The implementation of memristive synapses in neuromorphic computing is hindered by the limited reproducibility and high energy consumption of the switching behaviour of the devices. Typical filament-type memristors suffer, in particular, from temporal and spatial variation in the set voltage and resistance states due to stochastic filament formation. Here, we report memristors based on two-dimensional pentagonal palladium diselenide (PdSe₂) that can exhibit anomalous resistive switching behaviour with two interchangeable reset modes: total reset and quasi-reset. Heterophase grain boundaries are formed in the PdSe₂ via local phase transitions induced by electron-beam irradiation, which leads to residual filaments along the grain boundaries that can guide the formation of conductive filaments. When operated in the quasi-reset mode, the memristors show a sixfold improvement in switching variation compared with devices operating in the total-reset mode, as well as a low set voltage (0.6 V), long retention times and programmable multilevel resistance states. We also show that the devices can emulate synaptic plasticity and that multipattern memorization can be implemented using a crossbar array architecture.

忆阻突触在神经形态计算中的实现受到设备切换行为的有限可重复性和高能耗的阻碍。典型的灯丝型忆阻器尤其会因随机灯丝形成而受到设定电压和电阻状态的时间和空间变化的影响。在这里，我们报告了基于二维五边形二硒化钯 (PdSe ₂ ) 的忆阻器，它可以通过两种可互换的复位模式表现出异常的电阻切换行为：完全复位和准复位。PdSe _{2中形成异相晶界}通过电子束辐照引起的局部相变，导致沿晶界的残余细丝可以引导导电细丝的形成。在准复位模式下工作时，忆阻器的开关变化比在全复位模式下工作的器件提高了六倍，并且具有低设置电压 (0.6 V)、长保持时间和可编程多级电阻状态。我们还表明，这些设备可以模拟突触可塑性，并且可以使用交叉阵列架构实现多模式记忆。