Zhu Yuan, Nyberg Tomas, Nyholm Leif, Primetzhofer Daniel, Shi Xun, Zhang Zhen
Division of Solid-State Electronics, Department of Electrical Engineering, Uppsala University, 75121, Uppsala, Sweden.
Department of Chemistry, Uppsala University, Uppsala, Sweden.
Nanomicro Lett. 2024 Nov 22;17(1):69. doi: 10.1007/s40820-024-01559-2.
Memristive crossbar arrays (MCAs) offer parallel data storage and processing for energy-efficient neuromorphic computing. However, most wafer-scale MCAs that are compatible with complementary metal-oxide-semiconductor (CMOS) technology still suffer from substantially larger energy consumption than biological synapses, due to the slow kinetics of forming conductive paths inside the memristive units. Here we report wafer-scale AgS-based MCAs realized using CMOS-compatible processes at temperatures below 160 °C. AgS electrolytes supply highly mobile Ag ions, and provide the Ag/AgS interface with low silver nucleation barrier to form silver filaments at low energy costs. By further enhancing Ag migration in AgS electrolytes via microstructure modulation, the integrated memristors exhibit a record low threshold of approximately - 0.1 V, and demonstrate ultra-low switching-energies reaching femtojoule values as observed in biological synapses. The low-temperature process also enables MCA integration on polyimide substrates for applications in flexible electronics. Moreover, the intrinsic nonidealities of the memristive units for deep learning can be compensated by employing an advanced training algorithm. An impressive accuracy of 92.6% in image recognition simulations is demonstrated with the MCAs after the compensation. The demonstrated MCAs provide a promising device option for neuromorphic computing with ultra-high energy-efficiency.
忆阻交叉阵列（MCAs）为节能神经形态计算提供了并行数据存储和处理功能。然而，大多数与互补金属氧化物半导体（CMOS）技术兼容的晶圆级MCAs，由于忆阻单元内部形成导电路径的动力学缓慢，其能耗仍远高于生物突触。在此，我们报告了在低于160°C的温度下使用CMOS兼容工艺实现的晶圆级基于AgS的MCAs。AgS电解质提供高迁移率的Ag离子，并为Ag/AgS界面提供低银成核势垒，从而以低能量成本形成银细丝。通过微观结构调制进一步增强AgS电解质中的Ag迁移，集成忆阻器表现出创纪录的低阈值，约为 -0.1 V，并展示出达到飞焦耳值的超低开关能量，这与生物突触中观察到的情况相同。低温工艺还使MCA能够集成在聚酰亚胺基板上，用于柔性电子器件。此外，通过采用先进的训练算法，可以补偿用于深度学习的忆阻单元的固有非理想性。补偿后的MCAs在图像识别模拟中展示出92.6%的令人印象深刻的准确率。所展示的MCAs为具有超高能量效率的神经形态计算提供了一个有前景的器件选择。
