Scholarlink

<p>To achieve energy-efficient neuromorphic hardware, there is a need for compact and physics-based models that can model the coupled electrical, thermal, and phase transition dynamics in analog phase-change synapses. The paper focuses on the novel state-variable model that is specifically designed to capture the GST467 phase-change synapse dynamics by coupling the electro-thermal-kinetic dynamics with the crystalline fraction evolution and a localized thermal memory variable. Unlike other compact models based on threshold logic or binary PCM, the proposed model allows for gradual, repeatable, and variability-resilient conductance modulation in response to the same sub-volt programming pulses. The proposed model is implemented in MATLAB and is computationally efficient. Multi-level operation with more than 8 stable conductance states is achieved in the microampere current regime, and Monte Carlo simulations show that the synaptic weight precision is robust against ± 10% variability of the model parameters. Owing to its physical interpretability, scalability, and low computational overhead, the proposed framework provides a practical and extensible compact-modeling platform for large-scale neuromorphic crossbar simulations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A coupled electro-thermal kinetic state variable model for GST467 phase change synapses

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月6日

查看全文
  • Vikas Bhatnagar,
  • Adesh Kumar,
  • Manish H. Bilgaye

摘要

To achieve energy-efficient neuromorphic hardware, there is a need for compact and physics-based models that can model the coupled electrical, thermal, and phase transition dynamics in analog phase-change synapses. The paper focuses on the novel state-variable model that is specifically designed to capture the GST467 phase-change synapse dynamics by coupling the electro-thermal-kinetic dynamics with the crystalline fraction evolution and a localized thermal memory variable. Unlike other compact models based on threshold logic or binary PCM, the proposed model allows for gradual, repeatable, and variability-resilient conductance modulation in response to the same sub-volt programming pulses. The proposed model is implemented in MATLAB and is computationally efficient. Multi-level operation with more than 8 stable conductance states is achieved in the microampere current regime, and Monte Carlo simulations show that the synaptic weight precision is robust against ± 10% variability of the model parameters. Owing to its physical interpretability, scalability, and low computational overhead, the proposed framework provides a practical and extensible compact-modeling platform for large-scale neuromorphic crossbar simulations.