Computational elastodynamics for large-scale time dependent problems: study cases from seismology to structured materials
Andrea Colombi
http://www.imperial.ac.uk/people/a.colombi
Imperial College - London Faculty of Natural Sciences, Department of Mathematics
Research Associate Marie Skłodowska-Curie Individual Fellow
November, 17th
h. 12.00
Room Trasporti, Viale Risorgimento 2, Bologna
Abstract: The study of wave phenomena in physics and engineering problems is undergoing drastic transformations after the broad diffusion and improved capabilities of numerical methods and large scale parallel computing. In elastodynamic this transition is of topical importance: From the small to the large scale, today’s dynamics problems, foresee the evaluation of the modal response or the time transient analysis of complex systems. As the field moves towards full 3D models in multiscale and heterogeneous media, the quest for analytical solutions is very difficult or simply not possible, making computational methods on obliged choice. Typical examples are found in the study of waves and vibrations in structured materials (composites and metamaterials), multiscattering problems in heterogeneous media (non-destructive evaluation and acoustic imaging), and in Earthquake Engineering where design techniques are moving from the isolated building approach to more complete analysis encompassing the coupling between the subsurface, the surrounding structures and the seismic properties of the site.
Today's best computational methods for elastodynamics rely on high-order polynomial approximations (spectrally accurate) and fully exploit the tremendous potential of parallel computing. However they are far from being a black-box tool for researchers and engineers. The discretization process (meshing), the choice of the space and time integration scheme, the parallelization methods (CPU vs. GPU), and ultimately the post-processing of terabytes of data demand a truly multidisciplinary approach to the elastodynamic problem.
The aim of this seminar is to provide an overview of the advantages and challenges of high performance numerical methods (serial or parallel) for elasticity using examples taken from numerical seismology and elastic vibrations in structured media.
In particular we will review the basic theory underpinning high order spectral element methods, the steps that lead to the parallelization for a simple time transient or frequency domain problem and the tedious meshing steps. Then we will review the basic paradigm of parallel computing and the advantages of an access to centralized or national supercomputing facilities. Finally some tips and tricks for an efficient management of simulation results and data parallel post-processing.