based on neural network for pattern recognition Diffusion behavior study of model diesel components in polymermembranes Technology Journal of Applied Researchand

(1)

Availableonlineatwww.sciencedirect.com

Journal of Applied Research and Technology

www.jart.ccadet.unam.mx JournalofAppliedResearchandTechnology14(2016)367–374

Original

Diffusion behavior study of model diesel components in polymer membranes based on neural network for pattern recognition

Xiaoyi Liang

^a,b

, Xingsheng Gu

^c

, Changjian Ling

^a,b

, Zhen Yang

^a,b,d,∗

aSchoolofChemicalEngineering,StateKeyLaboratoryofChemicalEngineering,EastChinaUniversityofScienceandTechnology,Shanghai200237,China

bKeyLaboratoryforSpecialFunctionalPolymerMaterialsandTheirRelatedTechnologies,MinistryofEducation,EastChinaUniversityofScienceand Technology,Shanghai200237,China

cInformationScienceInstitute,EastChinaUniversityofScienceandTechnology,Shanghai200237,China

dSchoolofMaterialsScienceandEngineering,GeorgiaInstituteofTechnology,Atlanta,GA30332-0405,USA Received7December2015;accepted20June2016

Availableonline2December2016

Abstract

Aneuralnetworkforapatternrecognitionmodelisdevelopedforthefirsttimetopredictthediffusionbehaviorofthemodeldieselcomponents (dibenzothiopheneandquinolone)inthreedifferentmembranesofpolyvinylalcohol,polyvinylchlorideandpolymethylacrylate.Thesimulation resultsshowthattheexcellentperformancetargetparameteroptimizationareacanbeobtainedandtheeffectivedesulfurizationanddenitrification agentcanbefound.Comparedwiththeadvancedmoleculardynamicsimulationmethodandverifiedbyadsorptionexperiments,thesimulation valuesareingoodagreementwiththeexperimentaldataandmoleculardynamicsimulationdata.Theresultsrevealthatthepolyvinylchloride membranecanimprove thediffusionselectivityofdibenzothiopheneandit isselectedasthe mosteffectivedesulfurizationagent,whilethe polyvinylalcoholmembraneisselectedasthemosteffectivedenitrificationagenttoremovethenitrogencompounds.Developmenttimeandeffort ofscreeningdesulfurizationagentanddenitrificationagenttestsarealsoreducedbecausetheneuralnetworkforthepatternrecognitionmodel providesready-madedecisions.Therefore,theneuralnetworkforpatternrecognitionisaprospectandpracticabletheoreticalmethodtoresearch thediffusionbehaviorofmodeldieselcomponentsinpolymermembranes.

Keywords:Dieselcomponents;Desulfurization;Denitrification;Prediction;Neuralnetwork;Patternrecognition

1. Introduction

Sulfur compounds indiesel fuel can generateSOx during combustion, which would result in health damage and seri- ousenvironmentalproblems.Desulfurizationofdieselfuelhas increasinglygainedimportancesincemostofthecountries,par- ticularlythedevelopedones,haveimplementedmorestringent legislationtoregulatethesulfurcontentoftransportationfuels (Song& Ma,2003). Meanwhile,nitrides contained indiesel mayproduceasignificant effectontheir stability andhavea

∗Correspondingauthor.

E-mailaddresses:[email protected],[email protected] (Z.Yang).

PeerReviewundertheresponsibilityofUniversidadNacionalAutónomade México.

dominanteffectonthecolorofoilandcreationofresins(Sano, Choi, Korai, &Mochida,2004). Also,organic nitrides inoil productcauseairpollutionduringburning.Refiningdieselfuel andimprovingitsstabilityareofconsiderableinterestcurrently tomeettheincreasingneedsofhigherqualitydieselfuel.There- fore,itisurgenttodevelopthedesulfurizationanddenitrification technologiesofdieselfuel.Itisthereforeparticularlymeaningful tosearchfortheidealmembranematerials,whichcanfacilitate the diffusionof sulfurcompoundsandnitriccompounds,and restrainthediffusionof non-sulfurcompoundsandnon-nitric compounds.

Inthispaper,desulfurizationanddenitrificationtechnologies based on the activated carbon adsorption belong to non- hydrogenationdesulfurizationanddenitrificationtechnologies.

Inordertoimprovetheactivatedcarbonselectiveadsorptionof sulfurcompoundsandnitrogencompounds,alayerofpolymer filmmaterialwithhighselectivityiscoatedonthesurfaceofthe

http://dx.doi.org/10.1016/j.jart.2016.06.007

1665-6423/©2016UniversidadNacionalAutónomadeMéxico,CentrodeCienciasAplicadasyDesarrolloTecnológico.Thisisanopenaccessarticleunderthe CCBY-NC-NDlicense(http://creativecommons.org/licenses/by-nc-nd/4.0/).

(2)

activatedcarbon,whichensuresthesulfurcontentandnitrogen contentcannotbechanged.

