High-precision and linear weight updates by subnanosecond pulses in ferroelectric tunnel junction for neuro-inspired computing

Hf0.5Zr0.5O2-Based Ferroelectric Tunnel Junction as an Artificial Synapse for Speech Recognition

HfO2-based ferroelectric tunnel junctions (FTJs) are currently receiving significant attention in the fields of nonvolatile memory and neuromorphic computing. Here, an FTJ memristor utilizing a Pt/Hf0.5Zr0.5O2/TiO2/TiN architecture with a TiO2 interlayer is prepared, and it exhibits stable resistive switching with an ON/OFF ratio reaching 5.8 × 102 at an operational speed of 50 ns, along with good retention exceeding 105 s (extended to over 10 years by linear extrapolation) at high temperatures up to 160 °C. Additionally, the TiO2 interlayer improves the interface between HZO and TiN, resulting in superior resistive switching endurance of 2 × 108 cycles with an ON/OFF ratio greater than 50, compared to other HfO2-based FTJ devices. As an artificial synapse, the FTJ attains highly symmetric 128-state conductance manipulation with a low cycle-to-cycle variation of 2.75%. When leveraged in a simulated convolutional neural network for speech recognition tasks, the system achieves high accuracy ∼97.6%. Remarkably, even with a signal-to-noise ratio of 10 dB, the recognition accuracy remains at 90.2%, highlighting the remarkable noise immunity. These results underscore the significant potential of FTJs in applications involving multistate nonvolatile memory and artificial synapses, heralding advance in the field of neuromorphic computing and beyond.

Ferroelectric Hf0.5Zr0.5O2 with Enhanced Intermediate Polarization: A Platform for Neuromorphic and Logic-in-Memory Computing

Ferroelectric materials, known for their nonvolatile and reversible polarization states, are emerging as promising candidates for innovative computing paradigms such as neuromorphic computing and logic-in-memory (LiM) architectures. Their polarization dynamics in response to external stimuli closely emulates biological synapses, a feature crucial for learning and adaptation in neural networks. Achieving multiple intermediate states between fully polarized states is critical for energy-efficient computation. However, the exploitation of switching properties for weight modulation in ferroelectric materials remain underexplored. In this study, we demonstrate improved intermediate and cumulative polarization levels in Hf0.5Zr0.5O2 (HZO) through phase engineering. Consequently, HZO-based synaptic ferroelectric field-effect transistors (FeFETs) achieve a wide range of synaptic weights (up to 8 bits) with remarkable linearity, resulting in high classification accuracies of 98% for MNIST and 88% for Fashion-MNIST in neuromorphic computing tasks. Additionally, we present reconfigurable in-memory NOR and NAND logic functions along with 3-bit logic state generation using a multigate FeFET, demonstrating the potential for LiM operations. This work underscores the successful cointegration of neuromorphic and LiM computing functionalities within a unified platform, addressing key challenges in developing efficient and versatile computing architectures. Our findings highlight the potential of HZO to enable next-generation computing systems that seamlessly integrate learning and logic capabilities.

Subnanosecond flash memory enabled by 2D-enhanced hot-carrier injection

The pursuit of non-volatile memory with program speeds below one nanosecond, beyond the capabilities of non-volatile flash and high-speed volatile static random-access memory, remains a longstanding challenge in the field of memory technology1. Utilizing fundamental physics innovation enabled by advanced materials, series of emerging memories2-5 are being developed to overcome the speed bottleneck of non-volatile memory. As the most extensively applied non-volatile memory, the speed of flash is limited by the low efficiency of the electric-field-assisted program, with reported speeds6-10 much slower than sub-one nanosecond. Here we report a two-dimensional Dirac graphene-channel flash memory based on a two-dimensional-enhanced hot-carrier-injection mechanism, supporting both electron and hole injection. The Dirac channel flash shows a program speed of 400 picoseconds, non-volatile storage and robust endurance over 5.5 × 106 cycles. Our results confirm that the thin-body channel can optimize the horizontal electric-field (Ey) distribution, and the improved Ey-assisted program efficiency increases the injection current to 60.4 pA μm-1 at |VDS| = 3.7 V. We also find that the two-dimensional semiconductor tungsten diselenide has two-dimensional-enhanced hot-hole injection, but with different injection behaviour. This work demonstrates that the speed of non-volatile flash memory can exceed that of the fastest volatile static random-access memory with the same channel length.

Ferroelectric tunnel junctions integrated on semiconductors with enhanced fatigue resistance

Oxide-based ferroelectric tunnel junctions (FTJs) show promise for nonvolatile memory and neuromorphic applications, making their integration with existing semiconductor technologies highly desirable. Furthermore, resistance fatigue in current silicon-based integration remains a critical issue. Understanding this fatigue mechanism in semiconductor-integrated FTJ is essential yet unresolved. Here, we systematically investigate the fatigue performance of ultrathin bismuth ferrite BiFeO3 (BFO)-based FTJs integrated with various semiconductors. Notably, the BFO/gallium arsenide FTJ exhibits superior fatigue resistance characteristics (>108 cycles), surpassing the BFO/silicon FTJ (>106 cycles) and even approaching epitaxial oxide FTJs (>109 cycles). The atomic-scale fatigue mechanism is revealed as lattice structure collapse caused by oxygen vacancy accumulation in BFO near semiconductors after repeated switching. The enhanced fatigue-resistant behavior in BFO/gallium arsenide FTJ is due to gallium arsenide's weak oxygen affinity, resulting in fewer oxygen vacancies. These findings provide deeper insights into the atomic-scale fatigue mechanism of semiconductor-integrated FTJs and pave the way for fabricating fatigue-resistant oxide FTJs for practical applications.

Implementation of Multiply Accumulate Operation and Convolutional Neural Network Based on Ferroelectric Tunnel Junction Memristors

In the era of big data, traditional Von Neumann computers suffer from inefficiencies in terms of energy consumption and speed associated with data transfer between storage and processing. In-memory computing using ferroelectric tunnel junction (FTJ) memristors offers a potential solution to this challenge. Here, Hf0.5Zr0.5O2-based FTJs on a silicon substrate are fabricated, which demonstrates 32 conductance states (5-bit), low cycle-to-cycle variation (1.6%) and highly linear (nonlinearity <1) conductance manipulation. Based on an FTJ array with multiple FTJ devices, a custom-designed board with a field programmable gate array is utilized to perform accurate multiply accumulate operations and for image processing as various convolution operators. Notably, using FTJ devices as a convolutional layer, the convolutional neural network achieves a high accuracy of 92.5% for handwritten digit recognition, and exhibits orders of magnitude better energy efficiency compared to traditional CPU and GPU implementations. These findings highlight the promising potential of FTJs for realizing in-memory computing at the hardware level.

