An Ultrathin, Cyano-Functionalized Copolymeric Memristor by iCVD Process for Driving Convolutional Neural Networks of...

Area of Science:

• Materials Science
• Nanotechnology
• Computer Engineering

Background:

• On-chip learning requires weight storage elements with scalability, data retention, symmetry, linear conductance modulation, and fine-tuning.
• Memristors are crucial for artificial intelligence (AI) hardware acceleration.

Purpose of the Study:

• To fabricate high-performance memristors using cyano-based ultrathin copolymer films for convolutional neural networks (CNNs).
• To achieve linear, symmetric, and multi-level conductance modulation for efficient weight storage.

Main Methods:

• Fabrication of memristors using initiated chemical vapor deposition (iCVD) with 2-cyanoethyl acrylate (CEA) and di(ethylene glycol) divinyl ether (DEGDVE) copolymer films (<10 nm).
• Controlled polymer composition engineering to tune switching characteristics and filament formation.
• Electrical manipulation using ramp pulse series (RPS) to study filament dynamics and device reliability.
• Image classification tasks on datasets (Oxford 102 Flowers, Food-101, Stanford Cars) using various CNN architectures (VGG-X, ResNet-X, DenseNet).

Main Results:

• Achieved highly linear, symmetric, and multi-level conductance modulation by controlling CEA and DEGDVE ratios.
• Demonstrated reliable operation and studied conducting filament dynamics via RPS.
• Attained up to 88.39% classification accuracy on image datasets, validating the memristor-based CNN architecture.

Conclusions:

• The developed memristors show excellent performance for on-chip learning applications.
• Precise control over polymer composition enables tailored device characteristics for AI hardware.
• The memristor-based CNN architecture is efficient for real-world AI applications.