1138 Volume 55, Number 9, 2001 APPLIED SPECTROSCOPY0003-7028 / 01 / 5509-1138$2.00 / 0q 2001 Society for Applied Spectroscopy

On-Line Analysis Using Raman Spectroscopy for ProcessControl during the Manufacture of Titanium Dioxide

IAN M. CLEGG,* NEIL J. EVERALL, BERT KING, HUGH MELVIN, andCOLIN NORTONABB Eutech, Ltd., Daresbury Park, Warrington, Cheshire WA4 4BT, U.K. (I.M.C., C.N.); Retired, formerly of ICI Runcorn (B.K.);ICI PLC, Measurement Science Group, P.O. Box 90, Wilton, Middlesbrough, Cleveland TS90 8JE, U.K. (N.J.E.); and HuntsmanTioxide, Tioxide Europe Ltd., Haverton Hill Rd., Billingham, Stockton-on-Tees TS23 1PS, U.K. (H.M.)

A system based on Raman spectroscopy has been developed forclosed-loop control during the manufacture of titanium dioxide(TiO2). Spectra were obtained in the 50 to 700 cm21 range usingexcitation at ;532 nm and a � xed transmission-grating spectro-graph coupled to a CCD detector. Quantitation has been achievedby linear interpolation based on the integrated intensity of phononsat 142 and 610 cm21 due to anatase and rutile phases, respectively.The spectroscopic instrumentation has been integrated with a pow-der sampling system and installed on four full-scale manufacturingplants. Fiber coupling (up to 100 m) has been employed, and eachspectrometer system is multiplexed to three sampling heads. Spectraare corrected for changes in sample temperature, and the systemsare designed to be fully automatic and require no user intervention.

Index Headings: On-line; Raman spectroscopy; Process control; Ti-tanium dioxide.

INTRODUCTION

Raman spectroscopy is rapidly emerging as a leadingtechnique for on-line analysis and measurement. Thetechnique is very versatile and can be used to produceinformation, for example, on chemical composition andphysical form in a wide range of media. This versatilityensures that the technique is applicable to a broad rangeof industrial sectors, and it can be employed in manydifferent situations. In addition, the technique can be ap-plied in a very practical manner (see Table I) and manyof the problems traditionally associated with on-line an-alyzers can be avoided or their severity reduced.

A number of industrially based examples have beenreported regarding polymerization monitoring,1,2 polymer� lm production,3,4 distillation,5,6 reaction monitoring,7 andpharmaceuticals.8 A useful general review of process Ra-man applications appeared in 1997,9 and activity in thearea is also covered in the biannual reviews which appearin Analytical Chemistry. However, it is usually impossi-ble to gauge the actual industrial relevance of such ap-plications and the business value they can possess. Endusers are understandably reticent about revealing detailsof installations, their business impact as well as their owndegree of commitment to this technology.

Among practitioners in this � eld, end users and man-ufacturers alike, there is a � rm conviction that the tech-nique is currently being widely evaluated and employedand that in the future it will have a pervasive in� uenceindeed. Unfortunately, the overall perception of the tech-nique is hampered by the lack of a signi� cant number of

Received 9 March 2001; accepted 18 May 2001.* Author to whom correspondence should be sent.

real-world examples that have been detailed in the openliterature. The direct consequence of this is that there isa natural reluctance to accept the current value and in� u-ence of the technique, especially among those who are alittle removed from the details of the technology. Thishas probably limited the degree to which this technologyhas been embraced to date. Although this position can bepartially alleviated by production of further applicationnotes from instrument manufacturers and by research pa-pers, it is unlikely that there will be signi� cant changeunless more real-world applications are revealed andthoroughly described.

Here we report in some detail on the development andapplication of Raman spectroscopy to on-line control dur-ing the manufacture of titanium dioxide (TiO2). Thiswould appear to be one of very few de� nitively reportedexamples of on-line Raman spectroscopy being used forclosed-loop process control, i.e., where the output of theanalyzer directly controls the operating conditions withinthe manufacturing process.

Adaptation of the basic measurement technology tothis particular task could be used to illustrate the way inwhich such technologies can be developed and deployed,but this work is also signi� cant in a number of otherimportant respects: (1) The sample is a powder that isdif� cult to handle; it is hot, heterogeneous, and abrasive.(2) The development began in 1993, and control loopsbegan to be closed in 1994. Since then: (3) Four instal-lations have been made worldwide and each system ismultiplexed 3-way. (4) The systems have proven to bereliable, and the accrued bene� ts are considerable, bothin terms of product quality and commercial rewards. Apatent application has been � led.10

PROCESS DESCRIPTION

Titanium dioxide is a ubiquitous commodity pigmentthat combines good durability with high opacity and hasa pure white color. These features ensure that this mate-rial is incorporated into a wide range of products includ-ing paints, plastics, ceramics, paper, textiles, food, andpharmaceuticals. Worldwide annual capacity is over 3million metric tons. There are two main synthetic routesto TiO2 (known as sulfate and chloride), and this paperrefers to the sulfate route, which begins with either blastfurnace slag (a mixture of iron and titanium oxides) orIlmenite (FeTiO3, iron titanium oxide). The manufactur-ing process consists of a large number of processing stepsincluding digestion with sulfuric acid, precipitation, andcalcination; these are carried out in a range of processing

APPLIED SPECTROSCOPY 1139

TABLE I. Valuable practical aspects of on-line Raman spectros-copy.

