2007 Seminar at Lawrence Livermore National Laboratory: Atomistic Continuum Alison Kubota

This is a presentation from several years ago. The first part of this presentation involves work done mostly with Dr. Wilhelm G. Wolfer. This part involves developing continuum models by first developing nonparametric estimations of continuum properties from extremely large simulations on some of the largest supercomputers in the world.

The second part of the presentation deals with potential development which involves parameterizing complex models using some heuristic training rules for continuum thermoelastic calculations from atomistic simulations. Each thermoelastic calculation uses a 20 million iteration Metropolist Monte Carlo simulation in order to transform the static interatomic potential into thermoelastic properties using the fluctuation techniques by Ray and also by Parinello and Rahman.

The goal is thus to find the right parameterization for the potentials that will generate thermoelastic properties as a function of temperature, that most closely matches experimentally observed data. This is a technique that might be appropriate for Bayesian inference methods. However, we have some very special limitations. Any determination (prediction) require multiple 20 million iteration Metropolis Monte Carlo simulations, each requiring 12-24 hours on a processor node. Furthermore, because we have a high-dimensionality parameter space, and no pre-existing training data to rely on, we had to rely on developing very specific heuristic techniques in order to find the optimal parameter set with an economy of time and resources.