Memristive Artificial Synapses Based on Brownmillerite for Endurable Weight Modulation

Area of Science:

• Materials Science
• Neuroscience
• Computer Engineering

Background:

• Artificial synapses are crucial for computing paradigms that merge memory and computation.
• While memristors offer energy efficiency for artificial synapses, their long-term synaptic modulation is limited by random filament conduction.
• Enhancing the endurability and reliability of synaptic weight updates is essential for advanced neural network applications.

Purpose of the Study:

• To leverage topotactic phase transition (TPT) in brownmillerite-phased SrCoO2.5 (SCO2.5) for improved artificial synapse endurability.
• To demonstrate a novel memristive synapse design utilizing TPT for reversible oxygen ion migration.
• To validate the performance of TPT-based artificial synapses in deep and convolutional neural networks.

Main Methods:

• Fabrication of a heteroepitaxial Au/SCO2.5/SrRuO3/SrTiO3 2-terminal device.
• Utilizing density-functional theory (DFT) calculations and experimental Raman spectroscopy to demonstrate TPT behavior.
• Applying voltage pulses to induce TPT and characterize synaptic plasticity (long-term potentiation and depression).

Main Results:

• Demonstrated reliable, linear, and symmetric long-term potentiation and depression via voltage pulse-driven TPT.
• Achieved consistent and noise-free synaptic weight updates over 32,000 iterations and 640 cycles, showcasing high durability.
• Attained very high recognition accuracy in deep neural networks and convolutional neural networks for image datasets.

Conclusions:

• Topotactic phase transition in brownmillerite SCO2.5 provides a robust mechanism for reliable and endurable artificial synaptic weight updates.
• The developed TPT-based memristive synapse shows significant potential for advancing neuromorphic computing and artificial intelligence.
• This work offers critical insights for designing next-generation memristive devices for efficient and dependable neural network implementations.