The interface to the process can be easily accomplished. In normalpractice it is a non-contacting backscattering technique, i.e., access isonly needed to one side of a sample. Therefore, measurement probescan be simply inserted into process lines or vessels. Conversely, a probecan be inserted in a sampling loop or can be made available in ananalysis station located near to the source of samples.

Many sample types can be analyzed. Solids, liquids, and gases canbe analyzed directly as they are produced so that little or no preparationis required. Mixed phase systems such as emulsions, suspensions, dis-persions, and slurries can also be analyzed directly.

There is minimal intrusion to the process. Glass is a weak Ramanscatterer, so we can easily obtain spectra through windows or sightglasses or even through the wall of glass vessels such as reaction � asks.

The spectrometer can be easily protected. The signals can be readilyconducted through standard optical � bers so the spectrometer can behoused some distance from the process where it can be protected in aclosely controlled environment.

Hazardous samples can be readily sampled. The ability to ‘‘look’’through a window means that it is environmentally robust. Samplesunder vacuum, at high pressure, at low or high temperatures, or inchemically hazardous environments pose no insurmountable problems.

FIG. 1. Raman spectra of pure anatase and rutile pigments. The spectral differences are due to the differing lattice structures, hence, modes ofvibration, of the two phases. Note that even in pure rutile there is a weak band coincident with the main anatase peak position of 142 cm21, thisis NOT due to anatase contamination, but rather to a vibrational mode of the rutile lattice itself.

vessels and pieces of equipment which are operated ineither batch or continuous modes. There are numerouspossible variations in the design and operation of thesesteps, and there are therefore many subtleties and con-comitant complexities. It is, however, widely understood

that a key processing step occurs in the calcination stage,where the following sequence of reactions takes place:

TiO ·H SO ·H O ® TiO ·H SO ® TiO (nanophase)2 2 4 2 2 2 4 2

® TiO (anatase) ® TiO (rutile)2 2

This process step is carried out in a calcination kiln, arotary furnace that is similar to a cement kiln in physicalscale and general layout. It consists of a horizontal steelcylinder (with refractory lining), typically some 5 m indiameter and 70 m long. Overall, this sequence of reac-tions is endothermic and the kiln is heated using a largegas jet located at one end. The kiln, which is mounted ata slight incline, is rotated slowly so that feed materialentering the top of the kiln slowly travels down to thelower end of the kiln, going through a number of chem-ical reactions along the way, i.e., drying, acid removal,and phase transition. Manufacturing sites may operateseveral kilns, each with a capacity of up to 100 metrictons of product per day. Near the end of the kiln, theTiO2 reaches a temperature of up to 1200 8C; it then fallsinto a cooling vessel before being discharged into otherequipment for further processing. At this point in the pro-cess, the product, which is known as calciner discharge(CD), is a binary mixture of the rutile and anatase crystalforms. For most applications, including that describedhere, the CD must contain only a small fraction of theanatase phase (although some applications, notably tex-tiles and food products, require that the softer anatase

1140 Volume 55, Number 9, 2001

FIG. 2. Raman spectra showing the effect of increasing anatase level in an anatase/rutile mixture. The annotated number indicates the % anatasepresent. Note that the anatase band at 142 cm21 is relatively strong compared with the rutile bands, making the method very sensitive to detectinglow levels of anatase.

phase dominate the pigment). Therefore, an important ob-jective is to operate the calciner while maintaining theanatase at a low value (typically 1–1.5%) since this en-sures that material of consistently high quality is pro-duced. If the anatase level is too high, the product willbe soft and have poor optical properties, but it must notbe too low because this is likely to lead to a sinteredproduct that is dif� cult to process.

