Ab initio Molecular Dynamics

Size: px

Start display at page:

Download "Ab initio Molecular Dynamics"

Lester Hensley
5 years ago
Views:

1 Ab initio Molecular Dynamics Jürg Hutter Department of Chemistry, University of Zurich

2 Overview Equations of motion and Integrators General Lagrangian for AIMD Car Parrinello MD and Born Oppenheimer MD ab initio Langevin MD Stability and efficiency Thermostats External Drivers

3 Equations of Motion (EOM) Newton s EOM for a set of classical point particles in a potential. dv (R) M I RI = dr I These EOM generate for a given number of particles N in a volume V the micro canonical ensemble (NVE ensemble). The total energy E is a constant of motion!

4 Total Energy Total Energy = Kinetic energy + Potential energy Kinetic energy = T (Ṙ) = N Potential energy = V (R) I=1 M I 2 Ṙ2 We will use the total energy as an indicator for the numerical accuracy of simulations.

5 Lagrange Equation with d L dt Ṙ = L R L(R, Ṙ) = T (Ṙ) V (R) Equivalent to Newton s EOM in Cartesian coordinates, but is more general and flexible. Extended systems Constraints

6 Integration of EOM Discretization of time R(t) R(t + τ) R(t + 2τ) R(t + mτ) V (t) V (t + τ) V (t + 2τ) V (t + mτ) Efficiency: minimal number of force evaluations, minimal number of stored quantities Stability: minimal drift in constant of motion (energy) Accuracy: minimal distance to exact trajectory

7 Sources of Errors Type of integrator predictor-corrector, time-reversible, symplectic Time step τ short time accuracy measured as O(τ n ) Consistency of forces and energy e.g. cutoffs leading to non-smooth energy surfaces Accuracy of forces e.g. convergence of iterative force calculations (SCF, constraints)

8 Velocity Verlet Integrator R(t + τ) = R(t) + τv (t) + τ 2 V (t + τ) = V (t) + 2M f (t) τ [f (t) + f (t + τ) 2M Efficiency: 1 force evaluation, 3 storage vectors Stability: time reversible Accuracy: O(τ 2 ) Simple adaptation for constraints (shake, rattle, roll) Simple adaptation for multiple time steps and thermostats

9 Test on Required Accuracy of Forces Classical Force Field Calculations, 64 molecules, 330 K TIP3P (flexible), SPME (α = 0.44, GMAX = 25),

10 Stability: Accuracy of Forces Stdev. f Stdev. Energy Drift Hartree/Bohr µhartree µhartree/ns Kelvin/ns

11 Born Oppenheimer MD: The Easy Way dv (R) M I RI = dr I EOM V (R) = min KS({Φ(r)}; R) + const.] Φ Kohn Sham BO potential Forces f KS (R) = d min Φ E KS ({Φ(r)}; R) dr = E KS R + const. R + i (E KS + const.) Φ i } Φ {{ i R } =0

12 Computational Details System 64 water molecules density 1gcm 3 Temperature 330K Timestep 0.5fs DFT Calculations GPW, TZV2P basis (2560 bsf), PBE functional Cutoff 280 Rydberg, ɛ default = OT-DIIS, Preconditioner FULL_SINGLE_INVERSE Reference trajectory (1ps), ɛ SCF = 10 10

13 Stability in BOMD Unbiased initial guess; Φ(t) = Φ 0 (R(t)) ɛ SCF MAE E KS MAE f Drift Hartree Hartree/Bohr Kelvin/ns Consistent with results from classical MD Note accuracy of forces!

14 Efficiency: Initial Guess of Wavefunction 4th order Gear predictor (PS extrapolation in CP2K) Method ɛ SCF Iterations Drift (Kelvin/ns) Guess Gear(4) Gear(4) Gear(4) What is the problem? Time reversibility has been broken!

15 Generalized Lagrangian L(q, q, x, ẋ) = 1 2 M q µẋ2 E(q, y) + kµg( x y ) y = F (q, x) G( x y ) wavefunction optimization wavefunction retention potential Equations of motion M q = E q E y µẍ = E y F x + kµ F q + kµ G F y q [ G x + G y F x Extension of Niklasson Lagrangian, PRL (2008) ]

16 Dynamical System

17 Car Parrinello Molecular Dynamics y = x G( x y ) = 0 Lagrangian L(q, q, x, ẋ) = 1 2 M q µẋ2 E(q, x) Equations of motion M q = E q µẍ = E x

18 Properties of CPMD Accuracy: Medium Distance from BO surface controlled by mass µ Requires renormalization of dynamic quantities (e.g. vibrational spectra) Stability: Excellent All forces can be calculated to machine precision easily Efficiency: Good Efficiency is strongly system dependent (electronic gap) Requires many nuclear gradient calculations Not implemented in CP2K!

