Acceleration of Semiempirical Quantum Mechanical Calculations by Extended Lagrangian Molecular Dynamics Approach

The implementation and performance of the atom-centered density matrix propagation (ADMP) [J. Chem. Phys. 2001, 114, 9758] and the curvy-steps (CURV) methods [J. Chem. Phys. 2004, 121, 1152] are described. These methods solve the electronic Schrödinger equation approximately by propagating the electronic degrees of freedom using the extended Lagrangian molecular dynamics (ELMD) simulation approach. The ADMP and CURV methods are implemented and parallelized to accelerate semiempirical quantum mechanical (QM) methods (such as the MNDO, AM1, PM3, MNDO/d, and AM1/d methods). Test calculations show that both the ADMP and the CURV methods are 2∼4 times faster than the Born-Oppenheimer molecular dynamics (BOMD) method and conserve the total energy well. The accuracy of the ADMP and CURV simulations is comparable to the BOMD simulations. The parallel implementation accelerates the MD simulation by up to 28 fold for the ADMP method and 25 fold for the CURV method, respectively, relative to the speed of the single core BOMD. In addition, a multiple time scale (MTS) approach is introduced to further speed up the semiempirical QM and QM/MM ELMD simulations. Since a larger integration time step is used for the propagation of the nuclear coordinates than that for the electronic degrees of freedom, the MTS approach allows the ELMD simulation to be carried out with a time step that is larger than the time step accessible by the original ADMP and CURV methods. It renders MD simulation to be carried out about 20 times faster than the BOMD simulation, and yields results that are comparable to the single time scale simulation results. The use of the methods introduced in the present work provides an efficient way to extend the length of the QM and QM/MM molecular dynamics simulations beyond the length accessible by BOMD simulation.