Optimal prediction in molecular dynamics

Optimal prediction approximates the average solution of a large system of ordinary differential equations by a smaller system. We present how optimal prediction can be applied to a typical problem in the field of molecular dynamics, in order to reduc…

Authors: Benjamin Seibold

Optimal prediction in molecular dynamics
Optimal Predition in Moleular Dynamis Benjamin Seib old Departmen t of Mathematis T e hnial Univ ersit y of Kaiserslautern Gottlieb-Daimler-Straÿe, 67653 Kaiserslautern, German y Email: seiboldmathematik.uni-kl. de Abstrat Optimal predition appro ximates the a v erage solution of a large system of ordinary dieren tial equations b y a smaller system. W e presen t ho w optimal predition an b e applied to a t ypial problem in the eld of moleular dynamis, in order to redue the n um b er of partiles to b e tra k ed in the omputations. W e onsider a mo del problem, whi h desrib es a surfae oating pro ess, and sho w ho w asymptoti metho ds an b e emplo y ed to appro ximate the high dimensional onditional exp etations, whi h arise in optimal predition. The th us deriv ed smaller system is ompared to the original system in terms of statistial quan tities, su h as diusion onstan ts. The omparison is arried out b y Mon te-Carlo sim ulations, and it is sho wn under whi h onditions optimal predition yields a v alid appro ximation to the original system. Keyw ords: Optimal predition, moleular dynamis, surfae oating, hopping, Laplae's metho d, lo w temp erature asymptotis, Mon te-Carlo 1 In tro dution Computations in the eld of moleular dynamis t ypially require a large omputational eort due to t w o fators: 1. Small time steps are required to resolv e the fast atomi osillations. 2. Large systems are obtained due to the large amoun t of atoms whi h ha v e to b e omputed. A wide v ariet y of metho ds has b een dev elop ed to remedy these problems. Larger time steps are admitted e.g. b y smo othing algorithms, whi h a v erage in time o v er the fast osillations. V arious other metho ds redue the degrees of freedom, e.g. m ultip ole metho ds [14℄ in the on text of long range partile in terations. In this pap er w e in v estigate whether the metho d of optimal predition, as presen ted and analyzed in [1, 4 11, 15, 19℄, an in priniple b e applied to problems in moleular dynamis in order to redue the n um b er of atoms to b e omputed. As a rst step in this in v estigation w e onne to a one dimensional mo del problem whi h inherits partiular prop erties from a real moleular dynamis problem. In Setion 2 w e presen t the real problem as it arises in appliations, and deriv e the simplied mo del problem. The onsidered problem is Hamiltonian, hene in Setion 3 w e presen t the metho d of optimal predition in the sp eial ase of Hamiltonian systems. In Setion 4 1 2 PR OBLEM DESCRIPTION 2 w e apply the metho d of optimal predition to the mo del problem. This yields expressions in v olving high dimensional in tegrals. W e ev aluate these in tegrals b y asymptoti metho ds, emplo ying the fat that the pro ess is running at a lo w temp erature, whi h yields a new and smaller system. Setion 5 deals with the n umerial sp eed-up. In Setion 6 w e dene riteria, ho w to in v estigate whether the new system is a v alid appro ximation to the original system. W e sho w ho w imp ortan t statistial quan tities, su h as diusion onstan ts and energy utuations, an b e obtained b y n umerial exp erimen ts. These are presen ted in Setion 7, and w e in v estigate under whi h onditions optimal predition preserv es the relev an t statistial quan tities. 