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

For on-chip learning, ideal weight storage elements should have scalability, data retention, symmetry and linear conductance modulation, and weight fine-tuning capabilities. In this study, memristors are fabricated by employing cyano-based ultrathin copolymer films (<10 nm) using 2-cyanoethyl acrylate (CEA) and di(ethylene glycol) divinyl ether (DEGDVE) as functional monomers via an initiated chemical vapor deposition (iCVD) process, optimized to serve as a high-performance device for convolutional neural networks (CNNs). The device achieves highly linear, symmetric, and multi-level conductance modulation through precise control of polymer composition engineering. The switching characteristics and filament formation are controlled by varying the ratio of CEA and DEGDVE. In addition, the reliability and operation mechanism of the device are studied through non-invasive observation of the conducting filament dynamics via electrical manipulation using ramp pulse series (RPS). Finally, image classification tasks ares performed on high-resolution datasets such as Oxford 102 Flowers, Food-101, and Stanford Cars by varying pulse amplitudes and durations to simulate conductance modulation such as potentiation and depression of weights in memristors. Utilizing various networks such as VGG-X, ResNet-X, and DenseNet, the proposed system demonstrated robust performance, achieving up to 88.39% classification accuracy, validating the efficiency of the memristor-based CNN architecture in real-world AI applications.