Thursday, November 02, 20172:00 PM - 3:00 PMCNLS Conference Room (TA-3, Bldg 1690)|
Analyzing Molecular Dynamics Trajectories using Strain Functionals and Machine Learning
Extracting relevant information from atomic-scale simulations and upscaling to meso-/engineering-scale models relies on characterization and classification of the local environment of atoms. I will demonstrate the use of a complete and symmetry-adapted approach, referred to as strain functional analysis, for characterizing atomistic geometries. This approach defines a Gaussian-weighted neighborhood around each atom and characterizes the local geometry in terms of nth order central-moments of the neighborhood. Rotationally invariant metrics, referred to as strain functional descriptors, are derived from these nth order moments using concepts from group theory. Descriptors from a 6th order moment expansion can distinguish between different crystal structures and identify defects such as dislocations, stacking faults, and twins at finite-temperature. These descriptors will be used in conjunction with principal component analysis (PCA) and clustering algorithms (Gaussian Mixture Model) to analyze molecular dynamics trajectories of high strain-rate compression of Cu, Ta, and Ti. I will also demonstrate that the strain functional analysis can be used as a basis to develop and analyze inter-atomic potentials.
Host: Chris Neale