Back-End, CMOS-Compatible Ferroelectric Field-Effect Transistor for Synaptic Weights

Area of Science:

• Materials Science
• Electrical Engineering
• Computer Science

Background:

• Neuromorphic computing aims to mimic the brain's efficiency by co-locating memory and processing.
• Traditional von Neumann architectures face data transfer bottlenecks, limiting performance in data-intensive tasks like machine learning.
• Nonvolatile synaptic elements, such as memristors, are crucial for building effective neuromorphic systems.

Purpose of the Study:

• To demonstrate a ferroelectric Hf_0.57Zr_0.43O₂ (HZO) field-effect transistor (FeFET) as a viable artificial synapse.
• To investigate the device's analog potentiation and depression characteristics, linearity, and switching symmetry.
• To assess the FeFET's performance in terms of energy consumption, programming speed, and data retention.

Main Methods:

• Fabrication of HZO-based FeFETs with tunable channel thickness.
• Voltage-controlled programming to induce analog potentiation and depression.
• Characterization of device linearity, switching symmetry, writing energy, and programming time.
• Retention measurements using a tungsten oxide (WOx) readout channel with varying bit depth and noise levels.

Main Results:

• Demonstrated voltage-controlled, symmetric analog potentiation and depression with good linearity in HZO FeFETs.
• Achieved low writing energy (fJ) and fast programming times (40 ns).
• Engineered the on/off ratio (1-200%) and on-resistance (>100 kΩ) by adjusting channel thickness (8-15 nm).
• Successful retention measurements over 4 bit depth with low noise (1%).

Conclusions:

• The developed HZO FeFET shows excellent properties for artificial analog synapses, including continuous linear modulation and symmetric switching.
• The device's low power consumption, fast operation, and tunable characteristics make it suitable for large-scale neuromorphic systems.
• Utilizing earth-abundant materials and CMOS compatibility (BEOL integration) enhances its potential for widespread adoption in neuromorphic computing.