Abstract
We develop a new algorithm for database optimization of interatomic
potential models with Bayesian statistics. Conventional classical
potential fitting schemes generates a best fit parameter set, but do not
show inadequacies of the potential model nor give insight into viability
of the fitting database. Our algorithm generates an ensemble of
potential fits with Markov Chain Monte Carlo and make predictions based
on Bayesian error estimation according to the ensemble. We consider a
fitting database to be optimal when the sum of relative errors for all
entries of the database is minimized. A specific objective function is
proposed and an optimized database of the interatomic potential model
can be obtained by modifying the relative importance (weights) of
different structures in the database. We test the algorithm with a
Lennard-Jones potential fitting of Ti, which shows specific limitations
of this simple potential model. We also show that the derivative of the
objective function with respect to weight determines whether a structure
should be added to or removed from the database.
| Original language | English (US) |
|---|---|
| Pages (from-to) | 25007 |
| Journal | American Physical Society, APS March Meeting 2013, March 18-22, 2013, abstract #W25.007 |
| State | Published - Mar 1 2013 |