Stochastic Molecular Dynamics Derived from the Time-independent Schrödinger Equation Anders Szepessy Schrödinger observables approximated by: • Ehrenfest dynamics O(M −1) • Born-Oppenheimer dynamics O(M −1/2) • Langevin and Smoluchowski dynamics O(M −1 + gap−1) Which Molecular Dynamics? M Ẍ = −V 0(X) + ? • What is V ? • Are there fluctuations? How model temperature T ? • Is there dissipation? The Usual Derivation and the New Usual start from time-dependent Schrödinger equation: • self consistent field equation: wave function is product of nuclei and electron function • Ehrenfest dynamics: nuclei wave function becomes point measure • Born-Oppenheimer approximation: electron wave function is the ground state. The Usual Derivation and the New Usual start from time-dependent Schrödinger equation: • self consistent field equation: wave function is product of nuclei and electron function • Ehrenfest dynamics: nuclei wave function becomes point measure • Born-Oppenheimer approximation: electron wave function is the ground state. New start from time-independent Schrödinger equation: • use WKB-Ansatz • introduce time by momentum from characteristics of Eikonal eq. • introduce integrating factor from divergence of momentum • stability from consistency with Schrödinger equation. Difference Between the Two + experimental evidence of time-independent Schrödinger + nuclei paths behave classically without separation and small support + long time stability + stochastic perturbation of ground state leads to Langevin + colliding characteristic paths – only equilibrium situations Dynamics from the Time-independent Schrödinger Schrödinger: H(x, X)Φ(x, X) = EΦ(x, X) N 1 X H(x, X) = V (x, X) − ∆X n , 2M n=1 J 1X ∆xj + V (x, X) = − 2 j=1 − N X J X n=1 j=1 X 1≤k<j≤J Zn + |xj − X n| M 1 1 |xk − xj | X 1≤n<m≤N ZnZm |X n − X m| The WKB-Ansatz Φ(x, X) = ψ(x, X)eiM 1/2 θ(X) implies iM 1/2 θ(X) 0 = (H − E)ψe |θ0|2 = ( + Vn − E) ψ | 2 {z } =0 i 1 00 0 0 − 1/2 (ψ ◦ θ + ψθ ) + (V − Vn)ψ 2 M 1 00 iM 1/2θ − ψ e , 2M ψ·Vψ Vn := , ψ·ψ Z v · w := R3J v ∗(x, X, t)w(x, X, t) dx The Time-dependence The Time-dependence1 dXt dψ(x, Xt) ψ ◦θ =ψ ◦ = dt dt |{z} 0 0 0 =:p 1 Mott 1931; Briggs and Rost 2001 The Time-dependence2 dXt dψ(x, Xt) ψ ◦θ =ψ ◦ = dt dt |{z} 0 0 0 =:p d θ = div p = log |p| dt yields the Eikonal and transport equation for ϕ := |p|1/2ψ 00 0 = (H − E)Φ |θ0|2 + Vn − E) ψ = ( | 2 {z } =0 1/2 i 1 + − 1/2 ϕ̇ + (V − Vn)ϕ − |p|1/2∆X (ϕ|p|−1/2) |p|−1/2 eiM θ . 