(377h) An Evaluation of Learning-Augmented Monte Carlo

A common yet difficult problem encountered in molecular simulations is given a set of N molecules and their interactions, what configuration minimizes their potential energy? In 2005, Troisi, Ratner and Wong developed a novel algorithm for finding energy minimizing configurations of particles. Rigid, lattice-based molecules are able to learn about local clusters of molecules that minimize potential energy and bias the formation of these clusters. By moving these clusters as whole units, the algorithm finds configurations of lower potential energy in smaller amounts of time than a traditional Metropolis Monte Carlo simulation. We improve upon the Troisi algorithm and quantify its performance gains vs. Monte Carlo with cluster moves, a more meaningful comparison. We find that the learning algorithm does not sample the NVT ensemble and performs no better than cluster Monte Carlo in finding energy-minimizing configurations.

Jankowski, E. & Glotzer, S.C. (2007) Preprint.

Troisi, A., Ratner, M. & Wong, V., (2005) Proc. Natl. Acad. Sci. USA 102 255-260.