Ultrahigh-power-density flexible piezoelectric energy harvester based on freestanding ferroelectric oxide thin films

Flexible piezoelectric nanogenerators are emerging as a promising solution for powering next-generation flexible electronics by converting mechanical energy into electrical energy. However, traditional ferroelectric ceramics, despite their excellent piezoelectric properties, lack flexibility; while piezoelectric polymers, although highly flexible, have low piezoelectricity. The quest to develop materials that combine high piezoelectricity with exceptional flexibility has thus become a research focus. Herein, we present a breakthrough in this field with the fabrication of freestanding (111)-oriented PbZr0.52Ti0.48O3 single crystalline thin films, which exhibit remarkable flexibility and a high converse piezoelectric coefficient (~585 pm/V). This is achieved through water-soluble sacrificial layer to relieve substrate clamping and controlling the crystal orientation to further enhance the piezoelectric response. Our nanogenerators, constructed using these freestanding nanoscale membranes, demonstrate a record-high output power density (~63.5 mW/cm3), excellent flexibility (with a strain tolerance >3.4%), and superior mechanical stability in cycling tests (>60,000 cycles). These advancements pave the way for high-performance, flexible electronic devices utilizing ferroelectric oxide thin films.

In-memory ferroelectric differentiator

Differential calculus is the cornerstone of many disciplines, spanning the breadth of modern mathematics, physics, computer science, and engineering. Its applications are fundamental to theoretical progress and practical solutions. However, the current state of digital differential technology often requires complex implementations, which struggle to meet the extensive demands of the ubiquitous edge computing in the intelligence age. To face these challenges, we propose an in-memory differential computation that capitalizes on the dynamic behavior of ferroelectric domain reversal to efficiently extract information differences. This strategy produces differential information directly within the memory itself, which considerably reduces the volume of data transmission and operational energy consumption. We successfully illustrate the effectiveness of this technique in a variety of tasks, including derivative function solving, the moving object extraction and image discrepancy identification, using an in-memory differentiator constructed with a crossbar array of 1600-unit ferroelectric polymer capacitors. Our research offers an efficient hardware analogue differential computing, which is crucial for accelerating mathematical processing and real-time visual feedback systems.

Ultralow Power Optoelectronic Memtransistors Based on Vertical WS2/In2Se3 van der Waals Heterostructures

Memtransistors composed of 2D van der Waals (vdW) heterostructures are crucial for constructing artificial synaptic devices and realizing neuromorphic computing. The functional integration containing ultralow power, nonvolatile memory, and biomimetic synaptic behavior endows such devices with broad prospects. Here, we develop an optoelectronic memtransistor based on the WS2/In2Se3 vdW heterostructure and realize significant optical and electrical synaptic properties, which can simulate both short-range plasticity (STP) and long-range plasticity (LTP) of biological synapses. Under optical stimulation, the device demonstrates an ultralow power consumption (only 7.7 aJ per spike) significantly lower than biological synapses, indicating the application potential in large-scale neuromorphic hardware. Combining optical and electrical stimuli, we can perform multiple logic operations by controlling the optical and electrical inputs of the WS2/In2Se3-based memtransistor. Besides, simulated recognition utilizing the Modified National Institute of Standards and Technology data set can achieve a recognition accuracy of 85.41%. Notably, this accuracy can remain above 80% even with the introduction of Gaussian noise. These results demonstrate the promising potential of WS2/In2Se3-based memtransistors in future neuromorphic computing.

Weighted Echo State Graph Neural Networks Based on Robust and Epitaxial Film Memristors

Hardware system customized toward the demands of graph neural network learning would promote efficiency and strong temporal processing for graph-structured data. However, most amorphous/polycrystalline oxides-based memristors commonly have unstable conductance regulation due to random growth of conductive filaments. And graph neural networks based on robust and epitaxial film memristors can especially improve energy efficiency due to their high endurance and ultra-low power consumption. Here, robust and epitaxial Gd: HfO2-based film memristors are reported and construct a weighted echo state graph neural network (WESGNN). Benefiting from the optimized epitaxial films, the high switching speed (20 ns), low energy consumption (2.07 fJ), multi-value storage (4 bits), and high endurance (109) outperform most memristors. Notably, thanks to the appropriately dispersed conductance distribution (standard deviation = 7.68 nS), the WESGNN finely regulates the relative weights of input nodes and recursive matrix to realize state-of-the-art performance using the MUTAG and COLLAB datasets for graph classification tasks. Overall, robust and epitaxial film memristors offer nanoscale scalability, high reliability, and low energy consumption, making them energy-efficient hardware solutions for graph learning applications.

In situ training of an in-sensor artificial neural network based on ferroelectric photosensors

In-sensor computing has emerged as an ultrafast and low-power technique for next-generation machine vision. However, in situ training of in-sensor computing systems remains challenging due to the demands for both high-performance devices and efficient programming schemes. Here, we experimentally demonstrate the in situ training of an in-sensor artificial neural network (ANN) based on ferroelectric photosensors (FE-PSs). Our FE-PS exhibits self-powered, fast (<30 μs), and multilevel (>4 bits) photoresponses, as well as long retention (50 days), high endurance (109), high write speed (100 ns), and small cycle-to-cycle and device-to-device variations (~0.66% and ~2.72%, respectively), all of which are desirable for the in situ training. Additionally, a bi-directional closed-loop programming scheme is developed, achieving a precise and efficient weight update for the FE-PS. Using this programming scheme, an in-sensor ANN based on the FE-PSs is trained in situ to recognize traffic signs for commanding a prototype autonomous vehicle. Moreover, this in-sensor ANN operates 50 times faster than a von Neumann machine vision system. This study paves the way for the development of in-sensor computing systems with in situ training capability, which may find applications in new data-streaming machine vision tasks.