Automatic control of the kiln is dif� cult because thequality limits are tight, there are very few controlled pro-cess variables (e.g., the gas burner conditions), multiplechemical reactions take place, and there is a very longprocess time lag (;12 h) associated with the process. Thedesign of an effective and robust mathematical model tocontrol the calciner proved to be a considerable chal-lenge. The details of the model cannot be discussed herefor commercial reasons. However, a critical input param-eter for the model is the concentration of anatase in theproduct. The traditional approach to this problem wasregular, but infrequent, off-line analysis of product usingX-ray diffraction (XRD). Typically, samples were takenfrom the product cooler, transported to the plant labora-tory, pressed, and then analyzed using a conventionalXRD instrument with the results being manually enteredinto the process control computer. This analytical methodis slow but has proven to be effective and has providedthe basis of control during manufacturing over manyyears. However, there has been a succession of projects

attempting to raise the production rate per calciner as thistends to minimize cost per unit of production. Suchchanges made control of the process progressively moredif� cult and then impossible without rapid, high-qualitymeasurement of the composition of the product, and thiswas the driving force for the innovations described here.Development of an XRD technique to produce goodquality data at a much greater frequency than is currentlyavailable was considered to be impossible, and the onlyviable option was to develop a system based on Ramanspectroscopy.

EXPERIMENTAL

Instrumentation. All Raman spectra were measuredusing ‘‘Holoprobe 532’’ instruments from Kaiser OpticalSystems, Inc. (KOSI) (Ann Arbor, MI). They were con-� gured in a number of different ways according to localproduction requirements, but a typical description fol-lows. The spectrometers were essentially standard unitsequipped with a 100 mW, frequency doubled diode-pumped Nd:YAG laser source (Coherent Inc.) operatingat ;532 nm and matched to a four-way proprietary beamsplitter (KOSI) mounted within the Holoprobe chassis.All beams, which were balanced so as to be of approxi-mately equal power, were connected to standard Mk2probe head units (KOSI) mounted in the � eld (see below)via optical cables (of variable length, up to 100 meters).

APPLIED SPECTROSCOPY 1141

FIG. 3. Correlation of XRD values for anatase/rutile ratio versus the ratio of the integrated intensity of the Raman bands at 142 and 610 cm21. Agood linear correlation was obtained, the slope and intercept of which form the basis of the Raman calibration for % rutile determination.

The Mk2 probeheads are pseudo-confocal coaxial imag-ing probes. Typically, the � ber cables consisted of highquality silica � bers, with a 50 mm core for laser deliveryand 60 mm core for the return, clad and heavily armoredfor installation in the � eld. The return � bers were con-nected to a 4-way injection unit (KOSI) located at theinlet to the spectrograph, enabling simultaneous monitor-ing of up to four separate measurement channels by im-aging the � ber outputs onto different tracks of the CCDdetector. Spectral dispersion was achieved with proprie-tary holographic transmission gratings (KOSI). These(static) gratings allow the region from 21000 to 11460cm21 to be acquired from four channels simultaneouslywith no moving parts. In the � eld, the � ber probes were� tted with a variety of objectives to focus the laser ontothe sample. The working distance and f/number dependedon where they were employed, although most work wascarried out using an ; f /3, ;75 mm working distancelens (supplied by KOSI as the standard � t to the Mk2probe).

Typically, the laser power illuminating the sample, asmeasured at the focal point of the probe, was between 5and 10 mW. Acquisition conditions were normally set to10 averaged accumulations of 30 s each. Spectra can infact be acquired much more rapidly, but this regimematched the optimum data feed rate for the process con-trol model. The spectrometers were always calibrated forRaman shift using a neon emission source. No correctionfor broad band intensity response (� at � elding) wasmade.

Instrument control and spectral acquisition were

achieved using the Hologramsy and DataSentryy soft-ware packages (KOSI). Spectra were exported to Gramsversion 3.1 or 32 (Galactic Industries, Salem, MA) formanipulation, data extraction, communication with theplant control computer, and display.

Calibration Standards. Raman spectroscopy is aquantitative technique but is not, per se, an absolute tech-nique, hence the need to calibrate the spectra using sam-ples of known composition. Suitable reference materialswere produced by the Tioxide R&T group using knownmaterials containing ;0% and ;100% rutile, which weremixed together in various proportions to manufacturestandards in the range of 80–100% rutile.

Spectroscopic Aspects. The Raman spectra of anataseand rutile are well known and have been fully assigned.11

The measurement described here is based upon the factthat the two crystal forms have grossly different vibra-tional spectra in terms of band frequencies and intensities,and therefore, the composition of mixed-phase powderscan be readily assessed from their Raman spectra.12 Thisis illustrated by Fig. 1, which compares the Raman spec-tra of anatase and rutile. The differences are obvious,with the differing bands corresponding to speci� c pho-nons (vibrations of the crystal lattice).

Figure 2 compares the Raman spectra of samples ofcalciner discharge containing ;1%, 5%, and 10% ana-tase; the growth of the anatase band at 142 cm21 relativeto the rutile bands is dominant, and its intensity providesthe basis for quanti� cation using Raman spectroscopy. Itis also apparent that the anatase is readily detected evenat low percentage levels; this is because, per unit weight,

1142 Volume 55, Number 9, 2001

the peak intensity of the anatase band is intrinsically moreintense (;20-fold) than any of the rutile bands. Thus,Raman spectroscopy is inherently suited to the detectionof low levels of anatase in rutile. (Coincidentally, thereis also a very small band at 142 cm21 in the spectrum ofpure rutile, but the intrinsic intensity is too weak to ob-scure signi� cant levels of anatase. It only serves to intro-duce an offset into the calibration regression).

