Abstract

Genuinely new knowledge and scientific insight can be obtained about matter by combining random numbers with reliable and efficient first principles methods. Diverse ensembles of initial structures can be generated, and structurally optimised. The resulting low energy structures are candidates for stable, and metastable, phases and/or defects that might be experimentally realised. This, of course, depends on a sufficiently broad and thorough sampling of configuration space.

Algorithms which attempt to learn from (computational) experience are necessarily sequential, and correlated. A purely random strategy, as employed by Ab Initio Random Structure Searching (AIRSS),[1,2] is entirely parallel, and a natural fit to the high throughput computation (HTC) paradigm. The absence of correlation between the independent random samples ensures that it is possible to estimate when a sufficiently dense sampling has been achieved (or at least, has not been achieved). Challenging cases can be tackled by designing the initial random structures so that they focus the search in regions of configuration space that are anticipated to yield success.

The design of these random “sensible” structures will be explored, along with some new directions which promise to accelerate random search,[3] and recent applications to materials.

[1] C. J. Pickard, and R. J. Needs, Phys. Rev. Lett., 97 (4), 045504 (2006) & Journal of Physics-Condensed Matter, 23(5), 053201 (2011) [2] Released under the GPL2 license: http://www.mtg.msm.cam.ac.uk/Codes/AIRSS [3] C. J. Pickard, “Hyperspatial optimization of structures”, Phys. Rev. B, 99, 054102 (2019)