Ferroelectric properties of HfAlOx-based ferroelectric memristor devices for neuromorphic applications: Influence of top electrode deposition method

In this study, we compare the performance of ferroelectric memristor devices based on the fabrication method for the top electrode, focusing on atomic layer deposition (ALD) and physical vapor deposition techniques. We investigate the effects of these methods on the formation of the orthorhombic phase (o-phase) in HfAlOx (HAO) ferroelectric films, which is crucial for ferroelectric properties. The devices were fabricated with HAO films doped with 3.4% aluminum, followed by rapid thermal annealing at 700 °C. Our results demonstrate that the atomic layer deposition process forms a TiOxNy capping layer at the interface between the HAO film and the TiN top electrode, which promotes the o-phase formation. This capping layer effect leads to enhanced polarization characteristics, as evidenced by higher remnant polarization and tunneling electroresistance (TER) in the ALD-fabricated devices. The ALD method also results in a better interfacial layer condition, confirmed by a lower interfacial non-ferroelectric capacitance (Ci). Characterization techniques, including transmission electron microscopy, energy dispersive x-ray spectroscopy, and x-ray diffraction. These structural advantages contribute to enhanced electrical performance, demonstrating neuromorphic applications. Here, our study highlights the significant impact of the ALD deposition method on enhancing the ferroelectric properties and overall performance of ferroelectric memristor devices, making it a promising approach for advanced memory and neuromorphic computing applications.

Advancing Interpretable Regression Analysis for Binary Data: A Novel Distributed Algorithm Approach

Sparse data bias, where there is a lack of sufficient cases, is a common problem in data analysis, particularly when studying rare binary outcomes. Although a two-step meta-analysis approach may be used to lessen the bias by combining the summary statistics to increase the number of cases from multiple studies, this method does not completely eliminate bias in effect estimation. In this paper, we propose a one-shot distributed algorithm for estimating relative risk using a modified Poisson regression for binary data, named ODAP-B. We evaluate the performance of our method through both simulation studies and real-world case analyses of postacute sequelae of SARS-CoV-2 infection in children using data from 184 501 children across eight national academic medical centers. Compared with the meta-analysis method, our method provides closer estimates of the relative risk for all outcomes considered including syndromic and systemic outcomes. Our method is communication-efficient and privacy-preserving, requiring only aggregated data to obtain relatively unbiased effect estimates compared with two-step meta-analysis methods. Overall, ODAP-B is an effective distributed learning algorithm for Poisson regression to study rare binary outcomes. The method provides inference on adjusted relative risk with a robust variance estimator.

Reservoir Computing System with Diverse Input Patterns in HfAlO-Based Ferroelectric Memristor

Ferroelectric memristors, particularly those based on hafnia, are gaining attention as potential candidates for neuromorphic computing. These devices offer advantages over perovskite-based ferroelectric memristors owing to their simpler structures, compatibility with complementary metal-oxide semiconductor technology, and low-power consumption characteristics. Additionally, improvements in ferroelectric memristor's performance, such as enhancing tunneling electro resistance (TER) and polarization retention, can be achieved using methods like aluminum doping and insulating film deposition. In this study, we implement a physical reservoir computing (RC) system utilizing the metal-ferroelectric-insulator-semiconductor-structured ferroelectric memristor based on Al-doped HfO2 as an artificial synapse. Specifically, we ensure the universality and diversity of the system by experimentally demonstrating a robust reservoir layer capable of handling various types of input pulses. To utilize the ferroelectric memristor in the reservoir layer of the RC system, we employ partial polarization switching of ferroelectric materials. We measure the retention loss characteristics of the device for pulse amplitude, interval, and width, and quantify the time constant values by fitting them to a stretched exponential function. Additionally, we validate the suitability of the fabricated device as an artificial synapse by mimicking various short-term plasticity functions of biological synapses. Furthermore, we experimentally demonstrate various applications related to learning and memory of the brain, such as image training and Pavlov's experiment, utilizing the short-term memory characteristics of the fabricated device. Lastly, we evaluate the robustness of the RC system under various input conditions by employing the fabricated device as a reservoir layer.

Analog reservoir computing via ferroelectric mixed phase boundary transistors

Analog reservoir computing (ARC) systems have attracted attention owing to their efficiency in processing temporal information. However, the distinct functionalities of the system components pose challenges for hardware implementation. Herein, we report a fully integrated ARC system that leverages material versatility of the ferroelectric-to-mixed phase boundary (MPB) hafnium zirconium oxides integrated onto indium-gallium-zinc oxide thin-film transistors (TFTs). MPB-based TFTs (MPBTFTs) with nonlinear short-term memory characteristics are utilized for physical reservoirs and artificial neuron, while nonvolatile ferroelectric TFTs mimic synaptic behavior for readout networks. Furthermore, double-gate configuration of MPBTFTs enhances reservoir state differentiation and state expansion for physical reservoir and processes both excitatory and inhibitory pulses for neuronal functionality with minimal hardware burden. The seamless integration of ARC components on a single wafer executes complex real-world time-series predictions with a low normalized root mean squared error of 0.28. The material-device co-optimization proposed in this study paves the way for the development of area- and energy-efficient ARC systems.

Ultralow Energy Consumption and Fast Neuromorphic Computing Based on La0.1Bi0.9FeO3 Ferroelectric Tunnel Junctions

Low-power and fast artificial neural network devices represent the direction in developing analogue neural networks. Here, an ultralow power consumption (0.8 fJ) and rapid (100 ns) La0.1Bi0.9FeO3/La0.7Sr0.3MnO3 ferroelectric tunnel junction artificial synapse has been developed to emulate the biological neural networks. The visual memory and forgetting functionalities have been emulated based on long-term potentiation and depression with good linearity. Moreover, with a single device, logical operations of "AND" and "OR" are implemented, and an artificial neural network was constructed with a recognition accuracy of 96%. Especially for noisy data sets, the recognition speed is faster after preprocessing by the device in the present work. This sets the stage for highly reliable and repeatable unsupervised learning.

Ferroelectric materials for neuroinspired computing applications

In recent years, the emergence of numerous applications of artificial intelligence (AI) has sparked a new technological revolution. These applications include facial recognition, autonomous driving, intelligent robotics, and image restoration. However, the data processing and storage procedures in the conventional von Neumann architecture are discrete, which leads to the "memory wall" problem. As a result, such architecture is incompatible with AI requirements for efficient and sustainable processing. Exploring new computing architectures and material bases is therefore imperative. Inspired by neurobiological systems, in-memory and in-sensor computing techniques provide a new means of overcoming the limitations inherent in the von Neumann architecture. The basis of neural morphological computation is a crossbar array of high-density, high-efficiency non-volatile memory devices. Among the numerous candidate memory devices, ferroelectric memory devices with non-volatile polarization states, low power consumption and strong endurance are expected to be ideal candidates for neuromorphic computing. Further research on the complementary metal-oxide-semiconductor (CMOS) compatibility for these devices is underway and has yielded favorable results. Herein, we first introduce the development of ferroelectric materials as well as their mechanisms of polarization reversal and detail the applications of ferroelectric synaptic devices in artificial neural networks. Subsequently, we introduce the latest developments in ferroelectrics-based in-memory and in-sensor computing. Finally, we review recent works on hafnium-based ferroelectric memory devices with CMOS process compatibility and give a perspective for future developments.

