Hydrogen-Stabilized Self-Rectifying Memristor Arrays for Reliable Multilevel Synapses in Transformer-Based Keyword Spotting

초록

This study proposes a strategy to simultaneously improve conductance uniformity and data retention characteristics by introducing the incremental step pulse with verify algorithm (ISPVA) technique and hydrogen (H2) annealing into a non-filamentary TiN/Ti/HfO2/TiOx/TiN resistive switching memory device. The high Schottky barrier formed at the Ti/HfO2 interface induces asymmetric electron injection and limits reverse current flow, resulting in a rectifying ratio of approximately 1442. This self-rectifying characteristic provides an intrinsic advantage in suppressing sneak currents in crossbar arrays. The ISPVA technique improves the linearity and uniformity of conductance modulation, enabling the implementation of up to 6-bit multilevel states within a few-µA current range. In addition, H2 annealing stabilized conduction by forming hydrogen bonds with oxygen vacancies in the oxide layer and suppressing oxygen ion–vacancy recombination. As a result, data retention over 104 s and endurance exceeding 104 cycles were achieved even under a low energy consumption of 36.3 pJ. Furthermore, the experimentally obtained long-term potentiation and depression characteristics were implemented in a Transformer-based keyword spotting (KWS) model, achieving a recognition accuracy of 92.5%. These results suggest that the proposed device enables controlled analog conductance modulation with improved stability, showing its potential for Transformer-based neuromorphic computing applications. © 2026 The Author(s). Advanced Science published by Wiley-VCH GmbH.

키워드