In order to quantify composition of TiO2 from Ramandata, it is � rst necessary to calibrate the spectra usingsamples of known composition. The Raman band inten-sities are related to the material composition by Eq. 1.

%A I (142)5 K 1 C (1)1 1[ ]%R I (610)

where I (142) and I (610) refer to the intensities of theanatase and rutile bands located at 142 and 610 cm21,respectively, and %A and %R are the anatase and rutileconcentrations in wt %. The 142 cm21 band was chosenbecause it is the most sensitive to anatase, whereas eitherof the main rutile bands could have been used for refer-ence purposes. The intensity of the anatase band relativeto the rutile band is, to a � rst approximation, linearlyrelated to the mass ratio of anatase to rutile in the sampleaccording to Eq. 1.

The values of K1 and C1 in Eq. 1 depend on severalfactors. These include the relative Raman scattering ef-� ciencies of the two phases, the sample temperature, theinstrumental response as a function of wavelength, thepresence of dopants, which can be incorporated into thecrystal structure, and possibly the crystal/particle size.For these reasons, it is important to determine K1 and C1

for a given instrument con� guration using standards ofknown composition. It is essential to use standards thatare as similar in chemistry and morphology as possibleto the actual material to be analyzed on-line.

Figure 3 shows a plot of the XRD data correlated withthe Raman data for a set of reference materials. It is notedthat the y-intercept (C 1 in Eq. 1) is actually negative ow-ing to the overlapping rutile band (Fig. 1). Therefore, themeasured intensity ratio never reaches zero even at 0%anatase level.

For this calibration, a simple binary mixture is as-sumed, i.e., %A 1 %R 5 100. Therefore, if we measurethe anatase/rutile ratio using Eq. 1, we can infer the total% rutile in the mixture.

Fundamental Precision. Equation 1 was used to pre-dict the % R levels obtained from 22 sequential Ramanmeasurements of a static sample in order to determinethe fundamental precision of the method. The sample wasknown to be ;95% rutile, based on XRD measurements.The % R as predicted from Raman measurements was95.01, with standard deviation of 0.27%. Hence, the stan-dard error on the mean was 0.057, giving an uncertaintyin the mean of 6 0.11% at the 95% con� dence level.This is an excellent level of precision. The level deter-mined by XRD (10 measurements) was 94.89 6 0.4%,illustrating both the good agreement between Raman andXRD results and the fact that Raman is actually moreprecise than the primary standard in this case.

The Effect of Temperature. It should be noted thatin this case, using an off-line calibration with on-line Ra-

man data initially gave an underestimate of about 1%rutile compared to the XRD value. This was because theoff-line model used room temperature samples, whereasthe in situ measurements were made at 100–200 oC. Un-der the harmonic approximation, the intensity of a Ramanband increases exponentially with temperature as givenby Eq. 2.

AI 5 (2)

(1 2 exp(hn /kT ))

Here, A is a constant, h, k, and T have their usual mean-ing, and n is the wavenumber shift of the Raman band.Thus the intensity increase of low wavenumber Ramanbands will be proportionally higher than that of higherwavenumber bands, and this accounts for the initial over-estimate of anatase content from on-line measurements.The temperature of the TiO2 near to the focal point ofthe probe was measured and fed to the control macro,and, by using Eq. 2, the results were corrected and theeffects of temperature eliminated. It would, of course,also be possible to measure temperature directly using astokes–antistokes band ratio, but direct measurement wasfelt to be simpler and more reliable.

Possible Effect of Crystal Size and Shape. It is wellknown that in certain circumstances the Raman spectraof crystals are sensitive to changes in their size andshape. Absolute scattering intensity, relative intensities,and band shapes within a spectrum can all be affected.These effects can arise for several reasons, which havebeen summarized elsewhere.13 One of the most importanteffects is termed ‘‘phonon con� nement,’’ which relatesto the fact that with very small crystals, phonons areinherently localized to a small region, the uncertainty inmomentum becomes relatively large, and the Raman ef-fect can therefore excite fundamental phonons with non-zero momentum. Thus, we can sample regions of thephonon dispersion curve away from the zero wavevectorlimit (i.e., away from the center of the Brillouin zone).If the phonon dispersion curve is not � at, this will intro-duce shifts in band positions and shapes. If the shape ofthe phonon dispersion curve is known, we can calculatethe effect of phonon con� nement on the Raman spec-trum.14 As an example, Kelly et al. have shown how theshape of the anatase Eg band near 140 cm21 is sensitiveto crystal size and can be modeled by the phonon con-� nement model.15

