Crystal structure prediction by data mining

Abstract

The ever increasing number of experimentally determined crystal structures allows for the use of data mining methods to address crystallographic questions. Here we study the application of data mining for predicting the arrangement of molecules in unit cells of unknown dimensions (crystal structure prediction) as well as in unit cells of predetermined dimensions (fractional coordinate prediction). In this work, data mining is used to derive an atom-pair potential, which is then compared to known force fields. It is shown that the potential is physically reasonable when the data are sufficient in quality and quantity. For validation the energy function is applied to the problems of crystal structure prediction and fractional coordinate prediction. In both cases a large number of structures was generated and the structures were ranked according to their energies. Structure prediction was considered successful if a structure similar to the experimentally observed one was ranked highest. For crystal structure prediction the energy function is tested on an independent set of crystal structures taken from the P1 and P 1 ̄ space groups. We show that approximately 76% of the 218 molecules tested in space group P1 are predicted correctly. For the more complex space group P 1 ̄ the success rate is 24%. If the powder diffraction can be indexed, the problem simplifies to fractional coordinate prediction. With the assumption of known cell parameters the structure has been resolved in 92% of the test cases for P1 and 29% for P 1 ̄ .