Optimized particle-mesh Ewald/multiple-time step integration for molecular dynamics simulations

We develop an efficient multiple time step (NITS) force splitting scheme for biological applications in the AMBER program in the context of the particle-mesh Ewald (PME) algorithm, Our method applies a symmetric Trotter factorization of the Liouville operator based on the position-Verlet scheme to Newtonian and Langevin dynamics. Following a brief review of the NITS and PME algorithms, we discuss performance speedup and the force balancing involved to maximize accuracy, maintain long-time stability, and accelerate computational times. Compared to prior NITS efforts in the context of the AMBER program, advances are possible by optimizing PME parameters for NITS applications and by using the position-Verlet, rather than velocity-Verlet, scheme for the inner loop. Moreover, ideas from the Langevin/MTS algorithm LN are applied to Newtonian formulations here. The algorithm's performance is optimized and tested on water, solvated DNA, and solvated protein systems. We find CPU speedup ratios of over 3 for Newtonian formulations when compared to a 1 fs single-step Verlet algorithm using outer time steps of 6 fs in a three-class splitting scheme; accurate conservation of energies is demonstrated over simulations of length several hundred ps. With modest Langevin forces, we obtain stable trajectories for outer time steps up to 12 fs and corresponding speedup ratios approaching 5. We end by suggesting that modified Ewald formulations, using tailored alternatives to the Gaussian screening functions for the Coulombic terms, may allow larger time steps and thus further speedups for both Newtonian and Langevin protocols;, such developments are reported separately. (C) 2001 American Institute Of Physics.