However,itisquitedifficultandtimeconsumingtoexplore thepropermembranesdirectlybymeansofmostoftheexisting experimentaltechniques.Inviewofthefactthatexperimental methodofselecting thepropermembranematerialsisexpen- sive,timeconsumingandrequiresalotofpuresamples,thenew methodusingcomputersimulationtechniquetoscreentheopti- malmembranematerialsispresentedinthispaper.Wedevelop aneuralnetwork(NN)patternrecognition(PR)modelthatcan predict polymer membranediffusion propertiesdirectly from adsorbentexperimentaldata.

Patternrecognitionisarapidlyevolvingandapromisingsub- jectinartificialintelligencesciences.Inrecentyears,thetheory andmethodofpatternrecognitionhavebeengettingmoreand moreattention globally.PR isbased onthe truth “Birdsof a featherflocktogether”torecognize,discriminateandclassify theunknownattributeof things.Thebasicassumptionofpat- ternrecognitionisthatcertainfeaturescausedifferenttypesof observedthings,thentheyshouldhavesomeverysimilarfea- turesandsimilaritiesmake themtogetherinthesamepattern space.Whenthedirectconnectionbetweentheresearchobjec- tiveandthe influencingfactorisunknown,or itisdifficultto findoutthedirectconnectionwithmostoftheexistingtheories, pattern recognitionisaneffective methodwhichcansumma- rize the empirical data and searchfor the direct andindirect connections between the target anda number of factors and findouttheoptimizedtargetareasoroptimizeddirection(Jain, Murty, & Flynn, 1999). Pattern recognition technology was firstlyusedinchemistryandchemicalengineeringbyKowal- ski,BenderandWoldinthe1970s(Kowalski&Bender,1972;

Wold&Sjostrom,1977).Withtherapiddevelopmentofcom- putertechnology,patternrecognitionproducedpositiveeffects inthefieldofmaterialoptimaldesign.Twomaintypesofstatisti- calpatternrecognitionmethods(nearestneighborclassification andprincipalcomponentanalysis)havebeenusedformaterial optimaldesign;however,theyhavemanydifficultiesinselect- ing the discriminant function in the multidimensional space (Cardoso,1999;Chu&Plemmons,2005;Guan,Tao,Luo,&

Yuan, 2012;Kramer,1991).Asaresult,the greatdifferences betweentheclassificationresultsandtherealexperimentaldata mayleadtoerroneousconclusionsinmaterialoptimaldesign (Kalidindi & Graef, 2015; Panchal, Kalidindi, &McDowell, 2013).

Neuralnetworkisaneffectivemodelingmethodformaterial optimaldesign(Kou,Parks,&Tan,2012;Shabani&Mazahery, 2012;Srinivasan&Saghir,2013).Theneuralnetworkforpattern recognition(NN-PR)cangetridofthenoiseinthesampledata, reduceredundantstructure,improvetheprecisionandspeedof calculationandreduceharmtonetworkerrorforecasts.Thus,the simulationresultswillbemoreaccurateandreliable(Chattoraj, Mondal,Das,Roy,&Sadhukhan,2014;Chen,2015;Porrazzo, Cipollina, Galluzzo,&Micale, 2013).NN-PR isexpected to beamorescientific,advancedandfeasiblemethodformaterial optimaldesignresearch.ForNN-PR,optimizingandselecting themainparametersaretwokeyfactorsthateffecttherecognize results.

ThefirstaimofthepresentworkistodevelopaNN-PRmodel topredictthediffusionbehaviorofthemodeldieselcomponents (dibenzothiophene(DBT) and quinoline(QL)) in three differ- entmembranesofpolyvinylalcohol(PVA),polyvinylchloride (PVC)andpolymethylacrylate(PMA)whicharecoatedwith activecarbon;thesecondoneistoselecttheeffectivedesulfuri- zationanddenitrificationagent.Firstly,GAisusedtooptimize thestructureofNN.Then,theNN-PRmodelisbuilttogetexcel- lent desulfurization and denitrification agent discrimination results.Then,thevalidityoftheNN-PRmethodiscomparedand verifiedwiththemoleculardynamic(MD)simulationmethod, whichwehavedonebefore.Finally,thevalidityandpracticality oftheNN-PRmethodarefurtherdeterminedbycomparingthe resultsofthesimulatingdatawiththeexperimentaldata.

2. Materialsandexperimentaldesign

2.1. Polymermembraneactivatedcarbondesulfurization anddenitriﬁcationsorbents

Two main methods for diesel desulfurization are non- hydrodesulfurization (non-HDS) and hydrodesulfurization (HDS).Asanon-HDSmethod,adsorptivedesulfurizationhas theadvantagesofbeinglowcost,easytooperateandenviron- mentallyfriendlyaswellasnotreducingthecetanenumberof dieselfuel,amongothers.

Also,therearetwomainmethodsfordieseldenitrification:

non-hydrodenitrogenation (non-HDN) and hydrodenitrogenation(HDN).Asanon-HDNmethod,adsorptivedenitrogenation hastheadvantagesofbeinglowcostandthesimplicityofequip- mentandoperation;however,thecapacityoftheadsorbentis limited.

