557r 2009 American Chemical Society pubs.acs.org/EF

Energy Fuels 2010, 24, 557–562 : DOI:10.1021/ef900908pPublished on Web 10/28/2009

Sulfur Speciation of Crude Oils by Partial Least Squares Regression Modeling of Their

Infrared Spectra

Peter de Peinder,*,†,‡ TomVisser,*,‡ RudyWagemans,§ JanBlomberg,§HassanChaabani,§ Fouad Soulimani,‡

and Bert M. Weckhuysen‡

†VibSpec, Tiel, The Netherlands, ‡Inorganic Chemistry and Catalysis Group, Department of Chemistry, Utrecht University,The Netherlands, and §Shell Global Solutions International B.V., Amsterdam, The Netherlands

Received August 20, 2009. Revised Manuscript Received October 9, 2009

Research has been carried out to determine the feasibility of partial least-squares regression (PLS)modeling of infrared (IR) spectra of crude oils as a tool for fast sulfur speciation. The study is acontinuation of a previously developed method to predict long and short residue properties of crude oilsfrom IR and near-infrared (NIR) spectra. Retention data of two-dimensional gas chromatography (GC�GC) of 47 crude oil samples have been used as input for modeling the corresponding IR spectra. A total of10 different PLS prediction models have been built: 1 for the total sulfur content and 9 for the sulfurcompound classes (1) sulfides, thiols, disulfides, and thiophenes, (2) aryl-sulfides, (3) benzothiophenes, (4)naphthenic-benzothiophenes, (5) dibenzothiophenes, (6) naphthenic-dibenzothiophenes, (7) benzo-naphthothiophenes, (8) naphthenic-benzo-naphthothiophenes, and (9) dinaphthothiophenes. Researchwas carried out on a set of 47 IR spectra of which 28 were selected for calibration by means of a principalcomponent analysis. The remaining 19 spectrawere used as a test set to validate the PLS regressionmodels.The results confirm the conclusion from previous studies that PLS modeling of IR spectra to predict thetotal sulfur concentration of a crude oil is a valuable alternative for the commonly applied physicochemicalASTM method D2622. Besides, the concentration of dibenzothiophenes and three different benzothio-phene classes can be predicted with reasonable accuracy. The corresponding models offer a valuable toolfor quick on-site screening on these compounds, which are potentially harmful for production plants. Themodels for the remaining sulfur compound classes are insufficiently accurate to be used as a method fordetailed sulfur speciation of crude oils.

Introduction

Crude oils are highly complex mixtures of organic com-pounds with a large variety in elemental composition andchemical structures. All crude oils contain sulfur in concen-trations between 0.1 wt % in light samples up to 10% in, forexample, bitumen and tar sands.1 Themajority of the sulfur ispresent as organic molecules in more than 10000 differentstructures, ranging from aliphatic sulfides, disulfides, andalkyl-substituted thiophenes to a variety of large polycyclicbenzothiophenes.2 The presence of sulfur species in crude oilshas a severe impact on oil production and refinery processes.Next to direct corrosive effects on the plant infrastructure andequipment, macromolecular sulfur compounds form a sub-stantial part of the solid asphaltenes and may cause cloggingof pipelines.3 Therefore, an important task at productionplatforms and refineries is to quickly identify the compoundsthat are harmful for the production plant.

Another, well-known drawback of sulfur in crude oils is therelease of sulfur oxides (SOx) upon combustion of crude oilbased fuels. This environmental effect has led to more andmore severe directives on SOx emission. As a result, novel orimproved hydrodesulfurization (HDS) catalysts have been

