Introduction to self-learning hybrid Monte Carlo

Shiga, Motoyuki; Kobayashi, Keita; Okumura, Masahiko; Nagai, Yuki

no journal, , 

We propose a self-learning hybrid Monte Carlo (SLHMC) method, which is a first-principles simulation that enables efficient sampling by an auxiliary use of machine learning potentials. A classical trajectory generated on an approximate machine learning potential is regarded as a trial move. The acceptance probability by the Metropolis algorithm is then given based on the difference of Hamiltonian of density functional theory (DFT) energies. It is gauranteed that the sampled configurations constitutes an exact ensemble at the DFT level for a given thermodynamic condition. It is possible to improve the quality of the machine learning potential on the fly by adding the DFT energies sequentially to the training set. It was found that the SLHMC is able to accelerate the sampling by several tens of times more than first-principles molecular dynamics.