Unveiling Defect Motifs in Amorphous GeSe Using Machine Learning Interatomic Potentials

Area of Science:

• Materials Science
• Computational Materials Science
• Solid State Physics

Background:

• Ovonic Threshold Switching (OTS) selectors are key components in nonvolatile memory.
• Their nonlinear electrical characteristics and polarity-dependent threshold voltages are critical.
• The atomic-scale origins of defect states driving OTS remain poorly understood.

Purpose of the Study:

• To systematically investigate defect states in amorphous GeSe.
• To understand the atomic-scale origins of OTS behavior.
• To correlate electronic defect levels with specific structural features.

Main Methods:

• Utilized molecular dynamics simulations accelerated by machine learning interatomic potentials (MLIPs).
• Benchmarked various MLIP architectures, including descriptor-based and graph neural network (GNN)-based potentials.
• Employed an optimized GNN model to analyze 20 independent amorphous GeSe structures.

Main Results:

• Identified that higher-order interactions (≥4-body) and medium-range order are essential for accurate amorphous GeSe modeling.
• GNN architectures effectively capture these complex interactions, preventing spurious defects.
• Discovered two distinct defect motifs: aligned Ge chains (conduction band defects) and overcoordinated Ge chains (valence band defects).
• Correlated electronic defect levels with bond angle alignment and Peierls distortion.

Conclusions:

• Optimized GNN-based MLIPs provide accurate simulations of amorphous GeSe.
• Two specific defect types and their structural origins in amorphous GeSe are identified.
• These findings offer a theoretical basis for understanding defect-driven OTS phenomena.