Innon-HDSandnon-HDN,sphericalactivatedcarbonsare oftenusedasadsorbents.Inthisstudy,wemainlyimprovethe selectivityofsulfurcompoundsandnitrogencompoundsunder thepremiseofnochangingtheactivatedcarbonadsorptionsul- furcapacityandnitrogencapacity.Forthispurpose,thesurface of activatedcarbon iscoated withalayer of highlyselective polymermembranematerials.Thediffusionbehavior,dynamic selectiveadsorptionandadsorptioncapacityofDBTandQLare researchedinthreedifferentpolymermembranesystems(PVA, PVC,andPMA). Inordertoobtainthe optimizingdesulfuri- zationanddenitrification properties, coatedactive carbonare evaluatedbyNN-PRmodel.Meanwhile,theMDmethodandthe adsorptionexperimentalmethodarediscussedandcompared.

2.2. Preparationofsimulateddiesel

Petroleum fractions mainly contain DBT and substituted derivativesunderatemperaturehigherthan300^◦C.Inthisstudy, DBTwasselectedasamodeldieselsulfurcompoundcompo- nent.QLisatoxic,biodegradableorganicnitrogencompound anditwasselectedasnitrogencompoundindieselcomponents.

N-octanewasselectedasthemainsolventofthemodeldiesel.

Inordertomakeacomparisonandtakeverificationtests,two

(3)

Table1

TheequilibriumadsorptiondataofDBTandQLontheACStyrene.

Types qe,exp(␮mol/g) K(h(␮mol/g)⁻¹)

ACStyrene-DBT 557.3–588.2 0.0527–0.0678

ACStyrene-PVA-DBT 564.1–592.5 0.0638–0.0772

ACStyrene-PMA-DBT 556.9–590.4 0.0636–0.0779

ACStyrene-PVC-DBT 574.7–593.7 0.0634–0.0781

ACStyrene-QL 317.2–336.4 0.0438–0.0512

ACStyrene-PVA-QL 316.8–359.0 0.0457–0.0602

ACStyrene-PMA-QL 328.4–353.1 0.0425–0.0556

ACStyrene-PVC-QL 312.2–352.6 0.0453–0.0581

ACStyrene-DBTmeanstheadsorption ofDBTontheACStyrene;qe,exp isthe experimentalvalueofadsorptioncapacity(␮mol/g);Kistheconstantofpseudo- secondorderkineticadsorptionrate(h(␮mol/g)⁻¹).

typesofmodeldiesel(single-componentandtwo-component) wereprepared.

(1) Preparationofthesingle-componentmodeldiesel

PutacertainamountofDBTorQLinavolumetricflask, thenaddacertainamountofn-octane.Thesinglecomponent solutionconcentrationis10.0␮mol/g,andthenshakethe mixturewell.

(2) Preparationof thedouble-component modeldieselof the sameinitialconcentration

WeighacertainamountofDBTandQLandputthemin thevolumetricflask,thenaddacertainamountofn-octane.

Two of the solution concentrations are 10.0␮mol/g, and thenshakethemixturewell.

2.3. Single-componentanddouble-componentadsorption experimentaldata

Theequilibriumadsorption experimentdataincluding two importantparametersofadsorptionrateconstantandadsorption capacityareshowninTable1.

3. Neuralnetworkforpatternrecognitiontheoretical methods

3.1. Patternrecognition

Inthisstudy,screeningtechnologyfordesulfurizationagent and denitrification agent are based on the intelligent pattern recognitiontheory.Patternrecognitionestablishesthepercep- tionformedfromexperienceandanalyzesthedataaccumulated byameasurementsystem.Itisausefulmethodtoextractinfor- mationfromdata(Nianyi,Wencong,Ruiliang,Chonghe,&Pei, 1999).

Feature selection and extraction are two keys for pattern recognition.In general,the moredifferenttypesof candidate featuresweobtain,thebettertherecognitionresultsshouldbe.

Nevertheless,it isdifficulttodistinguishhighdimensionfea- turesbecauseofthedimensiondisasters.Inthisstudy,theneural network feature extraction method is used for feature selec- tionandextraction,whichisanonlinearfilteringcharacteristic

algorithmbycalculatingcontributionratioofdifferentindepen- dentvariablesforecastingability.

Theframeworkoftheneuralnetworkforthepatternrecog- nitionmodelisshowninFig.1.

3.2. Artiﬁcialneuralnetworks

ThefirstartificialneuralnetworkwasdevelopedbyMcCul- loch and coworkers in the early 1940s (McCulloch & Pitt, 1943).ANNisclassifiedasanon-symbolicartificialintelligence technique, whichsimulates the human brainandits underly- ingprocessisbasedongraphsandalgorithms(Papadopoulos, Beligiannis, & Hela, 2011). In the last few decades, some importantresearcheshavebeenreportedinthefieldof ANNs and their efficiency has been proven in various applications in the fields of engineering. The learning modes of neural networksfromtheinputsamplesmakethemverysuitablefor solving pattern recognition problems, which are one of the successful application areasof neural networks(Christopher, 1995).

ThemostwidelyusedtrainingalgorithmofNNsistheback- propagation(BP)algorithm,whichisagradient-basedmethod.

ThebasicstructureofBPNNisshowninFig.2.

Usually,BPNNisrepresentedbythefollowingmodel:

⎧⎪

⎪⎪

⎨

⎪⎪

⎪⎩

out¹_j =xj (j =0,1,...,N) ...

net^l_j=

pot(l−1)

i=0

ω^l_jiout_l^l−1 (l=2,3,...,L) out^l_j=f(net^l_j) (j =1,2,...,pot(l)) ...

