Optoelectronic Memristors With Tunable Photoconductive Response Supporting Mixed-Precision and Anti-Disturbance Neuromorphic Computing

Optoelectronic memristors with tunable features are the building blocks for hardware photonic electronic neural networks. In this article, a process parameter control strategy is proposed for HfOx-based memristors during atomic layer deposition (ALD). By adjusting the amount of oxygen source in HfOx layer, the device’s photoconductive response can be controlled in ITO/HfOx/TiO2/ITO memristors. The optoelectronic memristor with the lowest oxygen content exhibits the fastest photoconductive speed. Under the optical stimuli with various laser wavelengths and power densities, the storage capability is enhanced with multi-level characteristics (MLCs) compared to that under electronic stimuli. A hardware neural network architecture is constructed by using a complementary unit with oxygen-rich and oxygen-poor memristors, which can support mixed-precision recognition accuracy and improve internal anti-disturbance ability in the Fashion-MNIST dataset. Moreover, it shows strong robustness against external interference via a dynamic hidden layer design. Our findings may provide the potential to develop highly energy-efficient and anti-disturbance neuromorphic computing based on optoelectronic memristors.