Science Arts & Métiers (SAM)
is an open access repository that collects the work of Arts et Métiers Institute of
Technology researchers and makes it freely available over the web where possible.
This is an author-deposited version published in:
https://sam.ensam.eu
Handle ID: .
http://hdl.handle.net/10985/18396
To cite this version :
Amine AMMAR, Francisco CHINESTA - Direct numerical simulation of flexible molecules and
datadriven molecular conformation Comptes Rendus Mecanique Vol. 347, n°11, p.743753
-2019
Any correspondence concerning this service should be sent to the repository
Administrator :
archiveouverte@ensam.eu
Direct
numerical
simulation
of
flexible
molecules
and
data-driven
molecular
conformation
Amine
Ammar
a,
∗
,
Francisco
Chinesta
baLAMPA@ArtsetMétiersParisTech,2,boulevardduRonceray,BP93525,49035Angerscedex01,France
bESIGroupChair@PIMM,ArtsetMétiersInstituteofTechnology,CNRS,CNAM,HESAMUniversity,151boulevarddel’Hôpital,75013Paris, France
a
b
s
t
r
a
c
t
Keywords:
Flexiblemolecules Directnumericalsimulation Moleculardynamics Suspension Deep-learning
The present work aims at performing a molecular dynamics modeling of suspensions composedofflexiblelinearmolecules.Moleculesarerepresentedbyaseriesofconnected beads,whosedynamicsisgovernedbythreepotentials:theextensionpotentialaffecting theelongationofsegmentsconnectingconsecutivebeads,theonegoverningthemolecule bendingandfinallytheLennard-Jonesdescribingtheinteractionofnon-consecutivebeads. Apopulationofnon-interactingmoleculesissimulatedinelongationandshearflowsfor different flow and molecule parameters. The flow-induced conformation is analyzed in the different considered situations. Finally a model for predicting the evolution ofthe populationconformationwillbeobtainedbyusingdeep-learning.
1. Introduction
Polymersandviscoelasticfluidscanbedescribed atdifferentscales,deeplydiscussedinmanybooks,amongthem [1] [2]. The molecularscale (see [3] [4] and the referencestherein) precedesthe one ofkinetic theory wherethemolecules individualityisreplacedbyascalardistributionfunction–DF–involvingaseriesofconformationalcoordinates[2].Most ofkinetic theory modelsrepresentmolecules asmulti-bead-rod ormulti-bead-springs idealizations,usually coarsenedby using theend-to-end vector to reduce thedimensionality ofthe conformational space[5]. The resulting modelswere in generalsolvednumericallybyusingBrownianandstochasticnumericaltechniquesforcircumventingtheso-calledcurseof dimensionalitythatthehigh-dimensionalityoftheconformationspaceentails[6] [7].
In[8] and[9] authorsconsiderGaussianchainscombinedwiththeLangevinequationforbetteraccountforthemolecular conformation.Worksattainingtheatomicscalearequitescare,beingtheresultingmodelsdifficulttoextendtolargerscales [10–14].
The presentworkaims atperforming a moleculardynamics modelingofsuspensions composedof flexiblemolecules. Suchadescriptionscalehasasmainadvantagethefinerepresentationofthemolecularconfigurations,finerthantheones attainablewhenusingcoarserdescriptions(e.g.kinetictheorydescriptions).Themaindrawbackwhenusingtheseextremely finedescriptions concerns thecomputational cost.However, tohaveaccessto therealbehavior andevaluate thevalidity of coarserdescriptions, high fidelitysimulations atthe finestscale seems compulsory. Moreover, assoon as an accurate approximationofthemicroscopicphysicsisattainedandavailable,itcouldbethenintroducedintocoarserdescriptions.
*
Correspondingauthor.Fig. 1. Molecule schema.
When describingfinely moleculesasa seriesofconnectedbeads,thewholemolecular conformationstronglydepends on thechoiceofthepotentialsusedforderiving theforcestobe introducedintothemotionequation atthebeadslevel. Different potentials are in general considered: (i) the one describing the extension of the inter-beads segment; (ii) the one describing themolecule bending mechanism,that is expressedateach beadfromits two neighbors;andfinally (iii) the weak interaction (Lennard-Jones) between non-neighborbeads. The evaluation of the population conformation with respecttothedifferentavailablechoiceofpotentialsandinparticulartheirintensityisofmajorinteresttobetterdescribe rheological features. Inthepresentwork andwithoutloss ofgeneralityof theproposed methodologywe considerdilute suspensions,thatallowsneglectingtheinteractionofbeadsbelongingtodifferentmolecules.