19 BOMD y = Min x E(q, x) and µ = 0 Lagrangian L(q, q) = 1 2 M q 2 + E(q, y) L(x, ẋ) = 1 2ẋ2 + kg( x y ) M q ẍ Equations of motion = E q decoupled equations = k G x

20 BOMD with Incomplete SCF Convergence y = F(q, x) Min x E(q, x) and µ = 0 Equations of motion M q = E q E F y q ẍ = k G x G F y x SCF Error: Neglect of force terms E F y q G F and y x. EOM are coupled through terms neglected!

21 ASPC Integrator Integration of electronic DOF (x) has to be accurate: good wavefunction guess gives improved efficiency stable: do not destroy time-reversibility of nuclear trajectory ASPC: Always Stable Predictor Corrector J. Kolafa, J. Comput Chem. 25: (2004) ASPC(k): time-reversible to order 2k + 1

22 Orthogonality Constraint Wavefunction extrapolation for non-orthogonal basis sets C init = K B j+1 C t jτ C T t jτ S t jτ }{{} PS j=0 C t τ Coefficients B i are given by ASPC algorithm x = C init y[t] = C t S overlap matrix J. VandeVondele et al. Comp. Phys. Comm. 167: (2005)

23 Importance of Time-Reversibility Method ɛ SCF Iterations Drift (Kelvin/ns) Guess ASPC(3) ASPC(3) Gear(4) Gear(4) Gear(4)

24 Efficiency and Drift Method ɛ SCF Iterations Drift (Kelvin/ns) Guess ASPC(4) ASPC(4) ASPC(4) ASPC(4)

25 Efficiency and Drift Method ɛ SCF Iterations Drift (Kelvin/ns) ASPC(4) ASPC(5) ASPC(6) ASPC(7) ASPC(8)

26 Langevin BOMD Starting point: BOMD with ASPC(k) extrapolation Analysis of forces f BO (R) = f HF (R) + f Pulay (R) +f nsc (R) }{{} f (R) f BO : correct BO force f HF : Hellmann Feynman force f Pulay : Pulay force f nsc : non-self consistency error force f : approximate BO force

27 Forces in Approximate BOMD Approximate f nsc by fnsc = [( Vxc (ρ i ) ] ) dr ρ i ρ + V H ( ρ) I ρ i with ρ = ρ o ρ i, ρ o final (output) density, ρ i initial (predicted) density. Now assume f (R) + f nsc (R) = f BO (R) γ D Ṙ where γ D is a constant friction parameter.

28 Langevin EOM M R = f BO (R) (γ D + γ L )Ṙ + Θ with Θ a Gaussian random noise term and Θ(0)Θ(t) = 6(γ D + γ L )Mk B T δ(t) Given temperature T = k B 3 MṘ2 and an arbitrary γ L this determines γ D

29 Langevin BOMD 2nd Generation Car Parrinello (SGCP) T. D. Kühne et al., Phys. Rev. Lett (2007) single SCF step plus force correction needed extremely efficient for systems with slow SCF convergence γ D is small: correct statistics and dynamics difficult to stabilize in complex systems

30 SGCP: Examples T. Musso et al. Eur. Phys. J. B, (2018) Analysis of Force Errors Liquid Silicon at 3000 K

31 SGCP: Nanomesh h-bn on Rh(111)

32 SGCP: Force Errors

33 SGCP: h-bn on Rh(111) Comparision of MD Results

34 SGCP: h-bn on Rh(111) Performance

36 Micro-Canonical Ensemble System variables: N, V, E Microstate definition: Γ = {r N, p N } Energy H(Γ) = i p 2 i 2m i + U(r N ) All states have the same weight δ(h(γ) E) Partitionfunction Ω(N, V, E) = Γ δ(h(γ) E) = 1 N 1 h 3N dr N dp N δ(h(γ) E)

37 Molecular Dynamics Fixed number of particles: N Fixed volume: V Periodic boundary condition to avoid surface effects Total energy E(t) = i m i 2 ṙ 2 i (t) + U(r(t))

38 Constant of Motion: E de(t) dt = i F i = du(r) dr i Ė = i m i 2 2ṙ i r i + i = m i r i m i ṙ i F i m i + i du(r) ṙ i dr i ( F i )ṙ i = i (ṙ i F i ṙ i F i ) = 0 The energy is conserved. MD samples correct microstates.

