Enhanced temporal variation of resistive switching memories by insertion of rough titanium oxide thin film

Artificial intelligence (AI) possesses high adaptability and potential to replace human mental labor. However, only environments with high-performance computing resources and large power supplies can handle AI processing. Current computing technology is based on digital logic devices, leading to the inevitability of endless fetching of data among processors and memories. Moreover, acceleration of AI has been mainly studied at the software level, e.g., pruning of neural networks, which is insufficient for overcoming processing environment restrictions. Meanwhile, in-memory computing by physically composed neural networks is an emerging field. Resistive switching memory (RRAM) is a promising option, which is yet to be implemented because of the stochastic nature of the switching process. In this work, the temporal reliability of tantalum oxide-based RRAM was dramatically enhanced (similar to 1%) by the insertion of a rough titanium oxide thin film. The enhanced devices exhibited a classification accuracy of similar to 88%, showing superior performance and application potential for neuromorphic computing.