Exascale algorithm and software development for computational chemistry and other sciences

Abstract

High-performance computing (HPC) plays a vital role in addressing the computational problems posed by large-scale quantum chemistry simulations. Quantum chemical calculations provide fundamental insights into molecular interactions, facilitate the design of novel materials, and contribute to advancements in drug discovery. This thesis develops HPC-centric methods and algorithms that enhance the efficiency, scalability, and accuracy of these computations, enabling their application to large and complex molecular systems on modern supercomputing infrastructures. Traditional quantum chemistry methods, such as Hartree-Fock (HF), M\o ller-Plesset perturbation theory (MP2), and Coupled Cluster (CC), have excessively high computational costs that scale polynomially with the system size. To mitigate the computational cost of traditional quantum chemistry methods, fragmentation methods divide large molecular systems into smaller, computationally manageable units while preserving inter-fragment interactions. This work focuses on the Fragment Molecular Orbital (FMO) method and introduces a high-performance variant, the Coulomb-Perturbed Fragmentation (CPF) method, which replaces iterative electrostatic potential (ESP) updates with a single-step electrostatic embedding, significantly increasing computational efficiency while maintaining accuracy.

The Extreme-Scale Electronic Structure System (EXESS) serves as the computational foundation for implementing these methods. This framework integrates heterogeneous GPU-based architectures to accelerate quantum chemistry computations, employs a multi-layer dynamic load balancing scheme to distribute fragment calculations efficiently across computing nodes, and optimises memory access patterns to reduce data movement overhead. These advancements collectively enhance parallel efficiency, ensuring scalable performance across large-scale molecular simulations. To validate the effectiveness of these optimisations, this work introduces a benchmarking approach that evaluates scalability and computational efficiency across a range of molecular systems. The implemented algorithms demonstrate near-linear parallel scaling, achieving over 97\% efficiency in computations involving up to 146,592 atoms. These improvements significantly reduce the computational cost of large-scale electronic structure energy. While traditional quantum chemistry calculations rely on full double-precision arithmetic to ensure accuracy, this approach can be computationally expensive. To address this, this study investigates a mixed-precision arithmetic method, which dynamically adjusts floating-point precision based on a fixed threshold. This strategy enhances computational performance while maintaining numerical stability, offering a balance between efficiency and accuracy.

This study improves the scalability, efficiency, and feasibility of large-scale electronic structure simulations on exascale systems by combining HPC optimisation approaches with modern quantum chemistry methods. Not only do the suggested methods improve computational chemistry performance, but they also contribute to the development of scalable parallel algorithms, load-balancing strategies, and precision management techniques in heterogeneous HPC environments. These developments have significant implications for scientific computing, molecular modelling, and large-scale simulations in computational research, highlighting HPC's important role in expediting scientific discovery. Show more