🔬 materials science

On the origin of in-gap states in amorphous Ge $_2$ Sb $_2$ Te $_5$

By combining machine learning interatomic potentials with hybrid density functional theory, this study identifies that in-gap states in amorphous Ge $_2$ Sb $_2$ Te $_5$ primarily arise from structural defects like wrong bonds and specific coordination anomalies, which are depleted during glass aging to explain the mechanism of resistance drift in phase change memory devices.

Original authors: Omar Abou El Kheir, Marco Bernasconi

Published 2026-02-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Omar Abou El Kheir, Marco Bernasconi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Drifting" Memory Problem

Imagine you have a digital notebook (a Phase Change Memory device) that stores data by switching a special material between two states: a "crystal" state (easy to read, low resistance) and a "glass" state (harder to read, high resistance). This glass state is where your data lives when the power is off.

However, there is a glitch. Over time, even if you don't touch the device, the "glass" state slowly changes. It gets "stiffer" and harder to read. In the world of electronics, this is called resistance drift. It's like writing a note in a notebook, and over the next few days, the ink slowly fades or the paper shrinks, making the note harder to read. This makes it very difficult to store multiple levels of data (like storing a whole library of shades of gray instead of just black and white).

The scientists in this paper wanted to answer one question: Why does this drift happen? They suspected it was caused by tiny, invisible "defects" inside the atomic structure of the glass.

The Material: Ge2Sb2Te5 (The "GST" Cocktail)

The material they studied is called Ge2Sb2Te5 (or GST for short). Think of this material as a chaotic cocktail party of three types of atoms: Germanium (Ge), Antimony (Sb), and Tellurium (Te).

In a perfect crystal, everyone stands in neat rows. But in the "glass" state (the amorphous phase), it's a crowded, messy dance floor. Everyone is bumping into each other, and sometimes, people are standing in the wrong spots or holding hands with the wrong partners.

The Culprits: "Wrong Bonds" and "Bad Posture"

The researchers used powerful computers to simulate this atomic dance floor and found that the "drift" is caused by specific types of chaos in the crowd. They identified two main troublemakers:

Wrong Bonds (The Wrong Handshakes): In a perfect party, Germanium should only hold hands with Tellurium, and Antimony with Tellurium. But in the messy glass, sometimes Germanium holds hands with Germanium, or Germanium holds hands with Antimony. These are "wrong bonds."
- Analogy: Imagine a dance where partners are supposed to be opposite genders. If two men or two women accidentally grab hands, the dance gets awkward and unstable. These awkward handshakes create "traps" for electricity.
Overcoordinated Atoms (The Over-enthusiastic Dancers): Some atoms are trying to hold hands with too many people at once. A Germanium atom might be holding hands with 5 or 6 neighbors instead of the usual 3 or 4.
- Analogy: Imagine a person at a party trying to hug 6 people at the same time. They are in a weird, stretched-out posture. This "over-coordination" creates a weird spot where electricity gets stuck.

The "In-Gap" States: The Speed Bumps

The paper focuses on in-gap states. To understand this, imagine electricity flowing through the material like cars on a highway.

The Valence Band is the road going one way.
The Conduction Band is the road going the other way.
The Band Gap is the grassy median between the roads. Usually, cars (electrons) can't drive on the grass.

However, these "Wrong Bonds" and "Overcoordinated Atoms" act like speed bumps or potholes right in the middle of the grassy median.

When electricity tries to flow, it gets stuck in these potholes (trapped).
To get the car moving again, you need to give it a little push (voltage).
As the material "ages" (drifts), these potholes get filled in or smoothed out. Paradoxically, smoothing the road makes it harder for the cars to jump the gap, increasing the resistance and making the memory harder to read.

The Investigation: Using AI to See the Invisible

The researchers didn't just guess; they used a Machine Learning Interatomic Potential (MLIP).

Analogy: Think of this as a super-smart AI coach that has watched millions of hours of atomic dance videos. It knows exactly how these atoms move and interact without needing to calculate every single quantum physics equation from scratch (which would take too long).
They used this AI to generate 40 different "snapshots" of the messy glass.
Then, they used a high-precision microscope (Density Functional Theory) to look at the electronic structure of these snapshots.

