Hardware-Efficient Parallelized Optimization with COMSOL Multiphysics® and MATLAB®
Today, new processors provide increasing number of cores at rather constant clocking frequency. In sequential optimization algorithms, the forward model simulation is typically accelerated by multiple cores (shared-memory parallelization, SMP), which provides only limited speed-up and hardware efficiency. However, the Comsol Multiphysics® license includes parametric design capability allowing the application of population-based approaches with simultaneous simulation of multiple models with single cores. In this work, a simple optimizer based on latin hypercube sampling is presented to improve positioning of parts in an industrial-scale graphitization furnace. In a comparative study to sequential Nelder Mead Simplex (fminsearch), considerable performance advantages are demonstrated.
Téléchargement
- frommelt_presentation.pdf - 0.3MB
- frommelt_poster.pdf - 0.32MB
- frommelt_paper.pdf - 0.3MB
- frommelt_abstract.pdf - 0.33MB