During the last years, plasmonics has received an increasing attention due to its applications to imaging, microscopy, spectroscopy, optical interconnect and more. The interaction between nanometric metal structures and electromagnetic radiation at optical frequencies can be modeled by Maxwell’s equations without resort to quantum physics. However, the behavior of the permittivity function at these frequencies gives raise to a complex phenomenon: plasmonic resonances.
Plasmonic oscillations a...
During the last years, plasmonics has received an increasing attention due to its applications to imaging, microscopy, spectroscopy, optical interconnect and more. The interaction between nanometric metal structures and electromagnetic radiation at optical frequencies can be modeled by Maxwell’s equations without resort to quantum physics. However, the behavior of the permittivity function at these frequencies gives raise to a complex phenomenon: plasmonic resonances.
Plasmonic oscillations are characterized by their very high localization. The wavelength of the plasmonic oscillation is much shorter than the wavelength of the incident field that originates it. Besides, the plasmonic field rapidly decays away from the metal structure. All of this enables the
possibility of manipulating light at a subwavelength scale.
Unfortunately, simulating realistic plasmonic systems involves solving very expensive computational problems. Even with the use of supercomputers it is not an easy task, since parallelizing the routines solving the system presents complications. One of the main challenges in this case, and in parallel computing in general, is managing communication overhead. Often, the cost of communicating information between different nodes prevents the computational system to scale, which means that that
using a very large number of cores will not grant a fast execution of the program.
Here, we propose a novel approach based on Monte Carlo methods. It is possible to run a randomized simulation that obtains an approximate solution of the system. The expected value of this approximate solution corresponds to the actual value of the solution. Although this method converges slowly when run in a single processor, it is easy to parallelize, since different executions of the randomized algorithm can run independently without communication. Finally, the processors share their
information only once to obtain an average of all the approximate solutions. The final results approximates well enough the solution of the problem.
Share