However, it is important to realize that crystal size andshape only strongly in� uence Raman spectral bandshapesfor very small crystals (typically nanophase material).Under these conditions a reasonable fraction of atoms ormolecules (;10% or more for a 10 nm diameter crystal)are located at the crystal surface and hence in a differentenvironment, and the phonons are con� ned to a suf� -ciently small region for the momentum uncertainty to be-come important. For example, in the work of Kelly etal.,15 the anatase crystal size varied between 5 and 25 nmand caused signi� cant shifts (;10 cm21) and broadening(;100%) of the Eg band. In our own work13 we haveobserved signi� cant changes in the relative intensitiesand widths of the rutile bands for crystal sizes between5 and 50 nm. We also observed a new band near 110cm21, presumed to be due to the B1u mode, which is in-

APPLIED SPECTROSCOPY 1143

FIG. 4. Photograph of an early demonstrator installation with the probe suspended over a screw conveyor on a plant. Note the laser spot on thepowder (in the circle); the projection lens is shrouded in a cylindrical collar at the top of the photograph. For the sake of clarity, the cover used toshield the system from ambient light has been removed. The trough holding the powder and the conveying screw was approximately 40 cm wide.

active for perfect large crystals. Similar effects are ob-served for non-stoichiometric crystals, which contain lat-tice defects that reduce the translational symmetry.16

Fortunately, the above phenomena do not complicateour analysis of the composition of calciner discharge, asunder normal conditions the CD has been calcined for upto 12 hours at 1000 8C. Under these conditions the ana-tase and rutile crystals have grown to a size that is typicalfor this process (between 100 and 200 nm depending onthe grade of material being produced), and these crystalsare of an order of magnitude too large to convey strongsize effects to the Raman spectrum. We checked this bycarrying out a detailed study of the effect of crystal sizefor rutile and anatase in the 100–200 nm size range; nosigni� cant effects were detected compared with those dueto compositional or temperature changes.17 Furthermore,we sought to minimise the effects of band shape changesby measuring integrated band intensities rather than peakheights.

We have therefore developed an analytical model thatis able to cope with the changes in temperature of theprocess sample that can occur during normal operation,and we have also veri� ed that the results are not in� u-enced by changes in crystal size.

DEVELOPMENT AND APPLICATION

Sequence. If the development of a new on-line ana-lyzer system is to be fruitful, cost-effective, and timely,then a logical and methodical sequence of events mustbe followed. This is especially true in situations such asthe one described here, i.e., where the technique is rela-tively new to the � eld of on-line analysis, the system isto be used for closed-loop process control, and the en-vironmental conditions are fairly harsh. Some simpleguidelines in this area have recently been suggested.18 Inthe case reported here, a substantial family of laboratory-based, and then plant-based, proving trials were carriedout. The sequence was logical and methodical; the tech-nology was tested under conditions that were progres-sively more demanding and representative of those ex-isting in a full scale process plant. Photographs from anearly on-line trial and a � nal installation are shown inFigs. 4 and 5, respectively. The initial trial was particu-larly simple, using a long working-distance (;40 cm)objective to collect Raman scatter from powder movingalong a large screw conveyor (Fig. 4). Despite the factthat the powder was continually moving in and out ofthe focus (vertical motion was several inches), good Ra-

1144 Volume 55, Number 9, 2001

FIG. 5. A photograph of a � nal installation with the sampler located in the end of a large production vessel. The black rectangular ‘can’ is identicalto that shown sectioned in Fig. 6. The controller for the sampler is the stainless steel box (with red push button) on the left-hand side. CD isreturned to the process via the vertical stainless steel tube just to the left of the ‘can’ containing the probe.

man data and a reliable correlation with off-line XRDresults were still obtained. This data was suf� cient tojustify moving toward a robustly engineered, permanentinstallation (see below).

Process Interface and Implementation Details. Theinterface to the process is worthy of explanation since themajor weakness from which the Raman technique suffersis the requirement to gain, and maintain, good opticalaccess to the process sample. The sample analyzed mustbe presented reliably to the Raman probe and be in focus;it must also be representative of the composition of thewhole process � ow.

It was decided to sample the product using a smallscrew conveyor, the feed end of which intercepted ma-terial falling from the cooler. At this point, the CD istypically at a temperature of 150 to 200 8C. The conveyor,and its associated control unit, was based on a proprietarypackaged unit (Axiom plc, Ashton in Maker� eld, U.K.),heavily adapted for this application. An optical window(silica, 2 mm thick, 10 mm in diameter) was positionednear the end of the conveyor and was immersed in thepowder � owing through. The Raman probe was posi-tioned outside this window, and the system was adjustedso that the focal point of the probe was located just onthe interface of the CD and the window. Figure 6 shows

a sectioned, isometric projection of the equipment asso-ciated with the probe and also shows the physical rela-tionship with the conveyor. Careful optimization of bothscrew pitch and clearance from the window was neces-sary to optimize sample � ow over the window to preventfouling or damage.