2M | M {z } =0 2 Mott 1931; Briggs and Rost 2001 Characteristics of the Schrödinger Equation i 1 |p|1/2∆X (ϕ|p|−1/2) − 1/2 ϕ̇ + V (Xt) − Vn ϕ = 2M M Characteristics of Eikonal equation: Ẋt = pt ṗt = −Vn0(Xt) ϕ·Vϕ Vn = ϕ·ϕ The Ehrenfest Approximation Ẋt = pt ṗt = −φt · V 0(Xt)φt i φ̇t = V (Xt)φt M 1/2 is Hamiltonian system for HJ HE := |p|2/2 + φ · V (X)φ = E with characteristics (X, ϕr ; p, ϕi) and ϕ := 21/2M −1/4φ; and ψ̂t := φte R iM 1/2 0t φs ·V (Xs )φs ds implies i ˙ ψ̂t = (V − Vn)ψ̂t M 1/2 Ehrenfest Accuracy Φ̂ := ρ̂ 1/2 iM 1/2 θ̂ ψ̂e implies 1 ∆X (ρ̂1/2ψ̂) 2M = O(M −1) (H − E)Φ̂ = so that Z Z g(X) ρ̂ − ρ)dX = g(X) Φ̂ · Φ̂ − Φ · Φ dX = O(M −1) Motivation for Stable Eigenstate Perturbation Orthonormal eigenpairs {λn, Φn}, satisfying HΦn = λnΦn, X Φ̂ =: αnΦn n yields X n 1 (λn − E)αnΦn = − v, 2M which establishes Z 1 (λn − E)αn = − Φn · v dX . 2M | T3N {z } =: v̂n We have v̂n = 0, when λn = E, and let |v̂n0 (E)| := lim sup δ→0+ |v̂n(E + δ)| , δ which implies X |v̂n(λn)|2 ≤ 2 |λ − E| n n X |v̂n0 (E)|2 + X |v̂n|2. n {n:|λn −E|<1} Assume that X {n:|λn −E|<1} |v̂n0 (E)|2 + X |v̂n|2 = O(1). (1) n Motivation for bounded |v̂ 0|: Large number of nuclei N M δ perturbation of λn = |p|2/2 + Vn yields O(δN −1) perturbation of paths (X, p, ψ) and also θ change negligible (with appropriate time) so eigenstate change small 1/2 Φ = ψeiM θ The Born-Oppenheimer Approximation An electron eigenstate: ψ̂ = Ψn V (X)Ψn = λn(X)Ψn implies Ẋt = pt ṗt = −λ0n(Xt) (H − E)Φ̂ = O(M −1/2) and Z Z g(X) ρ̂ − ρ)dX = g(X) Φ̂ · Φ̂ − Φ · Φ dX = O(M −1/2) Stochastic Molecular Dynamics Approximation Improve Born-Oppenheimer Ehrenfest solution ψ̂n = Ψn + ψn⊥ ψn⊥(t) = St,0ψn⊥(0) − iM −1/2 Z t St,sΨ̇n(s)ds 0 so perturbation of ground state Ψ0 yields • fluctuation from stochastic initial data ψ̂n, n > 0 • dissipation from residual Ψ̇0(s) = Ψ00 ◦ Ẋ Which Initial Data for ψ̂n ? Liouville eq. for Ehrenfest ∂tf + ∂pE HE ∂rE f − ∂rE HE ∂pE f = 0 has many time-independent solutions f = h(HE ) Which Initial Data for ψ̂n ? Liouville eq. for Ehrenfest ∂tf + ∂pE HE ∂rE f − ∂rE HE ∂pE f = 0 has many time-independent solutions f = h(HE ) Idea: Nuclei act as heat bath for electrons independent Ẋn imply h(HE ) = e−HE /T (nuclei has this (unique) invariant SDE probability density) so Prob(electron configuration ψ̂) ∼ e−ψ̂·(V −λ0)ψ̂/T dψ̂ r dψ̂ i. Stochastic Ehrenfest Dynamics Let r(X) := YZ j≥0 −λ̄j |γj |2 /(T e 2 j≥0 |γj | ) P |γj |2 <C dγjr dγji Y T ∼ ( )1/2, λ̄ j≥0 j then observable in the Ehrenfest dynamics is R R −λ0 (X)/T r(X) −λ0 (X)/T g(X)e 3N g(X)e r(X)dX R r(0) dX R3N R . = R −λ (X)/T r(X) 0 −λ (X)/T r(X)dX 0 dX 3N e R3N e R r(0) The Spectral Gap Condition A large spectral gap α := −1 j>0 λ̄j P 1 yields r(X) = 1 + O(α), r(0) which implies R R −λ0 (X)/T −λ0 (X)/T g(X)e r(X)dX g(X)e dX 3N 3N RR RR = + O(α). −λ0 (X)/T r(X)dX −λ0 (X)/T dX e e R3N R3N Langevin and Smoluchowski Dynamics The stochastic Langevin dynamics dXt = ptdt √ dpt = −∂X λ0(Xt)dt − Kptdt + 2T KdWt and the Smoluchowski dynamics dXs = −∂X λ0(Xs)ds + √ 2T dWs has the unique invariant probability density e−(p◦p/2+λ0(X))/T dp dX R −(p◦p/2+λ0 (X))/T dp dX e 6N R respectively e−λ0(X)/T dX R −λ0 (X)/T dX R3N e Stochastic Approximation Theorem Langevin and Smoluchowski dynamics approximate Schrödinger observables with error O(M −1 + α) provided the assumption in EhrenfestP approximation holds together with the spectral gap condition α = j>0 λ̄−1 j 1. Conditions for Colliding Characteristics Paths (Caustics) At a point X of two colliding characteristic paths, (X, p−, z −) and (X, p+, z +), we need: • the phase θ is continuous, i.e. z − = z +, • a stable ϕ, i |p|1/2 X −1/2 ∆ (|p| ϕ̇ = (V − V )ϕ − ϕ), j n X 1/2 2M j M so take |p+| = |p−|, • θ is max-norm stable towards perturbations of the initial data, which implies the irreversible viscosity solution θ. SPDE from Smoluchowski MD with Erik von Schwerin Energy conservation: ∂t(cv T + m) = div(k∇T ) Phase field for m = g(φ): V k0∂tφ = div(k1∇φ) − V 0(φ) + k2T + noise Why noise? m SPDE from Smoluchowski MD with Erik von Schwerin Energy conservation: ∂t(cv T + m) = div(k∇T ) Phase field for m = g(φ): V k0∂tφ = div(k1∇φ) − V 0(φ) + k2T + noise Why noise? m Which Noise and Phase Field Equation? 1. Stochastic Smoluchowski molecular dynamics 2. Quantitative atomistic definition of phase field 3. Numerical computation of coarse-grained model functions 1. Stochastic Molecular Dynamics Energy: X |vi|2 + V (X1, . . . , XN ) | {z } 2 m | i {z } cv T Smoluchowski dynamics in diffusion time scale (γ = kB T ): p t t dX = −∂V (X )dt + 2γ dW t 2. Molecular Potential Energy Liquid Subcooled Liquid m Latent heat, L m Superheated Solid Solid TM T Pair interactions 1 XX V (X) = Ṽ (Xi − Xj ) 2 i j6=i Localized average (as SPH) 1 XX m(X, x) := Ṽ (Xi − Xj )η(x − Xi) 2 i j6=i {z } | observable 3. Coarse-Grained Stochastic MD Want coarse-grained approximation m̄ X t t bk (m̄t)dW̄kt dm̄ = a(m̄ )dt + k such that T T min E g m(X , ·) − g(m̄ ) . a,b 1. Ito implies t t dm(X , ·) = α(X )dt + X βj (X t)dWjt. j 2. Kolmogorov equation for ū(n, t) := E[g(m̄T ) | m̄t = n] T