Phase-Change Memory: Amorphization and Recrystallization of GST

MLPotential Materials Science Molecular Dynamics Phase-change memory GST Crystallization Amorphization

This tutorial will teach you how to:

• build a rocksalt-like GeSb₂Te₄ (GST124) crystal from scratch,

• make an amorphous GST region with a melt–quench molecular dynamics run,

• recrystallize the amorphous region at 600 K, and

• compare crystallization kinetics across candidate materials.

Production result: structural ordering at 515 K

Phase-change memory relies on rapid crystallization for a fast SET operation. This production example compares seeded recrystallization of GST124 (GeSb₂Te₄) and GST147 (GeSb₄Te₇) at 515 K. For each composition, 12 independently melt–quenched samples retain two fixed crystalline seed planes at the periodic boundary. The mobile ions and rectangular cell are relaxed before a 2 ps temperature ramp and a 400 ps growth stage using the QET-MatQ potential. Machine-learned interatomic potentials make it possible to investigate GST crystallization kinetics over atomistic time and length scales [4].

The figure compares two measures of how quickly the mobile region becomes more ordered. Each dot is one simulation; bars show the replica means, error bars show the sample standard deviations, and the vertical density curves show the spread of individual rates.

Fig. 56 Seeded structural-ordering rates at 515 K from 12 replicas per composition.

GST124 has the higher mean rate for both the near-90° angle count (\(3255 \pm 955\) versus \(2446 \pm 962\) angles per 1,000 mobile atoms per ns) and the \(q_4\)-classified crystalline-atom count (\(688 \pm 146\) versus \(539 \pm 182\) atoms per 1,000 mobile atoms per ns). The uncertainties are replica standard deviations, and the rate distributions overlap. These are structural-ordering rates, not measured growth-front velocities in m/s.

The GUI exercise below develops the GST124 construction, melt–quench, and seeded-growth steps used in this type of comparison.

Introduction

GST alloys can reversibly switch between an amorphous, high-resistance state and a crystalline, low-resistance state. A RESET-like thermal history melts and rapidly quenches the active region. A SET-like treatment heats the amorphous material below the melting point so that it crystallizes [1]. Atomistic simulation does not reproduce the complete electrothermal device, but it can isolate the structural kinetics that underpin these operations.

In this tutorial, the starting crystal has the metastable rocksalt-like GST124 structure. Te occupies one rocksalt sublattice. On the other sublattice, each conventional cell contains one Ge, two Sb, and one vacancy. The resulting seven-atom cell has exactly the Ge:Sb:Te ratio 1:2:4. Replicating it 4 x 4 x 6 gives 96 formula units:

Table 16 Tutorial system

Quantity	Value
Composition	Ge96Sb192Te384
Number of atoms	672
Starting lattice constant	6.10 Å
Supercell	4 x 4 x 6
Boundary condition	three-dimensional periodic
Fixed seed	two atomic planes across the z periodic boundary
Mobile atoms	all atoms outside the fixed seed

The ordered vacancy pattern is a convenient reproducible construction, not a claim that real metastable GST124 has perfect long-range vacancy order. The mobile part loses this artificial ordering during the melt. The two fixed planes provide the crystalline template from which the quenched region can regrow. Seeded growth is both reproducible and accessible on molecular-dynamics timescales [2] [3].

The simulation sequence is:

1. build and replicate crystalline GST124,

2. define fixed_seed and mobile regions,

3. heat, melt, and quench the mobile region at constant volume,

4. start a new 600 K trajectory from the quenched coordinates, and

5. inspect the growth front and quantify crystallization.

Part A: Build Crystalline GST124

Create a conventional rocksalt cell

1. Open AMSinput: SCM → New Input

2. Switch to a periodic engine, for example →

3. Open the crystal structures tool Builders → Crystal Structure and choose Cubic → NaCl

4. In the builder, set the two elements to Ge and Te

5. Set the cubic lattice constant to 6.10 Å

6. Click OK

7. Edit → Crystal → Convert To Conventional Cell

8. Choose View → Periodic → Periodic View Type → Repeat Unit Cell if the conventional cell is not clear

The conventional rocksalt cell contains four Ge sites and four Te sites. We will turn two Ge sites into Sb and delete one Ge site to make a vacancy.

Edit the cation sublattice

1. Select one Ge atom, then Select → Atoms of the Same → Type; confirm that exactly four Ge atoms are selected

2. Clear the selection and select any two of the four Ge atoms

3. Open Atoms → Change Atom Type and change their atom type to Sb

4. Select one of the two remaining Ge atoms and press Delete to create the vacancy

5. Confirm that the cell contains one Ge, two Sb, and four Te atoms

6. Save the input as GST124_unit_cell

Fig. 57 Building the rocksalt-like GST structure.

