Artificial Neural Networks (ANN) Applied to Monte Carlo Simulation of Helium.

The computational expense to accurately simulate large atomistic systems is enormous.  In order to overcome this computational cost empirical potentials are used, which rely on parameters that are adjusted to reproduce either experimental or first principle references.  Furthermore, careful parameterization of such potentials can be challenging and must be validated for each system.  Artificial Neural Networks (ANN) approaches such as those proposed by Behler and Parrinello circumvent this parameterization problem by employing machine learning techniques to calculate the potentials.^1,2  The Behler-Parrinello approach calculates the potential energy surface based off a set of reference calculations. This approach has the advantage that it can be well automated, new potentials can be constructed easily, permits application of potentials to arbitrary structures, and reaches accuracies comparable to first principle methods with a much lower computational cost.  We used ANN’s to calculate the potential energy surfaces (force field) for an atomistic Monte Carlo simulations.³  In particular, we are interested in simulating helium to obtain a structure factor to compare to experimental results. 

 

1.  Behler, M. Parrinello, Phys. Rev. Lett. 98 (2007) 146401.

2.  J. Behler, Int. J. Quant. Chem. 115 (2015) 1032–1050.

3.  N. Artrith, A. Urban, Computational Materials Science 114 (2016) 135–15.