T E g m(X , ·) − g(m̄ ) Z T X X 0 00 = E[ hū , α − ai + hū , βj ⊗ βj − bk ⊗ bk i dt] 0 j k 3. Expansion in α − a Z T 1 a= E α dt], T 0 Z X 1 TX b k ⊗ bk = E βj ⊗ βj dt]. T 0 j k Coarse-Grained Variables Density Phase Field : ρloc(*;x) Potential Energy Phase Field : m(*;x) 0.6 1.31 0.6 T d2m/dx2 (x) m(x) 0.4 0.4 1.3 0.2 0.2 1.29 0 0 1.28 −0.2 −0.2 −0.4 1.27 −0.4 −0.6 1.26 −0.6 −0.8 1.25 −0.8 −1 −1.2 −40 −20 0 20 40 60 x 80 100 120 140 1.24 −40 −20 0 20 1 1. α = γ∂xxm + ∂xA1 + A0 RT 2. a = 0 αdt/T → 0. 40 60 x 1 80 100 120 140 −1 −10 0 10 20 x1 30 40 50 60 Drift 0.5 f(m) 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 −1 f(m;1.0) 1.492 f(m;2.0) 2 0.885 f(m;0.7) 0.4 Orientation 1 Orientation 2 0 −0.5 0 m 0.5 −0.1 −1 −0.8 −0.6 −0.4 −0.2 m 0 0.2 0.4 0.6 Diffusion 35 3 3 30 25 30 2.5 2.5 2 2 20 1.5 20 10 1 0 100 1.5 15 0.5 1 50 100 10 0 100 50 y1 0 0 0.5 5 50 x1 0 y1 0 0 20 40 x1 60 80 100 0 Orientation dependence Let qi := A0(Oi)/A0(O1) ri := τ (Oi)/τ (O1) Then dm̄i = riqi γqi−1m̄00i 1/2 + A0(m̄i; O1) dt + ri b̄(m̄i; Oi)dW t becomes −1 (riqi) dm̄ = γqi−1m̄00 −1/2 −1 qi b̄(m̄; Oi)dW t + A0(m̄; O1) dt + ri Extensions • more directions • constant pressure • undercooled melt • real material • hybrid simulations 3. Expansion in α − a m(X t, x) ≈ m̄(x) leads to E Z T 0 t ū m(X , x), t , α(x) − a m̄(x) dt ≈ 0 Z 0 T ū m̄(x), T , E α(x) − a m̄(x) dt 0 2. Derivation of the Error T T E g m(X , ·) − g(m̄ ) 2. Derivation of the Error T T E g m(X , ·) − g(m̄ ) = E[ū(mT , T ) − ū(m0, 0)] 2. Derivation of the Error T T E g m(X , ·) − g(m̄ ) = E[ū(mT , T ) − ū(m0, 0)] RT = E[ 0 dū(mt, t)] 2. Derivation of the Error T T E g m(X , ·) − g(m̄ ) = E[ū(mT , T ) − ū(m0, 0)] RT = E[ 0 dū(mt, t)] hR i P T = E 0 hū0, αi + hū00, j βj ⊗ βj i + ∂tū dt 2. Derivation of the Error T T E g m(X , ·) − g(m̄ ) = E[ū(mT , T ) − ū(m0, 0)] RT = E[ 0 dū(mt, t)] hR i P T = E 0 hū0, αi + hū00, j βj ⊗ βj i + ∂tū dt hR i P P T = E 0 hū0, α − ai + hū00, j βj ⊗ βj − k bk ⊗ bk i dt Factorization of the density The Ehrenfest equation |θ̂0|2 1 i 0= + V − E ψ̌ − 1/2 (ψ̌ 0θ̂0 + ψ̌ θ̂00) 2 2 M yields Z XZ 0= (∂X j ψ̌ ∗ψ̌ + ψ̌ ∗∂X j ψ̌)dx ∂X j θ̂ + ψ̌ ∗ψ̌dx ∂X j X j θ̂ j = X T3J T3J ∂X j (ρ̂∂X j θ̂) j and X j ∂X j X j θ = div p = ∂p11 ∂X11 = ∂p11 ∂t ∂X11 ∂t ṗ11 dtd |p|2 d = 1= = log |p| 2 p1 2|p| dt implies ˙ X̂t) = ρ̂( X = X ˙ ∂X̂ j ρ̂(X̂t)X̂ j j ∂X̂ j ρ̂(X̂t)∂X̂ j θ̂n j = −ρ̂(X̂t) X ∂X̂ j X̂ j θ̂n j = −ρ̂(X̂t) div p̂ d = −ρ̂(X̂t) log |p̂t| dt with the solution C , |p̂t| where C is a positive constant for each characteristic. Coordinates