2 Problem Desription 2.1 The Ph ysial Problem In the pro dution of semiondutors a thin la y er (a few atomi monola y ers) of opp er has to b e oated (sputtered) on to a silion rystal. A te hnial desription of the pro ess an b e found in [18℄. Imp ortan t for the qualit y of the pro dut is that opp er atoms m ust not p enetrate to o deeply in to the silion rystal. The opp er atoms p enetrate rstly b y their kineti energy when hitting the rystal surfae, seondly b y the pro ess of atomi hopping , whi h will b e desrib ed in the follo wing. In order to obtain sp ei kno wledge ab out these pro esses, moleular dynamis sim ulations ha v e to b e arried out, as desrib ed in [13℄. One imp ortan t asp et of the oating pro ess is that the system is out of its thermo- dynamial equilibrium only for v ery short times, namely for ab out 10 − 11 seonds after one opp er atom has hit the surfae of the silion rystal. During this time the opp er atom p enetrates in to the rystal and soni w a v es transp ort the impat energy a w a y to the b ottom of the rystal, whi h is onstan tly b eing o oled. Hene, after 10 − 11 seonds the whole rystal is in equilibrium again. On the other hand, the time b et w een t w o opp er atoms hitting the rystal surfae is on a time sale of 10 − 4 seonds, i.e. the system is in equilibrium nearly all the time, in partiular the temp erature is onstan t with resp et to spae and time. Ho w ev er, ev en in the state of thermo dynamial equilibrium single opp er atoms an  hange their p osition in the silion rystal b y hopping ev en ts, i.e. a opp er atom gains b y aiden t enough energy to o v erome the p oten tial barrier b et w een t w o la y ers in the silion rystal and th us hops to a neigh b oring ell. By atomi hopping opp er atoms an p enetrate m u h deep er in to the silion rystal as their impat energy w ould allo w, hene the pro ess in equilibrium annot b e omitted from the omputation. The a v erage time b et w een t w o hopping ev en ts is on a time sale of 10 − 10 seonds, while the fast atomi osillations happ en on a time sale of 10 − 14 seonds. In this pap er w e sho w ho w the metho d of optimal predition an b e applied to the system in equilibrium. Only the atoms at the top la y ers of the rystal, where opp er atoms an b e found, are omputed exatly , while the silion atoms in the lo w er la y ers are k ept tra k of only in an a v erage sense. In order to k eep the te hnial diulties at a minim um, w e set up a one dimensional mo del problem whi h sim ulates atomi hopping. 2.2 The Mo del Problem In the mo del problem, w e assume t w o ma jor simpliations: 2 PR OBLEM DESCRIPTION 3 • F o us on a one dimensional problem, i.e. w e onsider n atoms lined up lik e b eads on a ord. A single opp er atom is inserted in to a line of n − 1 silion atoms. • The p oten tial V ( q ) dep ends only on the pairwise distanes of the atoms, i.e. V ( q 1 , . . . , q n ) = n X i,j =1 i k 6 = 0 else (4.21) In blo  k form it an b e written as ∂ 2 V ∂ ˆ q ∂ ˜ q ( ˆ q , r ( ˆ q )) =  0 B 12 0 0  , (4.22) where B 12 is an upp er triangular k × k matrix. Substituting these sp eial v etors and matries in to the equation for the virtual atoms in (4.17 ) yields the follo wing equation of motion in blo  k form  A 11 A 12 A 21 A 22  ·  ˙ r V ˙ r m + k e  =  0 B 12 0 0  · M − 1 · ˆ p , (4.23) 5 COMPUT A TIONAL SPEED-UP 11 whi h implies the k -dimensional system for r V ¯ A 11 · ˙ r V = B 12 · M − 1 l · ˆ p l . (4.24) Here ¯ A 11 = A 11 + ( A 12 · e ) · e T k , where e k = (0 , . . . , 0 , 1) T . The momen ta of the last k real atoms are denoted b y ˆ p l = ( p m − k + 1 , . . . , p m ) T , and M l is the diagonal matrix on taining the orresp onding masses. This relation an b e in terpreted as a b oundary la y er ondition whi h ats as if the rystal of silion atoms w as on tin ued to innit y , although it is atually ut o after the m -th atom. With the ab o v e mo diations the zero temp erature limit optimal predition system (4.17 ) b eomes an expliit (2 m + k ) -dimensional system of equations. Hene, one an exp et this system to yield a omputational sp eed-up, dep ending on the v alues of n , m and k . The question of sp eed-up will b e onsidered in Setion 5. Due to the v arious appro ximations in a hieving the smaller system it is not at all lear whether the smaller system yields the same dynamis