He Qian, Wang Hailiang, Zhang Yishu, Chen Anzhe, Fu Yu, Xue Guodong, Liu Kaihui, Huang Shiman, Xu Yang, Yu Bin
College of Integrated Circuits, Zhejiang University, Hangzhou, Zhejiang, China.
ZJU-Hangzhou Global Scientific and Technological Innovation Center, Hangzhou, Zhejiang, China.
Nat Commun. 2025 May 9;16(1):4340. doi: 10.1038/s41467-025-59815-x.
Neuromorphic computing based on two-dimensional materials represents a promising hardware approach for data-intensive applications. Central to this new paradigm are memristive devices, which serve as the essential components in synaptic kernels. However, large-scale implementation of synaptic matrix using two-dimensional materials is hindered by challenges related to random component variation and array-level integration. Here, we develop a 16 × 16 computing kernel based on two-transistor-two-resistor unit with three-dimensional heterogeneous integration compatibility to boost energy efficiency and computing performance. We demonstrate the 4-bit weight characteristics of artificial synapses with low stochasticity. The synaptic array demonstration validates the practicality of utilizing emerging two-dimensional materials for monolithic three-dimensional heterogeneous integration. Additionally, we introduce the Gaussian noise quantization weight-training scheme alongside the ConvMixer convolution architecture to achieve image dataset identification with high accuracy. Our findings indicate that the synaptic kernel can significantly improve detection accuracy and inference performance on the CIFAR-10 dataset.
基于二维材料的神经形态计算为数据密集型应用提供了一种很有前景的硬件方法。这种新范式的核心是忆阻器件，它是突触内核的基本组成部分。然而，使用二维材料大规模实现突触矩阵受到与随机元件变化和阵列级集成相关的挑战的阻碍。在此，我们基于具有三维异构集成兼容性的双晶体管双电阻单元开发了一个16×16的计算内核，以提高能源效率和计算性能。我们展示了具有低随机性的人工突触的4位权重特性。突触阵列演示验证了利用新兴二维材料进行单片三维异构集成的实用性。此外，我们引入了高斯噪声量化权重训练方案以及ConvMixer卷积架构，以实现高精度的图像数据集识别。我们的研究结果表明，突触内核可以显著提高在CIFAR-10数据集上的检测精度和推理性能。
