KR-20260066684-A - Phase-Change Memory-Based Neuromorphic Device Including 2D Nanomaterial-Based Inter-Cell Diffusion Barrier

Abstract

The present invention provides a phase change memory-based neuromorphic device that blocks the movement of chalcogenide constituent atoms and segregation of components occurring during the phase change process of a phase change unit (120) from affecting the electrical characteristics of adjacent cells (100) by placing a cell-to-cell diffusion barrier (200) containing a two-dimensional nanomaterial between adjacent unit cells (100) in a phase change memory array (300) in which a plurality of unit cells (100) are arranged. The cell-to-cell diffusion barrier (200) is formed as a heterojunction vertical partition structure in which a graphene layer (atomic diffusion blocking layer (210)) and an h-BN layer (electrical insulating layer (220)) are alternately stacked, thereby completely enclosing the side of the phase change unit (120), and simultaneously exhibiting a physical atomic blocking function by a size exclusion mechanism and a heat dissipation function by utilizing thermal anisotropy. As a result, the distribution overlap of analog multi-valued synaptic weights is suppressed, and the long-term inference accuracy of neuromorphic vector-matrix multiplication operations is preserved.

Inventors

• 안범주

Assignees

• 안범주

Dates

Publication Date

20260512

Application Date

20260415

Claims (1)

1. A phase change memory array having a plurality of unit cells arranged therein, wherein each unit cell comprises an upper electrode, a lower electrode, and a phase change portion disposed between the upper electrode and the lower electrode; It includes an inter-cell diffusion barrier disposed between adjacent unit cells to block the phase change constituent atoms from diffusing into adjacent cells, and A phase change memory-based neuromorphic device characterized in that the inter-cell diffusion barrier comprises a two-dimensional nanomaterial to block material movement or component segregation occurring during the phase change process of the phase change portion from affecting the electrical characteristics of adjacent cells.

Description

Phase-Change Memory-Based Neuromorphic Device Including 2D Nanomaterial-Based Inter-Cell Diffusion Barrier The present invention relates to a Phase Change Memory (PCM)-based neuromorphic device, and more specifically, to a neuromorphic device that fundamentally blocks the adverse effects on the synapse weight accuracy of adjacent cells (100) caused by material transfer and component segregation of chalcogenide constituent atoms occurring during the phase change operation of a phase change unit (120) by placing a cell-to-cell diffusion barrier (200) containing a two-dimensional nanomaterial between adjacent unit cells (100) in a phase change memory array (300) in which a plurality of unit cells (100) are arranged at high density. With the rapid proliferation of artificial intelligence computation, the importance of neuromorphic computing technology is being significantly highlighted to overcome the data bottleneck between memory and processor in the existing Von Neumann computing architecture, known as the "Memory Wall" problem. Neuromorphic computing mimics the neural synaptic mechanisms of the human brain at the hardware level, realizing the In-Memory Computing (IMC) paradigm where memory devices simultaneously perform computation and memory. Phase Change Memory (PCM) can store multiple analog resistance states by utilizing the phenomenon of reversible phase transition between amorphous and crystalline phases in chalcogenide alloy materials, such as Ge₂Sb₂Te₅ (GST) alloy. This analog multi-level resistance characteristic serves as the core technical basis for utilizing PCM as a neuromorphic synapse device, enabling Vector-Matrix Multiplication (VMM) operations to be performed directly through the summing of currents according to Ohm's Law. Major global semiconductor companies have already demonstrated PCM-based IMC chips at the 256×256 crossbar array level, which strongly supports the feasibility of realistic implementation of PCM neuromorphic technology. However, for PCM neuromorphic devices to maintain practical-level inference accuracy over the long term, the analog resistance state—that is, the synaptic weight—must remain stable over repetitive programming cycles and long-term operation. While slight variations in the resistance distribution may be within an acceptable margin in binary memory, in neuromorphic analog multi-valued PCMs that must distinguish eight or more resistance levels, even a few percent of resistance variation directly leads to weight errors, degrading neural network inference accuracy. Such stringent weight stability requirements unique to neuromorphic devices impose much higher reliability standards on device design than those for general data storage PCMs. One of the typical reliability degradation mechanisms of GST-based PCMs is the phenomenon of segregation. When a programming current pulse is applied to the GST phase change section (120), constituent elements Ge, Sb, and Te atoms are deflected in a specific direction by the electric field and thermal gradient. Specifically, positively charged Sb + ions are attracted by the electric field toward the lower electrode (110), which is the cathode direction, and Te atoms, which have relatively high electronegativity, tend to move toward the upper electrode (130). This deflected movement results in compositional inhomogeneity, in which an Sb-enriched core and a Te-enriched shell are formed within the active region, and recent studies have reported that these separated atoms diffuse beyond the boundary of the active region into adjacent crystalline regions. In particular, in the latest high-density PCM array (300) in which the spacing between unit cells (100) is reduced to tens of nanometers (nm) or less as integration density improves, there is a high probability that segregation atoms emitted during the phase change operation of adjacent cells (100) will flow into the interior of the phase change section (120) of the adjacent cells (100) or into the interface region thereof. These introduced heterogeneous atoms change the composition of the phase change section (120) of the cell (100), causing phase change characteristics such as crystallization temperature, melting point, and resistivity to fluctuate unpredictably. As a result, as iterative programming cycles accumulate, the accuracy of synapse weights gradually deteriorates, ultimately leading to a continuous decrease in neuromorphic computation accuracy. Existing technologies have primarily addressed these segregation problems at the internal level of a single cell, specifically through interface management at the electrode-GST interface (graphene thermal barriers, alloy doping, nitride capping layers, etc.), or through thermal engineering to prevent unintended phase transitions caused by thermal crosstalk between cells. However, these existing approaches do not provide direct physical barriers to the problem of lateral diffusion of segregation atoms into adjacent cells. Furtherm