as the original system. In Setion 6 w e will ompare the new system to the original system and in v estigate under whi h onditions the new system reets the orret dynamis, and under whi h onditions it do es not. 5 Computational Sp eed-Up Besides the ab o v e ph ysial in terpretations, the atual in ten tion w as to use optimal pre- dition as a metho d to redue the omputational eort. In this setion, w e onsider b oth the v ersion with l virtual atoms and the b oundary la y er v ersion. W e ompare the CPU times for omputations of the t w o optimal predition systems with the CPU time for the orresp onding omputations of the original system. The omparison is arried out in de- p endene of the sizes n and m . Sine the original v ersion of optimal predition do es only replae real atoms b y virtual ones, while on the other hand the b oundary la y er ondition v ersion allo ws to really omit atoms, a signian t sp eed-up an only b e exp eted from the b oundary la y er ondition v ersion. The omputations w ere p erformed on a net w ork of AMD A thlon-6 1.4 GHz omputers. The b oundary la y er ondition v ersion w as omputed with k = 10 virtual atoms. Figure 3 sho ws the sp eed-up fators with resp et to the original system for the t w o v ersions of opti- mal predition. The original v ersion of optimal predition do es not yield an y aeleration (the sp eed-up fators are less than 1). Apparen tly , for our mo del problem setting up and solving the linear system in (4.17 ) is more exp ensiv e than omputing the full system of equations. The b oundary la y er ondition v ersion, on the other hand, yields signian t sp eed-up fators for small m . In priniple, one an a hiev e arbitrarily high sp eed-up fa- tors b y k eeping m and k xed and inreasing n . In most ases, ho w ev er, the original rystal size n is giv en a priori, and suitable v alues for m are giv en b y the requiremen t that the new system m ust ha v e the same dynamis as the original one, as disussed in Setion 6. A  hoie of m whi h is reasonable in man y ases, is utting the whole rystal of n atoms in to t w o halv es, i.e. m = ⌊ n 2 ⌋ . The sp eed-up fators for the t w o optimal predition systems are sho wn in Figure 4. While the original v ersion do es not redue the omputational eort, the b oundary la y er ondition v ersion yields a go o d sp eed-up for larger n . Care should b e tak en when generalizing these results. In our mo del problem the p oten tial funtions are fairly  heap to ev aluate. In more ompliated systems it ould 6 COMP ARING TW O MOLECULAR D YNAMICS SYSTEMS 12 0 20 40 60 0 20 40 60 0 0.5 1 m optimal prediction speed−up n speed−up factor 0 20 40 60 0 20 40 60 0 5 10 m boundary layer condition speed−up n speed−up factor Figure 3: Sp eed-up fators dep ending on n and m . 0 50 100 150 0 1 2 3 4 5 6 7 8 9 10 n = 2m speed−up factor optimal prediction boundary layer condition Figure 4: Sp eed-up fators for n = 2 m v ery w ell pa y o to solv e the linear system in ( 4.17 ) instead, or to solv e the minimization problem diretly . On the other hand, the b oundary la y er ondition v ersion of optimal predition ould p ossibly fail to w ork in other appliations, e.g. in three spae dimensions. Su h questions will ha v e to b e in v estigated when applying optimal predition to a new problem. 6 Comparing T w o Moleular Dynamis Systems In moleular dynamis tra jetories in high dimensional phase spae are no appropriate means for omparing t w o systems, sine initial p ositions and momen ta an nev er b e kno wn exatly , and moleular dynamis is t ypially  haoti. Instead, omparing means to test whether b oth systems ha v e similar dynamis. This is represen ted b y statistial quan tities su h as time orrelation funtions, diusion onstan ts, utuations of energy , et. W e onsider the follo wing statistial pro esses in order to ompare the t w o systems: • The distribution of the p osition the opp er atom. A opp er atom, whi h is initially lo ated alw a ys at the same p osition, is