386496 Development of a Dynamic Cutoff Method and Interface Detection Algorithm for Long-Range Dispersion Interactions

Tuesday, November 18, 2014: 10:10 AM
212 (Hilton Atlanta)
Paul Springer1, Paolo Bientinesi1, Benjamin Berkels1 and Ahmed E. Ismail1,2, (1)AICES Graduate School, RWTH Aachen University, Aachen, Germany, (2)Aachener Verfahrenstechnik, Faculty of Mechanical Engineering, RWTH Aachen University, Aachen, Germany

We discuss the development and implementation of a dynamic cutoff method (DCM) for molecular simulations which exhibits a linear-time complexity as well as good scalability properties.  The DCM, which is intended for simulations of large interfacial systems to be performed on modern supercomputers, adjusts the cutoff for dispersion potentials for atoms as a function how far removed the atoms are from an interface. This design choice allows particles in a bulk phase to benefit from the reduced computation with limited loss of accuracy associated with traditional dispersion cutoffs, while incorporating more interactions for particles near an interface without the need for communications-intensive Ewald-like methods.

Moreover, the dynamic cutoff method requires rapid computation of the location of the interface throughout the course of a simulation. We have therefore also developed a real-time interface detection method which is perfectly parallelizable and scales linearly with the number of particles. This method is based on the work of Berkels et al., who developed an efficient image segmentation algorithm.

We have implemented the dynamic cutoff method as part of the open-source LAMMPS molecular dynamics package. Our results for a liquid/vapor system of Lennard-Jones particles, shows that the accuracy of the DCM is comparable to that of PPPM. A performance comparison indicates that the DC method is preferable for large systems because of the poor scaling of FFTs within the PPPM algorithm. We also demonstrate the good weak- and strong-scaling behavior of this method on the BlueGene/Q supercomputer at the Forschungszentrum Jülich.

Extended Abstract: File Not Uploaded