Eˆj =out_j^L (j =1,2,...,M)

(1)

wherepot(l)(l=1,2,...,L)isthenodenumberofeachlayer. ˆE_j istheestimatedvalueofthedesignobjective.Thehiddenlayer errorisdeterminedbythebackpropagatingalgorithm.TheBP algorithmisrepeatedseveraltimesuntilobtainingthereasonable residualerror.Thestepsoftheoptimizationprocedureandthe learningalgorithmaredetailedinPerlovsky(2001).

3.3. Neuralnetworkforpatternrecognition

Intheneuralnetworkforpatternrecognition,firstly,known samplesareusedtotraintheneuralnetwork,whichcangivedif- ferentoutputscorrespondingtodifferentkindsofinputs.Then, theunknownsamplescanbedistinguishedbythetrainedNN- PR model. The flow chart of the neural network for pattern recognitionisshowninFig.3.

TherolesofthemainpartsofNN-PRareasfollows:

(1) Sampleacquisition

Thisstepismainlyusedtogetacertainamountoftraining andrecognizingsamples.

(4)

Data preprocessing

Features selection and

extraction

Controlling of performance characteristics,

classification, optimization

Output results Experimental data

(pattern space)

Neural network for pattern recognition

Fig.1.Theframeworkofneuralnetworkforpatternrecognitionmodel.

Output target Output layer(NO.L)

Hidden layer

Input layer (NO.1) Input factor

…

Fig.2.BasicstructureofBPNN.

(2) Conventionalprocessing

The role of thisstep is equivalenttothe dataacquisi- tionandconventionalprocessingofthetraditionalpattern recognition, namely, by effective observing and quanti- tative sampling to obtain a series of data, and then the original samplesareobtained andsimulated byremoving noise.

(3) Featuretransform

In thisstep, feature transformationis presented bythe neuralnetworkbasedontheoriginalsamples.Theroleof thisstepissimilartofeatureselectionandextractionofthe traditionalpatternrecognition,buttheoriginalsamplesare processedbyneuralnetwork.

Theabovetwostepsarethedatapreprocessoftheneural networkforpatternrecognition.

(4) Neuralnetworkforpatternrecognition

Varying recognizedobjectsandresearch problems,the neural network istrained by choosing anappropriatenet- workstructureandlearningalgorithm.

Afterthetrainingprocess,thenetworkisequivalenttoa fixedmapping,andnewinputsamples(testsamples)canbe mappedtodifferenttypesaccordingtothedifferentdata.

4. Neuralnetworkforpatternrecognitionmodel

Inthisstudy,firstly,theoriginaladsorptionexperimentaldata werenormalizedto[01].Next,thenormalizeddataweretrained byBPnetworktogetthepredicteddata.Then,thepredicteddata wererecognizedandclassifiedbypatternrecognition.Finally, optimal desulfurization agent and denitrification agent were selectedbasedonNN-PRmodel.

4.1. Optimizingtheneuralnetworkforpatternrecognition Any continuousfunctioncan beapproximated witha sin- glehiddenlayerBPnetwork.Ingenerally,anymappingmode (fromM-dimensionaltoN-dimensional)canberecognizedby athree-layerBPneuralnetworkbasedonKolmogorovtheory (Hanafizadeh,Zavasan,&Khaki,2010;Kolmogorov,1963).In this network,theinputlayer includestwo adsorptionfactors:

adsorption rateconstantandadsorption capacity,andtheout- putlayeristheadsorptioncomprehensiveperformancemode.

NNishighlysensitivetothenumberofhiddenlayerneurons.

Therearenotheoreticalformulasusedtocomputethenumber ofhiddencells,anditisoftenselectedbyexperientialEq.(2):

n=√

m+k+a (2)

wherenisthenumberofhiddennodes.mandkarethenumbers of inputnodesandoutput nodes,respectively;a isaconstant between1 and10.Accordingtoexperience,the morehidden nodesintheNN,thebetterpredictionresultwillbeobtained.

However,theNNconfigurationwillbemorecomplictated.The numberofhiddennodesshouldbenolessthanthetotalnumber ofinputandoutputnodes.AccordingtoEq.(2),thenumberof hiddennodesisdesignatedas4.

TheoptimizingNN-PRmodelisimplementedbyusingMat- labprogrammingsoftware.Thebasicparametersareshownin

Data preprocess Sample

acquisition

Feature transform

Neural network for pattern recognition Conventional

processing

Fig.3.Flowchartofneuralnetworkforpatternrecognition.

(5)

Table2

Basicparametersoftheneuralnetworkforpatternrecognition.

Basicparameters Debuggedvalue

Targeterror 0.001

Transferfunction Sigmoid

Nodesofhiddenlayer 4

Activationfunction Tansig

Weight 0.5

Bias 0.5

Momentumfactor 0.92

Numberofepochs 170

10¹ 10⁰

10^–1

10^–2

10^–3

10^–4

0 50 100 150 200 250 300

Train

Fig.4.Networktrainingerrorcurve.

Table2.Theparametersaredebuggedtoimprovetheaccuracy andspeedofconvergence.