Manual Measurement Stage. The ability to multiplexthe measurement is useful in a number of respects but italso allows samples to be processed manually. Here, aspectrometer measurement channel was fed to a built-for-purpose sampling accessory designed to accept samplesdrawn from other parts of the process or to be used inthe event of failure of one of the continuous samples.This system was designed to cope with the environmentalconditions adjacent to the process and is completely au-tomatic in operation.

Data Analysis and System Control. As shown above,analysis of the spectroscopic data and calculation of re-sults is straightforward. When the system is used on-line,this is achieved by setting Array Basicy macros (Galac-tic) to auto-run immediately after GRAMS has acquiredspectra. The macro integrates the areas of the two peaks,corrects for the effect of temperature, and calculates thecrystalline mass fraction using pre-determined calibrationfactors in Eq. 1 by employing a linear model. The macro

APPLIED SPECTROSCOPY 1145

FIG. 6. Raman measurement probe attached to a screw conveyor. For the sake of clarity, the device is shown sectioned and has been separatedbetween the shutter and insulating spacer.

FIG. 7. This � gure shows the variation in measured rutile concentration as a function of time over a period of 15 h. Raman data is shown as acontinuous line; this was obtained on a 6 min cycle. XRD data is shown as discrete points (squares) and was obtained approximately every hour.

1146 Volume 55, Number 9, 2001

FIG. 8. This � gure shows the variation in signal area counts for the rutile peak as a function of time for four separate window clean cycles (A toD ) with each one covering a period of 4 h before the clean and 6 h afterwards. The cleaning process causes a zero area count (vertical spike)because the laser safety shutter needs to be closed. Note the large changes in peak area that occur before and after a cleaning cycle and thecharacteristic ‘recovery’ shape.

also converts the data to analogue output variables forautomatic transmission to the plant control computer,stores the output data, and also controls the manual stage.

Sample Heterogeneity. A concern with the use of Ra-man spectroscopy for analysis of heterogeneous solidssuch as TiO2 is the 50 mm laser spot size, which is com-parable to the size of the crystal agglomerates in theproduct. In the extreme case, and especially with shortspectral acquisition times, as product passes through thefocal volume of the probe there can be massive variationsin the relative proportion of scattering bands associatedwith the two crystalline phases. However, in normal op-eration, with a 6 min acquisition cycle, the coaddition ofsuccessive spectra averages out such variations, giving aresult that we have found to be robust in comparison withthe XRD data. With a standard spectral acquisition timeand the normal CD � ow through the sampler (;4 kg h21),the total area of sample illuminated by the laser is greaterthan the area of the sample disk analyzed by XRD.Therefore, the two techniques are similarly representativeand results from them can be readily compared.

A second concern relating to the use of Raman spec-troscopy with opaque solids is the fact that the depth ofpenetration into the material is small. The obvious con-sequence of this is that if product adheres to the window,then the analytical results become invalid because themeasured product is not representative of the bulk ma-terial. Attempts were made to solve this by calculating

differences between successive spectra to track small var-iations in the composition of the product; it was reasonedthat it would be a simple matter to differentiate betweenthis condition and a highly stable reading due to a con-stant sample adhering to the window. In practice, thisproved to be a very fragile diagnostic, especially in pe-riods of very good process control when small � uctua-tions are signi� cantly reduced in magnitude. This prob-lem was actually solved by modifying the sample systemso that the sampler was automatically set to reverse andempty at regular intervals, at which point the windowwas cleaned with a blast of air. The macro controllingthe whole system was modi� ed to incorporate a diagnos-tic check that detected differences in spectra obtained be-fore and after each air blast.

Spectrograph Calibration. Wavelength calibrationwas easily achieved by illuminating the entrance apertureof the spectrograph with a neon or argon line source. TheHolograms software automatically detects the position ofseveral discrete lines of known wavelength and uses theseto calibrate the wavelength of each pixel in the CCDtrack. Provided one knows the laser wavelength, the Ra-man shift can be readily determined for each pixel. Thisprocedure is fully automated, and once a calibration spec-trum has been acquired, the software can ‘‘self-calibrate’’with no manual intervention (such as manual peak pick-ing). In practice, because the spectrometer is temperaturecontrolled with air-conditioning and there are no moving

APPLIED SPECTROSCOPY 1147

FIG. 9. Variation in measured rutile content for 10 h periods around window cleaning cycles. Note that there is a negligible change in the measuredrutile content despite a near doubling of the rutile count area during this period. This data was obtained from cycle A shown in Fig. 8.

parts (aside from a camera shutter), the calibration sta-bility is excellent and the system rarely requires recali-bration (an annual adjustment is typical). Because we aremeasuring broad Raman bands, slight shifts in eitherspectrograph calibration or laser wavelength are not crit-ical. Calibration would potentially be more of an issue ifa multivariate model were necessary.

