Computational Improvements to Quantum Wave Packet ab Initio Molecular Dynamics Using a Potential-Adapted, Time-Dependent Deterministic Sampling Technique.

In a recent publication, we introduced a computational approach to treat the simultaneous dynamics of electrons and nuclei. The method is based on a synergy between quantum wave packet dynamics and ab initio molecular dynamics. Atom-centered density-matrix propagation or Born-Oppenheimer dynamics can be used to perform ab initio dynamics. In this paper, wave packet dynamics is conducted using a three-dimensional direct product implementation of the distributed approximating functional free-propagator. A fundamental computational difficulty in this approach is that the interaction potential between the two components of the methodology needs to be calculated frequently. Here, we overcome this problem through the use of a time-dependent deterministic sampling measure that predicts, at every step of the dynamics, regions of the potential which are important. The algorithm, when combined with an on-the-fly interpolation scheme, allows us to determine the quantum dynamical interaction potential and gradients at every dynamics step in an extremely efficient manner. Numerical demonstrations of our sampling algorithm are provided through several examples arranged in a cascading level of complexity. Starting from a simple one-dimensional quantum dynamical treatment of the shared proton in [Cl-H-Cl](-) and [CH3-H-Cl](-) along with simultaneous dynamical treatment of the electrons and classical nuclei, through a complete three-dimensional treatment of the shared proton in [Cl-H-Cl](-) as well as treatment of a hydrogen atom undergoing donor-acceptor transitions in the biological enzyme, soybean lipoxygenase-1 (SLO-1), we benchmark the algorithm thoroughly. Apart from computing various error estimates, we also compare vibrational density of states, inclusive of full quantum effects from the shared proton, using a novel unified velocity-velocity, flux-flux autocorrelation function. In all cases, the potential-adapted, time-dependent sampling procedure is seen to improve the computational scheme tremendously (by orders of magnitude) with minimal loss of accuracy.