X11 ∈ R parallel and X0 ∈ R3N −1 orthogonal to the characteristic ρ̂(X̂t) = direction Ẋ gives ρ̂(X)dX = ρ̂(X0) R dt dX11 = ρ̂(X0)dX0 , |p̂11| T dX0 X11 (T ) dX11 0 |p̂11 | using |dX̂11| dX̂11 = dt, = 1 1 |p̂1| |dX̂1 /dt| and the observable Z Z g(X̂, ψ̂, ρ̂)dX̂ = T3N 0 T Z A(Xt)ρ̂(X̂0)dX̂0 I dt . T (2) Dynamics from the time-dependent Schrödinger i∂tΦ(x, X, t) = H(x, X)Φ(x, X, t) N 1 X H(x, X) = V (x, X) − ∆X n , 2M n=1 J 1X ∆j+ V (x, X) = − 2 j=1 x − N X J X n=1 j=1 J X =: − 1 2 X 1≤k<j≤J j=1 1 |xk − xj | Zn + |xj − X n| ∆xj + HI M 1 X 1≤n<m≤N ZnZm |X n − X m| Z v ∗(x, X, t)w(x, X, t) dx ZR3J Z hv, wi := v ∗(x, X, t)w(x, X, t) dx dX v · w := R3N R3J Usual derivation: 1. self consistent field equation: wave function is product of nuclei and electron function 2. Ehrenfest dynamics: nuclei wave function becomes point measure 3. Born-Oppenheimer approximation: electron wave function is the ground state. 1. Time-dependent self consistent field equations Approximation Ansatz of separation Z t s s s s Φ(x, X, t) = ΨN (X, t)ΨE (x, t) exp i hΨN ΨE , HI ΨE ΨN i ds {z } | 0 H̄I satisfies time dependent self consistent field equation3 N X − (2M )−1 ∆Xn + ΨE · HI (X)ΨE ΨN , n=1 Z ∗ i∂tΨE = ΨN (X)V (X)ΨN (X) dX ΨE . i∂tΨN = R3N 3 Dirac P.A.M., Proc. Cambridge Phil. Soc. 26 (1930) 376–385. Φ solves perturbed full Schrödinger N X J 1X −1 i∂tΦ = − (2M ) ∆Xn − ∆xj + ΨE · HI ΨE 2 j=1 n=1 Z + Ψ∗N HI ΨN dX − H̄I Φ, R3N and compactly supported ΨN in δ small domain leads4 to O(δ) approximation of full Schrödinger in L2(dxdX). 4 Bornemann F.A., Nettesheim P. and Schütte C., J. Chem. Phys, 105 (1996) 1074–1083. 2. Ehrenfest dynamics from WKB ΨN = ψeiM 1/2 θ leads to Ehrenfest Ẍ = −ΨE · ∂X V (X)ΨE iM −1/2Ψ̇E = V ΨE (X, ΨE ) approximates5 TDSCF with error O(δ 2 + M −1/2) 5 Bornemann F.A., Nettesheim P. and Schütte C., J. Chem. Phys, 105 (1996) 1074–1083. Tully J.C., Faraday Discuss., 110 (1998) 407–419. Marx D. and Hutter J., Ab initio molecular dynamics: Theory and implementation, Modern Methods and Algorithms of Quantum Chemistry, J.Grotendorst(Ed.), John von Neumann Institute for Computing, Jülich, NIC Series, Vol. 1, ISBN 3-00-005618-1, pp. 301-449, 2000 3. Born-Oppenheimer approximation: ΨE = ground state An electron eigenstate: ΨE = Ψn V (X)Ψn = λn(X)Ψn implies Ẍ = −Ψn · ∂X V (X)Ψn = −∂X λn(X) Spectral gap can be used to prove6 O(M −1/4) approximation of Schrödinger in L2. 6 Hagedorn G.A., Commun. Math. Phys., 77 (1980) 1–19. Panati G., Spohn H. and Teufel S., Math. Mod. Numer. Anal.