What they found:
Almost all the "speed bumps" (in-gap states) were located exactly where the "wrong handshakes" (wrong bonds) and "over-enthusiastic dancers" (overcoordinated atoms) were. Specifically, they found that Germanium atoms in a tetrahedral shape (like a pyramid) were often the center of the chaos.

The Solution: Simulating "Aging" to Fix the Drift

The team wanted to see what happens if you let the glass "settle down" (age). They used a technique called Metadynamics.

Analogy: Imagine shaking a jar of marbles. At first, they are jumbled. If you shake them gently and let them settle, they eventually find a more stable, lower-energy arrangement.
They simulated this settling process on their computer models. They specifically targeted the areas with the "wrong bonds" and forced them to rearrange.

The Result:
When they forced the atoms to fix their "wrong handshakes" and stop "over-hugging," the speed bumps disappeared.

The "in-gap states" vanished.
The energy gap between the roads got wider.
The total energy of the system went down (it became more stable).

The Conclusion: Why This Matters

The paper concludes that the resistance drift in our memory devices is caused by the material slowly "healing" itself. As time passes, the atoms naturally fix their "wrong bonds" and stop being "overcoordinated."

The Good News: We now know exactly what the problem looks like at the atomic level.
The Bad News: The material wants to fix itself, which causes the drift.
The Future: If we can design new materials or manufacturing processes that prevent these "wrong bonds" from forming in the first place, or stop them from fixing themselves too quickly, we can build memory devices that don't drift. This would allow for faster, more reliable, and higher-capacity computers and phones.

In short: The memory drifts because the atomic "dance floor" is slowly cleaning up its messy handshakes. The scientists found the mess, and now they know how to stop the cleaning crew from ruining the party.

1. Problem Statement

Phase-change memory (PCM) devices rely on the reversible switching between crystalline and amorphous states of chalcogenide alloys, specifically Ge $_2$ Sb $_2$ Te $_5$ (GST). A critical limitation for PCM reliability and neuromorphic computing applications is resistance drift: the gradual increase in electrical resistance of the amorphous phase over time.

Mechanism: Drift is governed by the Poole-Frenkel mechanism, where electrical conduction is controlled by thermally activated hopping of carriers through localized electronic states within the mobility gap (in-gap states).
The Gap in Knowledge: While it is known that structural relaxations (aging) reduce the density of these states and increase resistance, the specific atomic origins of these in-gap states in GST remain debated. Previous studies suggested correlations with tetrahedral Ge, homopolar bonds, or overcoordinated atoms, but lacked a comprehensive statistical analysis linking specific structural motifs to electronic localization in realistic amorphous models.

2. Methodology

The authors employed a multi-scale computational approach combining Machine Learning Interatomic Potentials (MLIP) and Density Functional Theory (DFT) to generate and analyze a large ensemble of amorphous GST models.

Model Generation:
- MLIP (Neural Network): Used a DeePMD-based potential trained on DFT data to generate 30 amorphous models (297 atoms each) and 1 large model (999 atoms) via melt-quenching molecular dynamics. This potential was specifically designed to accurately reproduce the fraction of "wrong bonds" (homopolar and Ge-Sb bonds) which are often underestimated in standard DFT simulations.
- DFT Models: Generated 10 independent models (216 atoms) directly via DFT-MD quenching and utilized 4 existing models (459 atoms) from literature for comparison.
Electronic Structure Analysis:
- All models were geometry-optimized and analyzed using DFT with the HSE06 hybrid functional to accurately reproduce the band gap.
- Inverse Participation Ratio (IPR): Calculated to quantify the localization of Kohn-Sham orbitals. States with high IPR were identified as in-gap states (traps).
- Statistical Classification: In-gap states were classified as deep (mid-gap) or shallow (Urbach tails near band edges).
Structural Correlation: The local atomic environments of the most localized states were analyzed to identify specific structural motifs (e.g., coordination numbers, bond types).
Aging Simulation: Metadynamics simulations were performed on a selected model to mimic the aging process. Collective variables were biased to induce structural relaxations specifically in regions hosting in-gap states, observing the resulting changes in electronic structure and total energy.

