Data Mining in Chemistry exemplified for Crystal Structure Prediction

Nowadays two methods are available for simulations in chemistry, ab-initio calculations based on the quantum physics of the last century and methods based on knowledge retrieval from data bases. The importance of expert knowledge is known since the early stage of chemistry and crystallography. Even if computational chemistry based on scientific laws did make tremendous progress, since the seventies the development of informational technology allows for the automated knowledge retrieval from databases. Together with the rapid increase of collected data, data mining as a main tool for knowledge retrieval is an attractive alternative, if the problem is hardly accessible by other methods. The main idea of data mining on experimental structures is that nature is always right and any simulation has small errors. This approach allows us to parameterize any model as good as possible imposing that the model has to separate the experimental and the simulated structures in two classes. (Sometimes nature might be right, but experimentalist not, then we observe outliers and the approach does not work perfect) In this talk we will demonstrate the power of data mining on the example of crystal structure prediction. The successful prediction of crystal structures requires very accurate models. Already tiny errors in the free energy cause a wrong order in the ranking of different packing’s of molecules in the crystals structures and predict wrongly the region of stability for the different structures. We will show that training of the model on the different crystal structures allow us to predict correctly the structures as function of temperature and pressure. A very interesting aspect of the result is that the scoring function is identical to the free energy, which is hardly accessible to other methods. Since our model is based on atom pair energies, the crystal structure can be analyzed quantitative in terms of free atom pair energies in dependence of temperature and pressure. This quantification of the interactions insight the crystals might help in future for further understanding of crystal structures and might contribute to crystal engineering.

@InProceedings{Hof12,
  author       = {Hofmann, D.},
  title        = {Data Mining in Chemistry exemplified for Crystal Structure Prediction},
  booktitle    = {Collana seminari interni 2012},
  number       = {20120516},
  month        = {may},
  year         = {2012},
  url          = {https://publications.crs4.it/pubdocs/2012/Hof12},
}