Artificial synapse based on a tri-layer AlN/AlScN/AlN stacked memristor for neuromorphic computing

Neuromorphic devices have garnered significant attention for their potential to revolutionize conventional computing architecture and drive advancements in artificial neural systems. Aluminum nitride-based (AlN-based) memristors have emerged as particularly noteworthy due to their exceptional properties, including ultrafast switching speed, small switching current, substantial on/off ratio, controllable material growth, and compatibility with complementary metal-oxide-semiconductor (CMOS) processes. These remarkable characteristics hold immense significance in the fabrication of novel neuromorphic devices, specifically for artificial synapses. However, the commonly observed abrupt resistive switching behavior in AlN-based memristors poses a challenge to the recognition accuracy of artificial neural networks (ANNs) at the system level. Thus, achieving a gradual switching behavior with multi-level conductance becomes highly desirable for artificial synapses. Here, an interfacial engineering approach is introduced to optimize the Ag/AlN/Pt memristor by incorporating an aluminum scandium nitride (AlScN) layer within the AlN layer. The tri-layer AlN/AlScN/AlN stacked memristor (ASAM) demonstrates a notable achievement of gradual switching behavior at the RESET operation, attributed to the alleviation of abrupt conductive filament formation resulting from the ferroelectric polarization effect of the AlScN layer. Additionally, the ASAM shows excellent resistive switching performance with ultrafast switching speed (<5 ns), low operation voltage (<0.5 V), and ultralow power consumption as small as 0.2 pJ. By appropriately adjusting the current compliance and resetting stop voltage, the ASAM exhibits remarkable characteristics of controlled gradual switching and multi-level conductance. Moreover, the ASAM successfully mimics biological synaptic functions, such as long-term potentiation (LTP), long-term depression (LTD), paired-pulse facilitation (PPF), and Spike-Timing-Dependent Plasticity (STDP). Leveraging the near linearity of conductance modulation provided by the ASAM, the MNIST handwritten digits recognition task, which is simulated using experimental data by constructing a convolutional neural network (CNN), can achieve a high accuracy of 93% after 150 epochs. Importantly, the design scheme of the stacked device structure with polarization effect holds significant promise for the fabrication of nitride-based memristors with highly linear and symmetrical conductance regulation. Such characteristics are crucial for the development of analog computing architecture, enabling more efficient and accurate neuromorphic systems.