Build the simulation supercell

1. Choose Edit → Crystal → Generate Super Cell…

2. Set the diagonal replication factors to 4, 4, and 6; keep all off-diagonal entries zero

3. Click OK

4. Confirm that the structure contains 672 atoms: 96 Ge, 192 Sb, and 384 Te

5. Choose Edit → Crystal → Map Atoms To (0..1)

6. Save the input as GST124_4x4x6

Part B: Define the Crystalline Seed

The two planes adjacent across the z periodic boundary form one thin crystal slab with an interface on each side. They remain fixed during both the melt–quench and recrystallization jobs.

Assign the fixed and mobile regions

1. Choose View → View Direction → Along X-Axis to see the z layers from the side

2. Select all atoms in the lowest visible atomic plane near fractional z = 0

3. Hold Shift and add all atoms in the highest visible atomic plane; across the periodic boundary these are neighboring planes

4. Choose Model → Regions, create a region and name it fixed_seed

5. With fixed_seed selected, choose Select → Invert Selection

6. Create a new region and name it mobile

7. Visually confirm that the regions are disjoint and together cover all 672 atoms

8. Save the input

Fig. 58 The fixed_seed planes at the periodic boundary and the mobile interior.

Part C: Set the Machine-Learning Potential

Use the same potential for amorphization, recrystallization, and all candidate comparisons.

1. Switch to the machine-learning potential engine: →

2. Set Model to QET-MatQ

3. Set Task to Molecular Dynamics

4. If the model package is not installed, open SCM → Packages and install it before continuing

5. Save the input

Part D: Amorphize the Mobile GST

We use a single MD job with a piecewise-linear temperature program. The seed remains fixed, so this produces an amorphous mobile region next to a crystalline template rather than a fully amorphous cell.

Set the MD length and output

1. Go to Model → MD

2. Set Number of steps to 130000

3. Set Time step to 1.0 fs

4. Set Initial velocities to Random at 300 K

5. Set Trajectory sampling frequency to 500 steps

6. Set Checkpoint frequency to 10000 steps

Configure the melt–quench thermostat

1. Open Model → MD… → Thermostat and add a thermostat

2. Set Type to NHC

3. Set Temperature(s) to 300 1200 1200 300 300 K

4. Set Duration(s) to 10000 40000 60000 20000 steps

5. Set Damping constant (Tau) to 100 fs

6. Set NHC chain length to 10

7. Set Region to mobile

This schedule heats from 300 to 1200 K in 10 ps, holds at 1200 K for 40 ps, quenches to 300 K over 60 ps, and holds at 300 K for 20 ps. Repeated temperature entries create the constant-temperature segments.

Fix the crystalline seed

1. Open Model → Geometry Constraints and PES Scan

2. Select one atom that belongs to fixed_seed

3. Click + and add a Fixed position constraint for the complete fixed_seed region

4. Confirm that the thermostat acts only on mobile and that fixed_seed is not thermostatted

5. Save the job as GST124_melt_quench

Run the job. This takes about 45 min on an NVIDIA GeForce RTX 4090.

Inspect the amorphous structure

1. Open the completed job in AMSmovie

2. Go to the final trajectory frame

3. Confirm that the mobile region lacks the straight, repeating rocksalt planes of the starting structure

4. Confirm that both fixed seed planes retain their initial coordinates

5. Compare radial distribution functions for the starting and final mobile regions; the quenched structure should show broadened peaks and reduced long-range order

6. Inspect the temperature and potential energy curves for discontinuities or drift

Tip

To perform the RDF analysis, open the job in AMSmovie and choose MD Properties → RDF. Set the desired frame interval, then select the atoms to include; for example, choose Select → Select Region → mobile. Add the selected atoms to the RDF From/To selection with the +/- button, then define the range and bin width.

You can also perform the RDF analysis with a Python script. Download the PLAMS RDF script and run it with the path to your completed AMS results folder:

"$AMSBIN/amspython" make_gst124_rdf.py path_to_your_results_folder

Warning

Visual disorder alone is not sufficient validation. The final density, coordination statistics, partial radial distribution functions, and absence of atomic overlaps should all be checked before the structure is used as an amorphous model.

Fig. 59 AMSmovie analysis of the melt–quenched structure showing the potential energy and temperature evolution.

Fig. 60 RDFs of the initial and melt–quenched structures.

Part E: Recrystallize the Amorphous Region

Start from coordinates rather than an exact MD restart. This makes it possible to assign a fresh 600 K velocity distribution while keeping the same regions and constraints.