3. Key Contributions

Large-Scale Statistical Analysis: Unlike previous studies focusing on single models, this work analyzes 44 distinct amorphous models (totaling over 10,000 atoms), providing robust statistical evidence regarding the nature of in-gap states.
Refined Structural-Electronic Link: The study definitively links the presence of in-gap states to a combination of specific structural defects, moving beyond single-feature explanations.
Validation of MLIP for Electronic Properties: The authors demonstrate that their MLIP, which better captures the fraction of wrong bonds compared to standard DFT quenching, yields electronic structures consistent with DFT when analyzed with hybrid functionals, validating its use for large-scale electronic property prediction.
Direct Simulation of Drift: The use of metadynamics to actively "heal" the glass structure and observe the simultaneous removal of in-gap states and widening of the band gap provides a direct mechanistic link between structural relaxation and resistance drift.

4. Key Results

A. Nature of In-Gap States

Dominant Features: The vast majority of in-gap states (approx. 82%) are localized on atoms possessing at least one "wrong bond" (homopolar Ge-Ge, Sb-Sb, Te-Te, or heteropolar Ge-Sb bonds).
Co-occurring Motifs: Wrong bonds rarely act alone. They are strongly correlated with:
- Tetrahedral Ge (Ge $_T$ ): Found in ~40% of in-gap states.
- Overcoordinated Atoms: 5- or 6-coordinated Ge and 6-coordinated Sb atoms (resembling crystal-like environments with antisite defects) are present in ~58% of in-gap states.
Synergy: The most critical finding is that these features are synergistic. Tetrahedral Ge is stabilized by wrong bonds, and overcoordinated atoms often reside in clusters containing wrong bonds.
State Classification:
- Deep States: Primarily localized on chains involving wrong bonds, overcoordinated Ge/Sb, and tetrahedral Ge.
- Shallow States (Urbach Tails): Localized on similar motifs but also involve long Te-Te contacts.
Density: The calculated density of deep acceptor states ( $\sim 1.8 \times 10^{20}$ cm $^{-3}$ ) aligns well with experimental photoconductivity measurements.

B. Aging and Metadynamics Simulations

Structural Relaxation: Simulating the aging process by biasing the system to relax regions hosting in-gap states resulted in:
- Removal of Defects: Transformation of tetrahedral Ge to pyramidal 3-coordinated Ge, and reduction of overcoordinated atoms.
- Elimination of States: The targeted in-gap states (LUMO, LUMO+1, etc.) disappeared from the density of states (DOS).
- Energy Gain: The system gained $\sim 0.6$ eV of total energy, confirming that these defects represent high-energy metastable configurations.
- Band Gap Widening: The Tauc gap increased by 50 meV, consistent with experimental observations of band gap widening during aging.

C. Comparison with Previous Work

The results confirm and extend findings from GeTe studies, showing that while homopolar bonds are crucial, the specific combination of wrong bonds + tetrahedral/overcoordinated geometries is the primary driver of localization in GST.
The study clarifies discrepancies in literature regarding the fraction of Ge-Ge bonds, attributing previous underestimations to the limitations of standard DFT quenching in capturing the correct fraction of wrong bonds.

5. Significance

Mechanism of Resistance Drift: The paper provides a definitive atomistic explanation for resistance drift: it is caused by the structural relaxation of the amorphous glass, which eliminates high-energy structural motifs (wrong bonds, tetrahedra, overcoordinated atoms). This depletion reduces the density of trap states, thereby increasing the activation energy for conduction and raising the resistance.
Material Design: Understanding that "wrong bonds" and specific coordination defects are the culprits offers a clear pathway for mitigating drift. Strategies could include:
- Doping to suppress the formation of tetrahedral Ge or homopolar bonds.
- Engineering the thermal history or device architecture to stabilize the glass against these specific structural relaxations.
Computational Framework: The successful integration of MLIP for generating realistic structural models with hybrid DFT for electronic analysis sets a new standard for studying complex disordered materials where both structural accuracy and electronic precision are required.

In conclusion, this work establishes that the resistance drift in GST is a direct consequence of the annihilation of specific structural defects (wrong bonds coupled with tetrahedral/overcoordinated geometries) during the aging process, offering a fundamental basis for improving the stability of phase-change memory devices.