developed and as a consequence the sulfur content of fuels isdramatically reduced. Nowadays, the maximum sulfur concen-tration in Europe is 10 ppm (w/w S-total) for gasoline anddiesel4 and 1000 ppm for marine diesel.5 Desulfurization istherefore a big topic in oil industries. The current method ofchoice in refineries is HDS by means of, e.g., a cobalt-molybdenum based catalyst. This method is expected to staythe dominant technology for the coming years, even though it isstill not possible to eliminate the sulfur completely.6 On theother hand, HDS is an expensive treatment for deep desulfur-ization, while the removal of heterocyclic aromatic sulfurcompounds is not very effective. This is particularly relevantsince the exploration of the tar sand fields in Canada hasbrought large amounts of crude oils with high sulfur concentra-tions onto the world market. For that reason, research foralternative methods and ways to enhance the efficiency of theHDS process is ongoing.4 Obviously, also in this process,detailed knowledge of the qualitative and quantitative composi-tion of sulfur compounds in crude oils is essential. Besides, thetype and molecular structure of the sulfur compounds arefound to affect the crackability and detachability,7 while the

*To whom correspondence should be addressed. E-mail: info@vibspec.com (P.d.P.), t.visser@uu.nl (T.V.).(1) Hua, R.; Li, Y.; Liu, W.; Zheng, J.; Wei, H.; Wang, J.; Lu, X.;

Kong, H.; Xu, G. J. Chromatogr., A 2003, 1019, 101.(2) Beens, J.; Thijssen, R. J. High Resolut. Chromatogr. 1997, 20, 131.(3) Marshal, A. G.; Rodgers, R. P. Acc. Chem. Res. 2004, 37, 53.

(4) Ali, M. F.; Al-Malki, A.; El-Ali, B.; Martinie, G.; Siddiqui, M. N.Fuel 2006, 85, 1354.

(5) Directive on fuel quality, 98/70/EC as amended by 2003/17/EC.(6) Ring, Z.; Chen, J.; Yang, H.; Du, H.; Briker, Y.Proceedings of the

AIChESpringNationalMeeting, NewOrleans, LA, April 25-29, 2004; pp1386-1406.

(7) Xialolan, Z.; Jun, J.; Jianhua, L.; Yongtan, Y.; Chin, J. Anal.Chem. 2006, 34, 1546.

558

Energy Fuels 2010, 24, 557–562 : DOI:10.1021/ef900908p de Peinder et al.

desulfurization efficiency for an individual sulfur compounddiffers with the type of crude.8 Evidently, sulfur speciation ofcrude oils, either into detail or indicative and fast, is animportant task in oil industries.

Many analytical techniques have been explored for thispurpose, ranging from square wave voltametry9 and liquidchromatography10 to conventional gas chromatography(GC) and two-dimensional GC (GC � GC).11-13 A varietyof sulfur selective detectors have been used in combinationwith GC, such as atomic emission detection (GC-AED),14

sulfur chemiluminescence detection (GC-SCD),1,7,15-18 andmass spectrometric detection (MSD).19-24 Other detectiontechniques have been based on X-ray spectroscopy includingX-ray fluorescence (XRF)25-28 and X- ray absorption near-edge structure (XANES) spectroscopy.29-32 Furthermore,potential of temperature programmed reduction and oxida-tion methods has been studied27,33 as well as the newbut powerful technique of Fourier transform ion cyclotron

resonance (FT-ICR)-MS.3,34-38 Occasionally, infrared (IR)spectroscopy has been used, either including an oxidationpretreatment39 or without it.40 The advantage of IR is that itcan be easily performed on location without any preparationof the sample. In previous papers,41-43 we have demonstratedthe viability of chemometric modeling IR spectra of crude oilsto predict long and short residues properties of crude oilsstraightforward from their spectra. This method, based onpartial least squares (PLS) regression models, has been pa-tented.44 It is currently tested on-site as a fast alternative forthe much more elaborate physicochemical American Societyfor Testing andMaterials (ASTM) and Institute of Petroleum(IP) methods used so far. Also, the method turned out to beable to predict the sulfur content with high accuracy. For thatreason, a study to the potentials of PLS modeling of IRspectra as a tool for sulfur speciation is a logical next step.This article describes the results of that study using thespeciation data obtained from standard GC � GC analysisas reference values.

