Accelerating Dynamic Correlation Algorithms for Quantum Chemistry

Abstract

Relativistic quantum mechanics provides a fundamental description of virtually all chemical phenomena, though in practice quantum chemistry is formulated through approximate solutions of the Dirac equation.Among systematic approaches, configuration interaction (CI) yields the exact solution in the complete-basis-set limit but suffers from factorial growth in computer memory requirements, restricting its application to the smallest systems. The recently developed small tensor product distributed active space (STP-DAS) CI framework significantly shrinks these resource requirements, yet large correlated calculations remain out of reach. In contrast, density functional theory (DFT) offers an approximate solution with much more favorable scaling. While not systematically improvable, DFT is widely used for large molecular systems where CI is not tractable. The landscape of scientific computing today is moving away from traditional processors in favor of graphics processing units (GPUs). GPUs offer the potential of tremendous speedup, but their unique computing paradigm requires careful algorithmic design to maximize their effectiveness. In this dissertation, we present two GPU-accelerated algorithms that accelerate critical rate-limiting steps in DFT and STP-DAS CI calculations, enabling faster correlated methods across length scales. We describe the design considerations of these algorithms, contextualize their role in chemical applications, and demonstrate their performance on large systems.