GraPhAI: Neural Networks for Solving Centrosymmetric Crystal Structures

Crystal structure determination from low-resolution diffraction data is challenging due to the lack of a general-purpose method for solving the crystallographic phase problem. We have developed a graph neural network, GraPhAI, and propose a new and efficient diffraction data representation in a graph form for deep learning. The trained GraPhAI models are intended for ab initio phasing of down to 2 Å resolution data of typical unit-cell volume centrosymmetric crystal structures. The success rate is above 80% for structures containing an atom with Z ≥ 19 (e.g., metal-organic frameworks, coordination compounds, or inorganic structures). For purely organic crystal structures, the success so far is limited.