Computational cost severely limits the range of ab initio molecular dynamics simulations. Machine learning techniques are rapidly changing this state of affairs. Deep neural networks, that learn the interatomic potential energy surface from ab-initio data, make possible simulations with quantum mechanical accuracy at the cost of empirical force fields. These approaches can model not only atomistic dynamics but also dielectric response properties measured in experiments. I will discuss, in particular, the deep potential method developed at Princeton. In combination with incremental learning techniques, this approach makes possible to construct, with minimal learning cost, reactive potentials that are accurate over a vast range of thermodynamic conditions, such as the pressure and temperature regimes underlying molecular and ionic phases of water. The methodology will be illustrated with applications to some classic problems in physical chemistry.

About the Speaker

Roberto Car is the Ralph W. Dornte Professor in Chemistry at Princeton University, and Professor of Chemistry in the Princeton Institute for the Science and Technology of Materials. In 2016, Dr. Car was elected as a member of the National Academy of Sciences for his distinguished achievements in scientific research. Among his many awards he has received the A. Rahman Prize of the American Physical Society and the Dirac Medal and Prize of the International Centre for Theoretical Physics.

Learn more about Dr. Roberto Car in our recent news feature: https://www.physast.uga.edu/blogs/news/posts/2021-chhabra-landau-lecture-announced