| Summary: | Molecular dynamics simulation method is widely used to calculate and understand a wide range
of properties of materials. A lot of research efforts have been focused on simulation techniques
but relatively fewer works are done on methods for analyzing the simulation results. Large-scale
simulations usually generate massive amounts of data, which make manual analysis infeasible,
particularly when it is necessary to look into the details of the simulation results. In this dissertation,
we propose a system that uses computational method to automatically perform analysis of
simulation data, which represent atomic position-time series. The system identifies, in an automated
fashion, the micro-level events (such as the bond formation/breaking) that are connected to
large movements of the atoms, which is considered to be relevant to the diffusion property of the
material. The challenge is how to discover such interesting atomic activities which are the key to
understanding macro-level (bulk) properties of material. Furthermore, simply mining the structure
graph of a material (the graph where the constituent atoms form nodes and the bonds between the
atoms form edges) offers little help in this scenario. It is the patterns among the atomic dynamics
that may be good candidate for underlying mechanisms. We propose an event-graph model to
model the atomic dynamics and propose a graph mining algorithm to discover popular subgraphs
in the event graph. We also analyze such patterns in primitive ring mining process and calculate
the distributions of primitive rings during large and normal movement of atoms. Because the event
graph is a directed acyclic graph, our mining algorithm uses a new graph encoding scheme that
is based on topological- sorting. This encoding scheme also ensures that our algorithm enumerates
candidate subgraphs without any duplication. Our experiments using simulation data of silica
liquid show the effectiveness of the proposed mining system.
|
|---|
