The central objectives of the Partnership in Computational Sciences (PICS) ``First Principles Simulation of Materials Properties'' grand challenge project are the development of new scalable parallel first principles methods for large systems, the application of them to problems that are beyond the limits of present methods, and the demonstration that massively parallel supercomputers can play an essential role in enabling their development. We have exceeded our initial targets of being able to accurately treat large collections of atoms (>1,000,000 for classical potentials, >10,000 for tight binding molecular dynamics, and >500 for ab-initio methods) in both static and dynamical applications. The long range goal is to provide theoretical input that can guide the discovery and development of advanced materials and to make a new generation of materials simulation methods available to industrial, government and academic researchers.
A key to designing materials for structural, magnetic, optic, electrical, and high temperature applications is the understanding and ability to control the synthesis and processing at the atomic level. As many of the crucial macroscopic properties of materials actually depend on defects, clusters, and microscopic structures involving hundreds to thousands of atoms, it is only with the availability of modern high performance computing, provided by the HPCCP, that first principles modeling of these important structures and related materials properties can now be undertaken. To this end we have developed, tested and are now applying a hierarchy of increasingly accurate and computationally intensive techniques:
The melting and pressure/temperature phase diagram of carbon is an
example where conditions are too extreme for laboratory experiments,
but where accurate molecular dynamics simulations are leading to new
insights for understanding natural and artificial diamond synthesis.
Figure 10 shows a snapshot of 512 carbon atoms in a
diamond lattice in the process of melting (T > 4000K). In the picture,
the red atoms indicate four-fold bonded (diamond-like) atoms, the blue
atoms indicate three-fold (graphite-like) bonded atoms, and there are also
a number of two-fold and five-fold coordinated atoms. The large number
of three-fold atoms is an indication that the liquid phase is less
dense than the four-fold diamond phase. This is in contrast to
silicon, in which the liquid phase has a higher average coordination
than in the diamond structure. By running such simulations for the
coexistence conditions of the solid and liquid phases the melting
temperature of diamond (as a function of pressure) are determined.
Magnetic alloys are at the heart of a wide range of technological
applications from the oldest of structural materials to the next
generation of data storage and retrieval devices. However, a detailed
microscopic understanding of alloy magnetism is lacking, hindering
further development of these technologies. Using a new ab initio
method we have been able to study, for the first time, the nature of
the magnetic state in disordered alloys. In figure 11 Ni
(large blue spheres) and Cu (small red spheres) atoms occupy the
lattice sites of a 256-atom/unit cell model of a Ni rich disordered
NiCu alloy. The local Ni-site magnetic moment is distributed
inhomogeneously, varying from a minimum of approximately 0.29 Bohr
magnetons (short blue arrows) to a maximum of approximately 0.6 Bohr
magnetons (long red arrows). Interestingly, the magnetic moment on a
Ni-site correlates with the total magnetic moment on the nearest
neighbor shell of atoms surrounding it: large red arrows tend to be
surrounded by other reddish arrows, while small blue arrows are
surrounded by either Cu sites having no moment or other blue
arrows. The results of these simulations are being used to
re-interpret results of neutron scattering measurements of magnetic
correlations in these alloys and to provide new insights into the
properties of magnetic alloys.
What is needed for Atomistic Simulations of Large-Scale Defects:
This effort has been utilizing parallel computers for a number of years, and will soon be applying Teraflop computational speeds (and Gigabytes of memory) to outstanding problems in materials science. The properties of real materials are often dominated by defects. The strength of most metals, for example, is determined by the size and properties of the boundaries between differently oriented crystallites (``grains''). The Ames Laboratory Condensed Matter Theory Group is currently studying the structure and energetics of grain boundaries in diamond (for applications to thin-film growth), silicon (with applications to semiconductor properties), and in FCC metals (for understanding the strength and ductility of metals and alloys). The determination of these structures is very difficult, due to the many possible atomic arrangements near the grain boundaries. By combining the strengths of traditional techniques (such as molecular dynamic simulations, implemented on parallel architectures) with novel ``genetic algorithms'' that help systematically search different possible structures, we are able to efficiently explore and locate the lowest energy structures. The genetic algorithm provides a natural bridge to massively parallel machines, with multiple structures being simultaneously optimized. This requires the capability to perform 10-100 simulations concurrently, with each simulation requiring 10,000 atoms or more.
This project was accepted for part of the High Performance Computing Challenge at Supercomputing '95. Unlike many algorithms, our approach is not limited to a single computer, or to a single set of nodes on a parallel computer. Different initial structures may be optimized on different machines, with the different machines sending back optimized atomic structures. In this way, we can use many of the largest computers in the U.S., connected via ultra-fast communication lines, to combine all of the computational power for one project. We believe that our novel approach to distributing the problem will allow us to achieve Teraflop computing speeds on a real materials problem.