39 Definition of Temperature Instantaneous temperature T (t) = 2 1 3Nk B 2 i m i ṙ 2 i (t) Temperature T = T (t)

40 Equipartition Theorem The equipartition theorem is a general formula that relates the temperature of a system with its average energies. For a system described by the Hamiltonian (energy function) H(r, ṙ) r i H r j = p i H p j = δ ij k B T For an ideal gas H = H kin = 3 2 k BT

41 Canonical Ensemble System variables: N, V, T Closed system exchanging energy with a heat bath at temperature T. How to simulate on a computer? We can directly manipulate velocities (temperature). No extended heat bath needed. Molecular dynamics with periodic boundary conditions, but manipulate ṙ i during the integration.

42 Control of velocities/temperature Differential control: the temperature is fixed to the predescribed value and no fluctuations occur. Proportional control: velocities are corrected at each integration step through a coupling constant towards the prescribed temperature. The coupling constant determines the strength of the fluctuations around T. Integral control: the Hamiltonian is extended and variables are introduced wich reflect the effect of an external system which fix the temperature. Time evolution is derived from the extended Hamiltonian. Stochastic control: position and velocities are propagated according to modified equations of motion with friction and stochastic forces. The final equations give the correct mean value of the temperature.

43 NVE H(Γ) = Γ = {r N, p N } N i=1 1 2m i p i + U(r N ) microstate All accessible microstates are equally probable const. W NVE (Γ) = δ(h(γ) E) statistical weight Distribution of microstates Ω NVE = Γ δ(h(γ) E) 1 N! 1 h 3N dr N p N δ(h(γ) E) 5 / 1

44 Canonical Ensemble (NVT) Microstates follow the Boltzmann s distribution [ W NVT (Γ) = exp H(rN, p N ] ) k B T Canonical partition function: describes the statistical properties Q NVT = Γ Probability distribution [ exp H(Γ) ] 1 k B T N!h 3N P(Γ) = W NVT(Γ) Q NVT dr N dp N [ exp H(rN, p N ] ) k B T and the expectation value of the total energy is (β 1/k B T ) { E = ln Q NVT = 1 [ H(Γ) exp H(Γ) ] } β Q NVT k B T Γ

45 Velocity Rescaling At t and temperature T (t) all velocities are multiplied by λ T = 1 2 N i=1 To control the temperature 2 m i (λv i ) Nk B 2 N i=1 λ = T 0 /T (t) 2 m i vi 2 3 Nk B = (λ 2 1)T (t) Easy to implement Sampling does not correspond to any ensemble Not recommended for production MD Good for equilibration Not time-reversible, not deterministic 24 / 1

46 Perform one (or more) MD steps Compute the instantanous kinetic energy Rescale the velocities by a factor 25 / 1

47 Jumps by Rescaling 20 15!=1 K 10 5 K t 20!= t 26 / 1

48 Berendsen Thermostat Weak coupling to a heat bath: temperature correction dt (t) dt = 1 τ (T bath T (t)) τ coupling parameter: larger τ weaker coupling ) Exponential decay of the system towards the desired temperature T = δt τ (T bath T (t)) τ, thermostat is inactive too small τ, unrealistically low fluctuations τ = δt, standard rescaling Same problems as velocity rescaling T Time 27 / 1

49 Like a friction Continuous formulation of the velocity rescaling (1 ζδt) = ṗ i = f i ζp i 1 δt τ p 2 i i m i Nk B T 1 Friction positive when K > Nk B T /2, cooling. 1/ !=1 K t 20 15!=10 K t

50 Langevin Thermostat Random forces due to impacts exerted by molecules in the environment: irregular motion and friction. At each step all v i are corrected by a random force and a friction term m r i = U r i mγṙ i + W i (t) Frictional force and random motion are related, i.e. fluctuation dissipation theorem is satisfied W i (t), W j (t ) = δ ij δ(t t )6mΓk B T Gaussian distribution with infinitely short correlation time (great number of collisions) Transition probability P t = i ( p i P + U P + m i D 2 P m i r i r i p i p 2 + Γ p ) ip i p i

