Atomistic Simulations of PETN Detonation with a Machine Learning Potential
Accurate modeling of the detonation of explosives is currently limited by our inability to describe the release of chemical energy at the shock front. Continuum models are insufficiently accurate while the time required for ab initio molecular dynamics (AIMD) simulations scales with the cube of half the number of valence electrons in the system; the performance of a realistic simulation would require far too much time. There has been tremendous interest in constructing force fields for molecular dynamics (MD) using Machine Learning (ML). Recent developments have the potential to revolutionize computational materials science by overcoming the limitations of classical force fields, despite 40 years of rigorous work devoted to creating realistic classical models. We will describe attempts to train an ML interatomic potential (MLIP) using DeepMD, winner of the ACM Gordon Bell Award in 2020. The training data consist of one of the largest practical DFT simulations of PETN detonation ever performed: 7,424 atoms over 1594.25 fs, with 0.25 fs time steps. This calculation took two years to complete. We have trained neural network MLIPs with DeepMD on Nautilus and Raider utilizing data parallelism with multi-node training on both the A100 GPU nodes and the large-memory CPU nodes. Preliminary results strongly supported the importance of a long cutoff length, and this has placed severe memory requirements as the sizes of the neural networks was increased due to the large number of atoms in a single configuration. We compare the results of classical MD simulations with a MLIP using LAMMPS to the ab initio results obtained with CP2K.
PRESENTER
Lill, James
James.Lill@GDIT.com
937-255-1461
PET
CO-AUTHORS
Lill, James
James.Lill@GDIT.com
Boyer, Mathew
Mathew.Boyer@GDIT.com
Rice, Betsy M
betsy.rice.vol@army.mil
Byrd, Edward FC
edward.f.byrd2.civ@army.mil
CATEGORY
Computational Chemistry & Materials
SYSTEMS USED
Nautilus, Raider, Narwhal
SECRET
No