基于阻变器件动态模型的大规模脉冲神经网络仿真及多种应用功能实现算法的研究数据集
收藏资源简介：
随着人工智能、大数据和物联网等信息技术的快速发展，计算模式正由以算力为核心的计算密集型向以数据处理为核心的数据密集型转变。传统基于冯·诺依曼架构的计算系统由于计算单元与存储单元物理分离，在大规模数据搬运过程中面临显著的能耗和性能瓶颈，即“存储墙”和“功耗墙”问题。与此同时，随着集成电路制造工艺逼近物理极限，单纯依赖器件尺寸微缩已难以持续支撑计算性能的指数级提升，亟需发展基于新型器件和新计算范式的高能效计算体系。忆阻器（阻变存储器件）因其多阻态存储、类突触可塑性及阈值转变等仿生电学特性，被认为是构建存算一体脉冲神经网络系统的重要基础器件。本数据集面向基于忆阻器的脉冲神经网络研究，系统整合了单元器件模型、系统级仿真工具与算法评估结果，覆盖“器件—电路—算法—系统”多层级研究需求。数据集内容包括：可参数化的忆阻突触器件与忆阻神经元器件动态模型及其特性模拟；面向忆阻脉冲神经网络的可靠系统级模拟器源代码；基于该模拟器的算法开发与计算性能分析结果，以及对应规模指标的第三方测试报告。在单元器件层面，基于MATLAB和COMSOL建立了耦合随机微观导电细丝动力学及氧缺陷迁移机制的忆阻器动态模型，并通过实验数据拟合实现对器件非理想特性与可靠性的有效刻画。在系统与算法层面，结合PyTorch、SpikingJelly及Python自主开发工具，实现了大规模忆阻脉冲神经网络的建模、权重部署可视化与性能验证。该数据集为忆阻器件建模、脉冲神经网络算法研究以及高能效存算一体系统设计提供了系统、可复现的数据支撑与研究基础，数据量1.61GB。
With the rapid development of information technologies such as artificial intelligence (AI), big data, and the Internet of Things (IoT), computing paradigms are shifting from computation-intensive, computing-power-centric models to data-intensive, data-processing-centric models. Traditional computing systems based on the von Neumann architecture face significant energy consumption and performance bottlenecks during large-scale data movement due to the physical separation of computing and memory units, namely the "memory wall" and "power wall" issues. Meanwhile, as integrated circuit manufacturing processes approach physical limits, solely relying on device dimension scaling can no longer sustain exponential improvements in computing performance, creating an urgent need for high-efficiency computing architectures based on novel devices and new computing paradigms. Memristors (resistive switching memory devices) are regarded as critical fundamental devices for building in-memory computing spiking neural network (SNN) systems, owing to their biomimetic electrical properties such as multi-resistance state storage, synapse-like plasticity, and threshold switching. This dataset is targeted at research on memristor-based spiking neural networks, systematically integrating unit device models, system-level simulation tools, and algorithm evaluation results, covering multi-level research requirements ranging from "device-circuit-algorithm-system". The dataset content includes: parameterizable dynamic models of memristive synapse devices and memristive neuron devices along with their characteristic simulations; source code of a reliable system-level simulator for memristor-based spiking neural networks; algorithm development and computing performance analysis results based on this simulator, as well as third-party test reports corresponding to scale indicators. At the unit device level, dynamic memristor models coupling random microscopic conductive filament dynamics and oxygen vacancy migration mechanisms were established based on MATLAB and COMSOL, and effective characterization of the device's non-ideal characteristics and reliability was achieved through experimental data fitting. At the system and algorithm level, combined with PyTorch, SpikingJelly, and independently developed Python tools, modeling, weight deployment visualization, and performance verification of large-scale memristor-based spiking neural networks were realized. This dataset provides systematic and reproducible data support and research foundations for memristive device modeling, spiking neural network algorithm research, and high-efficiency in-memory computing system design, with a data volume of 1.61 GB.