Transfer the quenched coordinates

1. In AMSmovie, select the final frame of GST124_melt_quench

2. Choose File → Update Geometry in Input

3. In AMSinput, verify that fixed_seed and mobile still contain the intended atoms

4. Verify that the fixed-region constraint is still present

5. Save as GST124_recrystallization_600K

Configure the 600 K production MD

1. Keep Task set to Molecular Dynamics and Model set to QET-MatQ

2. In Model → MD, set Number of steps to 300000

3. Set Time step to 1.0 fs

4. Set Initial velocities to Random at 600 K

5. Set Trajectory sampling frequency to 1000 steps

6. Set Checkpoint frequency to 100000 steps

7. In the thermostat block, set Type to NHC and Temperature(s) to 600 K

8. Keep Tau at 100 fs, NHC chain length at 10, and Region at mobile

9. Confirm once more that fixed_seed is fixed and is excluded from the thermostat

10. Save the input

Run and monitor crystallization

1. Start the calculation with File → Run

2. Open SCM → Movie after the first trajectory frames are available

3. View the cell along x or y so that the growth direction is vertical

4. Play the trajectory and look for ordered layers propagating from both sides of the periodic seed slab

5. Plot potential energy and temperature versus time

This job takes about 2 h and 40 min on an NVIDIA GeForce RTX 4090.

Tip

A persistent fall in potential energy often accompanies crystallization, but energy alone is not a crystalline-fraction metric. Always compare it with a structural analysis.

Part F: Analyze the Recrystallization

GUI checks

1. In AMSmovie, compare frames at 0, 40, 80, and 120 ps

2. Confirm that ordering begins at the seed interfaces rather than appearing only as isolated crystallites in the interior

3. Graph MD Properties → Potential Energy and identify the main energy-drop interval

4. Graph MD Properties → Temperature and confirm stable sampling around 600 K

5. Compare the mobile-region radial distribution function at the start and end

6. Record the first frame at which the two growth fronts meet, if this occurs

Fig. 61 Potential energy and temperature along the recrystallization trajectory.

Quantitative production analysis

The production figure uses two complementary measures of ordering in the mobile region:

• Near-90° angular-ordering rate: Count angles from 80° to 100° around each mobile atom, using mobile neighbors within 3.6 Å. An increasing count indicates more of the approximately right-angle local geometry found in crystalline GST.

• Crystalline-atom ordering rate: Calculate local \(q_4\) order using neighbors within 3.6 Å, then count mobile atoms above a threshold halfway between crystalline and amorphous reference values. An increasing count indicates that more atoms have crystal-like local environments.

We can appreciate from the production case that, for this specific GST case, both analysis are consistent. The near-90° analysis applied to the single 600 K GUI trajectory show how the two mobile-region signals evolve during recrystallization. The dashed line is a fit over the same 50–200 ps interval used for the production comparison; the displayed slope describe this one trajectory, not a replica average.

Fig. 62 Near-90° angles in the 600 K GUI run. Points are unconnected; dashed line shows the 50–200 ps fit.

Tip

Download the angle-analysis script. Run it with the path to your completed recrystallization .results folder:

"$AMSBIN/amspython" 01-plot-near-90-angles.py GST124_recrystallization_600K.results

The script saves the PNG plot of the angle series.

Limitations

This tutorial models an idealized, periodic, seeded interface at fixed volume. It does not include electrodes, Joule heating, electronic transport, free surfaces, device-scale thermal gradients, or stress relaxation against a surrounding dielectric. The simulated interface speed is therefore a materials descriptor, not a direct prediction of a memory cell’s SET time.

Finite size, seed orientation, vacancy arrangement, quench rate, thermostat, and the chosen structural classifier can all affect the result. A production study should test these choices and should preserve all failed or censored runs in the statistical analysis.

Summary

In this tutorial, you learned how to use AMSinput and an ML potential to:

• construct a 672-atom rocksalt-like GST124 supercell from scratch,

• preserve two crystal planes as a seed,

• create a melt–quenched amorphous region,

• run seeded recrystallization at 600 K,

• inspect energy, temperature, structure, and growth-front motion, and

• quantify seeded growth-front velocities with a local \(q_4\) analysis.

The opening case study shows temperature-dependent seeded growth for GST124 and GST147.

References

[1]

J. H. Han, K. S. Jeong, M. Ahn, D. H. Lim, W. J. Yang, S. Jong Park, and M. H. Cho. Modulation of phase change characteristics in Ag-incorporated Ge2Sb2Te5 owing to changes in structural distortion and bond strength. Journal of Materials Chemistry C 5, 3973–3982 (2017).

[2]

J. Kalikka, J. Akola, J. Larrucea, and R. O. Jones. Nucleus-driven crystallization of amorphous Ge2Sb2Te5: A density functional study. Physical Review B 86, 144113 (2012).

[3]

Y. Zhou, W. Zhang, E. Ma, and V. L. Deringer. Device-scale atomistic modelling of phase-change memory materials. Nature Electronics 6, 746–754 (2023).

[4]

O. Abou El Kheir, L. Bonati, M. Parrinello, and M. Bernasconi. Unraveling the crystallization kinetics of the Ge2Sb2Te5 phase change compound with a machine-learned interatomic potential. npj Computational Materials 10, 33 (2024).