PyAMFF: A machine learning package for fitting and using atom-centered machine learning force field

Ab-initio calculations have been used commonly to study dynamics of complex systems and atomic-level understandings are significant to determine their barriers, reaction pathways and transition states. The high accuracy of these calculations, however, accompanies increase of computational cost and such tradeoff restricts their applicable areas. In an effort to overcome this, applying machine learning (ML) technique to learn potential energy surface (PES) of systems being studied has become one of the promising solutions for the last decades. The main advantages of using ML is that the accuracy is remained as high as ab-initio level while the computational cost is not. Especially, Behler and Parinello proposed a method of describing a system numerically in terms of atomic environment and feed the descriptor to the neural network (NN) that returns to the atomic energy of the system and allows force evaluation. PyAMFF (Python Atom-centered Machine Learning Force Field) is built off of this idea and provides user-friendly and high performance computations interfaced with the EON software, which supports long-timescale molecular dynamics. In the poster, I will present the PyAMFF code framework and its application for Li dynamics.