Ferroelectric Synapses in Neuromorphic Circuits—Integration of Perovskite- and Hafnia-Based Crossbars
Abstract
Training neural networks nowadays demands a large time and energy budget, slowing the spread of so-called “neuromorphic computing”. Efforts have been put in developing bio-inspired devices supporting neuromorphic computers, targeting a nano-sized $ (5x50x50 nm^{3}) $, low-power (~20 W/petaflop), non-volatile (>1 year), fast (100 ns set/read pulses) and analog (>5 bits) memory. In Mbit arrays, such devices will co-locate storage and computing. Today’s options rely on ion-motion (e.g. red-ox reactions or phase change materials) with intrinsic limitations in terms of endurance and stochasticity. In contrast, the ferroelectric memristor is a purely electronic concept. In this talk, we present integration schemes for crossbar arrays based on two ferroelectric memristors technology. The first one is based on epitaxial $ (BiFeO_{3}) $ on an oxide electrode $ (Ca_{0.96}Ce_{0.04}MnO_{3}) $, grown on $ 1x1cm^{2} $ single crystal substrates $ (YAlO_{3}) $. The synaptic functionality of such heterostructure was demonstrated earlier using simple nanocapacitors (shared bottom electrode)[1]. The second technology is on based polycrystalline $ HfZrO_{4} $, also on an oxide electrode $ (WO_{x}), $ but directly integrated on Silicium. It follows the Back-End-Of-Line compatible process-flow proposed by Bégon-Lours et al.[2]. Passive crossbar arrays are fabricated for both technologies, with different constraints on the processing conditions. The dynamics and the synaptic behavior of the devices are presented, and the functionality of small-scale cross-bar arrays are compared for both technologies. The memristors have comparable On/Off ratio (5-10) and resistance ranges (10-100 MOhms), although their footprint differ by two orders of magnitude. They both show gradual, long-term plasticity upon the application of pulses of increasing amplitude: despite a similar film thickness (~4 nm), the voltage required to operate the hafnia devices is 2V, compared to 5 to 8V for the perovskite. The conduction mechanisms across the ferroelectric thin films are interpreted from temperature dependent measurements. The reliability is also different, with a limited effect of fatigue and retention on the $ HfZrO_{4} $ devices[3], but a relatively strong aging effect on the $ BiFeO_{3} $. Remarkably, the $ BiFeO_{3} $ thin films have a single crystalline orientation, whereas the hafnia film is polycrystalline. The texture affects the coercive field distribution, with consequences on the multi-level functionality and on the cross-talk in passive crossbars. The prospects of these technologies for neuromorphic computing are discussed: if epitaxial perovskites are promising for unsupervised learning schemes in a passive crossbar array configuration, the back-end compatibility of hafnia synaptic weights offers the possibility of using selectors, with excellent predicted accuracy in tasks such as pattern recognition[4]. [1] H. Yamada, V. Garcia, S. Fusil, S. Boyn, M. Marinova, A. Gloter, S. Xavier, J. Grollier, E. Jacquet, C. Carrétéro, C. Deranlot, M. Bibes, A. Barthélémy, ACS Nano 2013, 7, 5385. [2] L. Begon-Lours, M. Halter, Y. Popoff, Z. Yu, D. F. Falcone, D. Davila, V. Bragaglia, A. La Porta, D. Jubin, J. Fompeyrine, B. J. Offrein, IEEE J. Electron Devices Soc. 2021, 9, 1275. [3] L. Bégon-Lours, M. Halter, M. Sousa, Y. Popoff, D. D. Pineda, D. F. Falcone, Z. Yu, S. Reidt, L. Benatti, F. M. Puglisi, B. Offrein, Neuromorphic Comput. Eng. 2022, 2, DOI 10.1088/2634-4386/ac5b2d. [4] L. Bégon-Lours, M. Halter, F. M. Puglisi, L. Benatti, D. F. Falcone, Y. Popoff, D. D. Pineda, M. Sousa, B. J. Offrein, Adv. Electron. Mater. 2022, 2101395.