Methods and Materials

A set of 47 crude oil samples, representing a wide range ofgeographical oil wells and hence a large variety of different sulfur

Table 1. Sulfur Compound Classes As Applied in This Study

1 STD sulfides, thiols, disulfides, thiophenes2 Ar-S aryl-sulfides3 BT benzothiophenes4 NBT naphthenic-benzothiophenes5 DBT di-benzothiophenes6 NDBT naphthenic-di-benzothiophenes7 BNaT benzo-naphthothiophenes8 NBNaT naphthenic-benzo-naphthothiophenes9 DNaT dinaphthothiophenes

10 S total total sulfur amount (including elemental S)

Table 2. Schematic Representation of the 18 Preprocessing

Methods Used

1 MC, 1800-6502 MC, 3500-6503 MSC, MC, 3500-6504 SNV, MC, 3500-6505 SNV, Detrend (2), MC, 3500-6506 SNV, Detrend (3), MC, 3500-6507 SG (25 2 0), MC, 3500-6508 SG (25 2 0), MC, 1800-6509 SG (25 2 1), MC, 3500-650

10 SG (25 2 1), MC, 1800-65011 SG (25 2 1), MSC, MC, 3500-65012 SG (25 2 1), MSC, MC, 1800-65013 SG (25 2 1), SNV, Detrend (2), MC, 3500-65014 SG (25 2 1), SNV, Detrend (2), MC, 1800-65015 SG (35 2 2), MC, 3500-65016 SG (25 2 2), MSC, MC, 3500-65017 SG (49 2 2), MSC, MC, 3500-65018 SG (49 2 2), MSC, MC, 1800-650

(8) Yang,Y.T.; Yang,H.Y.; Lu,W.Z.Chin. J. Chromatogr. 2002, 20,493.(9) Serafim, D. M.; Stradiotto, N. R. Fuel 2008, 87, 1007.(10) Sinkkonen, S. J. Chromatogr. 1989, 475, 421.(11) Beens, J.; Blomberg, J.; Schoenmakers, P. J. J. High Resolut.

Chromatogr. 2000, 23, 182.(12) Blomberg, J.; Schoenmakers, P. J.; Brinkman, U. A. Th.

J. Chromatogr., A 2002, 972, 137.(13) Blomberg, J.; Riemersma, T.; Van Zuijlen, M.; Chaabani, H.

J. Chromatogr., A 2004, 1050, 77.(14) Hegazi, A. H.; Andersson, J. T.; Abu-Elgheit, M. A.; El-Gayar,

M. Sh. Polycyclic Aromat. Compd. 2004, 24, 123.(15) Andari, M. K.; Behbehani, H. S. J.; Stanislaus, A. Fuel Sci.

Technol. Int. 1996, 14, 939.(16) Behbehani, H. S. J. 219th National Meeting of the American

Chemical Society, San Francisco, CA, March 26-30, 2000; AmericanChemical Society: Washington, DC, 2000; PETR-057.(17) Hua, R.; Wang, J.; Kong, H.; Liu, J.; Lu, X.; Xu, G. J. Sep. Sci.

2004, 27, 691.(18) Lee, I. C.; Ubanyionwu, H. C. Fuel 2008, 87, 312.(19) Glinzer, O.; Severin, D.; Beduerftig, C.; Czogalla, C. D.; Puttins,

U. Fresenius’ Z. Anal. Chem. 1983, 315, 208.(20) Dzidic, I.; Balicki, M. D.; Rhodes, I. A. L.; Haskell, I. A. L.

J. Chromatogr. Sci. 1988, 26, 236.(21) Payzant, J. D.; Montgomery, D. S.; Strausz, O. P. AOSTRA J.

Res. 1988, 4, 117.(22) Nishioka, M.; Tomich, R. S. Fuel 1993, 72, 1007.(23) Sinningh�e Damst�e, J.; Rijpstra, W. I. C.; de Leeuw, J. W.;

Lijmbach, G. W. M. J. High Res. Chromatogr. 1994, 17, 489.(24) Ma, X.; Sakanishi, K.; Isoda, T.; Mochida, I. Fuel 1997, 76, 329.(25) Waldo, G. S.; Mullins, O. S.; Penner-Hahn, J. E.; Cramer, S. P.