51 Stochastic Behavior The canonical distribution is then recovered as ( ) lim P = C exp 1 t k B T { mp 2 i 2 + U(rN )} i and friction and diffusion are related D = Γ m k BT Local thermostat: noise term depends on the particle Samples from canonical ensemble and ergodic. Allows larger time steps compared to non-stochastic thermostats. Momentum transfer is destroyed, i.e., no diffusion coefficients. 34 / 1

52 Nose Thermostat Heat bath integral part of the system: s, ṡ, and Q (coupling) L = i ms 2 ṙ 2 i 2 U(r N ) Qṡ2 gk B T ln s The conjugate momenta in the extended system are p i L ṙ i = m i s 2 ṙ i p s L ṡ = Qṡ where the auxiliary variable is a time-scaling parameter dt = sdt p i = p/s real system momentum Microcanonical ensemble in the extended system: for g = 3N + 1 canonical ensemble for the real system 38 / 1

53 Nose Hoover Using only differentiation in real time t : 1 s H Nose = i d dt ṙ i = p i m i ṗ i = U ξp i r i ξ = sp s /Q ṡ s ξ = 1 Q = d ln s dt ( i = ξ p 2 i m i g β ) p i + U(r N ) + ξ2 Q 2m i 2 + g ln s β conserved 39 / 1

54 Friction Coefficient Define the relaxation time as ν T 3Nk B T /Q dynamics of friction coefficient by (feed-back mechanism) p 2 [ ] i ξ = νt 2 i m i 3Nk B T 1 T = νt 2 T 1 Too large Q: canonical distribution after long simulation Too small Q: high-frequency T oscillations

55 Ergodicity Problems Microcanonic Andersen Nose-Hoover v r r r NVE: closed loop, constant energy shell Andersen: points corresponding to a Gaussian velocity distribution Nose Hoover: band trajectory determined by the initial conditions Not sufficiently chaotic to sample all phase-space; trapped in a subspace 41 / 1

56 Nose Hoover Chain Ergodicity is improved by thermostatting the thermostat variable: M Nose Hoover thermostats ṙ i = p i m i ṗ i = U ξ 1 p i r i ξ 1 = 1 Q 1 ( i p 2 i g m i β ) ξ i ξ 2 ξ j = 1 Q j ( Qj 1 ξ 2 j 1 k B T ) ξ j ξ j+1 ξ M = 1 ( QM 1 ξm 1 2 k B T ) Q M H NHC = i p i 2m i + U(r N ) + j Q j ξ 2 j 2 M + gk B Ts 1 + k B Ts j j=2 42 / 1

Thermostats Comparison 600 velocity velocity rescaling 300 0 0 600 position 2 4 6 8 velocity Berendsen

57 Thermostats Comparison 600 velocity velocity rescaling position velocity Berendsen position velocity Nose-Hoover chain position 2 4 Time [ps] / 1

58 Thermostats Table tune cont. L/G correct ergodic cons. q. determ. Velocity rescaling G? X Andersen X L X X Berendsen X X G? X Nosé-Hoover X X L/G X X X Nosé-Hoover chains X X L/G X X X X Langevin X X L X X X Stochastic velocity rescaling X X L/G X X X 44 / 1

59 Other Ensembles N, V, E microcanonical inner energy U N, V, T canonical Helmholtz free energy F F = U TS N, P, H isobaric-isoenthalpic enthalpy H H = U + pv N, P, T isobaric-isothermal Gibbs free energy G G = U TS + pv = H TS

60 How to simulate an isobaric ensemble? We can manipulate the volume of the system. This emulates an external piston used to keep the pressure constant. Instantaneous pressure p(t) fluctuates, forces on the container Averaging p(t) gives the observable internal pressure.

61 Pressure: Virial Theorem From equipartition theorem pi 2 H kin = = 1 U(r) r i 2m i 2 r i i The total potential U(r) is coming from in internal potential V int (r) and and a wall potential W (r). W (r) r i = p df r = p dv (divr) = 3pV r i i i The virial theorem is then 3pV = 2 H kin i V (r) r i r i

62 Isobaric-Isoenthalpic Ensemble The system exchange work with an external surface: Ω is a variable Cubic box Ω h = 0 Ω Ω 1 3 Scaled coordinates r i = Ω 1/3 s i Lagrangian in terms of scaled coordinates External applied pressure P L = 1 m i Ω 2 3 ṡ2 2 i U(s N, Ω) i P Ω work on the system to balance P: the internal pressure fluctuates 46 / 1