HfAlOx-based ferroelectric memristor for nociceptor and synapse functions

Efficient data processing is heavily reliant on prioritizing specific stimuli and categorizing incoming information. Within human biological systems, dorsal root ganglions (particularly nociceptors situated in the skin) perform a pivotal role in detecting external stimuli. These neurons send warnings to our brain, priming it to anticipate potential harm and prevent injury. In this study, we explore the potential of using a ferroelectric memristor device structured as a metal-ferroelectric-insulator-semiconductor as an artificial nociceptor. The aim of this device is to electrically receive external damage and interpret signals of danger. The TiN/HfAlOx (HAO)/HfSiOx (HSO)/n+ Si configuration of this device replicates the key functions of a biological nociceptor. The emulation includes crucial aspects, such as threshold reactivity, relaxation, no adaptation, and sensitization phenomena known as "allodynia" and "hyperalgesia." Moreover, we propose establishing a connection between nociceptors and synapses by training the Hebbian learning rule. This involves exposing the device to injurious stimuli and using this experience to enhance its responsiveness, replicating synaptic plasticity.

New-Generation Ferroelectric AlScN Materials

Ferroelectrics have great potential in the field of nonvolatile memory due to programmable polarization states by external electric field in nonvolatile manner. However, complementary metal oxide semiconductor compatibility and uniformity of ferroelectric performance after size scaling have always been two thorny issues hindering practical application of ferroelectric memory devices. The emerging ferroelectricity of wurtzite structure nitride offers opportunities to circumvent the dilemma. This review covers the mechanism of ferroelectricity and domain dynamics in ferroelectric AlScN films. The performance optimization of AlScN films grown by different techniques is summarized and their applications for memories and emerging in-memory computing are illustrated. Finally, the challenges and perspectives regarding the commercial avenue of ferroelectric AlScN are discussed.

The Investigation of Neuromimetic Dynamics in Ferroelectrics via In Situ TEM

The core task of neuromorphic devices is to effectively simulate the behavior of neurons and synapses. Based on the functionality of ferroelectric domains with the advantages of low power consumption and high-speed response, great progress has been made in realizing neuromimetic behaviors such as ferroelectric synaptic devices. However, the correlation between the ferroelectric domain dynamics and neuromimetic behavior remains unclear. Here, we reveal the correlation between domain/domain wall dynamics and neuromimetic behaviors from a microscopic perspective in real-time by using high temporal and spatial resolution in situ transmission electron microscopy. Furthermore, we propose utilizing ferroelectric microstructures for the simultaneous simulation of neuronal and synaptic plasticity, which is expected to improve the integration and performance of ferroelectric neuromorphic devices. We believe that this work to study neuromimetic behavior from the perspective of domain dynamics is instructive for the development of ferroelectric neuromorphic devices.

Pure ZrO2 Ferroelectric Thin Film for Nonvolatile Memory and Neural Network Computing

The recent discovery of ferroelectricity in pure ZrO2 has drawn much attention, but the information storage and processing performances of ferroelectric ZrO2-based nonvolatile devices remain open for further exploration. Here, a ZrO2 (∼8 nm)-based ferroelectric capacitor using RuO2 oxide electrodes is fabricated, and the ferroelectric orthorhombic phase evolution under electric field cycling is studied. A ferroelectric remnant polarization (2Pr) of >30 μC/cm2, leakage current density of ∼2.79 × 10-8 A/cm2 at 1 MV/cm, and estimated polarization retention of >10 years are achieved. When the ferroelectric capacitor is connected with a transistor, a memory window of ∼0.8 V and eight distinct states can be obtained in such a ferroelectric field-effect transistor (FeFET). Through the conductance manipulation of the FeFET, a high object image recognition accuracy of ∼93.32% is achieved on the basis of the CIFAR-10 dataset in the convolutional neural network (CNN) simulation, which is close to the result of ∼94.20% obtained by floating-point-based CNN software. These results demonstrate the potential of ferroelectric ZrO2 devices for nonvolatile memory and artificial neural network computing.

A Multistate Non-Volatile Photoelectronic Memory Device Based on Ferroelectric Tunnel Junction with Modulable Visible Light Photoresponse

Recently, certain ferroelectric tunnel junctions (FTJs) exhibit non-volatile modulations on photoresponse as well as tunneling electroresistance (TER) effects related to ferroelectric polarization states. From the opposite perspective, the corresponding polarization states can be read by detecting the levels of the photocurrent. In this study, we fabricate a novel amorphous selenium (a-Se)/PbZr0.2Ti0.8O3 (PZT)/Nb-doped SrTiO3 (NSTO) heterojunction, which exhibits a high TER of 3 × 106. Unlike perovskite oxide FTJs with a limited ultraviolet response, the introduction of a narrow bandgap semiconductor (a-Se) enables self-powered photoresponse within the visible light range. The self-powered photoresponse characteristics can be significantly modulated by ferroelectric polarization. The photocurrent after writing polarization voltages of +4 and -5 V exhibits a 1200% increase. Furthermore, the photocurrent could be clearly distinguished after writing stepwise polarization voltages, and then a multistate information storage is designed with nondestructive readout capacity under light illumination. This work holds great significance in advancing the development of ferroelectric multistate photoelectronic memories with high storage density and expanding the design possibilities for FTJs.