Fig.4isthenetworktrainingerrorcurve.Thenetworkhas reachedaccuracyafter170timeswhichissetasthenumberof epochs.

4.2. Denotingtwogrades

Todistinguishdifferentmodesbasedonpatternrecognition, asetofinputsampledatacanbemappedtoanoutputvector.By usingtheadsorptionrateconstantandadsorptioncapacityastwo criterions(superiorclassandinferiorclass)andthecomparison standards,wecangetthefollowingclassifications:ForDBT,the datarange DBTq>575␮mol/g andK>0.0636h(␮mol/g)⁻¹ are superiorclass points(the output values are setas 1) and denotedasgradeI;othersareinferiorclasspoints(theoutput valuesare setas0) anddenotedas gradeII.ForQL,thedata rangeq>320␮mol/gandK>0.0462h(␮mol/g)⁻¹aresuperior classpoints, anddenoted as grade I;othersare inferiorclass pointsanddenotedasgradeII.ThereareeighttypesofACStyrene

(theadsorptionperformance)datearedistinguishedfordifferent

modes,andeachACStyrenedatasetinclude10fortrainingand 5fortesting.Thereare80trainingdataand40testingdataare usedfortheimplementationoftheNN(shownasTable3).For any unknown samples,the gradecanbe recognizedcorrectly accordingtowhatkindofpointiscloseto.

4.3. Recognizedresults

Therecognizedresultsofactivatedcarboncoatingmaterials adsorptioncomprehensiveperformanceareshowninTable3.

FromTable3,DBTadsorbabilityofactivatedcarboncoated withPVCishigherthantheoriginalactivatedcarbon.Forexam- ple, sample No. 1, DBT adsorbability increased from grade II (q is 560␮mol/g and K is 0.0527h(␮mol/g)⁻¹) to grade I (q is 582␮mol/gand K is 0.0639h(␮mol/g)⁻¹),while the adsorptionrateofQLdecreasedfromgradedItogradeII(qis 555␮mol/gandKis0.0438h(␮mol/g)⁻¹).Themainreasonfor theirchangesisthatPVCcoatedonthesurfaceofthespherical activatedcarbonisgoodforthediffusionofDBT,whilebadfor thediffusionofQL,whichmayberelatedtotheinteractionof functionalgroups,whichindicatesthat PVCfilmhasabetter permeabilitytoDBT.

Also,itcanbeseenfromTable3 thattheactivatedcarbon coatedwithpolyvinylalcoholcouldincreasetheadsorptionrate of QLanddecreasetheadsorptionrate ofDBT,whichshows thatPVAfilmhasabetterpermeabilitytoQL.

5. Comparativestudyofmoleculardynamicsimulations method

ToverifytheaccuracyandeffectivenessofNN-PR,theNN- PRmodelresultswerecomparedanddiscussedwiththeresults oftheMDmethod.

5.1. Moleculardynamicsimulations

Moleculardynamicssimulations(Frenkel&Smit,2001)are widelyusedforcomputersimulationsoflargeandcomplexcal- culatingsystems.ThebasicprincipleofMDsimulationistoset upanappropriatesystembasedonNewton’smotionlaws,sta- tistical thermodynamics andstatistical physicsprinciple, and to calculate the macroscopic physical properties through the trajectoryofthesystem.

In the past 20 years, with the rapid development of the molecularsimulationtechnique,thetechniquehasbecome an

Table3

Recognizeresultsofactivatedcarboncoatingmaterialsadsorptioncomprehensiveperformance.

No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

ACStyrene-DBT II II II I II II I II II I I II II I I

ACStyrene-PVA-DBT II II II II II II II II II II II II II II II

ACStyrene-PMA-DBT II II II II II II II II II II II II II II II

ACStyrene-PVC-DBT I I I I I II II II II I I I II I II

ACStyrene-QL I II II II II II I II II I II I II II II

ACStyrene-PVA-QL II I I II I I I II I I II I II I I

ACStyrene-PMA-QL II II II II II II I II II I II II II II II

ACStyrene-PVC-QL II II II II II II II II II II II II II II II

(6)

importanttooltostudythemolecularstructureofthepolymer membranematerialandthemechanismofadsorptionanddiffu- sioninthepolymermembrane.Thedevelopmentofmolecular dynamics simulations providedhuge impetus for researching themicroscopicmechanismsforpermeationofmoleculesdis- solvedinthepolymerfilm.MDwouldexplainthewayinwhich avarietyoffactorsaffectedtheseparationperformanceonthe molecularscale;whilstthemolecularsimulationtechniquecan provide afeasibilitystudy, processoptimizationand material propertiesscreeningpredictionbeforetheexperiment(Tamai, Tanaka,&Nakanishi,1995).

In this study, moleculardynamics simulation was used to investigate the diffusion behavior of the model diesel components inthreedifferentmembranes.Inaccordance withthe real diesel environment,DBT andQL were selected as rep- resentativesof sulfur andnitrogencompounds.The diffusion coefficientsofDBTandQLinthreedifferentmembraneswere calculatedbyEinsteinequation.Then,inordertogetandver- ifythemostfavorabledieseldesulfurizationanddenitrification membranesystems,acomparisonofthediffusionselectivityof thethreedifferentmolecularmembraneswasperformed.

