Self-rectifying memristors based on epitaxial AlScN for neuromorphic computing

With the advancement of artificial intelligence, self-rectifying memristors have attracted increasing attention due to their potential for high-density integration in storage and neuromorphic computing systems. However, device stability still faces significant challenges. In this work, by using a CMOS-compatible process, we fabricate a high-performance memristor based on Pd/Al0.77Sc0.23N/TiN/Si devices on a silicon substrate. The crystallinity, surface roughness and ferroelectric properties of the epitaxially grown films were optimized by changing the doping ratio through dual-targeted nitrogen reactive magnetron sputtering. The device maintains good stability after 1000 consecutive scans of its I–V curve. The device can achieve switching ratios of about 100 and rectification ratios of 33. In addition, we simulated biological synapses and synaptic plasticity, such as long-term potentiation/inhibition, excitatory postsynaptic current, spike time-dependent plasticity (STDP), and double-pulse facilitation, and realized bidirectional control of conductance. More importantly, we designed a trajectory-based STDP circuit model by leveraging the amplitude characteristic of the device. This model was used to train spiking neural networks for the recognition of four flight markers: forward, landing, left turn, and right turn. Subsequently, the trained neural network was deployed on a drone, validating its effectiveness. This study demonstrates a feasible approach for the hardware implementation of unsupervised spiking neural networks based on AlScN ferroelectric memristors.