Our modelassumes linearmolecules composed ofbeads atwhich, other than the potentialsjustdiscussed, also acts Brownian events representingthe action ofthe moleculesof solventin which themolecule is assumedbe immersed. A population ofnon-interacting molecules is simulated in elongation and shear flows for different flow and molecule pa-rameters. The flow-induced conformation is analyzed in the different considered situations and some statistical outputs extracted.
Finally a modelfor predicting theevolution of the populationconformation willbe obtainedby using deep-learning, deeplyusedinavarietyofapplications,assystemcontrol[15,16] anddynamicalsystems[17] forcitingfew.
2. Fine-scalediscretesimulation
Thefinescalesimulationisperformedbyrepresentingthemoleculeasaseriesofconnectedbeadswhereforcesapply, thelastderivedfromdifferentpotentialsaswellasthebombardmentofsolventmolecules.
Aspreviously discussedweareconsideringthreedifferentpotentialsresponsibleofthreedifferentdeformation mecha-nisms:theLennard-Jonespotential VLJ isusedtodescribetheinter-beadsinteractions (ofnon-neighborbeads), andother
two potentials, VE and VB,are used to describe the molecule segments elongationsand chainbending respectively, see
Fig.1.
Thus,thetotalpotentialreads
V
=
M j=1 N−1 i=1 ViE,i+1;j+
M j=1 N−1 i=2 ViB−1,i,i+1;j+
M j=1 N i=1 N k=1 ViLJ,k;jχ
ik;j (1)where
M
isthenumberofmoleculesinthepopulation, j referstothemoleculechaininthatpopulation,andi referstothe consideredbeadcomposingit,i=
1,
. . . ,
N
(withN
thenumberofbeadscomposingthemolecularchain).Finallywedefine acharacteristicfunctionfordescribingthebeaddirectneighborhood:χ
ik=
1 if|
i−
k|
>
1 andχ
ik=
0 if|
i−
k|
≤
1TheLennard-Jonespotentialreads
ViLJ,k
=
4ε
σ
dik 12−
σ
dik 6 (2)withdik
= ||
xk−
xi||
thedistancebetweenbeadsB
i andB
k,located atpositions xi andxk respectively, andand
σ
the twousualparametersinvolvedintheLennard-Jonespotential.Theelongationpotentialisgivenby
ViE,i+1
=
KE 2 1−
di j deq 2 (3)wheredeq istheequilibriumdistancebetweentwosuccessivebeads,distanceatwhichthepotentialreachesits minimum
(andconsequentlytheassociatedforcevanishes).Inthepreviousexpression KE reflectsthepotentialintensity.
Thebendingpotentialbetweenthreesuccessivebeads
ViB−1,i,i+1
=
KB 2θ
i−1,i,i+1− θ
eq 2 (4)where
θ
i−1,i,i+1 istheangledefinedbyvectorsjoiningbeads i−
1,
i andi,
i+
1 withθ
eq theangleatequilibrium. Inourstudyweconsider
θ
eq=
70.
5◦ (thatcorrespondstotheanglebetweentwoconsecutivecarbone-carbonebonds).Again,thepotentialintensityisdescribedfromKB.
The force applyingat each bead related to thesethree molecule deformation mechanisms results from the potential derivativeaccordingto
FIi
= −
∂
V∂
xi(5)
Therearetwootherforcesactingonthebead.First,theonerelatedtothefluiddragFDi.Letv
(
x)
bethefluidvelocityat positionx,assumedunperturbedbytherodspresence,thedragforcereadsFiD
= ξ(
v(
xi)
− ˙
xi)
(6)wherexi
˙
denotesthevelocityofbeadB
i.FinallyaBrownianforceisalsoconsideredemulatingthesolventbombardment,andisexpressedfrom
FBrowi
=
√
2D dt Wi (7)whereD denotesthediffusioncoefficient,dt thecalculationtimestepandWiarandomvariablewithzeromeanandunit standarddeviation.
Thelinearmomentumbalanceinvolvingbead
B
ireadsFi
=
FDi+
FiI+
FBrowi=
m ai (8)withtheparticleaccelerationai
=
ddtx˙i=
d2xi
dt2.