5.2. Resultsofthemoleculardynamicssimulations

IntheMDmethod,diffusioncoefficientsareused tostudy diffusionaladsorptionperformanceofDBTandQLindifferent PVA,PVCandPMAmembranesystems.Meansquaredisplace- ment(MSD) curvesobtained by theMD methodare used to analyzeandcalculatethediffusioncoefficients.

Diffusionof moleculesinthe polymerfilmsystemcanbe calculatedbyMSDasshowninEq.(3)

MSD=

|r_i(t)−r_i(0)|²

(3) whereri(t)isthepositioncoordinateofthemoleculeiattpoint, andri(0)istheinitialpositioncoordinateofthemoleculei.

The relationshipbetweenthe diffusioncoefficient(D)and MSDcanbeexpressedasEq.(4)whichistheexponentialform ofEinsteinequation(Konduri&Nair,2007):

MSD=(2dD)t^α (4)

whereD is the diffusion coefficient;d is setas 3 andis the dimensionofmolecules’randommotion;αistheexponent;and tistime.

TheresultsforthediffusioncoefficientsareshowninTable4.

Toprovidescientificaltheoreticalsupport,thesingle-component system, which DBT and QL exist alone, and the double- component system, which DBT and QL exist together, are studiedtogether.

AccordingtoTable 4,when the DBTandQLexistalone, theDBTdiffusioncoefficientsarelowerthantheQLdiffusion coefficientsinthe PVA andPMA membrane systems; which meansthattherelativeconcentrationoftheDBTishigherthan thatofQL.Theyarethesameasinthedouble-componentPVA andPMAmembranesystems.WhileinthePVCmembranesys- tem,theDBTdiffusioncoefficientsarehigherthantheQL’sboth inthePVCsingle-componentorinthePVCdouble-component.

Table4

DiffusioncoefficientsofDBTandQLindifferentPVA,PVCandPAMmem- branesystems.

Membranesystems DDBT(D×10⁻⁹m²/s) DQL(D×10⁻⁹m²/s)

PVA(single-component) 0.068 0.182

PVA(double-component) 0.088 0.278

PVC(single-component) 0.210 0.147

PVC(double-component) 0.223 0.190

PAM(single-component) 0.143 0.250

PAM(double-component) 0.125 0.180

DDBTisdiffusioncoefficientinDBTmolecular;DQLisdiffusioncoefficientin QLmolecular.

Therefore,itcanbeconcludedthatthePVCisadvantageousfor removingsulfurcompoundsindieselfuel,whichisthesameas theresultoftheNN-PRmethod.

Also,accordingtothediffusioncoefficientofthemolecules inTable4,itcanbeseenthattheQLdiffusioncoefficientsare lower than the DBT diffusioncoefficients inthe PVC membrane systemswhen theDBT,or QL,existsalone;whichare thesameasinthedouble-componentPVCmembranesystems, whileinthePVAandPMAmembranesystems,theQLdiffusion coefficientsarehigherthanthoseoftheDBTs,whethertheyare insingle-componentorindouble-componentsolutions;thus,it canbeconcludedthatthePVAandPMAmembranesareadvan- tageousforremovingnitrogenouscompoundsindieselfuel.But, whichoneisbetterthanthatofthePVAorPMAmembrane?;

furthercomparisonsofthediffusionselectivityofthedifferent membranesareneeded.

In order to get the optimal denitrification membrane, a comparative study of the diffusion selectivity in the double- component solute systems of different molecular diffusion membranes is done by the MD method. The results of two typesofdiffusionselectivityareshowninFig.5,oneistheQL- DBTselectivityDDBT/DQL,andtheotheroneistheDBT-QL selectivityDQL/DDBT.

FromFig.5,itcanbeseenthatDDBT/DQLisgreaterthan1in thePVC;whileinPVAandPMA,DDBT/DQLislessthan1.Also, itcanbeseenthatDQL/DDBTisgreaterthan1inPVAandPMA membrane systems, especiallyfor the PVAmembrane, while

PAM PVC

PVA 0

1 2 3 4

Diffusion selectivity

DDBT/DQuinoline

DQuinoline/DDBT

Fig. 5.The diffusion selectivity of polymer membranes in the double- componentsolutesystems.

(7)

0 20 40 60 80 100 120 140 160 180 0.0

0.2 0.4 0.6 0.8 1.0

AC_Styrene-PVC-QL AC_Styrene-DBT AC_Styrene-QL AC_Styrene-PVC-DBT

C/C0

Amount of treated MDF (g-MDF/g-A)

Fig.6.DynamicadsorptionbreakthroughcurvesofDBTandQLofstyrene- basedactivatedcarbonspherescoatedwithPVC.

forthePVCmembrane,DQL/DDBT isless than1.Theresults suggestthat PVAisadvantageousfor dieseldenitrifying, and henceitwasselectedastheoptimaldenitrificationmembrane.

Again,itwasverifiedthatthePVCfilmwasadvantageousfor dieseldesulfurizing,whichfullyagreeswiththe resultof the NN-PRmethod.

