Large-Scale, Long-Term Nonadiabatic Electron Molecular Dynamics for Describing Material Properties and Phenomena in Extreme Environments

We describe the first principle-based electron force field (eFF) methodology for modeling the simultaneous dynamics of electrons and nuclei (eMD) evolving nonadiabatically under transient extreme conditions. We introduce the parallel implementation of eFF (pEFF) that makes it practical to perform simulations of the nonadiabatic dynamics of materials in extreme environments involving millions of nuclei and electrons, over multi-picoseconds time scales, and demonstrate its application to: (i) accurately determine density and predict percent ionization of hydrogen at high pressure (similar to 61 GPa) and temperatures up to 15,300K and (ii) determine, the single shock Hugoniot for lithium metal directly from the shock wave kinematics, i.e., mass velocities (U-p) and shock wave velocities (U-s), and shock density data. For (i), the density and ionization fractions of hydrogen atoms were calculated using the isobaric-isothermal ensemble at an isotropic pressure of 61.4 GPa and for temperatures between 300K and 15,300 K. The results at 15,300K describe a molecular fluid with density rho = 0.36 g/cm(3), in close agreement with existing experiments and theory, and similar to 0.5% ionization. This result provides no indication of the existence of a critical plasma phase-transition point at this particular temperature and pressure, as previously predicted by others. For (ii), the relationship between U-p and U-s was characterized to be linear and plastic in the range 1-20 km/s, and the single shock Hugoniot was determined in close agreement with published results for experimentally reported U(p)s. In addition to this, we provide a description of the materials' behavior for large U(p)s in terms of the appearance of a weak metallic plasma phase by U-p = 10 km/s, with similar or equal to 8% ionization, gradually transitioning to a denser plasma with an estimated similar or equal to 35% ionization by U-p = 15 km/s. Last but not least, we confirm the computational efficiency and scalability of pEFF by comparing its single processor performance against the fastest existing serial code, which results in a linear speedup similar to 10x faster for every 16,000 particles in favor of pEFF, and by evaluating its parallel performance in terms of its strong and weak scaling capabilities. Our results, on Los Alamos's Lobo supercomputer (a 38TFLOPSs Linux HPC with Quad-core AMD Opteron nodes interconnected with an Infiniband), show strong scaling with near ideal speedups for loads >62 particles per processor. Weak scaling is shown to be close to linear under the same per-processor load range. As an absolute reference, an NVT run with 2 million particle lithium bulk system (0.5 M nuclei and 1.5 M electrons) on Lobo takes similar to 0.44 s/timestep on 1024 processors (similar to 1 day/ps using an integration timestep of 0.005 fs). (C) 2010 Wiley Periodicals, Inc. J Comput Chem 32: 497-512, 2011