A comprehensive review of advanced trends: from artificial synapses to neuromorphic systems with consideration of non-ideal effects

A neuromorphic system is composed of hardware-based artificial neurons and synaptic devices, designed to improve the efficiency of neural computations inspired by energy-efficient and parallel operations of the biological nervous system. A synaptic device-based array can compute vector-matrix multiplication (VMM) with given input voltage signals, as a non-volatile memory device stores the weight information of the neural network in the form of conductance or capacitance. However, unlike software-based neural networks, the neuromorphic system unavoidably exhibits non-ideal characteristics that can have an adverse impact on overall system performance. In this study, the characteristics required for synaptic devices and their importance are discussed, depending on the targeted application. We categorize synaptic devices into two types: conductance-based and capacitance-based, and thoroughly explore the operations and characteristics of each device. The array structure according to the device structure and the VMM operation mechanism of each structure are analyzed, including recent advances in array-level implementation of synaptic devices. Furthermore, we reviewed studies to minimize the effect of hardware non-idealities, which degrades the performance of hardware neural networks. These studies introduce techniques in hardware and signal engineering, as well as software-hardware co-optimization, to address these non-idealities through compensation approaches.

Silicon-Compatible Ferroelectric Tunnel Junctions with a SiO2/Hf0.5Zr0.5O2 Composite Barrier as Low-Voltage and Ultra-High-Speed Memristors

The big data era requires ultrafast, low-power, and silicon-compatible materials and devices for information storage and processing. Here, ferroelectric tunnel junctions (FTJs) based on SiO2/Hf0.5Zr0.5O2 composite barrier and both conducting electrodes are designed and fabricated on Si substrates. The FTJ achieves the fastest write speed of 500 ps under 5 V (2 orders of magnitude faster than reported silicon-compatible FTJs) or 10 ns speed at a low voltage of 1.5 V (the lowest voltage among FTJs at similar speeds), low write current density of 1.3 × 104 A cm-2, 8 discrete states, good retention > 105 s at 85 °C, and endurance > 107. In addition, it provides a large read current (88 A cm-2) at 0.1 V, 2 orders of magnitude larger than reported FTJs. Interestingly, in FTJ-based synapses, gradually tunable conductance states (128 states) with high linearity (<1) are obtained by 10 ns pulses of <1.2 V, and a high accuracy of 91.8% in recognizing fashion product images is achieved by online neural network simulations. These results highlight that silicon-compatible HfO2-based FTJs are promising for high-performance nonvolatile memories and electrical synapses.

Realization of sextuple polarization states and interstate switching in antiferroelectric CuInP2S6

Realization of higher-order multistates with mutual interstate switching in ferroelectric materials is a perpetual drive for high-density storage devices and beyond-Moore technologies. Here we demonstrate experimentally that antiferroelectric van der Waals CuInP2S6 films can be controllably stabilized into double, quadruple, and sextuple polarization states, and a system harboring polarization order of six is also reversibly tunable into order of four or two. Furthermore, for a given polarization order, mutual interstate switching can be achieved via moderate electric field modulation. First-principles studies of CuInP2S6 multilayers help to reveal that the double, quadruple, and sextuple states are attributable to the existence of respective single, double, and triple ferroelectric domains with antiferroelectric interdomain coupling and Cu ion migration. These findings offer appealing platforms for developing multistate ferroelectric devices, while the underlining mechanism is transformative to other non-volatile material systems.

Physics, Structures, and Applications of Fluorite-Structured Ferroelectric Tunnel Junctions

The interest in ferroelectric tunnel junctions (FTJ) has been revitalized by the discovery of ferroelectricity in fluorite-structured oxides such as HfO2 and ZrO2 . In terms of thickness scaling, CMOS compatibility, and 3D integration, these fluorite-structured FTJs provide a number of benefits over conventional perovskite-based FTJs. Here, recent developments involving all FTJ devices with fluorite structures are examined. The transport mechanism of fluorite-structured FTJs is explored and contrasted with perovskite-based FTJs and other 2-terminal resistive switching devices starting with the operation principle and essential parameters of the tunneling electroresistance effect. The applications of FTJs, such as neuromorphic devices, logic-in-memory, and physically unclonable function, are then discussed, along with several structural approaches to fluorite-structure FTJs. Finally, the materials and device integration difficulties related to fluorite-structure FTJ devices are reviewed. The purpose of this review is to outline the theories, physics, fabrication processes, applications, and current difficulties associated with fluorite-structure FTJs while also describing potential future possibilities for optimization.

Giant tunnelling electroresistance in atomic-scale ferroelectric tunnel junctions

Ferroelectric tunnel junctions are promising towards high-reliability and low-power non-volatile memories and computing devices. Yet it is challenging to maintain a high tunnelling electroresistance when the ferroelectric layer is thinned down towards atomic scale because of the ferroelectric structural instability and large depolarization field. Here we report ferroelectric tunnel junctions based on samarium-substituted layered bismuth oxide, which can maintain tunnelling electroresistance of 7 × 105 with the samarium-substituted bismuth oxide film down to one nanometer, three orders of magnitude higher than previous reports with such thickness, owing to efficient barrier modulation by the large ferroelectric polarization. These ferroelectric tunnel junctions demonstrate up to 32 resistance states without any write-verify technique, high endurance (over 5 × 109), high linearity of conductance modulation, and long retention time (10 years). Furthermore, tunnelling electroresistance over 109 is achieved in ferroelectric tunnel junctions with 4.6-nanometer samarium-substituted bismuth oxide layer, which is higher than commercial flash memories. The results show high potential towards multi-level and reliable non-volatile memories.

Giant Electroresistance in Ferroelectric Tunnel Junctions via High-Throughput Designs: Toward High-Performance Neuromorphic Computing

Ferroelectric tunnel junctions (FTJs) have been regarded as one of the most promising candidates for next-generation devices for data storage and neuromorphic computing owing to their advantages such as fast operation speed, low energy consumption, convenient 3D stack ability, etc. Here, dramatically different from the conventional engineering approaches, we have developed a tunnel barrier decoration strategy to improve the ON/OFF ratio, where the ultrathin SrTiO3 (STO) dielectric layers are periodically mounted onto the BaTiO3 (BTO) ferroelectric tunnel layer using the high-throughput technique. The inserted STO enhances the local tetragonality of the BTO, resulting in a strengthened ferroelectricity in the tunnel layer, which greatly improves the OFF state and reduces the ON state. Combined with the optimized oxygen migration, which can further manipulate the tunneling barrier, a record-high ON/OFF ratio of ∼108 has been achieved. Furthermore, utilizing these FTJ-based artificial synapses, an artificial neural network has been simulated via back-propagation algorithms, and a classification accuracy as high as 92% has been achieved. This study screens out the prominent FTJ by the high-throughput technique, advancing the tunnel layer decoration at the atomic level in the FTJ design and offering a fundamental understanding of the multimechanisms in the tunnel barrier.

