Por Pedro Borlido (Departamento de Física, Faculdade de Ciências e Tecnologia, Universidade de Coimbra, Portugal).

Thanks to the increase in available computing resources, ab initio crystal structure prediction has become almost routine. However, although the most used underlying methods (such as Density Functional Theory) are quite accurate, they are still too expensive considering that the search space is vast. Nowadays it is possible to overcome this difficulty to some extent thanks to various machine learning methods, such as universal neural potentials. Here we'll talk a bit about how we've combined these two types of methods to discover new materials