Course Number: CSCI 699
Class Number: 29950D
Instructor: Aiichiro Nakano; office: VHE 610; phone: (213) 821-2657; email: email@example.com
Lecture: 5:00-6:50pm M W, GFS 223
Office Hour: 5:00-5:50pm F, VHE 610
HPC Office Hour: 2:30-5:00 pm T, LVL 3M
Prerequisites: (1) CSCI596 or basic experience in parallel computing; and (2) PHYS 516 or basic knowledge of numerical methods in computational sciences.
Textbooks: None; all course materials are provided online on the course Web page. See also:
Electronic structure, R. M. Martin (Cambridge Univ. Press, '04),
Ab initio molecular dynamics, D. Marx & J. Hutter (Cambridge Univ. Press, '09),
Time-dependent density-functional theory, C. A. Ullrich (Oxford Univ. Press, '12),
Effective computation in physics, A. Scopatz & K. D. Huff (O'Reilly, '15).
Computer simulation of quantum-mechanical dynamics has become an essential enabling technology for physical, chemical and biological sciences and engineering. Quantum-dynamics simulations on extreme-scale parallel supercomputers would provide unprecedented predictive power, but pose enormous challenges as well. This course surveys and projects algorithmic and computing technologies that will make quantum-dynamics simulations metascalable, i.e., "design once, continue to scale on future computer architectures".
The course first covers how the exponential time complexity for solving the quantum N-body problem is reduced to (1) O(N3) within the density functional theory (DFT), for which Walter Kohn received a Nobel chemistry prize in 1998, and (2) O(N) based on physical data-locality principles (e.g., Kohn's quantum nearsightedness principle). The course then introduces key abstractions (e.g., pseudopotentials and exchange-correlation functionals) and representation issues (e.g., plane-wave basis vs. real-space multigrids), which are necessary for efficient implementation of quantum molecular dynamics (QMD) simulations. This is followed by the design of QMD simulation algorithms on massively parallel supercomputers using message passing and multithreading, including our metascalable divide-conquer-recombine (DCR) algorithmic framework, as well as performance optimization on modern many-core processors and accelerators through memory hierarchies and vectorization. Advanced topics to be covered include (1) DCR approaches to excitation dynamics, (2) intersection of machine learning and quantum N-body problem, and (3) merger of quantum Monte Carlo (QMC) and QMD methods. The course ends with best software practices for co-developing extreme-scale QMD software for million-way parallelism.
Students will learn fundamental knowledge and gain hands-on experience in order to: (1) reduce the intractable quantum many-body problem to lower-complexity problems, while retaining the essential physics; (2) design scalable parallel algorithms for linear-scaling quantum-dynamics simulations; (3) develop metascalable quantum-dynamics software on current and future computer architectures.
Nonadiabatic quantum molecular dynamics simulation to study photoexcitation dynamics in MoSe2 bilayer [M.-F. Lin et al., Nature Commun. 8, 1745 (2017)].