Evolutionary atomistic methods for the simulation of surface processing with applications in micro engineering
CIC nanoGUNE Seminars
- Speaker
-
M. A. Gosálvez, DIPC and UPV, Donostia, Spain
- When
-
2011/11/14
12:00 - Place
- nanoGUNE seminar room, Tolosa Hiribidea 76, Donostia - San Sebastian
- Add to calendar
- iCal
Atomistic methods, such as Kinetic Monte Carlo (KMC) and Continuous Cellular
Automata (CCA) are often used to describe the time evolution of complex
surface processes. Traditionally, the reaction rates employed by these methods
are obtained via a bottom-up procedure, involving _ab initio_ calculations
of activation energies for the various competing atomistic processes as well
as physical insight about the mesoscopic and macroscopic behavior of the
system. As an alternative, a top-down approach is presented in this talk,
whereby an Evolutionary Algorithm (EA) is used to determine optimal values of
the reaction rates by directly comparing the macroscopic behavior of the
simulated system with that from the experiments. The deterministic nature of
the CCA method allows to illustrate the key aspects of the proposed procedure.
This includes the particular features of the evolutionary search and the
inherent need for extremely efficient simulations. As an example application,
anisotropic wet etching in aqueous solutions offers both the complex atomistic
nature and the relevant front propagation features, as well as a wide-spread
use in the microfabrication of complex three-dimensional structures. Based on
new, extensive experiments, the evolutionary CCA approach is validated by (i)
describing the correct macroscopic etch rate distribution for different
materials (silicon and quartz) in numerous etchants at different
concentrations and temperatures, and (ii) performing fast, accurate, full 3D
simulations of microengineering structures. This opens new opportunities for
knowledge transfer to companies around the world. Similar efforts for the
ongoing evolutionary KMC implementation are also described.
If the time allows for it, major aspects about the computational efficiency
will be described. A novel, fast, octree-based, parallel implementation of the
CCA method will be presented based on performing the calculations in
increasingly affordable, commodity Graphics Processing Units (GPUs), where we
achieve simulation speeds that are two orders of magnitude faster than the
equivalent, sequential implementation on a Central Processing Unit (CPU). For
purely sequential architectures, a novel, fast implementation of the method
will also be presented, based on keeping track of the front arrival time (or
Predicted Reaction Time, PRT) and using a self-balanced binary search tree
(SB-BST) for efficient access to the stored PRT values. This allows to draw
relations between the CCA and the Level Set Method.