Electrical Engineering Seminar
Nearby Skeleton Constrained Accelerated Recomputing, NSCAR, and the Bidirectional Adaptive Refinement Scheme, BARS: constrained Optimization meets Machine Learning to challenge the Von Neumann Computing Paradigm
ABSTRACT We discuss new algorithms for recomputing or computing families of cases with learning and adaptive algorithms. A variation approach is used and old solutions are used to constrain the new dynamics via Lagrange multipliers. Old system vs new system symmetries can be likewise incorporated. Applications in nonlinear kinetic plasma physics will be shown where Plasma waves and KEEN waves are modeled optimally using particle in cell methods. Ow put behind BARS.
Work supported by AFOSR and the DoE NNSA-FES Joint program on HEFLP.
BIO Bedros Afeyan earned his B.S. degree in Electrical Engineering from Concordia University, Montreal, Quebec, Canada in 1980. He then earned a Masters and a Ph.D. in Theoretical Plasma Physics from the University of Rochester working at LLE on modeling laser–plasma interaction problems. He moved to Lawrence Livermore National Laboratory in 1984 to finish his Ph.D. He subsequently worked at the University of Maryland as a post-doctoral researcher. He moved back to Livermore in 1993, working in inertial confinement fusion and magnetic fusion energy programs until 1996, followed by two years at the UC Davis-Livermore Applied Science Department when he also started his own company, Polymath Associates, which then became Polymath Research Inc. (PRI) in 1999. He has run that company for the last 20 years. PRI dedicates itself to research in laser–plasma interactions, photonics, harmonic multiresolution analysis, kinetic theory of plasmas and nonlinear self-organization in complex dynamical systems. He is the inventor of the Spike Trains of Uneven Duration and Delay (STUD) pulse program for LPI control, and the inventor or the Variational Approach to Parametric Instabilities in Inhomogeneous Plasmas, which was his Ph.D. research work. Afeyan also discovered kinetic electrostatic electron nonlinear (KEEN) waves, created the MDF (modified distribution function) paradigm to explain stimulated Raman scattering inflation and stimulated Brillouin scattering-filamentation entanglement, and coinvented image analysis tools such as MODEM (morphological diversity extraction method), a high road to denoising and feature extraction used in Z pinch, inertial confinement fusion, and target-fabrication modalities. He is also the inventor of recent algorithms that introduce machine learning into high-performance scientific computing, including nearby skeleton constrained accelerated recomputing (NSCAR), bidirectional adaptive refinement scheme (BARS), self-healing atomistic dynamics (SHAD), and adaptive particle orbit sampling technique for Lagrangian evolution (APOSTLE). These are methods that may largely address the ills of kinetic equation modeling methods that have hampered progress in the past.
Contact: Liliana Chavarria at 626-395-4715 email@example.com