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

Rongfeng Guan, Ang Liu, Yang SongResearch Square 2025 · DOI

TL;DR

Researchers have developed a symmetry-aware deep learning generative framework that successfully predicted the elusive high-pressure $\gamma$-phase of sodium amide ($NaNH_2$). By combining diffusion models with explicit crystallographic constraints and energy guidance, the team identified a complex 64-atom structure ($P2_1/c$) that had previously stumped conventional optimization-based search methods.

What Problem Does This Paper Address?

Predicting how crystals rearrange their atoms under extreme pressure—known as Crystal Structure Prediction (CSP)—is a "grand challenge" in materials science. For simple materials, conventional algorithms like USPEX (evolutionary) or AIRSS (random search) work well. However, they often fail when dealing with:

  1. Large Unit Cells: As the number of atoms increases, the search space grows exponentially.
  2. Complex Interactions: Ionic materials like $NaNH_2$ (a promising candidate for hydrogen storage) have long-range Coulombic forces and directional hydrogen bonding that create a "messy" potential energy landscape filled with local minima.

Before this study, the high-pressure $\gamma$-phase of $NaNH_2$ remained unresolved because conventional methods couldn't find a structure that matched experimental X-ray diffraction (XRD) data.

Methods

The authors propose a conditional deep learning generative framework based on Denoising Diffusion Probabilistic Models (DDPM). Key innovations include:

  • Direct-space Asymmetric Unit (DAU) Encoding: Instead of trying to predict the position of every single atom in a large cell, the model only predicts the "minimal set" of atoms (the asymmetric unit).
  • Symmetry-Consistent Reconstruction: The full crystal is reconstructed from the DAU using Wyckoff operations. This ensures the generated structures strictly obey the laws of crystallography.
  • Energy-Guided Diffusion: During the "denoising" process (where the model turns random noise into a structured crystal), a secondary neural network (an MLP) calculates the gradient of the predicted energy. This "steers" the generation toward thermodynamically stable, low-energy configurations.
  • Experimental Conditioning: The model uses lattice parameters derived directly from experimental XRD indexing as inputs, narrowing the search to physically relevant spaces.

Datasets

Because no public dataset existed for the $NaNH_2$ system, the authors built a custom one:

  • Training Data: 1,000 candidate $NaNH_2$ structures generated via AIRSS.
  • Labels: Each structure was labeled with total energy/enthalpy calculated using high-fidelity Density Functional Theory (DFT) via the CASTEP code.
  • Validation: The model's predictions were evaluated against new synchrotron X-ray diffraction patterns collected at the Advanced Photon Source.

Key Insights and Ideas

  • Symmetry as a Shortcut: By embedding symmetry (via DAUs and Wyckoff operations) directly into the model architecture, the researchers reduced the "dimensionality" of the problem. The model doesn't have to "learn" symmetry; it is forced to respect it.
  • The Failure of "Brute Force": The paper highlights that even the best conventional tools (USPEX/AIRSS) missed the $P2_1/c$ phase. This suggests that for complex ionic systems, data-driven "guidance" is more effective than purely stochastic or evolutionary searching.
  • Electronic Drivers of Stability: The stability of the newly found $\gamma$-phase is driven by a significant charge redistribution under pressure. As the lattice compresses, the ionic bonding between sodium and the $NH_2$ groups strengthens, and the Na-Na distances drop sharply, stabilizing the structure up to at least 14 GPa.

Key Results

  • Identification of the $\gamma$-phase: The model identified the high-pressure phase as a $P2_1/c$ structure with $Z = 16$ (64 atoms per unit cell).
  • Experimental Validation: Rietveld refinement showed an excellent match between the predicted structure and experimental XRD patterns at 10.06 GPa.
  • Phase Transition Timeline: The study confirmed two transitions: $\alpha \rightarrow \beta$ at ~1.7 GPa and $\beta \rightarrow \gamma$ at ~2.5 GPa.
  • Robustness: The $\gamma$-phase was found to be remarkably stable, maintaining its structure from 2.5 GPa up to the experimental limit of 14.0 GPa.

Limitations and Future Work

  • The "Ghost" $\beta$-phase: While the $\gamma$-phase was resolved, the $\beta$-phase remains difficult to pin down due to its very narrow stability range (1.7–2.3 GPa) and potential coexistence with the $\alpha$-phase.
  • Temperature Effects: The DFT simulations were conducted at 0 K, while experiments were at room temperature. This likely accounts for the discrepancy between the predicted transition pressure (~5.0 GPa) and the observed one (~2.5 GPa).
  • Generalization: While successful for $NaNH_2$, the authors suggest the next step is applying this framework to other complex ionic systems like $LiNH_2$ or $KNH_2$ to see if the generative approach holds across different chemical species.

Summary updated

Abstract

<title>Abstract</title> Understanding the high-pressure behavior of sodium amide (NaNH2) is essential for its applications in hydrogen storage and chemical synthesis. Conventional structure prediction methods have been shown to struggle in 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. In this study, we introduce a deep learning generative framework that enables conditional crystal structure prediction by incorporating lattice parameters and space group symmetry as explicit constraints, and applying energy-guided diffusion sampling during generation. The framework uses a direct-space asymmetric unit (DAU) to represent each structure, enabling symmetry-consistent encoding and efficient reconstruction of full crystals through Wyckoff operations. Applied to NaNH2, this method successfully predicts the high-pressure γ-phase as a P21/c structure (Z = 16, 64 atoms), which was previously missed by conventional approaches. This phase is confirmed by synchrotron X-ray diffraction and remains stable up to 14.0 GPa. Further analysis of charge density and atomic rearrangement reveals the mechanisms driving the phase transitions and explains the robust stability of the γ-phase under high pressure. This work demonstrates the potential of our deep learning generative framework to accelerate structure prediction in complex ionic materials and provides an effective framework for modeling crystals with large unit cells accurately.
BibTeX
@inproceedings{guan2025deep,
  author = {Rongfeng Guan and Ang Liu and Yang Song},
  title = {{Deep learning generative model for conditional crystal structure prediction of sodium amide}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-7103941/v1},
}
Figure