Asynchonous Swarm-Based Global Optimisation: Overcoming Premature Convergence to Produce Realistic Pt Nanoparticle Ensembles

Swarm intelligence-based algorithms have proven successful in exploring a potential energy surfaces (PESs) of chemical systems. One of the major limitations to recent implementations of these algorithms is that they have been implemented serially. To overcome this limitation we present our asynchronously parallel global optimisations (GO) artificial bee colony (ABC) methodology: pyGlobOpt. Furthermore, we demonstrate methodologies to properly tune and enhance pyGlobOpt for system specific character by developing assessment criteria that could be applied to any form of ensemble generation. Using this criteria we were able to demonstrate how to overcome premature convergence, an issue pervading the GO of some systems with the ABC algorithm, using a clustering-based methodology. We demonstrate that the inclusion of our clustering methodology and properly tuned pyGlobOpt parameters can lead to a 5-fold increase in the number of unique low-energy structures found as well as more than halving the average energetic distance from the global minimum. we apply our refined methods to Pt nanoparticles, an important catalytic material for the evolution and reduction of hydrogen. We produce ensembles of thermodynamically relevant structures at different degrees of hydrogenation and compare to carefully selected experimental metrics. We culminate in demonstrating the potential of our software to be leveraged by exascale high performance computers by performing linear-scaling DFT calculations within the pyGlobOpt framework.