Multi-GPU parallelization of Deep Potential Molecular Dynamics for high-performance computing

The recently developed Deep Potential Molecular Dynamics [1,2] (DPMD) builds many-body potentials for atomic systems based on the deep neural network and has been successfully applied to a variety of systems. The DPMD model owns the quantum mechanical accuracy and the linear growth computational complexity. In this work, we develop the multi-GPU parallelization for DeePMD-kit [3], an implementation of DPMD, and optimize the workflows when the package interfaces with LAMMPS and TensorFlow. We demonstrate that the resulting package is well-suited for high-performance computing with the aim of performing large-scale molecular simulations with quantum mechanical accuracy.

[1] L. Zhang, J. Han, H. Wang, R. Car, W. E, Physical Review Letters 120, 14301 (2018)
[2] L. Zhang, J. Han, H. Wang, W. Saidi, R. Car, W. E, In Advances in Neural Information Processing Systems, pp. 4441-4451 (2018)
[3] H. Wang, L. Zhang, J. Han, W. E, Computer Physics Communications 228, 178–184 (2018)