Nature-inspired crystal structure predictor
Scientists from Russia have reported a way to improve crystal structure prediction algorithms, making the discovery of new compounds multiple times faster. The results of the study were published in Computer Physics Communications.
Given the ever-increasing need for new technologies, chemists seek higher-performance materials with better strength, weight, stability and other properties. The search for new materials is a challenging task, and if performed experimentally, takes a lot of time and money, as it often requires trying a huge number of compounds at different conditions. Computers can assist with this, but they require good algorithms.
In 2005, Artem R. Oganov, now professor of Skoltech and Moscow Institute of Physics and Technology (MIPT), developed the evolutionary crystal structure prediction algorithm USPEX, perhaps the most successful algorithm in the field, now used by several thousand scientists worldwide. USPEX only needs to know which atoms the crystal is made of. Then it generates a small number of random structures whose stability is assessed based on the energy of interaction between the atoms. Next, an evolutionary mechanism accounts for natural selection, crossover and mutations of the structures and their descendants, resulting in particularly stable compounds.
In their recent study, scientists from Skoltech, MIPT and Samara State Technical University, led by Artem R. Oganov, improved USPEX's first step, which generates initial structures. Showing that purely random generation is not very effective, the researchers turned to nature for inspiration and developed a random structure generator based on a database of the topological types of crystal structures, amalgamating evolutionary approaches developed by Oganov and topological approaches developed by Professor Vladislav Blatov from Samara. Knowing that nearly all of the 200,000 inorganic crystal structures known to date belong to 3,000 topological types, one can very quickly generate an array of structures similar to the sought-for structure. The tests showed that thanks to the new generator, the evolutionary search copes with the prediction tasks 3 times faster compared to its previous version.
"The 3,000 topological types are the result of abstraction applied to real structures. Going the other way round, you can generate nearly all the known structures and an infinite number of unknown but reasonable structures from these 3,000 types. This is an excellent starting point for an evolutionary mechanism. Right from the start you most likely sample an area close to the optimal solution. You either get the optimal solution right in the beginning, or get somewhere near it and then get it by evolutionary improvement," explains Pavel Bushlanov, the first author of the study and a researcher at Oganov's laboratory at Skoltech.