in the ensem ble of man y exp erimen ts desrib ed b y a diusion pro ess, whose distribution an b e appro ximated b y Mon te- Carlo sampling. • The n um b er of hopping ev en ts up to a xed time. 6 COMP ARING TW O MOLECULAR D YNAMICS SYSTEMS 13 • The utuation of energy of the rst m atoms. The energy of the rst m atoms utuates around some xed a v erage. W e onsider the v ariane of the energy o v er a xed time in terv al. The rst t w o quan tities are related to the diusion of a single opp er atom in the silion rystal due to hopping ev en ts. Sine opp er hopping has b een the eet of in terest in the rst plae, it is a natural riterion for omparison. In moleular dynamis, statistial quan tities of ergo di systems an b e omputed b y long time a v eraging or Mon te-Carlo sampling. Long time a v eraging means running a single omputation and using limiting pro esses to appro ximate statistial quan tities. Im- p ortan t examples, e.g. for appro ximating self-diusion onstan ts [2℄, are the Gr e en-Kub o formula [17, 23℄ or the Einstein r elation , whi h b oth appro ximate ensem ble a v erages for ergo di systems b y long time a v eraging. In our appliation, ho w ev er, long-time omputa- tions are problemati, sine rstly the opp er atom ma y tra v el to the b oundaries of the silion rystal and seondly soni w a v es will ome in eet, as sho wn in Setion 7 . Hene, w e obtain the diusion onstan t b y Mon te-Carlo sampling instead, i.e. w e solv e the same system o v er and o v er again with short omputation times, where the initial onditions are sampled from the anonial measure. In other w ords: W e obtain the a v eraged quan tities not b y long-time a v eraging, but b y a v eraging o v er man y samples. Sampling b oth the initial p ositions q i and momen ta p i from the anonial measure Z − 1 e − H ( q,p ) w ould require exp ensiv e metho ds lik e a  eptan e-r eje tion metho ds or Metr op o- lis sampling [12, 16℄ due to the struture of the p oten tial V ( q ) . W e irum v en t su h prob- lems b y setting the initial p ositions q i in to the p oten tial minim um and sampling the initial momen ta p i from ˜ Z − 1 e − T ( p ) , whi h is just sampling indep enden tly from Gaussian distri- butions. In our sim ulations after ab out 5 · 10 − 14 s the Hamiltonian dynamis has driv en the system in to equilibrium. A dditionally , k eeping the initial p ositions xed automatially guaran tees to remain in the orret domain M L as giv en b y (4.2 ). 6.1 A Random W alk Mo del for the Copp er Diusion The onsidered opp er diusion is due to hopping ev en ts, while displaemen ts due to short osillations b et w een the same t w o silion atoms are not tak en in to aoun t. Hene, the or- resp onding diusion pro ess is disrete in spae on the spatial grid {− ν d 0 , . . . , 0 , . . . , ν d 0 } . Let us assume for the momen t, that the diusion pro ess is linear and hene desrib ed b y a (2 ν + 1) -dimensional  omp artment mo del ˙ u ( t ) = A · u ( t ) , u (0) = e m +1 = (0 , . . . , 0 , 1 , 0 , . . . , 0) T , (6.1) where A ∈ R (2 ν +1) × (2 ν +1) has olumn sums equal to zero to ensure mass onserv ation, i.e. d dt P i u i ( t ) = 0 . The analytial solution to (6.1 ) is u ( t ) = exp( tA ) · e m +1 . (6.2) One example for su h a pro ess is giv en b y the tridiagonal T o eplitz (up to the b oundary en tries) matrix with the stenil κ d 2 0 (1 , − 2 , 1) , whi h is a nite dierene appro ximation to the heat equation. Hene, the ompartmen t mo del (6.1 ) with this matrix on v erges to the heat equation as d 0 → 0 . Still, there are man y other matries A , whose disrete diusion pro esses all on v erge to the heat equation as d 0 → 0 . 