Schedule Nov 06, 2009
Using Dimensionality Reduction to Build a Potential Energy Surface for N2O/Cu(100) from Extremely Sparse Ab Initio Data
Sergei Manzhos (Univ. Tokyo)

Sergei Manzhos, Koichi Yamashita
Department of Chemical Systems Engineering, School of Engineering, University of Tokyo, 7-3-1, Hongo, Bunkyo-ku, Tokyo 113-8656, Japan


For the first time, we use a dimensionality reduction technique to build a continuous potential energy surface (PES) for a polyatomic molecule-surface reaction from extremely sparse ab initio samples.

DFT slab calculations are used to sample the configuration space of the system N2O/Cu(100), and the PES function is built with a method of Manzhos and Carrington [J. Chem. Phys. 127, 014103 (2007)] from only 4,300 single point energies. Physically motivated coordinates are used which expand the dimensionality to D=15 (resulting in sampling density of 1.75 data per dimension). A neural-network based algorithm finds a smaller set of new, adaptive coordinates in which the non-linear fit to the data is performed.

Molecular dynamics simulations are performed on the PES to calculate the probability of dissociative adsorption.

View poster as pdf.

Author entry (protected)