Asecond-ordertimeintegrationschemeisconsidered,consistingof
⎧
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎩
ani+1=
Fni m xni+1=
xni+ ˙
xnit
+
12ai(
t)(
t)
2˙
xni+1= ˙
xni+
12ani
+
ani+1t (9)
wherethesuperscriptn referstothetimestep,tn
=
nt.
Priortoproceedwiththetimeintegration,equationsarerewrittenindimensionlessform,byconsideringascharacteristic variables – length:
σ
– energy:ε
– mass:m – velocity:√
ε
/
m – acceleration:ε
/(
mσ
)
– time:σ
√
m/
ε
– diffusion:σ
−35 mwhichallowswritingthedimensionlessmodelinwhich,forthesakeofclarity,wedonotchangethenotationfor designat-ingthedimensionlessvariables.
Thus,weobtain
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎩
ViLJ,j=
4d1 i j 12
−
1 di j 6 ViE,i+1
=
KˆE 21
−
di j deq 2 VB i−1,i,i+1=
ˆ KB 2(θ
i−1,i,i+1− θ
eq)
2 FiI= −
∂∂xV i FiD= ˆξ(
v(
xi)
− ˙
xi)
FBrowi=
2Dˆ
dt W i Fi=
ai (10) whereKˆ
E=
Kε ,E Kˆ
B=
Kε ,Bˆξ = ξ√
ε
m/
σ
,Dˆ
=
Dσ
3 m 5 anddt=
dt σ m.Fig. 2. Initial conformation and its evolution when subjected to elongation. 3. Results
Inthe numericalresultsreportedinthepresentsection weassumedasequilibriumdistancebetweentwoconsecutive beadstheoneassociatedwiththeLennard-Jones(dimensionless)potential,i.e.deq
=
21/6.Inordertoavoidlargefluctuationsofthemoleculelengthweassumeaquitelargeelongationstiffness K
ˆ
E=
10000 andabendingpotential Kˆ
B=
1000.Inordertoderiverheologicalproperties(basedonthemolecularconformation),alargeenoughpopulationofmolecules must be considered, allowing one to compute accurate statistics. In the cases reported in the present paper
M
=
1000 chainsinvolvingN
beads,N
=
100,andconsequently99 inter-beadsegments areconsidered.Inallcasesthedimensionless diffusionanddragcoefficientswereassumedtakingunitvalues.Twosimplerheologicalflowswereaddressed,elongationalandsimpleshearflows,bothexpressedfromthe dimension-lesseffectivestrainrate
˙
,leadingtothefollowingdimensionlessvelocityfields:– elongationalflow, v
=
⎛
⎝
vvxy vz⎞
⎠ =
⎛
⎜
⎝
˙
x
−
1 2˙
y−
1 2z
˙
⎞
⎟
⎠
(11)– simpleshearflow,
v
=
⎛
⎝
vvxy vz⎞
⎠ =
⎛
⎝
z
˙
0 0⎞
⎠
(12)wherethreevaluesof
˙
willbeevaluated:(i)˙
=
5;(ii)˙
=
1,and(iii)˙
=
0.
2.Theinitialstateconsistsofthe
M
=
1000moleculesgeneratedrandomly.Forthatpurposethefirstbeadofeachchainis placeatpointx1.Thenthesecondbeadisrandomlyplacedatpointx2 byensuringx2−
x1=
deq.The thirdbeadisplaceatpointx3ensuringnowthetwoconditions:(i)
x3−
x2=
deqand(ii)L12L23= θ
eq,withLi,i+1=
xi+1−
xi.Theprocedure continuesuntilplacingthefinalbead.Byproceedinglikethis,thereistheriskthatbeadsinthesamemoleculeapproachtoomuch.Thus,beforeapplyingthe flow,themoleculespopulationissubjectedtotheLennard-Jonespotentialtoensurethefulfillmentoftheexcludedvolume. Fig.2depictstheinitialmoleculeconformationandthen,theonethatresultsfromitselongation.
3.1. Casestudies
Inwhatfollowsdifferentanalysesareperformedbyvaryingtheflowanditsintensity(elongationandsimpleshear,with
˙
=
5 and˙
=
0.