Intensity calibration is a different matter. The absoluteresponse of a spectrometer is a complex function of sev-eral variables (detector temperature, optical throughput,� lter ef� ciency, laser coupling ef� ciency, etc.). However,use of a band ratio for calibration means that the absoluteresponse is not important provided that any changes oc-cur equally at all wavelengths in the spectrum. Unfortu-nately, there is potential variability in the response of theinstrument as a function of Raman shift, i.e., the shapeof the broadband intensity/frequency response curve canchange. Such a change could arise due to instrumentaleffects such as the throughput of the � lters. For example,if an angle-tunable transmission � lter is moved, its trans-mission pro� le will change, and this could invalidate acalibration based on band ratios.

Calibration of the intensity response of dispersive orFT-Raman systems is a complex area that is beyond thescope of this paper; the reader is referred elsewhere fordetails.19–21 Ideally, one should normalize the frequencyresponse of each instrument by measuring the spectrumof a stable, traceable, broadband source of known emis-sion pro� le; then only one calibration curve would, in

principle, apply to all installations. To do this rigorously,one would need to illuminate the entire optical ‘‘returnchannel’’ with the source. Since the probehead is actuallymounted directly onto the process, this would necessitateeither removing the probehead for calibration or relayingthe white light into the probehead to illuminate the returnpath. The latter approach was attempted several times butwas ultimately abandoned; it was concluded that this pro-cedure added both cost and delay but was not necessaryfor this application. Instead, a pragmatic approach wasadopted which utilizes the fact that it is often necessaryto calibrate any analysis system installed on a plant usingreference materials. In other words, systems must beshown to function properly by providing the ‘‘correct’’result from known primary standards. To do this, a built-for-purpose calibration stage was installed at the mea-surement probe allowing reference samples to be placedat the focal point of the probe. The whole system couldtherefore be calibrated in situ. In this way, the require-ment of characterizing the optical response of the instru-ment was avoided.

RESULTS AND DISCUSSION

Figure 7 shows a typical plot of data on rutile contentobtained by both Raman and XRD methods over a periodof 15 h. Raman data was generated automatically at 6min intervals, and the XRD data was generated from grabsamples taken from the process at approximately 60 min

1148 Volume 55, Number 9, 2001

FIG. 10. This � gure shows a period of poor process control. Data is presented as % rutile as a function of time over a period of 24 h.

intervals. The data in Fig. 7 is typical of much of theoperation of these systems. The correlation between thetwo techniques is generally good with the occasional dis-crepancy being caused by factors such as inaccuracies inthe timing of grab sampling; these can reasonably be ig-nored.

An important requirement for on-line analyzer systemsis that they be robust and able to cope with variations inconditions that can occur during normal operation of themanufacturing process. Of particular concern with theprocess described here was the possibility that the mea-surement window could quickly become contaminatedwith a layer of CD, thus invalidating future data. Carefuldesign of the optical window and the mechanical arrange-ments for sampling the CD reduced the tendency of suchlayers to form. To test these design changes, the windowwas regularly cleaned with a soft cloth and returned topristine condition. Figure 8 shows four sets of data ob-tained over different 10 hour periods during which thesampling system was brie� y halted and the windowwiped clean. Interestingly, the system shows a ‘recovery’response, the immediate effect is an approximate dou-bling of the signal count area, which subsequently decaysback to the original value over a 4–6 hour period. Al-though some variation in detail is evident, this character-istic shape was frequently observed. The physical expla-

nation for this general effect, and for the variability inrecovery response, is incomplete, but it appears to becaused by a very thin coating of CD forming on the win-dow. This coating causes some of the illumination laserlight or the returning Raman light to be attenuated, thusreducing the overall signal level at the detector. The mag-nitude of the effect is variable because it is related to thefraction of � ne material in the product stream, which canvary as a function of process operating conditions. Thelayer formed is very thin and is out of the optimum focalplane of the probe and therefore has no measurable effecton the analytical result. This is veri� ed by Fig. 9, whichshows that there was no effect on the measured rutilelevel despite the large change in overall spectral intensitysignal level experienced as a result of event A (see Fig.8).

The Raman instrument is now installed in four Hunts-man Tioxide sulfate factories worldwide, demonstratingclearly that this technology can be installed in an inhos-pitable process environment. The ability to measure rutilecontent at the point of production has enhanced the con-trollability of the calcination process by providing a muchhigher quality stream of data, both in terms of accuracyand frequency. The problems of maintaining equipmentshould not be underestimated, however; the Raman ana-lyzers are supported by a sophisticated remote data ac-

APPLIED SPECTROSCOPY 1149

FIG. 11. This � gure shows a period of good process control. Data is presented as % rutile as a function of time over a period of 24 h. Note theexpanded % Rutile scale of this � gure relative to Fig. 10.

quisition system that permits a service group to monitorperformance and make annual predictive maintenancevisits.

