On phase separation and crystallization of Ge-rich GeSbTe alloys from atomistic simulations with a machine learning interatomic potential

This paper presents a highly transferable machine learning interatomic potential for Ge-rich GeSbTe alloys, which was used to simulate nanosecond-scale crystallization at 600 K, revealing that kinetic effects during memory set operations lead to metastable phase separation into Sb-doped cubic GeTe and amorphous GeSb/Ge rather than thermodynamically stable products.

The Gist

The Big Picture: The "Smart" Memory Chip

Imagine your computer or smartphone has a tiny, super-fast memory chip called Phase Change Memory (PCM). It works like a light switch, but instead of flipping a switch, it uses heat to change a special material from a "messy, disordered" state (amorphous) to an "organized, neat" state (crystalline).

• Messy state = 0 (Off)
• Organized state = 1 (On)

The material used is usually a mix of Germanium (Ge), Antimony (Sb), and Tellurium (Te), known as GST. For the memory to work in your phone, it needs to be stable even when the phone gets hot during manufacturing (soldering). To fix this, scientists make the material "Germanium-rich" (more Ge). This makes it tougher, but it also makes the "switching on" process (writing data) much more complicated.

The Problem: A Recipe Gone Wrong

When you try to turn on this "Germanium-rich" memory, you heat it up. Thermodynamics (the laws of physics) says the material should separate into two very specific, stable ingredients: pure Germanium and a perfect mix of GST.

However, in the real world, this process happens incredibly fast—like a blink of an eye (nanoseconds). Because it's so fast, the material doesn't have time to find the perfect, stable recipe. Instead, it gets stuck in a "half-baked" state, forming weird, unstable intermediate mixtures.

Scientists knew this was happening, but they couldn't see exactly how the atoms were moving and rearranging themselves during that split second. It's like trying to watch a magic trick happen in slow motion, but the camera is too slow.

The Solution: The "Crystal Ball" (Machine Learning Potential)

To solve this, the researchers built a Machine Learning Interatomic Potential (MLIP).

The Analogy:
Imagine you are trying to teach a robot how to bake a cake.

1. The Old Way (DFT): You calculate the chemistry of every single molecule from scratch for every step. It's incredibly accurate, but it takes so long that you can only bake a tiny crumb of cake before the sun sets. You can't see the whole cake rise.
2. The New Way (MLIP): You feed the robot a massive library of thousands of "perfect cake recipes" and "baking experiments" calculated by the slow method. The robot learns the patterns of how the ingredients interact. Now, the robot can bake a giant cake in seconds, and it's almost as accurate as the slow method.

The researchers created this "smart robot" (the MLIP) by feeding it data on every possible combination of Ge, Sb, and Te atoms. They trained it to understand how these atoms attract, repel, and bond with each other.

The Experiment: Watching the Atoms Dance

Once they had their "smart robot," they simulated what happens when they heat up three different types of Germanium-rich memory alloys to 600 Kelvin (about 320°C, the temperature inside a working memory chip). They watched the simulation for a few nanoseconds (billionths of a second).

What they found:
Instead of separating cleanly into the "perfect" ingredients (Pure Ge + Perfect GST), the material did something messy and fascinating:

1. The Great Separation: The atoms started sorting themselves out immediately.
2. The "Messy" Crystals: The Germanium and Tellurium atoms rushed together to form crystals of GeTe (Germanium Telluride), but they were slightly "doped" with Antimony. Think of this like making a chocolate chip cookie, but the chips are slightly melted and mixed with a bit of peanut butter. It's not the perfect cookie recipe, but it's what forms first.
3. The "Messy" Liquid: The remaining Antimony and Germanium stayed stuck together in a messy, liquid-like blob (Amorphous GeSb).

The Metaphor:
Imagine a party where everyone is mixed up (the amorphous state). When the music starts (heating up), the guests want to separate into two groups: the "Ge-Te" dancers and the "Sb" dancers.

