Parallel Implementation of Machine Learning-Based Many-Body Potentials on CPU and GPU

Authors: Yaoguang Zhai (University of California, San Diego), Nathaniel Danandeh (University of California, San Diego), Zhenye Tan (University of California, San Diego; Tongji University), Sicun Gao (University of California, San Diego), Francesco Paesani (University of California, San Diego), Andreas W. Goetz (San Diego Supercomputer Center)

Abstract: Machine learning models can be used to develop highly accurate and efficient many-body potentials for molecular simulations based on the many-body expansion of the total energy. A prominent example is the MB-pol water model that employs permutationally invariant polynomials (PIPs) to represent the 2-body and 3-body short-range energy terms.

We have recently shown that the PIPs can be replaced by Behler-Parinello neural networks (BP-NN). We present OpenMP parallel implementations of both PIP and BP-NN models as well as a CUDA implementation of the BP-NN model for GPUs. The OpenMP implementations achieve linear speedup with respect to the optimized single threaded code. The BP-NN GPU implementation outperforms the CPU implementation by a factor of almost 8. This opens the door for routine molecular dynamics simulations with highly accurate many-body potentials on a diverse set of hardware.

Best Poster Finalist (BP): no

Poster: pdf
Poster summary: PDF
Reproducibility Description Appendix: PDF


Back to Poster Archive Listing