1. 1 CE 530 Molecular Simulation Lecture 12 David A. Kofke Department of Chemical Engineering SUNY Buffalo kofke@eng.buffalo.edu

2. 2 Review  Several equivalent ways to formulate classical mechanics Newtonian, Lagrangian, Hamiltonian Lagrangian and Hamiltonian independent of coordinates Hamiltonian preferred because of central role of phase space to development  Molecular dynamics numerical integration of equations of motion for multibody system Verlet algorithms simple and popular

3. 3 Dynamical Properties  How does the system respond collectively when put in a state of non-equilibrium?  Conserved quantities mass, momentum, energy where does it go, and how quickly? relate to macroscopic transport coefficients  Non-conserved quantities how quickly do they appear and vanish? relate to spectroscopic measurements  What do we compute in simulation to measure the macroscopic property?

4. 4 Macroscopic Transport Phenomena  Dynamical behavior of conserved quantities densities change only by redistribution on macroscopic time scale  Differential balance  Constitutive equation  Our aim is to obtain the phenomenological transport coefficients by molecular simulation note that the “laws” are (often very good) approximations that apply to not-too-large gradients in principle coefficients depend on c, T, and v Fick’s lawFourier’s lawNewton’s law MassEnergyMomentum

5. 5 Approaches to Evaluating Transport Properties  Need a non-equilibrium condition  Method 1: Establish a non-equilibrium steady state “Non-equilibrium molecular dynamics” NEMD Requires continuous addition and removal of conserved quantities Usually involves application of work, so must apply thermostat Only one transport property measured at a time Gives good statistics (high “signal-to-noise ratio”) Requires extrapolation to “linear regime”  Method 2: Rely on natural fluctuations Any given configuration has natural anisotropy of mass, momentum, energy Observe how these natural fluctuations dissipate All transport properties measurable at once Poor signal-to-noise ratio

6. 6 Mass Transfer  Self-diffusion diffusion in a pure substance consider tagging molecules and watching how they migrate  Diffusion equations Combine mass balance with Fick’s law Take as boundary condition a point concentration at the origin  Solution  Second moment RMS displacement increases as t 1/2 Compare to ballistic r ~ t For given configuration, each molecule represents a point of high concentration (fluctuation)

7. 7 Interpretation  Right-hand side is macroscopic property applicable at macroscopic time scales  For any given configuration, each atom represents a point of high concentration (a weak fluctuation)  View left-hand side of formula as the movement of this atom ensemble average over all initial conditions asymptotic linear behavior of mean-square displacement gives diffusion constant independent data can be collected for each molecule Slope here gives D Einstein equation

8. 8 Time Correlation Function  Alternative but equivalent formulation is possible  Write position r at time t as sum of displacements dt r(t) v

9. 9 Time Correlation Function  Alternative but equivalent formulation is possible  Write position r at time t as sum of displacements  Then r 2 in terms of displacement integrals rearrange order of averages correlation depends only on time difference, not time origin Green-Kubo equation

10. 10 Velocity Autocorrelation Function Backscattering  Definition  Typical behavior Zero slope (soft potentials) Diffusion constant is area under the curve Asymptotic behavior (t  ) is nontrivial (depends on d)

11. 11 Other Transport Properties  Diffusivity  Shear viscosity  Thermal conductivity

12. 12 Evaluating Time Correlation Functions  Measure phase-space property A(r N,p N ) for a sequence of time intervals  Tabulate the TCF for the same intervals each time in simulation serves as a new time origin for the TCF A(t 0 )A(t 1 )A(t 2 )A(t 3 )A(t 4 )A(t 5 )A(t 6 )A(t 7 )A(t 8 )A(t 9 ) A0A1A0A1 A1A2A1A2 A2A3A2A3 A3A4A3A4 A4A5A4A5 A5A6A5A6 A6A7A6A7 A7A8A7A8 A8A9A8A9 A(t 0 )A(t 1 )A(t 2 )A(t 3 )A(t 4 )A(t 5 )A(t 6 )A(t 7 )A(t 8 )A(t 9 ) ++++++++ +... A0A2A0A2 A1A3A1A3 A2A4A2A4 A3A5A3A5 A4A6A4A6 A5A7A5A7 A6A8A6A8 A7A9A7A9 A(t 0 )A(t 1 )A(t 2 )A(t 3 )A(t 4 )A(t 5 )A(t 6 )A(t 7 )A(t 8 )A(t 9 ) +++++++ +... A0A5A0A5 A1A6A1A6 A2A7A2A7 A3A8A3A8 A4A9A4A9 A(t 0 )A(t 1 )A(t 2 )A(t 3 )A(t 4 )A(t 5 )A(t 6 )A(t 7 )A(t 8 )A(t 9 ) ++++ +... tt tt

13. 13 Direct Approach to TCF  Decide beforehand the time range (0,t max ) for evaluation of C(t) let n be the number of time steps in t max  At each simulation time step, update sums for all times in (0,t max ) n sums to update store values of A(t) for past n time steps at time step k:  Considerations trade off range of C(t) against storage of history of A(t) requires n 2 operations to evaluate TCF kk-n

14. 14 Fourier Transform Approach to TCF  Fourier transform  Convolution  Fourier convolution theorem  Application to TCF Sum over time origins With FFT, operation scales as n ln(n)

15. 15 Coarse-Graining Approach to TCF 1.  Evaluating long-time behavior can be expensive for time interval T, need to store T/  t values (perhaps for each atom)  But long-time behavior does not (usually) depend on details of short time motions resolved at simulation time step  Short time behaviors can be coarse-grained to retain information needed to compute long-time properties TCF is given approximately mean-square displacement can be computed without loss Method due to Frenkel and Smit

16. 16 Coarse-Graining Approach to TCF 2. n0

17. 17 Coarse-Graining Approach to TCF 2. n0

18. 18 Coarse-Graining Approach to TCF 2. n0n0

19. 19 Coarse-Graining Approach to TCF 2. n0

20. 20 Coarse-Graining Approach to TCF 2. n0

21. 21 Coarse-Graining Approach to TCF 2. n0n0

22. 22 Coarse-Graining Approach to TCF 2. n0

23. 23 Coarse-Graining Approach to TCF 2. n0

24. 24 Coarse-Graining Approach to TCF 2. n0n0 n0

25. 25 Coarse-Graining Approach to TCF 2. n0 et cetera

26. 26 Coarse-Graining Approach to TCF 2. n3n3 0 This term gives the net velocity over this interval Approximate TCF

27. 27 Coarse-Graining Approach: Resource Requirements  Memory for each level, store n sub-blocks for simulation of length T = n k  t requires k  n stored values compare to n k values for direct method  Computation each level j requires update (summing n terms) every 1/n j steps total scales linearly with length of total correlation time compare to T 2 or T ln(T) for other methods

28. 28 Summary  Dynamical properties describe the way collective behaviors cause macroscopic observables to redistribute or decay  Evaluation of transport coefficients requires non-equilibrium condition NEMD imposes macroscopic non-equilibrium steady state EMD approach uses natural fluctuations from equilibrium  Two formulations to connect macroscopic to microscopic Einstein relation describes long-time asymptotic behavior Green-Kubo relation connects to time correlation function  Several approaches to evaluation of correlation functions direct: simple but inefficient Fourier transform: less simple, more efficient coarse graining: least simple, most efficient, approximate