Skip to content

Workflow Overview

The amorphous-to-crystalline (a2c) workflow predicts crystalline structures that emerge from amorphous precursors.

Workflow Diagram

graph LR
    A[Random<br/>Structure] --> B[Melt-Quench<br/>MD]
    B --> C[Subcell<br/>Extraction]
    C --> D[Structure<br/>Optimization]
    D --> E[Analysis]

Five Stages

1. Initial Structure Generation

Generate a random atomic configuration with desired composition.

Function: random_packed_structure()

Purpose: Create physically reasonable starting point with no severe atomic overlaps

Key parameters: - composition: Chemical formula (e.g., "Si64") - cell: 3×3 unit cell matrix (Å) - auto_diameter: Auto-calculate atomic radii - seed: Random seed for reproducibility

2. Melt-Quench Molecular Dynamics

Create amorphous structure through thermal cycling.

Function: melt_quench_md()

Three phases:

  1. High-T Equilibration: Melt the structure
  2. Linear Cooling: Quench to trap amorphous state
  3. Low-T Equilibration: Relax at target temperature

Key parameters: - T_high: Melting temperature (typically 1.5-2x melting point) - T_low: Target temperature (e.g., 300K) - equi_steps, cool_steps, final_steps: Duration of each phase - time_step: Integration timestep (fs) - friction: Langevin thermostat friction

Temperature Selection

Choose T_high above the material's melting point to ensure complete melting

3. Subcell Extraction

Divide amorphous structure into overlapping regions.

Function: extract_crystallizable_subcells()

Method: Regular grid in fractional coordinates

Key parameters: - d_frac: Grid spacing (0.2 = 20% of cell) - n_min, n_max: Atom count range (e.g., 2-8) - cubic_only: Restrict to cubic subcells - allowed_atom_counts: Specific atom counts

Performance

Smaller d_frac creates more subcells. Use 0.2-0.33 for good balance.

4. Structure Optimization

Relax each subcell to find stable crystal structures.

Function: relax_unit_cell()

Method: FIRE algorithm with FrechetCellFilter

Optimizes: - Atomic positions - Cell shape and volume

Key parameters: - max_iter: Maximum optimization steps (e.g., 200) - fmax: Force convergence criterion (eV/Å)

5. Validation and Analysis

Identify and rank stable structures.

Validation: valid_subcell() - Energy decreased during optimization - Formation energy physically reasonable - No atomic fusion (atoms too close) - Forces converged

Analysis: Using pymatgen - Determine space group with SpacegroupAnalyzer - Compare to reference structures with StructureMatcher - Rank by energy per atom

Complete Example

Full working implementation: example/Si64.py

Performance

Stage Bottleneck Solution
Random packing Overlap removal Increase diameter
Melt-quench MD steps Reduce steps for testing
Subcell extraction Too many subcells Increase d_frac
Optimization Force calculations Use GPU, parallelize

See Also