6. Experimentalveriﬁcation

TofurthervalidatetheresultsoftheNN-PRmethod,theDBT andQLadsorption behaviorsof styrene-based spherical acti- vatedcarboncoatedwithPVCandPVAwereinvestigatedby experimentalresearch.Thebreakthroughcurveisthemaintool toresearchtheadsorptionbehaviorintheexperimentalmethod.

Byanalyzingfixedbeddynamicsadsorption experiments,the breakthroughcurvesobtainedwereasfollows:

6.1. BreakthroughcurvescoatedwithPVC

Dynamicadsorption breakthroughcurvesof DBTandQL ofstyrene-basedactivatedcarbonspherescoatedwithPVCare shownin Fig. 6. It can be seen that the adsorption volumes ofDBTandQLdecreasedatthepenetratingpointafterbeing coatedwithPVC.DBTis decreasedfrom77.0g-MDF/gA to 58.0g-MDF/gA,andQLisdecreasedfrom38.0g-MDF/gAto 7.0g-MDF/gA. Nevertheless, the penetrating time (from the penetratingpointtothesaturationpoint)ofQLislongerthan thatofDBT,whichindicatesthattheadsorptionrateofDBTwill increase,whiletheadsorptionrateofQLwilldecrease;hence, thePVCmembranecanenhancetheproliferationofDBT,which alsoshowsthat PVCisfit forremoving sulfurcompoundsin dieselfuel.

6.2. Dynamicadsorptionbreakthroughcurvescoatedwith PVA

Dynamicadsorption breakthroughcurvesof DBTandQL ofstyrene-basedactivatedcarbonspherescoatedwithPVAare showninFig.7.ItcanbeseenthatadsorptionvolumesofDBT andQLhavebeendecreaseddrasticallyatthepenetratingpoint afterbeingcoatedwithPVA.Also,thepenetratingtimeofDBT islongerthanthatofQL,whichindicatesthattheQLadsorp- tionrate is slightlyincreased, whilethe DBTadsorption rate

0.0 0.2 0.4 0.6 0.8 1.0

AC_Styrene-PVA-DBT AC_Styrene-QL AC_Styrene-DBT AC_Styrene-PVA-QL

C/C0

Amount of treated MDF (g-MDF/g-A)

0 20 40 60 80 100 120 140 160 180

Fig.7.DynamicadsorptionbreakthroughcurvesofDBTandQLofstyrene- basedactivatedcarbonspherescoatedwithPVA.

is significantly decreased.It was verified that PVA is fit for removingnitrificationcompoundsindieselfuel.

7. Conclusions

Inthispaper,thediffusionselectivitiesofDBTandQLinthe threepolymermembranesystems(PVA,PVC,andPMA) are investigatedbasedonNN-PR.Adsorptioncapacityconstantand adsorptionrateareselectedasdoublecriteriaandtwogradesare selectedasthecomparisonstandard.TheresultsshowthatPVC membranecanimprovethediffusionselectivityofDBT,which alsoshowsthatPVCmembraneisadvantageoustoremovesulfur compoundsindieselfuel,whilePVAmembranecanincreasethe diffusionselectivityofQLanditisadvantageousfortheremoval ofnitrogencompounds.

Then,MDisusedtostudythediffusionbehaviorofDBTand QL,anditsinfluencingfactors,inthreedifferentpolymermem- brane systems (PVA,PVC, andPMA). The sameconclusion canbedrawnbyinvestigatingthediffusioncoefficientandfree volumefraction ofdiffusionmoleculesatthemolecularlevel.

ThoughtheprincipleandmethodofNN-PRandMDaretotally different,thesameconclusionisreached,whichcanensurethe high accuracy of NN-PR. Also, NN-PR carries moreadvan- tages:ityieldsmoreintuitiveconclusionsanditisnotnecessary toestablishanadsorptionkineticsmodel.

BycomparingtheNN-PRmethodwiththeexperimentaldata, itcanbedeterminedthattheresultsoftheNN-PRmethodare consistentwith theexperimental results,whichalsoconfirms thattheNN-PRmethodcanbewellusedtopredictanddeter- minethediffusionbehaviorofthepolymermembranesinmodel diesel.Moreover,thisapproachcanbeusedasanalternativeto theexperimentalmethodtherebyreducingtherelatedcost.

Conﬂictofinterest

Theauthorshavenoconflictsofinteresttodeclare.

Acknowledgments

This work is partly supported by National Science Foun- dationof China(21177038),ChinaScholarshipCouncilFund (201406745031) and Material Informatics for Engineering Design Research Group of Woodruff School of Mechanical

(8)

Engineering, Georgia Institute of Technology, Atlanta, GA, USA. The authors would like to thank and the anonymous reviewersfortheirconstructivesuggestionsthathaveimproved thequalityofthiswork.

References

Cardoso,J.F.(1999).High-ordercontrastsforindependentcomponentanalysis.

NeuralComputation,11(1),157–192.

Chattoraj,S.,Mondal,N.K.,Das,B.,Roy,P.,&Sadhukhan,B.(2014).Carbaryl removalfromaqueoussolutionbyLemnamajorbiomassusingresponse surfacemethodologyandartificialneuralnetwork.JournalofEnvironmental ChemicalEngineering,2(4),1920–1928.

Chen,T.(2015).AnalyzingandforecastingtheglobalCO2 concentration– Acollaborativefuzzy-neuralagentnetworkapproach.JournalofApplied ResearchandTechnology,13(3),364–373.

