measurements of alumina single crystals. In: S.A. Asher, P. Stein (Ed.). In: Proceedings of the XVth International Conference On Raman Spectroscopy, Pittsburgh, PA. Wiley, New York, pp. 827±827. Jones, J.G., 1997. Intelligent process control of ®ber chemical vapor deposition. Ph.D. dissertation, University of Cincinnati. Kerans, R.J., private communication. Kerans, R.J., Hay, R.S., Pagano, N.J., Parthasarathy, T.A., 1989. The role of the ®ber-matrix interface in ceramic composites. Bull. A. Cer. Soc. 68 (2), 429±442. Klir, G.J., Yuan, B., 1995. Fuzzy Sets And Fuzzy Logic Theory and Applications. Prentice±Hall, Englewood Clis. Kosko, B., 1992. Fuzzy associative memories. Neural Networks And Fuzzy Systems A Dynamical Systems Approach To Machine Intelligence. Prentice-Hall, Englewood Clis, pp. 314±322. Marshall, D.B., Evans, A.G., 1985. Failure mechanisms in ceramic ®ber/ceramic matrix composites. J. Am. Cer. Soc. 68 (5), 225±231. Morgan, P.E.D., Marshall, D.B., 1993. Functional interfaces for oxide/oxide composites. Mat. Sci. Engng. A162, 15±25. Sambasivan, S., Moris, J.A., Petusky, W.T., 1996. Rb b-alumina as an interface coating in oxide CMCs. Cer. Engng. Sci. Proc. 17 (4), 250± 257. Sedgewick, T.O., Lydtin, H. ed, 1979. Proceedings of the 7th International Conference on Chemical Vapor Deposition, Los Angeles, CA. The Electrochemical Society, Princeton, NJ. Strife, J.R., Kee, R.J., 1996. Manufacturing science of silicon nitride chemical vapor deposition report WL-TR-96-4091. Wright Laboratory, Air Force Materiel Command, WPAFB, OH. Authors' Biographies Patrick H. Garrett, PhD, PE, is a professor at the University of Cincinnati and Visiting Scientist at the Materials Directorate, Wright-Patterson Air Force Base. He has 30 years of industrial and academic experience in instrumentation and control systems for manufacturing processes. His research includes the development of comprehensive new process performance accountability measures. John G. Jones received a Bachelor of Science in Electrical Engineering From Rose±Hulman Institute of Technology in 1991. He received a PhD from the Department of Electrical and Computer Engineering and Computer Science in 1997 from the University of Cincinnati with specialization in Control Systems. Graduate research has been in the development of a ®ber chemical vapor deposition process in conjunction with other researchers at the Materials Directorate located at Wright-Patterson AFB, OH. Dr Jones is currently a member of the process design group at the Materials Directorate. Dr Jones has been a member of the Institute of Electrical and Electronic Engineering for over 10 years, and is currently a member of both the Cincinnati Junior Chamber of Commerce and the Riverside Junior Chamber of Commerce. Paul Jero was born in Washington, Illinois on March 9, 1961. He attended the University of Illinois and earned a Bachelor of Science in Ceramic Engineering (highest honors) in 1983. Continuing at the University of Illinois, he earned a Master of Science in 1985 and a PhD in 1988, both in Ceramic Engineering. Dr. Jero joined the Materials Directorate, Wright-Patterson AFB, OH, in April 1988 and became a member of the ceramics group. His work has focused on characterization of ®ber/matrix interfaces in ceramic composites, and more recently on development of oxidation resistant interface coatings for CMC's. His work on micro-tests for measurement of interfacial properties in ceramic composites has led to establishment of a world class facility within the laboratory, as well as signi®cant extension of scienti®c knowledge in the area of interfacial phenomena. Dr. Jero is a member of the American Ceramic Society, and serves as a technical reviewer for the society. 626 J. Jones et al. / Engineering Applications of Arti®cial Intelligence 11 (1998) 619±626
