Improve the synaptic performance of resistive switching devices through interface engineering

Transition metal based resistive switching devices (like HfO₂, TiO₂₎ has been shown to be a good candidate for neuromorphic computing for its bio-inspired synaptic properties, however, the non-linear conductance change synaptic behaviour prohibits further improvement due to poor accuracy of neural network training. Here, we provide a way to eliminate the intrinsic non-linearity through electrode-oxide interface engineering, including oxygen profile control and oxide heterostructure stacking, which can improve the neural network training accuracy and shorten the training time.