Fuel 1992, 71, 53.(26) Snape, C. E.; Ismail, K.; Mitchel, S. C.; Bartle, K. Speciation of

Organic Sulfur Forms in Solid Fuels and Heavy Oils. In Composition,Geochemistry and Conversion of Oil Shales; Snape, C. E., Ed.; KluwerAcademic Publishers: Dordrecht, The Netherlands, 1995; pp 125-142.(27) Snape, C. E.; Yperman, J.; Franca, D.; Bartle, K. Eur. Comm.,

[Rep.] EUR 1998, EUR 17947, 1–100.(28) Barker, L.R.;Kelly,W.R.;Guthrie,W.F.EnergyFuels 2008, 22,

2488.(29) Waldo,G. S.; Carlson, R.M.K.;Moldowan, J.M.; Peters,K. E.;

Penner-Hahn, J. E. Geochim. Cosmochim. Acta 1991, 55, 801.(30) Kasrai, M.; Bancroft, G. M.; Brunner, R. W.; Jonasson, R. G.;

Brown, J. R.; Tan, K. H.; Feng, X.Geochim. Cosmochim. Acta 1994, 58,2865.(31) Sarret, G.; Connan, J.; Kasrai, M.; Eybert-Berard, L.; Bancroft,

G. M. J. Synchrotr. Rad. 1999, 6, 670.(32) Mijovilovich, A.; Pettersson, L. G. M.; Mangold, S.; Janousch,

M.; Susini, J.; Salom�e, M.; de Groot, F. M. F.; Weckhuysen, B. M.J. Phys. Chem. A 2009, 113, 2750.(33) Snape, C. E.; Mitchel, S. C.; Ismail, K.; Garcia, R. Rev. Anal.

Chem.-Euroanal. VIII 1994, 103.(34) Guan, S.;Marshall,A.G.; Scheppele, S. E.Anal.Chem. 1996, 68, 46.(35) Hughey, C. A.; Rodgers, P. R.; Marshall, A. G.; Qian, K.;

Robbins, W. K. Org. Geochem. 2002, 33, 743.(36) Klein, G. C.; Rodgers, R. P.;Marshall, A. G.Fuel 2006, 85, 2071.(37) Hughey, C.A.;Galasso, S.A.; Zumberge, J. E.Fuel 2007, 86, 758.(38) Panda, S. K.; Schrader, W.; Al-Hajji, A.; Anderson, J. T. Energy

Fuels 2007, 21, 1072.

(39) Saetre, R.; Somogyvari, A. Prepr.-Am. Chem. Soc., Div. Pet.Chem. 1989, 34, 268.

(40) Samedova, F. I.; Martynova, G. S.; Yusifov, Y. G.; Guseinova,B. A.; Ismailov, E. G. Azarb. Neft Tasarrufati 2008, 4, 39.

(41) De Peinder, P.; Petrauskas,D.D.; Singelenberg, F.; Salvatori, F.;Visser, T.; Soulimani, F.; Weckhuysen, B. M. Appl. Spectrosc. 2008, 62,414.

(42) De Peinder, P.; Petrauskas, D. D.; Singelenberg, F.; Salvatori, F.;Visser, T.; Soulimani, F.; Weckhuysen, B.M.Energy Fuels 2009, 23, 2164.

(43) De Peinder, P.; Petrauskas,D.D.; Singelenberg, F.; Salvatori, F.;Visser, T.; Soulimani, F.; Weckhuysen, B. M. Vib. Spectrosc. 2009, 51,205.

(44) De Peinder, P.; Petrauskas,D.D.; Singelenberg, F.; Salvatori, F.;Visser, T.; Weckhuysen, B. M. PCT Patent Application WO 2008/135411, 2008.

559

Energy Fuels 2010, 24, 557–562 : DOI:10.1021/ef900908p de Peinder et al.

compounds and concentrations, has been used. All samples werestored in a refrigerator at 4 �Candbrought to ambient conditions24 h prior to analysis. Next, samples were homogenized byshaking the sample can every 10 min for 1 h. Experimentalprotocols on further pretreatment, preparation, and spectralrecording have been used throughout the study to ensure theacquisition of reproducible, high quality data. Details on theseprotocols can be found in ref 41. Modeling of the IR spectra hasbeen carried out for 10 different sulfur classes, as listed in Table 1,i.e., the total sulfur content and 9 sulfur speciation groups,commonly used in GC � GC analysis.

