In-Sensor Computing Realization Using Fully CMOS-Compatible TiN/HfO-Based Neuristor Array

Area of Science:

• Materials Science and Engineering
• Neuromorphic Computing
• Artificial Intelligence Hardware

Background:

• The proliferation of AI and sensory data creates significant energy-efficiency challenges due to data movement between sensory, memory, and computing units.
• Existing heterogeneous integration methods like chiplets reduce data movement but do not solve the fundamental energy overheads from physically separated components.
• Brain-inspired in-sensor neuromorphic computing (ISNC) offers a solution for data-intensive applications, but material and manufacturing incompatibilities hinder development.

Purpose of the Study:

• To overcome the material and manufacturing incompatibility issues in developing in-sensor neuromorphic computing (ISNC) systems.
• To implement a fully CMOS-compatible neuristor array for stable and efficient neuromorphic processing directly within sensory devices.

Main Methods:

• Development and implementation of a fully CMOS-compatible TiN/HfO2-based neuristor array.
• Characterization of the neuristor array for multilevel analogue modulation, dispersion, and thermal stability.
• Integration of the neuristor array into an ISNC system for sensory data processing and information recognition.

Main Results:

• The TiN/HfO2 neuristor array exhibited stable, reproducible neuromorphic computing with multilevel analogue modulation and minimal dispersion.
• The device demonstrated no significant conductance degradation even at 125 °C, alongside modulatable sensory and multi-store memory functions.
• The ISNC system achieved a high information recognition accuracy of 93%, showcasing frequency selectivity and activity-dependent plasticity.

Conclusions:

• The developed CMOS-compatible TiN/HfO2 neuristor array effectively addresses key challenges in ISNC system development.
• This work presents a viable pathway towards affordable, highly efficient sensory neuromorphic systems for AI applications.
• The demonstrated performance highlights the potential of ISNC for advanced, energy-efficient data processing at the edge.