Master Thesis defense by Jonas Dornonville de la Cour

Title: A Massively Parallel Lockstep Pipeline For Full Isomerspace Optimisation

Abstract:
Fullerenes are carbon molecules with a hollow cage-like structure. Its surface is made up exclusively of pentagon and hexagon rings. The number of theoretically stable fullerenes is infinite, growing with $\mathcal{O}(N^9)$, only a small fraction of which have been synthesised. The ones that have, have shown great promise in a variety of uses, e.g. allergy and asthma medicine, cancer treatments solar cells, biosensors, printable electronics, and a host of other applications.
Therefore, it is of great interest to be able to compute the geometries and properties of these molecules.
However, full ab-initio calculations like density functional theory (DFT) is completely infeasible for more than single fullerene analysis. Thus, a more efficient approach is needed. Forcefield (FF) methods are far more computationally efficient, and have been shown to produce results that are in good agreement with DFT. Current state-of-the-art FF methods are able to compute optimal geometries for $C_{200}$ isomers in $100\mu s$. While this is certainly fast, it is not yet sufficient for full isomerspace exploration.
This thesis presents a fully lockstep parallel implementation of pipeline for isomerspace forcefield optimisation. Our lockstep parallel approach leverages thousands of compute cores to attain roughly 3 orders of magnitude (1000-1500x) faster performance than previous state-of-the-art FF implementation. Our implementation demonstrates essentially perfect scaling, consequently enabling further performance gains for larger isomerspaces, given sufficient compute resources.
The final pipeline allows us to exhaustively produce and optimise the entire $C_{200}$ isomerspace in a projected time of 6 hours, making what was previously completely impractical (247 days), enabling full isomerspace exploration, and setting the stage for molecular property analysis of billions of fullerenes.