Neuromorphic computing based on memristors could help meet the growing demand for data-intensive computing applications such as artificial intelligence. Analogue memristors with multiple conductance states are of particular use in high-efficiency neuromorphic computing, but their weight mapping capabilities are typically limited by small on/off ratios. Here we show that memristors with analogue resistive switching and large on/off ratios can be created using two-dimensional van der Waals metallic materials (graphene or platinum ditelluride) as the cathodes. The memristors use silver as the top anode and indium phosphorus sulfide as the switching medium. Previous approaches have focused on modulating ion motion using changes to the resistive switching layer or anode, which can lower the on/off ratios. In contrast, our approach relies on the van der Waals cathode, which allows silver ion intercalation/de-intercalation, creating a high diffusion barrier to modulate ion motion. The strategy can achieve analogue resistive switching with an on/off ratio up to 10⁸, over 8-bit conductance states and attojoule-level power consumption. We use the analogue properties to perform the chip-level simulation of a convolutional neural network that offers high recognition accuracy.

基于忆阻器的神经形态计算可以帮助满足对人工智能等数据密集型计算应用日益增长的需求。具有多种电导态的模拟忆阻器特别适用于高效神经形态计算，但它们的权重映射能力通常受到小开/关比的限制。在这里，我们展示了可以使用二维范德华金属材料（石墨烯或二碲化铂）作为阴极来创建具有模拟电阻开关和大开/关比的忆阻器。忆阻器使用银作为顶部阳极，硫化铟磷作为开关介质。以前的方法侧重于通过改变电阻开关层或阳极来调节离子运动，这可以降低开/关比。相比之下，我们的方法依赖于范德华阴极，它允许银离子嵌入/解嵌入，从而产生高扩散势垒来调节离子运动。该策略可以实现模拟电阻式开关，开/关比高达 10：⁸，超过 8 位电导状态和阿焦耳级功耗。我们使用模拟属性来执行卷积神经网络的芯片级仿真，该网络提供高识别精度。