2).Eachsimulationleadstoafigurecomposedof8subfigures,each onerepresenting(toptobottom,left toright):– representation ofthe molecules population wherethe molecule barycentre is placed at theorigin ofthe coordinate system;
– averageofthedifferentmoleculesegmentsjoiningconsecutivebeads(from1to
N
−
1); – theorientationtensorcalculatedbydefininguij,i+1
=
L j i,i+1 Lij,i+1(13)
where j referstheconsideredmoleculeinthepopulation, j
=
1,
. . . ,
M,
andi theconsideredbead,i=
1,
. . . ,
N
−
1,from whichtheorientationtensorwritesS
=
1M
M j=1 N−1 i=1 uij,i+1⊗
uij,i+1 (14)– theconformationtensorC computedfromtheend-to-endvectorL1j,N accordingto
C
=
1M
M j=1 L1j,N⊗
L1j,N (15)– the distributionofuij,i+1 ontheunit sphere(notedasDFinthefigure),eachvector beingassociatedwithaGaussian distributionforsmoothingtherepresentation;
– asimilarrepresentationbutnowconcerningonlytheend-to-endvectorsu1j,N; – averageofthesegmentsofallmoleculeswiththeassociatedstandarddeviation;
– averageofsegmentslocatedatthemoleculeendsandatthecenterwiththeirassociatedstandarddeviations.
3.2. Discussion
Figs.3and4concernelongation.Itcanbenoticedthatcentralsegmentsareveryelongated,withafinallengththatat highelongationdoublestheoneofsegmentslocatedatthemoleculesextremities.Whendecreasingtheelongationintensity central segments become lessextended,andthe end-to-end conformationremains alignedwiththe elongation, however the segments orientation distribution exhibits a particular annular distribution whose radius depends on the elongation intensity.
Figs.5and6concernshearflow.Segmentshaveapproximatelythesamelengthatthealmoststeadystate.Anovershoot is noticed at the central segments during the transient regime. At highshear rates the distribution concentrate at two regions(4bysymmetry)thatattenuateswhendecreasingtheshearrate.
3.3. Derivingamodeloftheconformationevolution
Asdiscussedintheintroductionsection directnumericalsimulations areveryexpensivefromthecomputationalpoint ofview,andinthatcaseequationsdescribingtheconformationevolutiondepending ontheconformationandtheapplied flowkinematicsseemspreferable.
Asthiskindofmodelisquitedifficulttoobtain,hereweproposeusingdeep-learningtechniquesforthatpurpose.The conformationevolutionisassumedevolvingaccordingto
dC
dt
=
f(
C,
∇
v)
(16)withthevelocitygradientgivenby
∇
v=
⎛
⎜
⎝
∇
v11∇
v12 0 0−
∇v11 2 0 0 0−
∇v11 2⎞
⎟
⎠
(17)Fig. 3. Elongation at˙=5.
Weelaboratedaneuralnetwork–NN–involvingtwohiddenlayersof20neurons,beingtheinputsthe6components ofthe(symmetric)conformationtensorandthetwocomponentsofthevelocitygradient
∇
v11and∇
v12,beingtheoutputsthesixcomponentsoftimederivativeoftheconformationtensor.
Werestricttomoleculescomposedof20beads,i.e.
N
=
20,withthelowerbendingstiffness Kˆ
R=
100.600flowinducedconformations aresimulatedbychoosingrandomlyboth gradientofvelocitycomponents,accordingwiththe 101.4(ω−0.5), where
ω
isarandomvariableuniformlydistributedintheinterval[
0,
1]
.Thischoiceallowsensuringvaluesoftheeffective strainrateintheinterval(
0.
2,
5)
,coveredalmostuniformlyinthelogarithmicscale.Fig. 4. Elongation at˙=0.2.
When theneural network extractthe internal weights,the solutionis computedagainfordifferentvalues ofvelocity gradient,randomly chosen,inorder tocomparethe directnumericalsolutionwiththeneural-network predictions.Fig.7 comparesboth conformationpredictions.The conformationalpredictions arequalitativelyinagreementwiththeones de-rivedfromdirectnumericalsimulations,howeverquantitativelysomegapsarenoticed.Forreducingthem,differentroutes exist: consideralternative networks, increasing thenumber ofhidden layers orthe numberof nodes(neurons) inthose layers.However,thosechoicesaredetrimentalwithrespecttothetrainingefficiency.AdeepercomprehensionofNNseems necessaryforbetterrepresentingthesubjacentphysics,tobetterperformpredictionswhilekeepingasreducedaspossible
Fig. 5. Shear at˙=5.
the trainingneeds (it isimportanttonote thatdirect numericalsimulationsthat servefortraining theNNareextremely expensivefromthecomputationalviewpoint).