Polarized Tunneling Transistor for Ultralow-Energy-Consumption Artificial Synapse toward Neuromorphic Computing

Neural networks based on low-power artificial synapses can significantly reduce energy consumption, which is of great importance in today's era of artificial intelligence. Two-dimensional (2D) material-based floating-gate transistors (FGTs) have emerged as compelling candidates for simulating artificial synapses owing to their multilevel and nonvolatile data storage capabilities. However, the low erasing/programming speed of FGTs renders them unsuitable for low-energy-consumption artificial synapses, thereby limiting their potential in high-energy-efficient neuromorphic computing. Here, we introduce a FGT-inspired MoS2/Trap/PZT heterostructure-based polarized tunneling transistor (PTT) with a simple fabrication process and significantly enhanced erasing/programming speed. Distinct from the FGT, the PTT lacks a tunnel layer, leading to a marked improvement in its erasing/programming speed. The PTT's highest erasing/programming (operation) speed can reach ∼20 ns, which outperforms the performance of most FGTs based on 2D heterostructures. Furthermore, the PTT has been utilized as an artificial synapse, and its weight-update energy consumption can be as low as 0.0002 femtojoule (fJ), which benefits from the PTT's ultrahigh operation speed. Additionally, PTT-based artificial synapses have been employed in constructing artificial neural network simulations, achieving facial-recognition accuracy (95%). This groundbreaking work makes it possible for fabricating future high-energy-efficient neuromorphic transistors utilizing 2D materials.

Filament-free memristors for computing

Memristors have attracted increasing attention due to their tremendous potential to accelerate data-centric computing systems. The dynamic reconfiguration of memristive devices in response to external electrical stimuli can provide highly desirable novel functionalities for computing applications when compared with conventional complementary-metal-oxide-semiconductor (CMOS)-based devices. Those most intensively studied and extensively reviewed memristors in the literature so far have been filamentary type memristors, which typically exhibit a relatively large variability from device to device and from switching cycle to cycle. On the other hand, filament-free switching memristors have shown a better uniformity and attractive dynamical properties, which can enable a variety of new computing paradigms but have rarely been reviewed. In this article, a wide range of filament-free switching memristors and their corresponding computing applications are reviewed. Various junction structures, switching properties, and switching principles of filament-free memristors are surveyed and discussed. Furthermore, we introduce recent advances in different computing schemes and their demonstrations based on non-filamentary memristors. This Review aims to present valuable insights and guidelines regarding the key computational primitives and implementations enabled by these filament-free switching memristors.

Artificial Neuron Devices

Efforts to design devices emulating complex cognitive abilities and response processes of biological systems have long been a coveted goal. Recent advancements in flexible electronics, mirroring human tissue's mechanical properties, hold significant promise. Artificial neuron devices, hinging on flexible artificial synapses, bioinspired sensors, and actuators, are meticulously engineered to mimic the biological systems. However, this field is in its infancy, requiring substantial groundwork to achieve autonomous systems with intelligent feedback, adaptability, and tangible problem-solving capabilities. This review provides a comprehensive overview of recent advancements in artificial neuron devices. It starts with fundamental principles of artificial synaptic devices and explores artificial sensory systems, integrating artificial synapses and bioinspired sensors to replicate all five human senses. A systematic presentation of artificial nervous systems follows, designed to emulate fundamental human nervous system functions. The review also discusses potential applications and outlines existing challenges, offering insights into future prospects. We aim for this review to illuminate the burgeoning field of artificial neuron devices, inspiring further innovation in this captivating area of research.

Reconfigurable Neuromorphic Computing: Materials, Devices, and Integration

Neuromorphic computing has been attracting ever-increasing attention due to superior energy efficiency, with great promise to promote the next wave of artificial general intelligence in the post-Moore era. Current approaches are, however, broadly designed for stationary and unitary assignments, thus encountering reluctant interconnections, power consumption, and data-intensive computing in that domain. Reconfigurable neuromorphic computing, an on-demand paradigm inspired by the inherent programmability of brain, can maximally reallocate finite resources to perform the proliferation of reproducibly brain-inspired functions, highlighting a disruptive framework for bridging the gap between different primitives. Although relevant research has flourished in diverse materials and devices with novel mechanisms and architectures, a precise overview remains blank and urgently desirable. Herein, the recent strides along this pursuit are systematically reviewed from material, device, and integration perspectives. At the material and device level, one comprehensively conclude the dominant mechanisms for reconfigurability, categorized into ion migration, carrier migration, phase transition, spintronics, and photonics. Integration-level developments for reconfigurable neuromorphic computing are also exhibited. Finally, a perspective on the future challenges for reconfigurable neuromorphic computing is discussed, definitely expanding its horizon for scientific communities.

Gradual conductance modulation by defect reorganization in amorphous oxide memristors

Amorphous oxides show great prospects in revolutionizing memristors benefiting from their abundant non-stoichiometric composition. However, an in-depth investigation of the memristive characteristics in amorphous oxides is inadequate and the resistive switching mechanism is still controversial. In this study, aiming to clearly understand the gradual conductance modulation that is deeply bound to the evolution of defects-mainly oxygen vacancies, forming-free memristors based on amorphous ZnAlSnO are fabricated, which exhibit high reproducibility with an initial low-resistance state. Pulse depression reveals the logarithmic-exponential mixed relaxation during RESET owing to the diffusion of oxygen vacancies in orthogonal directions. The remnants of conductive filaments formed through aggregation of oxygen vacancies induced by high-electric-field are identified using ex situ TEM. Especially, the conductance of the filament, including the remnant filament, is larger than that of the hopping conductive channel derived from the diffusion of oxygen vacancies. The Fermi level in the conduction band rationalizes the decay of the high resistance state. Rare oxidation-migration of Au occurs upon device failure, resulting in numerous gold nanoclusters in the functional layer. These comprehensive revelations on the reorganization of oxygen vacancies could provide original ideas for the design of memristors.

