Dr. Robert Q. Topper
Wednesday, April 18, 2018
11:00 to 11:50 am
C – 121
Stochastic Searching and Machine Learning Methods for Acceleration and Analysis of Atomistic Simulations
Abstract:
Monte Carlo and molecular dynamics simulation methods are widely used tools for atomistic simulations of systems ranging from individual molecules and nanoparticles through biomolecules in solution, surface adsorbates, bilipid layers and membrane channels, and bulk materials in the solid, liquid and superfluid states. This work requires the continued development of efficient algorithms for ergodic sampling and optimization of mega-dimensional surfaces, as well as methods which can analyze – even “learn” – the vast amount of data produced by simulations. We have been using the “magwalking” – sawtooth simulated annealing (MW-SSA) Monte Carlo method combined with density-functional theory, MP2, and CCSD(T) electronic structure theory calculations to predict the structures and thermodynamic properties of ammonium halide nanoparticles. These studies are relevant to atmospheric chemistry in polluted marine environments. I will also describe the use of machine learning methods such as artificial neural networks and Gaussian kernel ridge regressions to represent, store and analyze high-dimensional free energy surfaces obtained from “metadynamics” simulations, with applications to solid-solid phase transitions and conformational distributions of polypeptides in solution.