6 COMP ARING TW O MOLECULAR D YNAMICS SYSTEMS 14 The hopping b eha vior of the opp er atom an b e mo deled as a sp ei random w alk. One imp ortan t asp et of the opp er hopping is that hopping ev en ts are  orr elate d , in the sense that giv en the opp er atom has just hopp ed, it is m u h more lik ely than normally that a seond hopping ev en t to the same diretion follo ws, sine the kineti energy is not ompletely lost in one single jump. Hene, the opp er diusion is formally non-Mark o vian. Still, the hopping an b e desrib ed as a Mark o vian random w alk b y assuming hopping ev en ts to b e unorrelated, but to allo w the opp er atom to hop o v er more than one silion atom in one single hopping ev en t. These assumptions lead to a mo del whi h is t ypial in the on text of sto  hasti pro- esses, see [22℄. Let X t denote the p osition of the opp er atom at time t . F or the n th hopping ev en t let T n b e the hopping time, ∆ T n = T n − T n − 1 the time sine the previous hopping ev en t, and ∆ n ∈ Z \ { 0 } the n um b er of silion atoms whi h the opp er atoms hops o v er to the righ t. Here ∆ n < 0 means hopping to the left. Assume no w that b oth the ∆ T n and the ∆ n are indep enden t and distributed aording to P (∆ T n ∈ [ s, s + d s )) = α · exp( − αs ) d s, (6.3) P ( | ∆ n | = i ) = p i . (6.4) Here α is a parameter on trolling the hopping rate (see [21, 22℄ for a deriv ation), and ( p i ) i ∈ N is a non-negativ e sequene satisfying P i ∈ N p i = 1 2 and the onstan t γ = P k i =1 ( d 0 i ) 2 p i is nite. F or our mo del problem w e simply assume p = ( p 1 , . . . , p k ) . A dditionally , w e restrit to symmetri random w alks, whi h is reasonable as long as the opp er atom do es not approa h the rystal b oundaries. An analysis of the desrib ed random w alk, follo wing the analysis in [22℄, yields that the v ariane of the pro ess at time t equals V ar ( X t ) = 2 γ αt . A omparison with the v ariane giv en b y the heat equation yields the relation κ = γ α, (6.5) whi h allo ws us to sp eak onsisten tly of a diusion onstan t κ also for the disrete diusion pro ess (6.1). As deriv ed in [22℄ the probabilit y distribution of the random w alk X t is desrib ed b y (6.1), where the so alled innitesimal gener ator A is a band matrix with the stenil α · ( p k , . . . , p 1 , − 1 , p 1 , . . . , p k ) . Dene ˜ A as a orresp onding band matrix with the stenil γ − 1 · ( p k , . . . , p 1 , − 1 , p 1 , . . . , p k ) , hene A = κ ˜ A . Both matries ha v e to b e  hanged in the upp er and lo w er ro ws aording to the appropriate b oundary onditions. Assume no w, that the v alues p 1 , . . . , p k are kno wn. Hene, the matrix ˜ A is ompletely determined. Let v ( t ) ∈ R 2 ν +1 denote the n umerially obtained distribution v etors for the p osition of the opp er atom for all times t . Initially v (0) = e m +1 , as in (6.1 ). Sine it is unlear, whether the diusion parameter κ is onstan t in time, w e let it b e a time dep enden t parameter κ ( t ) , whi h w e ompute at times t 1 , . . . , t µ . F or obtaining the v alue κ j = κ ( t j ) w e onsider the diusion on the time in terv al I j = [ t j , t j ] , where t j = t j − ∆ t 2 and t j = t j + ∆ t 2 . Giv en ∆ t is not to o large, w e an appro ximately assume the diusion parameter to b e onstan t on the giv en in terv al: κ ( τ ) = κ j ∀ τ ∈ I j . On I j w e use the data pro vided b y the real ev olution v ( t ) to ompute an L 2 -t of the diusion parameter κ j with resp et to error funtional F ( κ j ) = Z t j t j k e ( τ − t j ) κ ( t j ) ˜ A · v ( t j ) − v ( τ ) k 2 2 d τ , (6.6) 7 NUMERICAL EXPERIMENTS 15 whi h is partiularly stable with resp et to errors in v due to the Mon te-Carlo sampling. The time ∆ t m ust on the one hand b e small enough to justify the appro ximation that the diusion parameter is onstan t on the in terv als I j , on the other hand it m ust b e large enough to ha v e already some diusion tak en plae. In Subsetion 7.1 w e ompute diusion parameters for the n umerial data obtained for our mo del problem. 