Predictive modeling and simulation of properties and multi-scale processes in materials science. Tasks for Exaflops-era supercomputers

The approach is developed which allows to find out the problems which need for their solution exaflops supercomputers. The approach is demonstrated at the examples of topical problems of material science, condense matter and dense plasma physics where atomistic modeling is necessary to apply. The correspondence is established for each problem between phenomena studied and computational cores number needed. Modeling parallel programs scalability is shown as well as perspective of the modeling methods predictive ability extension with the increase of computational cores number and / or use of special architecture (graphical processing units).
The following problems are considered: 1) surface modification at processing of metals by sub-picosecond laser pulses, 2) radiation-induced aging of nuclear reactors fuels, 3) phase transition kinetics in metastable liquids, 4) methane and hydrogen gas hydrates structures and computation of their properties, 5) polymers multiscale models, 6) dusty plasmas, 7)ion recombination in liquid and gaseous dielectric media at discharge break and relaxation, 8) electric double layer between graphite and electrolyte, influence of electron-hole electrode structure on capacity. Predictive modeling reliability is checked by comparisons with experiments.
The modeling methods hierarchy, which is necessary to describe properties of matter at different space and time scales, is considered in frames of the multiscale approach. Density functional theory (quantum molecular dynamics) is applied at the deepest nm/pm scale to model electron dynamics and to construct effective interaction potentials between particles. Classical molecular dynamics modeling is used to treat moving atoms systems up to micro-scale. Kinetic theory and continuum mechanics is used to proceed with micro-scale. Particular attention is paid to the exchange of information between different scales, i.e. to the unified description of systems from nano to micro levels. Parallelization efficiency comparison is performed for three classes of problems at fat tree and torus topologies (in Russian).