• Thermodynamics (The Ideal): Everyone should eventually sort perfectly into two distinct, quiet rooms.
• Kinetics (The Reality): Because the party is ending so fast (nanoseconds), the "Ge-Te" dancers form a dance circle immediately, but they grab a few "Sb" guests by mistake. The "Sb" guests are left standing in a huddle nearby, unable to leave the room in time. The result is a chaotic, intermediate party that looks different from the "perfect" ending.

Why This Matters

1. It Explains the Mystery: The simulation confirmed that the memory cells aren't forming the "perfect" crystals we thought they should. They are forming these metastable (temporary) GeTe crystals. This explains why experimental measurements (like looking at the chips under a microscope) show strange compositions that didn't match the simple theory.
2. It's Fast and Accurate: The new AI tool allows scientists to simulate millions of atoms over time scales that were previously impossible. This is like upgrading from a telescope that can only see one star to a wide-angle camera that can see the whole galaxy.
3. Better Memory Design: By understanding exactly how these atoms behave in the first few nanoseconds, engineers can design better memory chips that switch faster, last longer, and are more reliable.

The Takeaway

The researchers built a super-smart AI model that acts like a high-speed microscope for atoms. They used it to watch Germanium-rich memory chips "turn on." They discovered that the material doesn't follow the "perfect" textbook recipe; instead, it rushes to form a messy, temporary mix of crystals and blobs because it's in a huge hurry. This discovery helps us understand how our phones and computers actually store data at the atomic level.

Technical Summary

1. Problem Statement

Phase Change Memories (PCMs) rely on the reversible transformation between amorphous and crystalline states of chalcogenide alloys, typically Ge-Sb-Te (GST). While standard GST225 (Ge$_2$Sb$_2$Te$_5$) is used for standalone and neuromorphic devices, embedded memory applications (e.g., in microcontrollers) require alloys with higher thermal stability to withstand soldering temperatures (~550 K). This has led to the development of Ge-rich GST (GGST) alloys.

However, the crystallization mechanism of GGST is complex:

• Thermodynamic vs. Kinetic Conflict: Thermodynamics predicts that GGST should decompose into pure Ge and stoichiometric GST compounds (e.g., GST225 or Sb$_2$Te$_3$). However, experimental evidence suggests that kinetic effects during the short "set" operation (nanosecond timescale) lead to the formation of metastable intermediate phases (e.g., GeTe-rich crystals and GeSb amorphous regions) rather than the thermodynamically stable products.
• Simulation Limitations: Understanding these kinetics requires atomistic simulations over nanosecond timescales with large system sizes. Traditional Density Functional Theory (DFT) is too computationally expensive for these scales, while classical empirical potentials often lack the accuracy required to describe the complex bonding and phase separation in multi-component chalcogenides.

2. Methodology

The authors developed a Machine Learning Interatomic Potential (MLIP) based on the DeePMD (Deep Potential Molecular Dynamics) framework to bridge the gap between DFT accuracy and MD efficiency.

• Training Database Construction:

  • The MLIP was trained on a large database (~470,000 configurations) of energies, forces, and virial tensors computed via DFT (PBE functional).
  • The dataset covers a wide range of the Ge-Sb-Te phase diagram, including:
    • Elemental phases (Ge, Sb, Te).
    • Stoichiometric binaries (GeTe, Sb$_2$Te$_3$, Sb$_2$Te) and ternaries (GST225, GeSb$_2$Te$_4$).
    • Off-stoichiometric alloys (Ge-rich GexTe, GeSb, GST523).
    • Superlattices and interfaces.
  • Active Learning: The database was iteratively enriched using metadynamics simulations to specifically target Ge segregation and Ge-GeSbTe interfaces, ensuring the potential accurately describes phase separation barriers.

