| name | mat-amorphization |
| description | Generate amorphorized structures from crystalline starting points using a melt-quench MD protocol. |
| category | ["materials"] |
Amorphorization
Goal
To generate disordered, amorphous structures from crystalline inputs using molecular dynamics (MD). This is achieved through a "melt-quench" protocol, where the material is heated above its melting point and then rapidly cooled to "freeze" the liquid-like disorder.
Protocol: Melt-Quench
The standard Computational amorphization protocol involves:
- Supercell Setup: The system must be large enough to avoid spurious periodicity effects in the amorphous state. Generally, $>100$ atoms is recommended.
- Melting (Stage A): Heat the system to $T_{melt}$. $T_{melt}$ should be significantly higher than the experimental melting point (often 1000K higher) to ensure rapid loss of crystalline memory within MD timescales.
- Equilibration (Stage A/B): Maintain the liquid at $T_{melt}$ for several picoseconds to ensure structural randomized.
- Quenching (Stage B): Cool the system linearly to the target temperature (e.g., 300K).
- Cooling Rate: A critical parameter. Typical MD cooling rates are $1-10$ K/ps ($10^{12}-10^{13}$ K/s). Slower rates yield more stable, realistic amorphous structures but are computationally expensive.
- Annealing/Equilibration (Stage C): Relax the density and local structure at the target temperature.
- Quenched/Static Relaxation (Stage D): Perform a final geometry optimization (0K) to find the local energy minimum of the amorphous state.
Instructions
1. Preparation
- Supercell: Use the
prep_supercell.py helper script. By default, it generates an orthorhombic conventional supercell with approximately 100 atoms, ensuring a robust starting point for amorphization.
python .agents/skills/mat-amorphization/scripts/prep_supercell.py --input crystalline.cif --output supercell.cif
- Foundation Potential: Select a robust model like
MACE-MP-large or CHGNet using the mcp_mace_load_model (or similar) tool.
2. Execution (The Melt-Quench Cycle)
Amorphization is performed by calling the run_md tool in a sequence:
Stage 1: Melting
Heat the system to a high temperature (e.g., 3000K) to eliminate crystalline order.
- Tool:
mcp_mace_run_md
- Thermostat:
nvt_langevin (Robust for high-T dynamics).
- Parameters:
temperature=3000, steps=5000 (10 ps), ensemble="nvt_langevin", timestep=2.0.
Stage 2: Quenching
Cool the system rapidly to the target temperature (e.g., 300K).
- Tool:
mcp_mace_run_md
- Thermostat:
nvt_langevin (Supports specific set_temperature ramping).
- Monitor: Use
monitor_type="quenching" and monitor_params={"temperature_end": 300, "steps": 5000}.
- Parameters:
temperature=3000 (start), steps=5000 (10 ps), ensemble="nvt_langevin".
- Note: Ensure the input structure is the output of Stage 1.
Stage 3: Equilibration
Relax the structure at the target temperature to reach equilibrium distribution.
- Tool:
mcp_mace_run_md
- Thermostat:
nvt_bussi (Bussi-Donadio-Parrinello) - Provides correct canonical sampling.
- Parameters:
temperature=300, steps=2500 (5 ps), ensemble="nvt_bussi".
3. Analysis & Verification
Use the analyze_amorphous.py script to verify the results:
- RDF (Radial Distribution Function): Confirm the absence of long-range order.
- Coordination Number: Check local bonding environments.
Helper Scripts
prep_supercell.py: Expands a unit cell to a supercell.analyze_amorphous.py: Calculates RDF and coordination numbers from the final structure.- RDF (Radial Distribution Function): Crystalline structures show discrete, sharp peaks at long distances. Amorphous structures show a sharp first peak, a broader second peak, and then decay to 1.0 (no long-range order).
- Coordination Number: Check if the local coordination (e.g., 4 for Si) is maintained despite the global disorder.
Foundation Potential Selection
- ml-foundation-potentials
- MACE-MP-large or CHGNet are recommended for high-temperature MD as they are trained on diverse configurations.
Examples
See .agents/skills/mat-amorphization/examples/ for validated amorphous structures.
Author: Bowen Deng
Contact: GitHub @learningmatter-mit
