Back-end and Flexible Substrate Compatible Analog Ferroelectric Field Effect Transistors for Accurate Online Training in Deep Neural Network Accelerators

TL;DR

This paper introduces a flexible, low-cost ferroelectric FET device suitable for analog neural network training, demonstrating high accuracy and stable conductance states for scalable deep learning hardware.

Contribution

It presents a novel ferroelectric FET design compatible with flexible substrates, enabling precise analog weight updates for efficient online DNN training.

Findings

Achieved >96% accuracy on MNIST dataset.

Demonstrated >10^4 conductance states with reproducibility.

Showed linear and symmetric weight updates in hybrid devices.

Abstract

Online training of deep neural networks (DNN) can be significantly accelerated by performing in-situ vector matrix multiplication in a crossbar array of analog memories. However, training accuracies often suffer due to device non-idealities such as nonlinearity, asymmetry, limited bit precision and dynamic weight update range within constrained power budget. Here, we report a three-terminal Ferroelectric-Field-Effect-Transistor based on low thermal budget processes that can work efficiently as an analog synaptic transistor. Ferroelectric polymer P(VDF-TrFE) as the gate insulator and 2D semiconductor MoS2 as the n-type semiconducting channel material makes them suitable for flexible and wearable substrate integration. The analog conductance of the FeFETs can be precisely manipulated by employing a ferroelectric-dielectric layer as the gate stack. The ferroelectric-only devices show excellent performance as digital non-volatile memory operating at +-5V while the hybrid ferroelectric-dielectric devices show quasi-continuous resistive switching resulting from gradual ferroelectric domain rotation, important for their multibit operation. Analog conductance states of the hybrid devices allow linearity and symmetry of weight updates and produce a dynamic conductance range of 104 with >16 reproducible conducting states. Network training experiments of these FeFETs show >96% classification accuracy with MNIST handwritten datasets highlighting their potential for implementation in scaled DNN architectures.