Phys. Rev. Lett. (in the press); preprint at http://arxiv.org/abs/1109.2618 (2011)

Using Schrödinger's equation, the atomic and electronic structure of molecules can be calculated from first principles — that is, without the need to make assumptions that might bias or restrict the numerical accuracy of the calculation. Or, rather, in principle it can. In practice, the computational effort needed to solve Schrödinger's equation increases rapidly with the number of atoms involved, which constrains the size and complexity of the molecules that can be practically modelled. And so, assumptions must be made, limiting the accuracy with which the properties of all but the simplest of molecules can be calculated.

Matthias Rupp and colleagues have found what seems to be a short cut. Rather than calculating a given property of an unknown molecule from scratch, they have developed an algorithm that predicts its value by learning from a database of molecules whose properties are already known. To test their approach, the authors use it to predict the atomization energy of unknown molecules from a set of 7,165 known molecules. Their predictions achieve an order of magnitude greater precision than other approximate methods, in a billionth of the time required for a full quantum calculation.