Deep learning generative model for conditional crystal structure prediction of sodium amide

Understanding the high-pressure behavior of sodium amide (NaNH₂) is essential for its applications in hydrogen storage and chemical synthesis. Conventional structure prediction methods often struggle to accurately capture its pressure-induced phase transitions due to the complexity of its potential energy surface, which arises from strong ionic interactions, large unit cell size, and significant atomic rearrangements under compression. Here we introduce an experimental-informed deep learning generative framework for conditional crystal structure determination in NaNH₂ under crystallographic constraints. The framework conditions on experimentally indexed lattice parameters and candidate space-group symmetry represents each structure using a direct-space asymmetric unit (DAU), and applies energy-guided diffusion sampling to generate low-enthalpy candidates. Applied to NaNH₂, the workflow successfully identifies the high-pressure γ-phase as a P2₁/c structure (Z = 16, 64 atoms), which is validated by synchrotron X-ray diffraction and remains stable up to 14.0 GPa. Charge-density analysis and atomic rearrangement under compression elucidate the mechanisms driving the phase transitions and rationalize the stability of the γ-phase under high pressure. Overall, this work demonstrates a general strategy for resolving experimentally observed but structurally unsolved phases in complex ionic materials when lattice metrics and symmetry constraints are available.