Noam Bernstein, Michael J. Aziz, Div. of Engineering and Applied Sciences, Harvard University, Cambridge, MA; Efthimios Kaxiras, Physics Dept and Div. of Engineering and Applied Sciences, Harvard University, Cambridge, MA.
To understand the microscopic mechanism for cystallization we need to distill, from the atomic positions as a function of time that we obtain from MD, a description of the underlying network rearrangements. To conform to our picture of a ``mechanism'' this description should comprise a series of localized connectivity changes that lead to the crystallization of some atoms. To that end we calculate time averaged positions from an MD run (to smooth out thermal vibration), and find clusters of adjacent atoms that change network connectivity at the same time. We do this for each time step, and define ``mechanisms'' composed of a time series of clusters that overlap in some of their participating atoms.
By inspecting sequences of configurations from an MD run we pick out groups of atoms that crystallize and approximately when they move into crystal lattice positions. This information allows us to select the network connectivity change ``mechanisms'' that lead to crystallization. By playing back the configurations captured from the MD run with all motion outside of the crystallizing ``mechanism'' frozen out we generate a sequence of atomic positions. These are completely relaxed, and then used as an input to a program that finds low energy paths connecting the stable configurations. This program creates chains in configuration space coupled by harmonic springs and minimizes the total energy of the path by conjugate gradient relaxation. Using this procedure we have found seven events that led to the crystallization of a total of 16 atoms. These include some simple mechanisms with high activation energies, some complex ones with lower activation energies, and one very complex mechanism with a high activation energy.