US scientists have developed a way to solve crystal structures that combines powerful computational methods with data from experiments or databases – but that does not require much human input. Previous computational methods to predict structures rarely use experimental data, take a long time and are limited to compounds with small unit cells. They also give structures that generally have lower symmetry than those which have been experimentally determined, suggesting that the answers may not be quite right.
‘One of the dirty little secrets that people don’t generally talk about is that, with a lot of these methods, you always find structures with no symmetry. And you know that can’t really be right because not many structures in nature [have no symmetry],’ says Chris Wolverton of Northwestern University in Illinois, who led the work.
Wolverton and his colleague Bryce Meredig questioned why traditional crystal structure prediction methods did not take account of this fact. ‘Given that we know most crystals have symmetry, often quite high symmetry, why are we ignoring that?’ asked Wolverton. This led them to use information that can be easily determined from either powder diffraction experiments, or obtained from databases. ‘It started out as a very practical approach. We should just take the information we have and use it,’ he explains.
It’s really shaking up crystal structure prediction. It seems to be a lot more powerful than traditional methods
Using these known parameters means the computational part of their approach can be constrained to a smaller ‘search space’, greatly speeding things up. ‘It means these crystal structure prediction tools are much more efficient when you’re not searching over such a gigantic space,’ says Wolverton. This is important because density functional theory calculations can take a long time. Solving structures from experimental powder diffraction data is also difficult and time-consuming, often relying a lot on people’s chemical intuition and experience in the area.
As a proof of principle, Meredig and Wolverton applied their technique to four different materials, each of which has structures that are still debated. One of these, magnesium imide (MgNH), had been studied in the 1960s using x-ray diffraction and more recently using neutron diffraction, giving a total of four possible space groups. ‘This method can address these kind of controversies – if you want to call them that,’ says Wolverton. ‘We can run this method in all four space groups and find out which one both has the lowest energy and fits the diffraction patterns the best.’
In this case, the structure from the recent neutron diffraction experiment gave the best answer – and it turns out this structure also matched the older x-ray diffraction experiments better than any of the three space groups proposed at the time, according to Wolverton.
The right symmetry
As well as being quicker than traditional structure prediction methods, Wolverton and Meredig’s approach is also more likely to find the ‘right’ answer. ‘This method will find the lowest energy state seven out of 10 times, whereas often what we find in the tools with no symmetry is that we may run the code 50 times and we’re lucky if we find the lowest energy structure once,’ says Wolverton.
Maryjane Tremayne, of the University of Birmingham, UK, works on both evolutionary algorithms and powder diffraction and is impressed by Meredig and Wolverton’s method. ‘It’s really shaking up crystal structure prediction. It seems to be a lot more powerful [than traditional methods],’ she says. ‘Their application of the genetic algorithm is impressive because it has located lower energy minima than the conventional genetic algorithm.’
But Tremayne also thinks the paper will provoke debate: ‘It could divide the field; I feel it’s a controversial paper. What it implies is that you could have a computational technique that overrides the need for human chemical intuition.’ Tremayne believes that there is still a place for experienced chemists. ‘There are indeed a very many incomplete and incorrect refinements, but those are possibly incomplete or incorrect because the data is telling you that the model you are trying to impose on it is incorrect. You need that human intervention to interpret what is going wrong.’
Of course, debate can be what drives science forward, as Tremayne is happy to suggest: ‘It’s very interesting, very topical and exactly the kind of paper that needs to be out there to get these discussions going.’