IR Spectroscopy. IR measurements have been carried out atroom temperature on a Bruker Tensor-27 Fourier transforminfrared (FT-IR) spectrometer equipped with aDTGS detector.The sample compartment was flushed with dry air to reduceinterference of H2O. Spectra were recorded with a horizontalATR accessory (FastIR, Harrick Scientific Products) with aZnSe crystal as the internal reflection element. The spectralresolution was 4 cm-1 for all spectra, and 50 scans wereaccumulated with medium apodization for each spectrum.

ATR-intensity correction was not applied. Although the highviscosity of several of the crude oils wouldmake it reasonable toperform the IR measurements at elevated temperatures, all IRmeasurements were carried out at room temperature (20 �C) forpractical reasons and to obtain a high screening velocity. Acover plate was used to prevent evaporation of light ends duringmeasurement.

Gas Chromatography. GC � GC analysis was performedon a double column Hewlett-Packard P 6890 gas chromato-graph (Agilent Technologies) equipped with a CIS4 PTV in-jector, a sulfur chemiluminescence detector, and a liquidnitrogen cryogenicmodulation assembly (ZoexCorp.). The firstcolumn was a nonpolar DB-1, dimethylpolysiloxane, 10 m, 0.25mm i.d., 0.25 μm Df. (J&W Scientific) and the second one amediumpolarity stationary phaseBPX-50, 50%phenyl(equiv.)-polysilphenylene-siloxane, 2 m, 0.10 mm i.d., 0.10 μm Df.(SGE). The modulation capillary was comprised of DPTMDSfused silica tubing, 2 m (1 m in loop), 0.10 mm i.d. (BGBAnalytik Vertrieb, Germany). The initial oven temperature forthe first dimension column was 40 �C. After an initial hold of

Table 3. Crude Oil Samples for Calibration (C1-C28) and Validation (V1-V19) Used for Modeling of 10 Different Sulfur Classesa

concentration (ppm)

sample STD Ar-S BT NBT DBT NDBT BNaT NBNaT DNaT S total

C1 14 8 359 57 742 114 187 26 15 7 180C2 4 1 20 3 59 4 13 2 1 1 370C3 62 254 1083 759 1521 683 519 197 62 11 700C4 53 59 461 191 490 195 171 61 38 5 630C5 11 67 598 332 1038 442 436 164 111 8 350C6 26 15 98 24 85 18 24 9 6 2 530C7 586 272 811 308 412 116 63 18 9 9 900C8 1803 607 6791 929 5335 1068 2678 804 870 54 200C9 138 91 111 58 124 37 50 16 11 3 000C10 62 70 366 167 218 64 31 13 6 4 330C11 385 318 2458 600 2097 638 728 309 205 21 900C12 170 830 2129 2149 3312 1425 1529 632 527 30 500C13 2488 557 6340 1193 4357 1301 1943 726 589 50 900C14 130 612 2096 2020 3193 1270 1415 571 522 31 400C15 269 1412 2704 2306 4248 1345 1597 589 506 38 800C16 1848 421 4907 974 2810 757 773 270 120 47 500C17 1039 262 950 200 1371 250 322 74 78 10 800C18 1622 640 3400 867 2033 756 664 289 105 20 500C19 623 289 5689 1075 4872 1166 1816 438 336 45 500C20 1179 293 1052 209 1475 277 331 72 62 11 100C21 1228 451 2821 929 1694 608 540 268 158 21 500C22 71 239 1606 492 1790 533 705 297 246 14 700C23 1022 501 3070 875 2448 844 1053 471 436 32 200C24 333 182 5114 928 4264 1124 1577 471 315 31 700C25 1448 520 5452 1292 2652 599 566 112 38 30 900C26 57 35 160 58 208 51 57 29 0 4 580C27 871 332 4506 717 4290 1086 1987 714 670 48 700C28 757 247 4264 756 3323 681 1101 201 154 37 600

