On artificial crystal structure generation for solving the phase problem with deep learning

We discuss and present approaches for generating artificial crystal structures for training neural networks to solve the phase problem. Structure generation is considered as a two-step process involving sampling unit-cell parameters and filling the unit cell with atoms. The former step includes generating lattice basis vectors from randomly sampled unit-cell volume. Apart from randomly placing atoms, we use database data to guide fast and scalable generation of molecule-like fragments. The recently developed neural network PhAI is then used as a benchmark and retrained with various sets of training data to assess how the corresponding models perform on experimental crystal structure data. We found a significant improvement in PhAI retrained on a new kind of artificial data to generalize the phase problem solution for larger unit-cell structures.