1 Introduction

For inorganic crystals, in many cases it is already now possible to predict the stable structure at arbitrary external pressure. Towards the ambition of designing novel materials prior to their synthesis in the laboratory, reliable and efficient prediction of the structure of more complex (in particular, molecular) crystals becomes imperative.

Molecular crystals are extremely interesting because of their applications as pharmaceuticals, pigments, explosives, and metal-organic frameworks 69; 70. The periodically conducted blind tests of organic crystal structure prediction, organized by Cambridge Crystallographic Data Centre (CCDC) have been the focal point for this community and they reflect steady progress in the field 71; 72; 73; 74; 75. The tests show that it is now possible to predict the packing of a small number of rigid molecules, provided there are cheap force fields accurately describing the intermolecular interactions. In these cases, efficiency of search for the global minimum on the energy landscape is not crucial. However, if one has to use expensive ab initio total energy calculations or study systems with a large number of degrees of freedom (many molecules, especially if they have conformational flexibility), the number of possible structures becomes astronomically large and efficient search techniques become critically important.

Moreover, the nature of weak chemical interactions makes it common that a molecules have a wide variety of ways of packing with lattice energies within a few kJ/molecule of the most stable structure. Thus prediction of such large structures is certainly a challenge, especially if the number of trial structures has to be kept low to enable practical ab initio structure predictions. Recent pioneering works 76; 77; 78, in particular, using metadynamics 77 offer examples of this.

Compared to other methods, evolutionary algorithms have special advantages. Exploring the energy surface, such algorithms arrive at the global minimum by a series of intelligently designed moves, involving self-learning and self-improvement of the population of crystal structures 23. Our USPEX (Universal Structure Predictor: Evolutionary Xtallography) code 4; 26; 27; 28; 79, proved to be extremely efficient and reliable for atomic crystals, and here we present an extension of this algorithm to complex crystals made of well-defined units. In the following sections, we will mainly discuss molecular crystals. Crystals containing complex ions and clusters can be equally well studied using the methodology developed here.