Abstract

General consensus holds that the more empirical information needed by a structure prediction method, the less powerful it is, in the sense that it cannot predict a variety of structures from knowledge of chemical composition alone. However, for almost all global search techniques, incorporation of prior knowledge of some form improves performance by increasing speed, or where it can be evaluated, improving accuracy. Considering the computational complexity of the problem, I argue that to move towards being able to carry out routine predictions of the possible structures of a system, methods must be able to systematically take into account what prior information already exists for the system. By comparing different methods of energy landscape search methods in a common framework of general steps and considering (1) where and how different methods incorporate `prior knowledge’, and (2) how these inclusions affect predictive power, speed, and accuracy, we can look for explanations of existing strengths or means of potential improvement. This review focuses primarily on ab initio applications of simulated annealing, Monte Carlo basin-hopping, evolutionary algorithms, particle swarm optimisation, and random search.