LAMMPS is a classical molecular dynamics code with a materials focus. I'll discuss three efforts the development team is currently working on which attempt to leverage large-scale parallelism in different ways. The first is formulation and use of new classical interatomic potentials which can be automatically calibrated against large data sets of quantum electronic structure calculations (DFT). These SNAP potentials have linear computational cost in the size of the simulated system and can be more accurate than conventional potentials for important material properties, albeit at higher per-atom cost. A side benefit is this allows SNAP models to strong-scale more efficiently on large machines. The second is algorithmic work to extend the hyperdynamics (HD) method to support larger systems and more parallelism. HD can enable longer timescale simulations for certain kinds of systems. The third is methods for weighted geometric load-balancing which we have found particularly useful for coarse-grained models where the simulation domain can have large or time-dependent variations in particle density. Finally, I will briefly highlight how LAMMPS is addressing the challenge of achieving optimal computational performance for a wide variety of models and potentials running on the growing range of node architectures deployed in current and future HPC platforms.

Recent algorithmic work in LAMMPS for extending accuracy and time scales for materials modeling

254th National Meeting (2017)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.