Christopher,M.B.(1995).Neuralnetworksforpatternrecognition.pp.39–52.

NewYork:OxfordUniversityPress.

Chu,M.,&Plemmons,R.(2005).Nonnegativematrixfactorizationandappli- cations.BulletinofInternationalLinearAlgebraSociety,34,1–5.

Frenkel,D.,&Smit,B.(2001).Understandingmolecularsimulation:From algorithmstoapplications.pp.79–86.SanDiego:AcademicPress.

Guan,N.,Tao,D.,Luo,Z.,&Yuan,B.(2012).NeNMF:Anoptimalgradient methodfornonnegativematrixfactorization.IEEETransactionsonSignal Processing,60(6),2882–2898.

Hanafizadeh,P.,Zavasan,R.A.,&Khaki,H.R.(2010).Anexpertsystem forperfumeselectionusingartificialneuralnetwork.ExpertSystemswith Applications,37(12),8879–8887.

Jain,A.K.,Murty,M.N.,&Flynn,P.J.(1999).Dataclustering:Areview.ACM ComputingSurveys(CSUR),31(3),264–323.

Kalidindi,S.R.,&Graef,M.D.(2015).Materialsdatascience:Currentstatus andfutureoutlook.AnnualReviewofMaterialsResearch,45,171–193.

Kolmogorov,A.N.(1963).Ontherepresentationofcontinuousfunctionof manyvariablesbysuperpositionofcontinuousfunctionsofonevariableand addition.AmericanMathematicalSocietyTranslation,28,55–59.

Konduri,S.,&Nair,S.(2007).Acomputationalstudyofgasmoleculetrans- portinapolymer/nanoporouslayeredsilicatenanocompositemembrane material.TheJournalofPhysicalChemistryC,111(5),2017–2024.

Kou,X.Y.,Parks,G.T.,&Tan,S.T.(2012).Optimaldesignoffunctionally gradedmaterialsusingaproceduralmodelandparticleswarmoptimization.

TheJournalofComputer-AidedDesign,44,300–310.

Kowalski,B.R.,&Bender,C.F.(1972).Patternrecognition.Powerfulapproach tointerpretingchemicaldata.JournaloftheAmericanChemicalSociety, 94(16),5632–5639.

Kramer,M.A.(1991).Nonlinearprincipalcomponentanalysisusingautoas- sociativeneuralnetworks.JournaloftheAmericanInstituteofChemical Engineers,37(2),233–243.

McCulloch,W.,&Pitt,W.(1943).Alogicalcalculusoftheideasimmanent.

BulletinofMathematicalBiology,5(4),115–133.

Nianyi,C.,Wencong,L.,Ruiliang,C.,Chonghe,L.,&Pei,Q.(1999).Chemo- metricmethodsapplied toindustrialoptimizationand materialsoptimal design.ChemometricsandIntelligentLaboratorySystems,45,329–333.

Panchal,J.H.,Kalidindi,S.R.,&McDowell,D.L.(2013).Keycomputa- tionalmodelingissuesinintegratedcomputationalmaterialsengineering.

TheJournalofComputer-AidedDesign,45(1),4–25.

Papadopoulos,V.D.,Beligiannis,G.N.,&Hela,D.G.(2011).Combining experimentaldesignandartificialneuralnetworksforthedeterminationof chlorinatedcompoundsinfishusingmatrixsolid-phasedispersion.Applied SoftComputing,11(8),5155–5164.

Perlovsky,L.(2001).Neuralnetworksandintellection.pp.150–168.NewYork:

OxfordUniversityPress.

Porrazzo, R.,Cipollina, A.,Galluzzo, M.,& Micale, G.(2013). Aneural network-basedoptimizingcontrolsystemforaseawater-desalinationsolar- poweredmembranedistillationunit.Computers&ChemicalEngineering, 54,79–96.

Sano,Y.,Choi,K.H.,Korai,K.,&Mochida,I.(2004).Effectsofnitrogen andrefractorysulfurspeciesremovalonthedeepHDSofgasoil.Applied CatalysisB:Environmental,53(3),169–174.

Shabani,M.O.,&Mazahery,A.(2012).ArtificialIntelligenceinnumericalmod- elingofnanosizedceramicparticulatesreinforcedmetalmatrixcomposites.

AppliedMathematicalModelling,36(11),5455–5465.

Song,C.,&Ma,X.(2003).Newdesignapproachestoultra-cleandieselfuels bydeepdesulfurizationand deepdearomatization.AppliedCatalysisB:

Environmental,41(1),207–238.

Srinivasan, S., & Saghir, M. Z. (2013). Modeling of thermotransport phenomenoninmetalalloysusingartificialneuralnetworks.AppliedMath- ematicalModelling,37(5),2850–2869.

Tamai,Y.,Tanaka,H.,&Nakanishi,K.(1995).Moleculardesignofpolymer membranesusingmolecularsimulationtechnique.FluidPhaseEquilibria, 104,363–374.

Wold,S.,&Sjostrom,M.(1977).Chemometrics:Theoryandapplication,ACS symposiumseries,52(pp.243–282).Washington,DC:AmericanChemical Society.