Improving MNIST recognition accuracy through enhanced switching uniformity of nitrogen-doped CMOS process-compatible Ta₂O₅-based ReRAM device

Abstract In this study, we developed the complementary metal oxide semiconductor process-compatible Ta 2 O 5 -based resistive random access memory (ReRAM) with improved uniformity through 6% nitrogen (N) doping. The N-doped TaO X layer, acting as an oxygen reservoir, helped control the stochastic behavior of conductive filaments. As a result, the N-doped device showed a 42% and 46% reduction in current variation for the low resistance state and high resistance state, respectively, compared to the undoped device. Additionally, the set voltage ( V SET ) variation decreased by 50%, marking a significant improvement in performance. These results represent the good current and voltage variation values among ReRAM devices. The MNIST-based recognition accuracy of the N-doped device-based neural network was evaluated using the NeuroSim simulator, achieving about 90% accuracy, which is 10% higher than the undoped device. This N-doped ReRAM device technology is expected to be utilized for applications in neuromorphic computing requiring high uniformity of device.