Memristor Crossbar Array with Enhanced Device Yield for In-Memory Vector-Matrix Multiplication

In this work, we present a fabrication strategy for high-yield memristor crossbar arrays. Our approach uses an Al₂O₃/TiO_x-based bilayer memristor with a combination of a dielectric and an oxygen reservoir layer. The fabrication process is optimized by controlling the thickness of the Al₂O₃ layer to decrease the forming voltage, thus reducing the possibility of device failure due to excessive current during the forming process. We also investigate yield trends by controlling the thickness and oxygen concentration of the TiO_x layer, achieving a yield of over 98% under the optimal conditions. We then fabricate a memristor crossbar array under the optimized conditions and statistically characterize the devices in the array. As a compute-in-memory in-memory computing application, we develop a fully connected neural network for 5 × 5 image classification based on in-memory vector-matrix multiplication. By transferring the pretrained network to the crossbar array with an error of less than 5%, 100% classification accuracy can be experimentally achieved as a result of the inference measurement for 480 test images.