213种溶剂表面张力+粘度+热导率

408J.Chem.Inf.Comput.Sci.2001,41,408-418

PredictionofSurfaceTension,Viscosity,andThermalConductivityforCommon

OrganicSolventsUsingQuantitativeStructure-PropertyRelationships

GregoryW.KauffmanandPeterC.Jurs*

DepartmentofChemistry,ThePennsylvaniaStateUniversity,152DaveyLaboratory,

UniversityPark,Pennsylvania16802

ReceivedOctober2,2000

Predictivemodelsforthesurfacetension,viscosity,andthermalconductivityof213commonorganicsolventsarereported.Themodelsarederivedfromnumericaldescriptorswhichencodeinformationaboutthetopology,geometry,andelectronicsofeachcompoundinthedataset.Multiplelinearregressionandcomputationalneuralnetworksareusedtotrainandevaluatemodelsbasedonstatisticalindicesandoverallroot-mean-squareerror.Eight-descriptormodelsweredevelopedforbothsurfacetensionandviscosity,whileanine-descriptormodelwasdevelopedforthermalconductivity.Inaddition,asinglenine-descriptormodelwasdevelopedforpredictionofallthreeproperties.Theresultsofthisstudycomparefavorablytopreviouslyreportedpredictionmethodsforthesethreeproperties.

INTRODUCTION

Solventpropertiesareimportantinmanyareasofscientificresearch.ArecentbookbyAshetal.categorizedover1700organicliquidsdeemed“solvents”intotheirrespectiveareasofapplication.1Theindustriesrepresentedinthiscompilationincludetextile,agricultural,automotive,chemicalprocessing,cosmetic,inkmanufacturing,petroleum,andpharmaceutical.Inaddition,theselectiveuseofsolventsinacademicenvironmentsispracticedonadailybasisbyscientistsofeverydiscipline.

Surfacetension(σ),viscosity(η),andthermalconductivity(λ)areclassifiedassurfaceandtransportpropertiesofaliquid.Themagnitudesofthesepropertiesaredependentuponintermolecularinteractionsbetweenthesolventmol-ecules.Theseinteractionsdictatethesurfaceelasticityandtheflowandthemiscibilityoftwophasesaswellastheabilityoftheliquidtodissipatetemperaturefluctuationsduringareactionprocess.Inthecurrentwork,anattemptismadetousenumericaldescriptorstoencodefundamentalfeaturesof213commonorganicsolventsbasedsolelyonmolecularstructureandtoidentifythosefeatureswhichbestmodelthesurfacetension,viscosity,andthermalconductivityofeachsolventatasingletemperature.Ourquantitativestructure-propertyrelationship(QSPR)approachtomodelbuildingusesmultiplelinearregression(MLR)analysisandcomputationalneuralnetworks(CNNs).Eachmodelgener-atedisthenvalidatedusinganexternalpredictionsetofcompounds.

QSPRshavebeenreportedforavarietyofphysicalpropertieswhichhavebeenrecentlyandthoroughlyre-viewed.2Inparticular,manyrecentQSPRsforthepredictionofsurfacetensionandviscosityhaveappearedintheliterature.TwopapershavereportedQSPRstudiesforsurfacetension.Thefirst,byStantonandJurs,3employedtheADAPTsoftwarepackage4,5tofinda10-descriptormultiple

*Correspondingauthorphone:(814)865-3739;fax:(814)865-3314;e-mail:pcj@psu.edu.

linearregression(MLR)modelforalimitedsetof166alkanes,aliphaticesters,andalcohols.Thesecond,byKavunandco-workers,6employedESMA7tofindaseven-descriptorMLRmodelforadiversebutasmalldatasetof81compounds.

Since1997,fourpapersreportingQSPRmodelsforviscosityhavebeenpublished.Thefirst,reportedbySuzukiandco-workers,8describedaquantitativeproperty-propertyrelationshipforadiversesetof361compoundsusingthesoftwarepackageChemProp.9Nine-descriptorMLRandCNNmodelsusingfourphysicochemicalpropertiesandfiveindicatorvariablesweredeveloped.Aspointedoutinsubsequentpapers,theuseofphysicalpropertiesforthepredictionofanotherpropertycanhaveseriouslimita-tions.10,11TheremainingthreepapersemployedCODESSA12todevelopeitherMLRmodelsorPLSregressionmodels.Ivanciucandco-workers10developedafive-descriptorMLRmodelforadiversesetof337compoundsusingaleave20%outcross-validationmethod.Similarly,Katritzkyandco-workers11developedafive-descriptormodelforasetof361compoundsusingaone-thirddatasubsetcross-validation.AnothermodelwasreportedbyCocchiandco-workers.13Usingasmallsetof46compounds,a16-variablefour-componentPLSregressionmodelwasdeveloped.TheresultsofallthreeCODESSA-basedmodelswereconsistentwithoneanotherandemployedsimilardescriptors.

Thecommonthemeamongeachofthemodelspresentedintheliteratureforpredictingsurfacetensionandviscosityistheuseofdescriptorswhichencodeinformationabouthydrogen-bondingandpolarsurfaceinteractions.Thismakessenseduetotheinfluenceofsuchelectronicfeaturesinreal-systemintermolecularinteractions.Therefore,itwasexpectedthatseveralsuchdescriptorswouldappearinthemodelsgeneratedinthepresentwork.NoliteraturereportsofQSPRsforthermalconductivitycouldbefound;however,thecloserelationshiptoviscositywouldallowonetoconfidentlyassumethatsimilardescriptorswouldbeusefulforthepredictionofbothproperties.

10.1021/ci000139tCCC:$20.00©2001AmericanChemicalSociety

PublishedonWeb03/02/2001

第1页下一页