Determining neighborhood phases in hard-sphere systems using machine learning

A challenging problem in particle-based modeling is one of classifying the many structures which involve very large networks of bonds. Based on capacity to judge if a system is amorphous or solid from radial distribution functions, we set up two machine-learning systems able to identify local structures in mono-component hard-sphere simulations. The machines are constituted of logistic or support-vector regressions applied to linear model, second- and third-degree polynomial hypothesis. We labeled the sphere as solid or amorphous following a bond-order parameter and characterized them with radial structure functions. The features were enough to machine-learning systems predicting the labels with great accuracy.