A Pareto Approach to Parameter Optimization of Empirical Interatomic Potentials

Event Sponsor: 
Mathematics and Computer Science Division LANS Seminar
Start Date: 
Dec 14 2016 - 3:00pm
Building 240/Room 1404-1405
Argonne National Laboratory
Eugene Ragasa
Speaker(s) Title: 
University of Florida
Stefan Wild

The traditional process of selecting an optimal parameterization of an empirical potential is generally approached as a multiple-objective optimization problem. An initial guess is made of the potential parameters and gradient methods are used to minimize a weighted function of the errors between the predicted and target parameter values. Here, we discuss an approach centered around the Pareto surface, which represents the set of all "optimal" parameterizations, calculated from the differences of the predicted values of physical properties of interest with respect to reference values. The methodology is illustrated with example results for a Buckingham potential for MgO to generate a Pareto surface. This uses the approach in an iterative process to identify an ensemble of parameterizations and illustrates the choices in tradeoffs of fidelity that need to be made in simulating competing material properties.

Miscellaneous Information: 

Refreshments will be served.