Computer simulation techniques such as, molecular dynamics (MD) and Monte Carlo (MC) simulations play an important role in the fundamental understanding of chemical reactions and materials properties at the atomistic level that is difficult to obtain from experiments. Central to these atomistic simulations is the interatomic potential which describes the interactions between the atoms. An accurate description of interatomic interactions can be obtained using ab initio electronic structure methods. Such methods compute the potential energy and the corresponding force field from the first principles, but are also computationally expensive and hence, often limited to small systems.

A common approach relies on the use of empirical potential energy functions, which are relatively faster to evaluate. Such functions make it feasible to run simulations consisting of a few millions of atoms. In this study, a new method was developed for parameterization of the empirical potential energy functions using genetic algorithms (GA). GA is a stochastic, derivative-free global optimization technique which is particularly suitable for the optimization of multivariate, nonlinear functions exhibiting multiple local minima. The performance of GA was improved by implementing neural networks (NN) for evaluating the fitness function.

Although, empirical potential functions provide a simple and physically interpretable description for the interatomic interactions, their applicability is limited by the underlying functional form. A new method which combines the accuracy of ab initio electronic structure calculations and function approximation capabilities of neural networks was developed. Neural networks are trained to determine the functional relationship between configuration coordinates and the corresponding energies, forces and atomic charges. During training of the neural network, both the functional form as well as the parameters are optimized simultaneously and thus, obviating the requirement of a priori assumption about the functional form. Modified novelty sampling technique was implemented for identifying and sampling the important regions of multidimensional configuration space relevant in MD simulations. This method enables efficient sampling for developing ab initio-potential energy surfaces for new materials.