• Validation Strategy:

  • Accuracy: Root Mean Square Errors (RMSE) were calculated for energies (7.6 meV/atom) and forces (148 meV/Å), comparable to DFT test sets.
  • Transferability: The potential was validated against unseen compositions (e.g., GST725, Ge$_{9.6}$Sb$_2$Te$_5$) and various phases (liquid, amorphous, crystalline). Structural properties (pair correlation functions, coordination numbers, tetrahedral fractions) and dynamical properties (self-diffusion coefficients) showed excellent agreement with reference DFT data.
  • Dynamical Validation: Crystal growth velocities and melting points for GST225 were simulated and compared with previous NN-MD and experimental data.

• Simulation Protocol:

  • Three specific Ge-rich alloys were simulated at 600 K (representing memory operation conditions) on the nanosecond timescale:
    1. GST523 (Ge$_5$Sb$_2$Te$_3$) on the Ge-Sb$_2$Te$_3$ tie-line.
    2. GST725 (Ge$_7$Sb$_2$Te$_5$) on the Ge-GST225 tie-line.
    3. Ge$_{9.6}$Sb$_2$Te$_5$ (Ge-rich) near the intersection of tie-lines.
  • Analysis: The Steinhardt order parameter ($Q_4$) was used to identify crystalline atoms. Smooth Overlap of Atomic Positions (SOAP) combined with Principal Component Analysis (PCA) and K-means clustering was used to identify distinct chemical/structural environments and track the evolution of phase separation.

3. Key Contributions

1. Development of a Transferable MLIP: Creation of a highly accurate and transferable DeePMD potential for the Ge-Sb-Te system, capable of describing complex off-stoichiometric alloys and interfaces without retraining.
2. Resolution of Kinetic Pathways: The study successfully simulates the nanosecond-scale crystallization and phase separation of GGST alloys, a timescale inaccessible to DFT but critical for memory device operation.
3. Identification of Metastable Products: The simulations reveal that the "set" process does not yield thermodynamically stable pure Ge and GST225/Sb$_2$Te$_3$. Instead, it produces metastable intermediate phases driven by kinetics.

4. Key Results

• Phase Separation Mechanism: Upon cooling to 600 K, the amorphous GGST alloys rapidly undergo phase separation (within 1–3 ns).
  • The system segregates into GeTe-like regions and GeSb-like regions.
  • In the most Ge-rich alloy (Ge$_{9.6}$Sb$_2$Te$_5$), pure amorphous Ge regions also form.
• Crystallization Pathway:
  • Crystallization initiates within the GeTe-like regions.
  • The final product on the nanosecond scale is cubic GeTe slightly doped with Sb (forming a percolating crystalline network) surrounded by amorphous GeSb (and pure Ge in the richest alloy).
  • Contrast with Thermodynamics: The thermodynamically stable products (pure Ge + GST225 or Sb$_2$Te$_3$) are not formed within the simulation time. The system is kinetically trapped in the metastable GeTe/GeSb state.
• Agreement with Experiment: The simulated formation of Sb-doped cubic GeTe crystallites aligns with recent experimental observations (TEM/EELS) in PCM cells, which previously struggled to distinguish between Sb-doped GeTe and stoichiometric GST due to similar lattice constants.

5. Significance

• Device Optimization: The findings explain why Ge-rich alloys exhibit higher crystallization temperatures ($T_x$) and faster set speeds. The kinetic barrier to forming the thermodynamically stable phase (requiring long-range Sb diffusion) is bypassed by the rapid formation of metastable GeTe crystals.
• Interpretation of Experimental Data: The study provides a theoretical basis for interpreting complex experimental characterizations of PCM cells, clarifying that observed GeTe-like phases are likely kinetic intermediates rather than final equilibrium states.
• Methodological Advancement: This work demonstrates the power of MLIPs in investigating complex, multi-component phase change materials, enabling the simulation of realistic device-relevant timescales and compositions that were previously out of reach for first-principles methods. It paves the way for simulating full device-scale operations, including the effects of electric fields and thermal gradients.