PZT-Enabled MoS2 Floating Gate Transistors: Overcoming Boltzmann Tyranny and Achieving Ultralow Energy Consumption for High-Accuracy Neuromorphic Computing

Low-power electronic devices play a pivotal role in the burgeoning artificial intelligence era. The study of such devices encompasses low-subthreshold swing (SS) transistors and neuromorphic devices. However, conventional field-effect transistors (FETs) face the inherent limitation of the "Boltzmann tyranny", which restricts SS to 60 mV decade-1 at room temperature. Additionally, FET-based neuromorphic devices lack sufficient conductance states for highly accurate neuromorphic computing due to a narrow memory window. In this study, we propose a pioneering PZT-enabled MoS2 floating gate transistor (PFGT) configuration, demonstrating a low SS of 46 mV decade-1 and a wide memory window of 7.2 V in the dual-sweeping gate voltage range from -7 to 7 V. The wide memory window provides 112 distinct conductance states for PFGT. Moreover, the PFGT-based artificial neural network achieves an outstanding facial-recognition accuracy of 97.3%. This study lays the groundwork for the development of low-SS transistors and highly energy efficient artificial synapses utilizing two-dimensional materials.

High-Performance Neuromorphic Computing and Logic Operation Based on a Self-Assembled Vertically Aligned Nanocomposite SrTiO3:MgO Film Memristor

Neuromorphic computing based on memristors capable of in-memory computing is promising to break the energy and efficiency bottleneck of well-known von Neumann architectures. However, unstable and nonlinear conductance updates compromise the recognition accuracy and block the integration of neural network hardware. To this end, we present a highly stable memristor with self-assembled vertically aligned nanocomposite (VAN) SrTiO3:MgO films that achieve excellent resistive switching with low set/reset voltage variability (4.7%/-5.6%) and highly linear conductivity variation (nonlinearity = 0.34) by spatially limiting the conductive channels at the vertical interfaces. Various synaptic behaviors are simulated by continuously modulating the conductance. Especially, convolutional image processing using diverse crossbar kernels is demonstrated, and the artificial neural network achieves an overwhelming recognition accuracy of up to 97.50% for handwritten digits. Even under the perturbation of Poisson noise (λ = 10), 6% Salt and Pepper noise, and 5% Gaussian noise, the high recognition accuracies are retained at 95.43%, 94.56%, and 95.97%, respectively. Importantly, the logic memory function is proven experimentally based on the nonvolatile properties. This work provides a material system and design idea to achieve high-performance neuromorphic computing and logic operation.

All-ferroelectric implementation of reservoir computing

Reservoir computing (RC) offers efficient temporal information processing with low training cost. All-ferroelectric implementation of RC is appealing because it can fully exploit the merits of ferroelectric memristors (e.g., good controllability); however, this has been undemonstrated due to the challenge of developing ferroelectric memristors with distinctly different switching characteristics specific to the reservoir and readout network. Here, we experimentally demonstrate an all-ferroelectric RC system whose reservoir and readout network are implemented with volatile and nonvolatile ferroelectric diodes (FDs), respectively. The volatile and nonvolatile FDs are derived from the same Pt/BiFeO₃/SrRuO₃ structure via the manipulation of an imprint field (E_imp). It is shown that the volatile FD with E_imp exhibits short-term memory and nonlinearity while the nonvolatile FD with negligible E_imp displays long-term potentiation/depression, fulfilling the functional requirements of the reservoir and readout network, respectively. Hence, the all-ferroelectric RC system is competent for handling various temporal tasks. In particular, it achieves an ultralow normalized root mean square error of 0.017 in the Hénon map time-series prediction. Besides, both the volatile and nonvolatile FDs demonstrate long-term stability in ambient air, high endurance, and low power consumption, promising the all-ferroelectric RC system as a reliable and low-power neuromorphic hardware for temporal information processing.

Ferroelectric Transistors for Memory and Neuromorphic Device Applications

Ferroelectric materials have been intensively investigated for high-performance nonvolatile memory devices in the past decades, owing to their nonvolatile polarization characteristics. Ferroelectric memory devices are expected to exhibit lower power consumption and higher speed than conventional memory devices. However, non-complementary metal-oxide-semiconductor (CMOS) compatibility and degradation due to fatigue of traditional perovskite-based ferroelectric materials have hindered the development of high-density and high-performance ferroelectric memories in the past. The recently developed hafnia-based ferroelectric materials have attracted immense attention in the development of advanced semiconductor devices. Because hafnia is typically used in CMOS processes, it can be directly incorporated into current semiconductor technologies. Additionally, hafnia-based ferroelectrics show high scalability and large coercive fields that are advantageous for high-density memory devices. This review summarizes the recent developments in ferroelectric devices, especially ferroelectric transistors, for next-generation memory and neuromorphic applications. First, the types of ferroelectric memories and their operation mechanisms are reviewed. Then, issues limiting the realization of high-performance ferroelectric transistors and possible solutions are discussed. The experimental demonstration of ferroelectric transistor arrays, including 3D ferroelectric NAND and its operation characteristics, are also reviewed. Finally, challenges and strategies toward the development of next-generation memory and neuromorphic applications based on ferroelectric transistors are outlined.

An Optoelectronic Synapse Based on Two-Dimensional Violet Phosphorus Heterostructure

Neuromorphic computing can efficiently handle data-intensive tasks and address the redundant interaction required by von Neumann architectures. Synaptic devices are essential components for neuromorphic computation. 2D phosphorene, such as violet phosphorene, show great potential in optoelectronics due to their strong light-matter interactions, while current research is mainly focused on synthesis and characterization, its application in photoelectric devices is vacant. Here, the authors combined violet phosphorene and molybdenum disulfide to demonstrate an optoelectronic synapse with a light-to-dark ratio of 10⁶ , benefiting from a significant threshold shift due to charge transfer and trapping in the heterostructure. Remarkable synaptic properties are demonstrated, including a dynamic range (DR) of > 60 dB, 128 (7-bit) distinguishable conductance states, electro-optical dependent plasticity, short-term paired-pulse facilitation, and long-term potentiation/depression. Thanks to the excellent DR and multi-states, high-precision image classification with accuracies of 95.23% and 79.65% is achieved for the MNIST and complex Fashion-MNIST datasets, which is close to the ideal device (95.47%, 79.95%). This work opens the way for the use of emerging phosphorene in optoelectronics and provides a new strategy for building synaptic devices for high-precision neuromorphic computing.

