Song Jaesub, Moon Seokho, Byun Jinho, Kim Jiye, Choi Junyoung, Kwak Hyunjeong, Hwang Inyong, Ji Changuk, Kim Seyoung, Kim Jong Kyu
Department of Materials Science and Engineering, POSTECH, Pohang, 37673, Republic of Korea.
Small. 2025 Aug;21(32):e2504507. doi: 10.1002/smll.202504507. Epub 2025 Jun 16.
Memristors have emerged as a key building block for artificial neural networks (ANNs), offering energy efficiency and high scalability for hardware-based synaptic weight updates. As device miniaturization is crucial for enhancing memristor performance, hexagonal boron nitride (h-BN) stands out as a promising resistive switching medium due to its excellent insulating characteristics even at an atomically thin scale. However, conventional h-BN memristors suffer from abrupt switching behavior by uncontrollable filament formation, limiting their potential for ANN applications. Here, h-BN-based memristors exhibiting linear and symmetric analog switching by leveraging multiple nano-filament confinement is presented. The geometric confinement between suspended h-BN films and the apexes of GaN nano-cones facilitates analog switching behavior, reducing cycle-to-cycle variation and ensuring stable consecutive operations. Electrical analyses reveal that analog switching behavior originates from the controlled formation of multiple nano-filaments within the confined geometry. ANNs implemented with these nano-filaments confined to h-BN memristors exhibit highly linear and symmetric synaptic weight updates, enabling precise training with minimal accuracy degradation. This work establishes multiple nano-filament confinement as a universal design strategy for achieving reliable and linear analog switching in memristors, paving the way for advanced neuromorphic computing.
忆阻器已成为人工神经网络（ANNs）的关键构建模块，为基于硬件的突触权重更新提供了能源效率和高可扩展性。由于器件小型化对于提高忆阻器性能至关重要，六方氮化硼（h-BN）因其即使在原子级薄尺度下也具有出色的绝缘特性，成为一种很有前景的电阻开关介质。然而，传统的h-BN忆阻器因不可控的细丝形成而存在突然开关行为，限制了它们在ANN应用中的潜力。在此，提出了一种基于h-BN的忆阻器，通过利用多个纳米细丝限制来展现线性和对称的模拟开关。悬浮的h-BN薄膜与GaN纳米锥顶点之间的几何限制促进了模拟开关行为，减少了逐周期变化并确保了稳定的连续操作。电学分析表明，模拟开关行为源于在受限几何结构内多个纳米细丝的受控形成。用这些限制在h-BN忆阻器中的纳米细丝实现的人工神经网络展现出高度线性和对称的突触权重更新，能够以最小的精度下降进行精确训练。这项工作将多个纳米细丝限制确立为在忆阻器中实现可靠和线性模拟开关的通用设计策略，为先进的神经形态计算铺平了道路。
