Machine learning potentials for atomistic simulations

Computer simulations are now standard in condensed matter and materials physics, but there is an economic need to accurately simulate systems with increasingly more atoms and from more fundamental physics. Molecular dynamics (MD) simulations based on Density Functional Theory (DFT) are powerful, but are restricted to several hundred atoms and practical simulation times of less than 1 ns. Classical MD simulations using analytic potentials are orders of magnitude faster, but the potentials are based on more crude approximations and are therefore less accurate. Additionally, they are often qualitatively wrong when simulating phenomena in which the approximations potentially break down, such as phase transitions or at surfaces.

The field of machine learning (ML) has exploded in recent years. In particular, neural networks have been applied with great success in areas like autonomous driving to mastering Go. Many ML algorithms can be trained on potential energy surfaces generated using DFT as a function the atomic configuration of the system, which can include information about different crystal structures, surfaces, defects, etc. These ML potentials can be almost as accurate as the DFT calculations that generated the data but can the simulations can also be almost as fast as those performed with analytical potentials.

Our group is interested in developing methodologies and open source tools to easily and reliably generate ML potentials, and to use these potentials to study systems with a realistic number of atoms in which size effects are important, such as nanoparticles.

People

Jorge A. Muñoz
BS UTEP '07. MS, PhD Caltech '09, '13
Reynaldo Martinez
Reynaldo Martinez
Physics and Computer Science
Adrian De la Rocha A
Adrian De la Rocha
Physics