A SURROGATE MODEL OF MOLECULAR DYNAMICS SIMULATIONS FOR POLAR FLUIDS: SUPERVISED LEARNING METHODS FOR MOLECULAR POLARIZATION AND UNSUPERVISED METHODS FOR PHASE CLASSIFICATION

Abstract

Molecular Dynamic (MD) simulation is a standard computational tool in soft matter physics. While very powerful, it is computationally expensive, leading to some simulations taking days or even weeks to complete depending on the size of your computer cluster. Finding computationally cheap surrogate models which can learn the output features of MD simulation is therefore highly motivated. In this report I explore the use of deep neural network ensembles as well as support vector machine regressors as surrogate models for MD simulation. From the output of the surrogate models, we can then employ unsupervised learning methods to get insight into the physics of our system, and classify boundaries between phases. We will also show the potential of this method to uncover behavior not realized by other methods.