Realizing multiple non-volatile resistance states in a two-dimensional domain wall ferroelectric tunneling junction

Two-dimensional ferroelectric tunnel junctions (2D FTJs) with an ultrathin van der Waals ferroelectrics sandwiched by two electrodes have great applications in memory and synaptic devices. Domain walls (DWs), formed naturally in ferroelectrics, are being actively explored for their low energy consumption, reconfigurable, and non-volatile multi-resistance characteristics in memory, logic and neuromorphic devices. However, DWs with multiple resistance states in 2D FTJ have rarely been explored and reported. Here, we propose the formation of 2D FTJ with multiple non-volatile resistance states manipulated by neutral DWs in a nanostripe-ordered β'-In₂Se₃ monolayer. By combining density functional theory (DFT) calculations with nonequilibrium Green's function method, we found that a large TER ratio can be obtained due to the blocking effect of DWs on the electronic transmission. Multiple conductance states are readily obtained by introducing different numbers of the DWs. This work opens a new route to designing multiple non-volatile resistance states in 2D DW-FTJ.

Ultrathin Nitride Ferroic Memory with Large ON/OFF Ratios for Analog In-Memory Computing

Computing in the analog regime using nonlinear ferroelectric resistive memory arrays can potentially alleviate the energy constraints and complexity/footprint challenges imposed by digital von Neumann systems. Yet the current ferroelectric resistive memories suffer from either low ON/OFF ratios/imprint or limited compatibility with mainstream semiconductors. Here, for the first time, ferroelectric and analog resistive switching in an epitaxial nitride heterojunction comprising ultrathin (≈5 nm) nitride ferroelectrics, i.e., ScAlN, with potentiality to bridge the gap between performance and compatibility is demonstrated. High ON/OFF ratios (up to 10⁵ ), high uniformity, good retention, (<20% variation after > 10⁵  s) and cycling endurance (>10⁴ ) are simultaneously demonstrated in a metal/oxide/nitride ferroelectric junction. It is further demonstrated that the memristor can provide programmability to enable multistate operation and linear analogue computing as well as image processing with high accuracy. Neural network simulations based on the weight update characteristics of the nitride memory yielded an image recognition accuracy of 92.9% (baseline 96.2%) on the images from Modified National Institute of Standards and Technology. The non-volatile multi-level programmability and analog computing capability provide first-hand and landmark evidence for constructing advanced memory/computing architectures based on emerging nitride ferroelectrics, and promote homo and hybrid integrated functional edge devices beyond silicon.

A Gate Programmable van der Waals Metal-Ferroelectric-Semiconductor Vertical Heterojunction Memory

Ferroelectricity, one of the keys to realize non-volatile memories owing to the remanent electric polarization, is an emerging phenomenon in the 2D limit. Yet the demonstrations of van der Waals (vdW) memories using 2D ferroelectric materials as an ingredient are very limited. Especially, gate-tunable ferroelectric vdW memristive device, which holds promises in future multi-bit data storage applications, remains challenging. Here, a gate-programmable multi-state memory is shown by vertically assembling graphite, CuInP₂ S₆ , and MoS₂ layers into a metal(M)-ferroelectric(FE)-semiconductor(S) architecture. The resulted devices seamlessly integrate the functionality of both FE-memristor (with ON-OFF ratios exceeding 10⁵ and long-term retention) and metal-oxide-semiconductor field effect transistor (MOS-FET). Thus, it yields a prototype of gate tunable giant electroresistance with multi-levelled ON-states in the FE-memristor in the vertical vdW assembly. First-principles calculations further reveal that such behaviors originate from the specific band alignment between the FE-S interface. Our findings pave the way for the engineering of ferroelectricity-mediated memories in future implementations of 2D nanoelectronics.

A self-assembly growth strategy for a highly ordered ferroelectric nanoisland array

Ferroelectric nanoislands have attracted intensive research interest due to their size effect induced exotic physical properties and potential applications in non-volatile ferroelectric memories. However, the self-assembly growth of highly ordered ferroelectric nanoisland arrays is still a challenge. Here, by patterning a LaAlO₃ substrate with etched nanocavities to provide preferential nucleation sites, highly ordered self-assembled BiFeO₃ nanoisland arrays with robust ferroelectric topological quad-domain configurations were achieved. From the thermodynamic and kinetic perspectives, three factors are critical for achieving highly ordered self-assembled nanoisland arrays, that is, preferential nucleation sites, an appropriate relationship between the surface energy and the interface energy, and the growth rate difference of films. This approach can also be employed for the self-assembly growth of nanoisland arrays in other ferroelectric materials, which facilitates the design of ferroelectric nanostructure-based nanodevices.

Emerging Memristive Devices for Brain-Inspired Computing and Artificial Perception

Brain-inspired computing is an emerging field that aims at building a compact and massively parallel architecture, to reduce power consumption in conventional Von Neumann Architecture. Recently, memristive devices have gained great attention due to their immense potential in implementing brain-inspired computing and perception. The conductance of a memristor can be modulated by a voltage pulse, enabling emulations of both essential synaptic and neuronal functions, which are considered as the important building blocks for artificial neural networks. As a result, it is critical to review recent developments of memristive devices in terms of neuromorphic computing and perception applications, waiting for new thoughts and breakthroughs. The device structures, operation mechanisms, and materials are introduced sequentially in this review; additionally, late advances in emergent neuromorphic computing and perception based on memristive devices are summed up. Finally, the challenges that memristive devices toward high-performance brain-inspired computing and perception are also briefly discussed. We believe that the advances and challenges will lead to significant advancements in artificial neural networks and intelligent humanoid robots.

High-Speed Nanoscale Ferroelectric Tunnel Junction for Multilevel Memory and Neural Network Computing

Ferroelectric tunnel junction (FTJ) is one promising candidate for next-generation nonvolatile data storage and neural network computing systems. In this work, the high-performance 50 nm-diameter Au/Ti/PbZr_0.52Ti_0.48O₃ (∼3 nm, (111)-oriented)/Nb:SrTiO₃ (Nb: 0.7 wt %) FTJs are achieved to demonstrate the scaling down capability of FTJ. As a nonvolatile memory, the FTJ shows eight distinct resistance states (3 bits) with a large ON/OFF ratio (>10³), and these states can be switched at a fast speed of 10 ns. Intriguingly, the long-term potentiation/depression and spike timing-dependent plasticity, that is, fundamental functions of biological synapses, can be emulated in the nanoscale FTJ-based artificial synapse. A convolutional neural network (CNN) simulation is then carried out based on the experimental results, and a high recognition accuracy of ∼93.8% on fashion product images is obtained, which is very close to the result of ∼94.4% by a floating-point-based CNN software. In particular, the FTJ-based CNN simulation also exhibits robustness to input image noises. These results indicate the great potential of FTJ for high-density information storage and neural network computing.