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

Area of Science:

• Materials Science and Engineering
• Electrical Engineering
• Computer Science

Background:

• Resistive switching memory devices are crucial for next-generation computing.
• Improving conductance uniformity and data retention is essential for advanced applications.
• Non-filamentary devices offer advantages in stability and scalability.

Purpose of the Study:

• To enhance conductance uniformity and data retention in TiN/Ti/HfO2/TiOx/TiN resistive switching memory.
• To leverage self-rectifying characteristics for sneak current suppression in crossbar arrays.
• To demonstrate the potential for neuromorphic computing applications, specifically Transformer-based keyword spotting.

Main Methods:

• Introduction of the incremental step pulse with verify algorithm (ISPVA) technique.
• Application of hydrogen (H2) annealing to the memory device.
• Fabrication and characterization of non-filamentary TiN/Ti/HfO2/TiOx/TiN devices.

Main Results:

• Achieved a high rectifying ratio of approximately 1442 due to a high Schottky barrier.
• Demonstrated improved linearity and uniformity of conductance modulation with up to 6-bit multilevel states.
• H2 annealing stabilized conduction, leading to data retention over 10^4 s and endurance exceeding 10^4 cycles.
• Low energy consumption of 36.3 pJ was achieved.
• Implemented in a Transformer-based keyword spotting (KWS) model with 92.5% recognition accuracy.

Conclusions:

• The proposed strategy effectively improves conductance uniformity and data retention in resistive switching memory.
• The self-rectifying property is advantageous for dense memory arrays.
• The device shows significant potential for energy-efficient neuromorphic computing, particularly for Transformer-based models.