The FEAST eigenvalue algorithm and solver with new perspectives for first-principle electronic structure calculations

Professor Eric Polizzi
Abstract
FEAST is a new general purpose eigenvalue algorithm that takes its inspiration from the density-matrix representation and contour integration technique in quantum mechanics [Phys. Rev. B 79, 115112, (2009)], and it can be understood as a subspace iteration algorithm using approximate spectral projection [http://arxiv.org/abs/1302.0432 (2013)]. The algorithm combines simplicity and efficiency and offers many important capabilities for achieving high performance, robustness, accuracy, and multi-level parallelism on modern computing platforms. FEAST is also the name of a comprehensive numerical library package which currently (v2.1) focuses on solving the symmetric eigenvalue problems on both shared-memory architectures (i.e. FEAST-SMP version -- also integrated into Intel MKL since Feb 2013) and distributed architectures (i.e. FEAST-MPI version) including three levels of parallelism MPI-MPI-OpenMP. \\ \\ In this presentation, we aim at expanding the current capabilies of the FEAST eigenvalue algorithm and developing an unified numerical approach for solving linear, non-linear, symmetric and non-symmetric eigenvalue problems. The resulting algorithms retain many of the properties of the symmetric FEAST including the multi-level parallelism. It will also be outlined that the development strategy roadmap for FEAST is closely tied together with the needs to address the variety of eigenvalue problems arising in computational nanosciences. Consequently, FEAST will also be presented beyond the "black-box" solver as a fundamental modeling framework for electronic structure calculations. \\ \\ Three problems will be presented and discussed: (i) a highly efficient and robust FEAST-based alternative to traditional self-consistent field (SCF) procedure for solving the non-linear eigenvector problem (J. Chem. Phys. 138, p194101 (2013)]); (ii) a fundamental and practical solution of the exact muffin-tin problem for performing both accurate and scalable all-electron electronic structure calculations using FEAST on parallel architectures [Comp. Phys. Comm. 183, p2370 (2012)]; (iii) a FEAST-spectral-based time-domain propagation techniques for performing real-time TDDFT simulations. In order to illustrate the efficiency of the FEAST framework, numerical examples are provided for various molecules and carbon-based materials using our in-house all-electron real-space FEM implementation and both the DFT/Kohn-Sham/LDA and TDDFT/ALDA approaches.
  • Computational Mathematics and Applications Seminar