Cross-Temperature FeFETs Enabling Long- and Short-Term Memory for Reservoir Computing Network

Abstract

Hardware neural networks based on emerging nonvolatile memory are promising candidates to overcome the Von Neumann computing bottleneck. This study investigates the device characteristics and reliability of ferroelectric field-effect transistors (FeFETs) with a focus on their temperature-dependent performance. At 300 K, the FeFET demonstrates a 6.2 V memory window (MW) with 26.4% endurance degradation after 107 program/erase (P/E) cycles and 92.39% retention after 104 s. The accelerated charge trapping/detrapping dynamics enable superior short-term memory (STM) functionality. Remarkably, cryogenic operation at 77 K enhances the MW to 8 V while achieving exceptional stability with merely 0.4% degradation after 107 cycles and 99.02% retention at 104 seconds. The enhanced characteristics make it ideal for long-term memory (LTM) applications. Moreover, a reservoir computing (RC) network is proposed based on the cross-temperature FeFETs. By integrating the STM properties at 300 K and the LTM benefits at 77 K, the proposed RC network achieves a classification accuracy of 76.73% on the CIFAR-10 image recognition task. This surpasses the standalone results of 41.65% and 23.69% of 300 K and 77 K conditions, respectively. The findings highlight the potential to develop highly energy-efficient FeFET-based neuromorphic computing with varying temperature systems.