Hafnium oxide-based nonvolatile ferroelectric memcapacitor array for high energy-efficiency neuromorphic computing

Recent advancements in neuromorphic computing hardware have led to significant progress in image classification, speech recognition, and fuzzy computing, outperforming traditional von Neumann computing paradigm. However, the widely-investigated memristor-based neuromorphic computing hardware still suffers high writing/reading currents and serious variability issue as well as sneak path challenges, leading to high power consumption and peripheral circuit design complication. Memcapacitor-based neuromorphic computing is expected to alleviate these problems, while the limited memory windows and endurance hindered the practical applications. Here, we present a hafnium oxide-based ferroelectric memcapacitor developed through work function engineering. The memcapacitor demonstrates an overall excellent performance in memory windows (∼7.8 fF/μm²), endurance (>10⁹ cycles), retention (>10 years), dynamic energy consumption (31 fJ/inference), and near-zero standby static power consumption. The fabricated memcapacitor array shows high linearity and device-to-device variations, and can perform complete multiplication-accumulation (MAC) operation. The constructed artificial neural network (ANN) achieves 96.68 % accuracy on the MNIST data set after 200 epochs. Our findings underscore the potential of ferroelectric memcapacitor device as a robust candidate for high energy-efficiency neuromorphic computing applications in intelligent terminals.