 |
Center for Nonlinear Studies |
Molecular simulations are the largest consumer of supercomputing time worldwide. Molecular simulations rely on one of the two models: accurate and very computationally expensive quantum-mechanical models, most notably the density functional theory, and empirical interatomic potentials that postulate a simple functional form of interatomic interaction that is fast to compute. Machine learning interatomic potentials (MLIPs) has recently been put forward as a promising methodology of combining the quantum-mechanical accuracy and the computational efficiency of the empirical potentials. MLIPs postulate a functional form that is fast to compute, yet flexible enough to be able to represent arbitrary interatomic interactions. In my talk I will give an overview of the existing developments in the field of MLIPs, present the MLIPs developed in my group, and finally show how active learning can ensure reliability of such potentials. I will illustrate applications of such potentials in molecular dynamics, crystal structure prediction, prediction of alloy phase diagrams, and cheminformatics. If you would like to meet with Prof. Shapeev, please contact Sergei Tretiak (serg@lanl.gov)
Host: Sergei Tretiak