V1 23 22 119 54 175 44 42 24 0 4 200V2 31 48 1700 357 1534 351 364 65 32 15 400V3 684 217 5322 1063 4053 1005 1476 326 274 43 000V4 1759 687 4306 1055 1980 557 486 164 90 33 900V5 20 21 212 68 380 92 121 36 0 4 840V6 966 268 5603 915 3651 785 1247 266 137 41 200V7 252 403 2213 760 1717 415 430 83 42 25 200V8 242 130 5449 991 5010 1195 2070 581 456 32 300V9 2073 665 2902 994 1760 642 531 259 149 28 200V10 881 292 3981 760 3094 876 1298 538 466 44 700V11 1210 437 5146 862 3994 1010 1642 618 473 42 600V12 736 235 3537 600 2614 751 1183 518 444 43 100V13 1279 374 3191 705 1470 271 223 53 23 20 300V14 807 473 2715 668 1248 236 223 61 34 25 000V15 263 150 1121 338 805 191 174 79 38 11 300V16 638 235 608 124 370 77 73 18 9 8 130V17 1149 446 4123 862 3096 871 1097 294 175 31 400V18 174 90 418 150 362 93 99 41 35 7 010V19 675 518 3935 1109 3082 1030 1148 480 479 28 500

aCompound class abbreviations refer to names listed in Table 1. Concentrations (parts per million, ppm) have been determined with GC � GC.

560

Energy Fuels 2010, 24, 557–562 : DOI:10.1021/ef900908p de Peinder et al.

5 min, the oven was programmed at a rate of 2.5 �C/min to320 �C, which was maintained for 20 min. The secondary ovenchamber for the second dimension column had an initial tem-perature of 90 �C. After an initial hold of 5 min, it wasprogrammed at a rate of 2.5 �C/min up to 370 �C, which wasmaintained for 20 min. The hot-pulse duration was set to500ms, and themodulation timewas 10 s. Samples were injectedeither pure or, when viscosity at 60 �Cor S-content did not allowso, diluted with toluene and/or cyclohexane. Concentrations ofcomponents in parts per million sulfur (ppm S) were calculatedby means of a classified internal standard.

Chemometrics.Modeling was performed using the PLS Tool-box (EigenvectorResearch, Inc.) forMatLab (TheMathWorks,Inc.) on the IR spectra of the 47 crude oils. Principle componentanalysis (PCA) of the total data set was applied to obtainrepresentative subsets for calibration and validation. A groupof 28 spectra was selected for calibration (samples C1-C28).The remaining 19 spectra were used for validation (samplesV1-V19). As input for modeling of the nine different sulfurcompound classes, the concentrations as determined with GC�GC have been used. Modeling for the total sulfur content wascarried out on the data as determined according to ASTMmethod D2622.

Prior tomodeling, a baseline correction was applied to the IRspectra by subtracting a third degree polynomial fit usingthe regions 4000-3500, 2500-2000, 1900-1800, 1560-1520,1000-990, and 650-600 cm-1. Subsequently, the region2500-1800 cm-1 was removed from the spectra since no absor-bance bandswere observed in this region.Next, preprocessing ofthe IR spectra was optimized for all 10 sulfur classes bysystematic varying preprocess parameters like scaling, smooth-ing, region selection, and spectrum derivative options. Thisresulted in a selection of 18 different preprocessing methods,based on previous modeling experience with this data set,41-44

which are listed in Table 2.MC refers tomean centering andwasapplied in all cases. The spectral range was either 3500-650 or1800-650 cm-1. For scaling, either the option “none”, multi-plicative signal correction (MSC), or standard normal variate(SNV) with and without detrending, second, or third orderpolynomial, was applied. The Savitzky-Golay (SG) smoothingand differentiation parameters were varied from 25 to 49 points,using a second order polynomial and none, first, or secondderivative. As an example, preprocessingmethod 13 comprises aSG smoothing with 25 points using a second order polynomialand taking the first derivative followed by SNV, detrendingwitha second order polynomial, and MC on the 3500-650 cm-1

region. For each of the 18 preprocessingmethods, PLSmodelingwas carried out for the 10 sulfur classes, which resulted in 180models. From these, the 10 models with the lowest root-mean-square-error-of-prediction (RMSEP) value for each of thesulfur classes were selected for concentration prediction.

