Alan Nichol (Eng)
ESDG November 13th 2013
Intra- and intermolecular potential energy surfaces derived from ab initio data
by machine learning
In recent years it has become possible to predict the properties of
molecular materials using potential energy surfaces fitted only to ab initio
calculations. With current methods and hardware it feasible to solve the
Schroedinger equation at the CCSD(T) level of accuracy for at most a few
molecules at a time. Unfortunately this and other quantum chemistry methods
scale steeply with system size [O(N7) for CCSD(T)] so in many cases cannot
be applied. Much effort has been devoted to developing potential energy
surfaces which map the accurate quantum mechanical results to functional
forms which can be evaluated at vastly reduced expense. A number of methods
exist for fitting these potentials, most of which require a great deal of
expert knowledge, iterative improvement, or both. We use the recently
developed method of Gaussian Approximation Potentials (GAP) to improve upon
this process in several ways. GAP makes use of Gaussian Process regression,
a principled, automatic, nonparametric, Bayesian approach to function
fitting. We show how intra- and intermolecular potential energy surfaces for
arbitrary molecules can be made automatically using GAP, and how these can
be made systematically more accurate by including more training data.