63 Barostat Correct the internal pressure by scaling the voluem, i.e. the inter-particle distances The system is coupled to a barostat Isobaric-Isoenthalpic ensemble NPH P(r N, p N, Ω) δ(c K(p N ) U(r N, Ω) ΠΩ) where Π is the internal pressure Isobaric-Isothermal ensemble NPT ( P(r N, p N, Ω) exp K(pN ) + U(r N ), Ω) + ΠΩ k B T 47 / 1

64 Berendsen Barostat Deviation from desired pressure dp(t) dt = P 0 P(t) τ P τ P relaxation time Rescale volume by η the volume, η 1/3 coordinates γ is the compressibility. η(t) = 1 δt τ P γ(p 0 P(t)) Asotropic or anisotropic, τ P strength of coupling, not correct ensemble distribution

65 Andersen Barostat Associate a kinetic energy to the variable volume L = 1 m i Ω 2 3 ṡ2 2 i W Ω 2 U(s N, Ω) PΩ i The coupling mimics the action of a piston of mass W (strength of coupling) Equations of motion ( W Ω = 1 Ω 2 3 3Ω s i = 1 U 2 m i Ω 2 3 s i 3 m i ṡ 2 i i i Ω Ω s i ) U s i P = P int P s i Feedback loop to adjust the simulation volume: internal pressure fluctuates around the external pressure 49 / 1

66 Correct Distribution Conserved Hamiltonian of the 6N + 2 dimensional system H = 1 2 i π 2 i m i Ω Π2 2W + U(sN, Ω) + PΩ low W will result in rapid box size oscillations large W slow adjustmen of the volume Partition function (N, P, H) = dω dπ dπ N ) s N δ (H + Π2 2W H trajectories consistent with the NPH ensemble The combination with one of the constant temperature methods allows the simulation in the NPT ensemble. 50 / 1

67 Experiment with barostat and thermostat Cons. [kj/mol] V [A 3 ] P [katm] Berendsen Thermostat and Barostat τt =0.1 τp =1, 2, 4, time (ps) Cons. [kj/mol] V [A 3 ] P [katm] Nose--Hoover Thermostat and Barostat τt =0.1 τp =1, 2, 4, time (ps) 51 / 1

68 Multiple Time Step Method (r-respa) M.E. Tuckerman et al. JCP, (1992) Equation of motion ṙ = p m ṗ = F fast + F slow il = Liouville operator ( ) p m r + F fast + F slow p p + il ref + il slow Integrator x(δt) = exp(ilδt)x(0) = exp(il slow δt/2) [exp(il ref δt)] n exp(il slow δt/2) + O(δt 3 )

69 MTS in CP2K Two FORCE_EVAL sections defining methods Example DFT with hybrid functionals F ref = F GGA F slow = F hybrid

70 PLUMED: MD Driver and Enhanced Sampling Definition of many Collective Variables (CV) Analysis features for MD and MetaDynamics Bias methods for enhanced sampling MD Logarithmic Mean Force Dynamics Method Experiment Directed Simulation Methods

71 i-pi: External MD Driver Kapil et al., Comp. Phys. Comm. 236, (2018) Driver software communicating with force engines (e.g. CP2K) through socket interfaces

72 i-pi: Features MD and Path-Integral MD (PIMD) for NVE, NVT, NPT Ring Polymer Contraction (RPC) MD and Centroid MD Generalized Langevin Equation (GLE) thermostats PI+GLE and PIGLET Ring Polymer Instantons Thermodynamic Integration, Geomerty optimization, Saddle point search Harmonic vibrations Multiple Time Step algorithms Metadynamics (interface to PLUMED) Replica Exchage MD Second Generation CP like integration

73 i-pi: Multiple Time Steps

74 Summary: BOMD in CP2K Born Oppenheimer MD with ASPC(3) is default in CP2K SCF convergence criteria depends on system is a reasonable starting guess Best used together with OT and the FULL_SINGLE_INVERSE preconditioner, for large system an iterative update of the preconditioner should be used PRECOND_SOLVER INVERSE_UPDATE Langevin dynamics is an option needs some special care, analysis of forces

MD Thermodynamics. Lecture 12 3/26/18. Harvard SEAS AP 275 Atomistic Modeling of Materials Boris Kozinsky

MD Thermodynamics. Lecture 12 3/26/18. Harvard SEAS AP 275 Atomistic Modeling of Materials Boris Kozinsky MD Thermodynamics Lecture 1 3/6/18 1 Molecular dynamics The force depends on positions only (not velocities) Total energy is conserved (micro canonical evolution) Newton s equations of motion (second order