Results and Discussion

Gas Chromatography. The concentrations in ppm of the 9different sulfur compound classes and S total, as determinedwith GC � GC analysis and ASTM method 2622, respec-tively, are presented in Table 3. Calibration samplesC1-C28, used for building the models, as well as samplesapplied for validation (V1-V19) represent a wide range ofconcentrations for the different sulfur speciation classes,which validates PCA of the IR spectra for this study.To illustrate the results from GC � GC, a retention time-intensity plot for crude oil C21 is shown as an example inFigure 1. Note that the physico-chemically determinedamount of “total S” differs from the summed concentrationsas measured with GC � GC. This is due to the fact that thelatter method only covers compounds that elute in the boiling

point range from ambient to 370 �C, whereas the ASTMmethod includes also higher-boiling compounds as well aselemental and inorganic sulfur.

Infrared Spectroscopy. As reported before,41-43 the IRspectra of crude oils are very similar, particularly afterintensity normalization and preprocessing. This is illustratedin Figure 2, showing the overlay of the 28 baseline correctedcrude oil spectra of the calibration set C1-C28. All spectraare dominated by strong absorption bands of aliphatic C-Hstretching (3000-2800 cm-1) and bending (1470-1350cm-1) vibrations. Small differences are present in the finger-print region (1300-650 cm-1). The absorption bands in thisregion can be merely attributed to aromatic skeletal modes.In general, specific C-S, S-H, and/or S-S vibrations arenot very IR active because of the small dipole momentchange during the vibration of these structural elements.45

However, for example, thiophene rings exhibit several sharpbands related to ring stretching (1550-1350 cm-1) anddC-H out of plane vibrations (800-690 cm-1).45,46

Data Analysis. PLS modeling of the 10 sulfur concentra-tion classes, using 18 different preprocessing methods, re-sulted in 180 models. For these 180 models, the RMSEP

Figure 1. GC � GC plot of crude oil C21. S-compound classes andthe internal standard have been indicated. White colors representhigh concentrations and black colors low concentrations.

Figure 2. Overlay of 28 spectra of crude oils as used for calibrationof the PLS-models.

(45) Lin-Vien,D.,Colthup,N.B., Fateley,W.G.,Graselli, J.G., Eds.The Handbook of Infrared and Raman Characteristic Frequencies ofOrganic Molecules; Academic Press, Inc.: San Diego, CA, 1991.

(46) Infrared and Raman Interpretation Support software, IRIS 3.0,Thiophenes, http://www.vibspec.com.

561

Energy Fuels 2010, 24, 557–562 : DOI:10.1021/ef900908p de Peinder et al.

values obtained for each speciation class were divided by thestandard deviation of the calibration values to express therelative error. These relative errors have been plotted as afunction of the preprocessing method for each sulfur specia-tion class in Figure 3. The figure illustrates that, independentof the applied preprocessing method, some classes (e.g.,NBT) are better predicted than others (e.g., STD). In ouropinion, this demonstrates the ability of the models toextract structure related correlations from the IR spectra.Next, the models with the lowest RMSEP values foreach of the 10 classes were selected for further evaluation.This is summarized in Table 4, showing for each classthe applied preprocessing method, the mean concen-tration value for the calibration set, the correspondingstandard deviation, the root-mean-square-error-of-valida-tion (RMSECV) value, the number of latent variable (LVs)thatwas used for themodel, and theRMSEPvalues obtainedfor the validation set. In addition, the corresponding plots ofthe predicted versus the measured concentrations for thecalibration set (b) and the validation set (�) are shown inFigure 4.

