Back to Colloquia

Physics Colloquium
Friday, October 7th, 2005, 4:00 P.M.

E300 Math/Science Center; Refreshments at 3:30 P.M. in Room E200

Duane D. Johnson

University of Illinois Urbana-Champaign

Multiscaling Methods using Machine-Learning via Genetic Programming

Genetic Programming (GP) -- a genetic algorithm (GA) that evolves computer programs -- is a machine-learning strategy for optimization in complex engineering problems or regression of complex relationships. Here I discuss using GP concepts to eliminate the bottleneck for multi-timescale modeling computation of the entire potential energy surface. Dramatic computational saving is obtained by avoiding explicit calculation of all kinetic activation barriers, and we simulate dynamics in materials that span seconds (experimental times) via kinetic Monte Carlo simulations (~9 orders increase in time at 300 K over Molecular Dynamics). To exemplify the ideas, we apply a simple GP to a reasonably complex case vacancy-assisted migration on a surface of a phase-separating fcc binary alloy. Extending these ideas I will show our initial work on using GP to create empirical potentials that approach accuracy of ab initio quantum chemistry methods. Time permitting I will briefly mention the other disparate, but interesting, research areas pursued in my group.

Duane Johnson is Professor and Bliss Faculty Scholar in Materials Science & Engineering and Physics, and he is the Director of the Materials Computation Center at The Frederick Seitz Materials Research Laboratory at the University of Illinois, Urbana-Champaign

Emory College | Graduate School of Arts and Sciences | Emory University | Search | Index | Help