7 Numerial Exp erimen ts W e sim ulate a rystal of n = 70 atoms with 69 silion atoms and 1 opp er atom, whi h starts at the 22 nd p osition. F or the optimal predition system, w e  ho ose m = 50 . These are enough atoms su h that a v eraged quan tities mak e sense and one an sp eak of a ther- mo dynamial equilibrium. A dditionally , the rystal has a reasonable in terior region whi h is not aeted b y an y b oundary eets. The in tegration is p erformed b y the lassial ex- pliit fourth order Runge-Kutta metho d, whi h w e prefer o v er energy preserving metho ds in this on text, as it allo ws signian tly larger time steps, while still resolving the hopping ev en ts aurately . The in tegration time step is ∆ t = 2 . 5 · 10 − 15 s, and the omputation time is t ∗ = 4 . 0 · 10 − 13 s, whi h is short enough, su h that the  hange in total energy due to the in tegrator is insignian t. W e solv e the system N = 25000 times at a temp erature T = 7000 K , with the initial data sampled as desrib ed in Setion 6. F or our exp erimen ts, Mon te-Carlo estimates yield an exp eted error in the distribution of ab out 4 · 10 − 3 , whi h is signian tly smaller than the dierene of an y t w o quan tities in omparison. Note that the short omputation time t ∗ allo ws to exlude an y eets aused b y soni w a v es tra v eling through the rystal. The v elo it y of sound an b e estimated as deriv ed in [17℄: The equilibrium distane inside the silion rystal is appro ximately d 0 = 1 . 87 · 10 − 10 m . Using the Y oung mo dulus Y = d 0 · ∂ 2 H lo  ∂ d 2 0 ( d 0 ) = 5 . 05 · 10 − 9 kg · m s 2 w e obtain a v elo it y of sound of c = q Y · d 0 m S i = 4 . 50 · 10 3 m s . The shortest distane of the opp er atom to an y b oundary in the rystal is s left = q 22 − q 1 = 2 . 01 · 10 − 9 m in the original system and ˜ s righ t = q 50 − q 22 = 1 . 90 · 10 − 9 m in the optimal predition system. Hene, the minim um time a soni w a v e tak es to tra v el from a b oundary to the opp er atom is appro ximately t min = ˜ s righ t /c = 4 . 22 · 10 − 13 s, whi h is longer than the omputation time t ∗ = 4 . 0 · 10 − 13 s. In Subsetion 7.4 w e will deal with the ase when soni w a v es are presen t. Figure 5 sho ws the distribution of the p osition of the opp er atom in the silion rystal when solving the original system. The analogous distribution for the optimal predition system lo oks indistinguishably similar in this kind of plot. The pro ess is apparen tly of a diusiv e nature, but the diusion parameter  hanges with time, as already the ev olution of the maxim um indiates. The follo wing analysis will onrm this observ ation. The time-dep enden t relativ e error b et w een the distributions for the original and the optimal predition system e ( t ) = max x | v ( x, t ) − ˜ v ( x, t ) | max x | v ( x, t ) | (7.1) is less than 3% up to the time t = 3 . 0 · 10 − 13 s and still less than 9% o v er the omplete time in terv al, whi h is not o v erwhelmingly small, but do es indiate denite similarities b et w een the t w o distributions. 7 NUMERICAL EXPERIMENTS 16 0 1 2 3 4 x 10 −13 15 20 25 30 0 0.2 0.4 0.6 0.8 1 position in crystal t Figure 5: Distribution of the opp er atom's p osition 7.1 Diusion P arameters W e ompute the diusion parameters κ ( t ) for the original and the optimal predition system, using the metho d desrib ed in Subsetion 6.1 . The parameters ( p 1 , . . . , p k ) of the orresp onding random w alk mo del are obtained b y Mon te-Carlo sampling. W e use the exp erimen ts for the original system, whi h already yielded the distribution sho wn in Figure 5. In ea h Mon te-Carlo exp erimen t w e follo w the path of the opp er atom and luster onseutiv e hopping ev en ts whi h happ en in an in terv al of ∆ t 1 = 6 . 0 · 10 − 14 s to a single one. Prior to this lustering, fast double hopping ev en ts in to opp osing diretions, i.e. those happ ening inside of ∆ t 2 = 2 . 