⇦Chinese Physics Letters, 2023, Vol. 40, No. 12, Article code 127202Viewpoint Unleashing the Power of Moiré Materials in Neuromorphic Computing John Paul Strachan* Affiliations Peter Grünberg Institute (PGI-14), Neuromorphic Compute Nodes, Forschungszentrum Jülich, Jülich, Germany and RWTH Aachen University, Aachen, Germany Received 24 November 2023; accepted manuscript online 27 November 2023; published online 22 December 2023 *Corresponding author. Email: j.strachan@fz-juelich.de Citation Text: Strachan J P 2023 Chin. Phys. Lett. 40 127202    Abstract DOI:10.1088/0256-307X/40/12/127202 © 2023 Chinese Physics Society Article Text Reservoir computing has been an intriguing paradigm in the field of artificial intelligence and machine learning that draws inspiration from the complex dynamics of recurrent neural networks found in biological systems. Unlike traditional neural networks, reservoir computing separates the training of a fixed, randomly connected ‘reservoir’ layer from a simpler ‘readout’ layer. This distinctive architecture allows the reservoir to process information in a highly dynamic and nonlinear manner, making it adept at handling temporal data.[1] The reservoir's ability to capture intricate temporal dependencies could be leveraged for tasks such as time series prediction, speech recognition, and pattern classification. In turn, this could apply to diverse domains of applications, including natural language processing, signal processing, and computational neuroscience. To implement reservoir computing, several distinct types of emerging devices have been proposed before, including memristive,[2] optical,[3] magnetic,[4] and ferroelectric.[5] However, the reservoir node and the weight of readout layer in the reservoir computing system have not been implemented by the same device while also supporting tunable and reproducible dynamics. Reproducible dynamics, in particular, can be a challenge to achieve in ionic-based memristive systems. Now writing in Chinese Physics Letters,[6] Shi-Jun Liang and Feng Miao from Nanjing University, in collaboration with Bin Cheng from Nanjing University of Science and Technology, reported the Moiré synaptic transistor. This novel synaptic transistor has been constructed from an h-BN/bilayer graphene/h-BN heterostructure. By leveraging the unique properties of the Moiré quantum materials, the research team has created a versatile building block for reservoir computing systems and demonstrated the reservoir node and the weight readout layer can be implemented by a single device. Thus, the Moiré synaptic transistor reported in this work represents a significant step forward for neuromorphic computing, which endeavors to mimic functions observed in biological brains. One of the notable achievements in this study lies in the transistor's ability to faithfully emulate the dynamic behavior of biological synapses. Through a careful manipulation of stimuli, the researchers were able to induce both short-term plasticity (STP) and long-term (LTP) potentiation, i.e., key elements in the learning and memory processes of biological systems. This versatility enables the Moiré synaptic transistor to serve in both the reservoir and readout layers in a reservoir computing architecture. Another important innovation of this work is to show that the nuanced dynamics of the Moiré synaptic transistor can be used for the mixture sampling technique. This operates on the principle of selectively recording current states at specific timeframes during input processing, enabling the extraction of diverse temporal features with good precision. By integrating this into a full-Moiré physical neural network (MPNN), the researchers demonstrated improved success in classifying handwritten digits from the MNIST dataset. With a compact proof-of-concept network, the authors could yield a recognition accuracy of 90.8%, showcasing the promise of this technology. In this sense, the integration of a Moiré artificial synapse for reservoir and other dynamical computing paradigms is an exciting opportunity for high efficiency information processing. The work led by Liang, Cheng and Miao represents an initial foray of Moiré quantum materials into the frontier of neuromorphic computing. The research could pave the way for future opportunities in this field, offering a fresh approach to the use of quantum materials in advanced electronics and computing systems.[7] As we look ahead, the integration of Moiré quantum materials into dynamical and programmable networks of devices offers an extra tool in our toolbox for the long-term vision of more biologically faithful functionalities. There are a wide variety of behaviors and properties observed in intelligent biological systems that are not even remotely present in modern computing architectures. As our community builds up novel computing blocks, the attainment of new capabilities with improved efficiencies is an enticing goal. This study marks a nice milestone in the approach toward such intelligent computing.[8] There remain many opportunities for improvement in the future. One notable issue lies in scaling up the Moiré synaptic transistor to larger integrated arrays by employing CVD-grown large-area graphene and h-BN material. While the device demonstrates remarkable performance on an individual basis, ensuring uniformity and reliability across a broader network remains a crucial area for future research. Of course, reducing the operating voltages, power levels, and time-scales for operations of the Moiré synaptic transistor will be paramount for energy efficiency and compatibility with advanced CMOS nodes. While the device's capabilities are impressive, scaling down the device dimensions will warrant future research and development as well. References Dynamical memristors for higher-complexity neuromorphic computingReservoir computing using dynamic memristors for temporal information processingExperimental demonstration of reservoir computing on a silicon photonics chipNeuromorphic computing with nanoscale spintronic oscillatorsAll-ferroelectric implementation of reservoir computingMoiré Synaptic Transistor for Homogeneous-Architecture Reservoir ComputingMoiré heterostructures: highly tunable platforms for quantum simulation and future computing2D materials for intelligent devices
[1] Kumar S, Wang X, Strachan J P et al. 2022 Nat. Rev. Mater. 7 575
[2] Du C, Cai F, Zidan M A et al. 2017 Nat. Commun. 8 2204
[3] Vandoorne K, Mechet P, van Vaerenbergh T et al. 2014 Nat. Commun. 5 3541
[4] Torrejon J, Riou M, Araujo F A et al. 2017 Nature 547 428
[5] Chen Z W, Li W J, Fan Z et al. 2023 Nat. Commun. 14 3585
[6] Wang P, Chen M, Xie Y, Pan C, Watanabe K, Taniguchi T, Cheng B, Liang S J, and Miao F 2023 Chin. Phys. Lett. 40 117201
[7] Chen M Y, Chen F Q, Cheng B, Liang S J, and Miao F 2023 J. Semicond. 44 010301
[8] Pan X, Li Y, Cheng B et al. 2023 Sci. Chin. Phys. Mech. & Astron. 66 117504