Figure 10 shows calciner operation where, under man-ual control, regular drops in the rutile quality parameteroccur. It is a characteristic of this non-linear process thatrutile drops are sudden and catastrophic. Using the Ra-man, conditions leading to such drops can be detected bythe control system much more quickly than previouslyand the appropriate corrective action taken. Control isexercised by manipulating the energy entering the calci-ner, and Tioxide has a number of schemes for achievingthis. Figure 11 shows a period of stable operation usingRaman control. The differences between these two plotsare obvious and the improvement gained is signi� cant.

CONCLUSION

On-line Raman spectroscopy has some very useful fea-tures that enable it to be installed in a very practical man-ner, and, as has been shown in this paper, it can be usedto produce effective solutions to very dif� cult analysisproblems. Application to hot, abrasive, and heteroge-neous solids such as TiO2 perhaps represents an extremecase in terms of the collection of practical dif� culties thatmight be encountered during development and imple-mentation. However, application to solid materials is afrequent requirement of on-line analysis, and successfulapplication of Raman spectroscopy to TiO2 should enablethis technology to be applied elsewhere. Further, the work

described here encourages the view that this technique ismaturing very rapidly and successful applications willemerge from a range of other industries.

ACKNOWLEDGMENTS

The authors are indebted to a number of colleagues, especially J. P.Besson (quality and technical manager at Tioxide, Calais), for their help,support, and suggestions. We have been fortunate to work closely withHarry Owen, Joe Slater, Ian Lewis, and several of their colleagues atKaiser Optical Systems, Inc., Ann Arbor, MI, and we are particularlygrateful for the access they gave us to pre-commercial technology dur-ing the development phase of this work. We also thank ICI PLC andHuntsman Tioxide PLC for permission to publish this article.

1. A. Al-Khanabashi, M. Dhamdhere, and M. Hansen, Appl. Spec-trosc. Rev. 33, 115 (1998).

2. N. Everall and B. King, Macromol. Symp. 141, 103 (1999).3. N. Everall, Analytical Applications of Raman Spectroscopy, M. J.

Pelletier, Ed. (Blackwell Science, Oxford, 1999), pp. 127–192.4. S. Farquharson and S. F. Simpson, Proc. SPIE-Int. Soc. Opt. Eng.

1681, 276 (1992).5. E. D. Lipp and R. L. Grosse, Appl. Spectrosc. 52, 42 (1998).6. M. Z. Martin, A. A. Garrison, M. J. Roberts, P. D. Hall, and C. F.

Moore, Process Quality and Control 5, 187 (1993).7. J. J. Freeman, D. O. Fisher, and G. J. Gervasio, Appl. Spectrosc.

47, 1115 (1993).8. C. Frank, Analytical Applications of Raman Spectroscopy, M. J.

Pelletier, Ed. (Blackwell Science, Oxford, 1999), pp. 224–275.9. F. Adar, R. Geiger, and J. Noonan, Appl. Spectrosc. Rev. 32, 45

(1997).10. J. P. Besson et al., European Patent Application EP19960003

06583.11. U. Balachandran and N. G. Eror, J. Solid State Chem. 42, 276

(1982).

1150 Volume 55, Number 9, 2001

12. R. J. Capwell, F. Spagnolo, and M. A. DeSesa, Appl. Spectrosc.26, 537 (1972).

13. N. J. Everall, ‘‘Raman Spectroscopy of the Condensed Phase,’’ Vol-ume 1—Introduction to the Theory and Practice of VibrationalSpectroscopy, Handbook of Vibrational Spectroscopy, P. Grif� thsand J. Chalmers, Eds. (John Wiley and Sons), to be published Sep-tember 2001.

14. F. H. Pollak, Analytical Raman Spectroscopy, J. G. Grasselli andB. J. Bulkin, Eds. (John Wiley and Sons, New York, 1991),Chap. 6.

15. S. Kelly, F. H. Pollak, and M. Tomkiewicz, J. Phys. Chem. 101,2730 (1997).

16. J. C. Parker and R. W. Siegel, Appl. Phys. Lett. 57, 943 (1990).17. N. J. Everall, I. M. Clegg, and P. W. B. King, Appl. Spectrosc., to

be submitted.18. I. Clegg, N. Everall, and B. King, European Pharmaceutical Review

5, 27 (2000).19. J. M. Tedesco and K. Davis, Proc. SPIE-Int. Soc. Opt. Eng. 3537,

200 (1998).20. P. R. Graves, ‘‘Calibration and Data Processing’’, in Practical Ra-

man Spectroscopy, D. J. Gardiner and P. R. Graves, Eds. (Springer–Verlag, Berlin, 1989), pp. 77–101.

21. P. Hendra, C. Jones, and G. Warnes, FT-Raman Spectroscopy (EllisHorwood, New York, 1981), pp. 134–147.