0 · 10 − 14 s , whi h are solely due to osillations of silion atoms, m ust b e exluded. The times ∆ t 1 and ∆ t 2 are suitably  hosen for our mo del problem. The results of the ab o v e desrib ed Mon te-Carlo exp erimen ts are sho wn in Figure 6. The hopping probabilities are giv en b y the b o x histogram. Note that due to Mon te-Carlo errors and b oundary eets the resulting v alues are not exatly symmetri. Sine w e wish to onsider a symmetri random w alk, w e  ho ose as p 1 , . . . , p k the a v erage v alues of the obtained results. Here, only p 1 , . . . , p 11 are greater than zero. Hene, w e  ho ose k = 11 for the omputation of the diusion parameters. The urv e denotes the v alues i 2 p i (saled to t in to the same plot), whi h are relev an t, sine the v ariane P k i =1 i 2 p i = γ d 2 0 is prop ortional to κ (see relation (6.5)). Figure 7 sho ws the time-dep enden t diusion parameter κ ( t ) for the original system and the optimal predition appro ximation, omputed as desrib ed in Subsetion 6.1 . In b oth ases w e  ho ose ∆ t = 2 . 5 · 10 − 14 s . Imp ortan t observ ations and estimates are: • The diusion parameters start with a strong p eak, and after 1 · 10 − 13 s they utuate around a xed v alue of κ = 4 . 4 · 10 − 10 m 2 s . This b eha vior is to o pronouned to b e only due to Mon te-Carlo and appro ximation errors. The initial b eha vior is most lik ely due to the fat that the system do es not exatly start in its thermo dynamial equilibrium. 7 NUMERICAL EXPERIMENTS 17 −10 −5 0 5 10 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 hopping width probability p i p i i 2 Figure 6: Probabilities of hopping o v er more than a single atom 0 0.5 1 1.5 2 2.5 3 3.5 4 x 10 −13 0 1 2 3 4 5 6 7 8 x 10 −8 t / s κ / m 2 /s original system optimal prediction Figure 7: Diusion parameters o v er time • Considering that κ ( t ) sho ws su h pronouned b eha vior, the t w o funtions κ ( t ) for the original system, and ˜ κ ( t ) for the optimal predition appro ximation are remark ably lose to ea h other. Up to the time t = 1 . 4 · 10 − 13 s the the distane b et w een the t w o urv es is quite small. After that time the urv es dier more, but sho w the same features. After 3 . 0 · 10 − 13 s the error tak es its maxim um, whi h oinides with the errors observ ed in the distributions sho wn in Figure 5 . • In order to roughly estimate whether the obtained diusion parameter is reasonable, w e ompare it with a diusion o eien t measured in a real material. In [3℄ the diusion onstan t of opp er in a silion rystal at a temp erature of T 0 = 1273 K is giv en as κ 0 = 4 . 4 · 10 − 10 m 2 s . Assuming that the diusion onstan t dep ends linearly b oth on temp erature and the p oten tial barrier, as e.g. with the Einstein form ula [17℄, w e obtain an estimate κ ≈ κ 0 · T T 0 · ∆ E 0 ∆ E = 1 . 7 · 10 − 8 m 2 s for T = 7000 K , using the v alues ∆ E 0 = 3 e V and ∆ E = 0 . 43 e V . A lo ok at Figure 7 indiates that our exp erimen tally obtained diusion parameters are indeed in this region. 7.2 The Num b er of Hopping Ev en ts Besides diusion parameters, also the pure n um b er of hopping ev en ts whi h happ en up to a giv en time t ∗ should b e preserv ed b y optimal predition. Figure 8 sho ws the n um b er of hopping ev en ts for the ab o v e omputation plotted as histograms. The solid line represen ts the original system, and the dashed line stands for the optimal predition system. The t w o graphs dier only for the probabilities of zero and one hopping ev en t. Apart from that, one an sp eak of the same hopping b eha vior. 7 NUMERICAL EXPERIMENTS 18 0 2 4 6 8 10 12 14 16 18 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 number of hopping events N fraction of experiments with N hopping events original system optimal prediction Figure 8: Num b er of hopping ev en ts 0 0.5 1 1.5 2 x 10 −51 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 energy variance / kg 2 m 4 /s 3 fraction of experiments original system optimal prediction Figure 9: T otal energy utuation 7.3 Energy Flutuations While the total energy is onstan t for the original system as w ell as for the optimal predition appro ximation, the energy of the rst m atoms E left ( t ) = 1 2 m X i =1 p 2 i ( t ) m i + m X i,j =1 i

Original Paper

Loading high-quality paper...

Comments & Academic Discussion

Loading comments...

Leave a Comment