Dynamic FET-based memristor with relaxor antiferroelectric HfO2 gate dielectric for fast reservoir computing

Reservoir computing (RC), as a framework for artificial intelligence (AI) computation, is derived from recurrent neural networks, but with higher efficiency benefits from its much simpler training process. A hardware-based physical RC employs a reservoir that is a fixed non-linear system, and in a simple RC structure, the reservoir is ideally a short-term dynamic field effect transistor-based memristor. In this work, we demonstrate that reservoir computing employing a polarization-modulated transistor as a physical reservoir with a relaxor antiferroelectric back gate significantly reduces parameters, power consumption, and calculation steps of the neural network. Here, relaxor antiferroelectricity is realized in an Al-doped Hf0.5Zr0.5O2 (Al:HZO) thin film, and with the poling process-sensitive modulation of the Al:HZO gate dielectric, a dynamic and polarization field-modulated transistor is demonstrated as a short-term memristor. This short-term memory behavior is further used to encode binary images to form a reservoir for fast image processing, where the reservoir mimics small convolution operations to convert images. This reservoir computing is also applied to image classification, denoising, and similarity judg-ment, and very reliable results are obtained. This work provides a new approach to hardware-based physical reservoir computing.& COPY; 2023 Elsevier Ltd. All rights reserved.