4. Conclusion
In this paper, we proposed a direct numerical simulation of a suspension of non-interacting flexible molecules, and evaluated theflowinduced conformation.Simulationsprovethattheconformationcanexhibitunexpectedfeatures,asthe annulardistributionofthemolecularsegmentsorientation.
Fig. 6. Shear at˙=0.2.
Thesecondcontributionofthepresentworkconcernsthederivationofanevolutionequationfortheso-called conforma-tiontensor,that isexpecteddescribingtherheologicalfindings,byusingartificialintelligencetechniques,andinparticular deep-learning.Ithasbeenprovedthatthelearnedmodelallowsderivingveryaccuratepredictionsforflowsdifferingfrom theonesconsideredinthelearningstage(neuralnetworkconstruction).
The obtentionof richerexpressions of theconformation evolutionin generaltransient andcomplexflows, as well as the conformation / flow kinematics coupling through an adequate constitutive equation, constitute the main works in progress.
Fig. 7. Comparing direct numerical solutions (left) and neural-network based prediction (right). Acknowledgements
TheauthorsgratefullyacknowledgetheANR(“Agencenationaledelarecherche”,France)foritssupportthroughproject AAPG2018DataBEST.
References
[1]C.Binetruy,F.Chinesta,R.Keunings,FlowsinPolymers,ReinforcedPolymersandComposites.AMultiscaleApproach,SpringerBriefs,Springer,2015.
[2]F.Chinesta,E.Abisset,AJourneyAroundtheDifferentScalesInvolvedintheDescriptionofMatterandComplexSystems,SpringerBriefs,Springer, 2017.
[3]G.Ianniruberto,G.Marrucci,Flow-inducedorientationandstretchingofentangledpolymers,Philos.Trans.R.Soc.Lond.A361(2003)677–688.
[5]R.B.Bird,C.F.Curtiss,R.C.Armstrong,O.Hassager,DynamicsofPolymericLiquids,Vol.2:KineticTheory,JohnWiley&Sons,1987.
[6]M.Somasi,B.Khomami,N.J.Woo,J.S.Hur,E.S.G.Shaqfeh,Browniandynamicssimulationsofbead-rodandbead-springchains:numericalalgorithms andcoarse-grainingissues,J.Non-Newton.FluidMech.108 (1–3)(2002)227–255.
[7]G.Venkiteswaran,M.Junk,A QMCapproachfor highdimensionalFokker-Planckequationsmodellingpolymericliquids,Math.Comput.Simul.68 (2005)43–56.
[8]R.Winkler,P.Reineker,L.Harnau,Modelsandequilibriumpropertiesofstiffmolecularchains,J.Chem.Phys.101 (9)(1994)8119–8129.
[9]L. Harnau,R. Winkler, P. Reineker, Dynamicstructure factor ofsemiflexible macromolecules indilute solution,J. Chem. Phys. 104 (16)(1996) 6255–6258.
[10]W.Paul,G.Smith,D.Yoon,Staticanddynamicpropertiesofan-C100H202meltfrommoleculardynamicssimulations,Macromolecules30(1997) 7772–7780.
[11]T.Kreer,J.Baschnagel,M.Mueller,K. Binder,MonteCarlosimulationoflong chainpolymermelts:crossover fromRousetoreptationdynamics, Macromolecules34(2001)1105–1117.
[12]S.Krushev,W.Paul,G.Smith,Theroleofinternalrotationalbarriersinpolymermeltchaindynamics,Macromolecules35(2002)4198–4203.
[13]M.Bulacu,E.vanderGieesen,Effectofbendingandtorsionrigidityonself-diffusioninpolymermelts:amolecular-dynamicsstudy,J.Chem.Phys. 123(2005)114901.
[14]M.O.Steinhauser,Staticanddynamicscalingofsemiflexiblepolymerchains–amoleculardynamicssimulationstudyofsinglechainsandmeltsMech, Time-Depend.Mater.12(2008)291–312.
[15]M.Tanaskovic,L.Fagiano,C.Novara,M.Morari,Data-drivencontrolofnonlinearsystems:anon-linedirectapproach,Automatica75(2017)1–10.
[16]T.Duriez,S.Brunton,B.R.Noack,MachineLearningControl-TamingNonlinearDynamicsandTurbulence,Springer,2017.