Amorphous Boron Nitride Memristive Device for High-Density Memory and Neuromorphic Computing Applications

Abstract

Although two-dimensional (2D) nanomaterials are promising candidates for use in memory and synaptic devices owing to their unique physical, chemical, and electrical properties, the process compatibility, synthetic reliability, and cost-effectiveness of 2D materials must be enhanced. In this context, amorphous boron nitride (a-BN) has emerged as a potential material for future 2D nanoelectronics. Therefore, we explored the use of a-BN for multilevel resistive switching (MRS) and synaptic learning applications by fabricating a complementary metal-oxide-semiconductor (CMOS)-compatible Ag/a-BN/Pt memory device. The redox-active Ag and boron vacancies enhance the mixed electrochemical metallization and valence change conduction mechanism. The synthesized a-BN switching layer was characterized using several analyses. The fabricated memory devices exhibited bipolar resistive switching with low set and reset voltages (+0.8 and -2 V, respectively) and a small operating voltage distribution. In addition, the switching voltages of the device were modeled using a time-series analysis, for which the Holt's exponential smoothing technique provided good modeling and prediction results. According to the analytical calculations, the fabricated Ag/a-BN/Pt device was found to be memristive, and its MRS ability was investigated by varying the compliance current. The multilevel states demonstrated a uniform resistance distribution with a high endurance of up to 10(4) direct current (DC) cycles and memory retention characteristics of over 10(6) s. Conductive atomic force microscopy was performed to clarify the resistive switching mechanism of the device, and the likely mixed electrochemical metallization and valence change mechanisms involved therein were discussed based on experimental results. The Ag/a-BN/Pt memristive devices mimicked potentiation/depression and spike-timing-dependent plasticity-based Hebbian-learning rules with a high pattern accuracy (90.8%) when implemented in neural network simulations.