First of all, the results confirm the conclusion from ourprevious papers41-43 that the prediction of the total sulfurconcentration of crude oils by means of PLSmodeling of theIR spectra is a valuable alternative for ASTMmethod 2622.The models to predict the dibenzothiophenes (DBT) ispromising followed by the related benzothiophene com-pound classes BT, NBT, and NDBT. The correct predictionof DBT concentrations is particularly interesting in view ofthe fact that these compounds are the major sulfur contain-ing species left in fuels after hydrodesulfurization.Moreover,

the models for the speciation of DBT together with BT,NBT, and NDBT might be useful, as this type of compoundis known to hamper efficient crude oil processing and refin-ing. Furthermore, we conclude that the models for theremaining classes STD, Ar-S, BNaT, NBNaT, and DNaTare less useful for concentration prediction. The differencesin the predictive power of the models can be explained by theassumption that vibrations related to benzothiophene struc-tures are well represented in the IR spectra, whereas othersulfur containing functional groups lack specific sulfur re-lated absorption bands.

Figure 3. RMSEP values divided by the standard deviation for the10 sulfur speciation classes for 18 different preprocessing methods.

Table 4. Results of Optimized PLS Models to Predict the Concen-

trations of 10 Different Sulfur Compound Classes in Crude Oilsa

S class

preprocessing(methodnumber)

meanconcn(ppm)

STDEV(ppm)

RMSECV(ppm) LVs

RMSEP(ppm)

STD 15 586 585 448 6 537Ar-S 10 328 302 229 4 147BT 18 2161 1864 769 6 700NBT 15 666 625 306 6 228DBT 9 1976 1553 632 10 383NDBT 7 585 451 213 8 187BNaT 9 773 722 367 5 331NBNaT 11 281 253 95 8 194DNaT 13 227 243 131 4 199S total 9 21587 16351 5403 7 2520

aMean concentration, preprocessing method, STDEV, RMSECV,and LV values refer to the calibration set and RMSEP values to thevalidation set.

Figure 4. Prediction plots of PLS modeling the concentration of 10sulfur speciation classes of crude oils based on their IR spectra.Calibration spectra (b) and validation spectra (�). The correspond-ing preprocessing methods are listed in Table 4.

562

Energy Fuels 2010, 24, 557–562 : DOI:10.1021/ef900908p de Peinder et al.

Closer examination of the correlation plots in Figure 4reveals that the concentration distribution for some of thesulfur groups (e.g., Ar-S andNBT) is not homogeneous. Thisis due to the fact that the PCA to obtain calibration andvalidation subsetswas based on spectral variation andnot ondifferences in sulfur specie concentrations. It cannot beexcluded that, in some cases, PCA based on this parametermight lead to better models.

Finally, it should be noted that in refineries effluents havelower average sulfur concentrations. This requires the develop-ment of dedicated models. However, as demonstrated in thisstudy, several of the models developed for crude oils alsoperformwell at lowsulfur concentrations.Therefore,webelievethat PLSmodeling of the IR spectra of refinery effluents mighthave similar potential for sulfur speciation purposes.

Conclusions

PLSmodeling of the IR spectra of crude oils is a valuablealternative toASTMmethod 2622 to predict the total sulfurcontent of these materials. The application as a tool for

sulfur speciation, however, is limited. From the nine differ-ent sulfur compound classes that are usually determinedwith standard GC�GC analysis, the models to predict theconcentration of DBT and the related benzothiophenecompound classes BT, NBT, and NDBT perform reason-ably well. However, the models for the remaining classesSTD, Ar-S, BNaT, NBNaT, and DNaT are less useful. Assuch, PLS regression modeling is not as widely applicablefor sulfur speciation as GC�GC. On the other hand, it canbe a fast, clean, and nonelaborate method for qualitativeand quantitative on-site or even in situ screening of crudeoils on (di-) benzothiophenes, a class of compoundswhich isknown to be detrimental in crude oil processing and apredominant sulfur-residual in fuels.

Acknowledgment. This work was carried out by financialsupport of Shell Global Solutions International B.V., TheNetherlands.Dr.D.Petrauskas andDr.F. Salvatori are gratefullyacknowledged for providing the crude oil samples and permissionto use the corresponding data.