Atomic Nb-doping of WS2 for high-performance synaptic transistors in neuromorphic computing

Neuromorphic computing, an innovative technology inspired by the human brain, has attracted increasing attention as a promising technology for the development of artificial intelligence systems. This study proposes synaptic transistors with a Li1-xAlxTi2-x(PO4)3 (LATP) layer to analyze the conductance modulation linearity, which is essential for weight mapping and updating during on-chip learning processes. The high ionic conductivity of the LATP electrolyte provides a large hysteresis window and enables linear weight update in synaptic devices. The results demonstrate that optimizing the LATP layer thickness improves the conductance modulation and linearity of synaptic transistors during potentiation and degradation. A 20 nm-thick LATP layer results in the most nonlinear depression (αd = -6.59), whereas a 100 nm-thick LATP layer results in the smallest nonlinearity (αd = -2.22). Additionally, a device with the optimal 100 nm-thick LATP layer exhibits the highest average recognition accuracy of 94.8% and the smallest fluctuation, indicating that the linearity characteristics of a device play a crucial role in weight update during learning and can significantly affect the recognition accuracy.

Metal-oxide semiconductor (MOS)-based synaptic transistors are promising candidates for highly integrated neuromorphic chips. Ferroelectrics and electrolytes have been extensively studied, but they have not satisfied the scalability and integration of the chips. We report a synaptic thin-film transistor (TFT) based on polycrystalline InGaO (C-IGO) semiconductor, exhibiting a large counterclockwise hysteresis with a memory window (MW) ratio of 57%, attributed to positively charged oxygen vacancy (Vo+/++) migration within the C-IGO layer. The C-IGO TFT could be integrated with the back-end-of-line (BEOL) process. The scalable process could be achieved with self-aligned (SA) coplanar C-IGO TFT with conventional gate insulator (GI) materials. The C-IGO synaptic TFT exhibits high synaptic performance, including long-term potentiation/depression (LTP/LTD) characteristics. Modified National Institute of Standards and Technology (MNIST) handwritten digit image recognition simulation based on potentiation/depression curves yields high recognition accuracy (91.64%). The results validate the potential for high-density, neuromorphic, and artificial intelligence (AI) applications.

Nanoscale channel organic ferroelectric synaptic transistor array for high recognition accuracy neuromorphic computing

As layered materials beyond graphene discovered after 2004, two-dimensional (2D) transition metal dichalcogenides (TMDs) are vital to fundamental research and practical applications, owing to their unique crystal structures and excellent properties, as well as diversity of the electronic band structures. 2D TMDs have played an important role in electronics, optoelectronics, energy storage and catalysis. To meet with the increasing requirements of programmable and function-integrated devices, property modulation could be concerned as one of the most essential strategies for 2D TMDs. In comparison with conventional external electrical field induction, strain regulation could be much more efficient. In detail, external induction through the electrical field leads to the electron delocalization along the field direction and then induces the transformation of band structures, but electrical field exhibits small modulation for monolayer TMDs. On the contrary, the strain regulation shows excellent tuning efficiency for continuous and reversible modulation. As a result, the strain regulation has become a commonly-used strategy for property tuning of 2D TMDs. On the basis of the lattice transformation, the strain regulation of 2D TMDs can lead to different overlapping of d and p orbits of metal and chalcogen atoms, and then affect the electronic structures of 2D TMDs. Therefore, it can be applied to electronics, optoelectronics, magnetic devices and piezoelectronics. The strategies of introducing strains to 2D TMDs are classified into lattice induction, local deformation, macroscopic regulation and so on. Lattice induction is attributable to the structural distortions and mismatches, including atomic defect induction and lattice mismatch induction. The former one demonstrates that the microenvironment affected by atomic vacancies and doping atoms can introduce strain to 2D TMDs. The latter one means that the lattice mismatches between two materials (between two different TMDs in a heterostructure or between a TMD and the substrate) can result in lattice distortion and then induce the strain. However, strain introduced through lattice induction is fixed and difficult to achieve the reversible modulation. Local deformation refers to the morphological transformation at the scale of several micrometers, which can be induced by the bubbles and wrinkles of 2D TMDs, external forces of tips, as well as patterned substrates. Generally, large but nonuniform strain can be introduced to 2D TMDs through the local deformation, which results in the funnel effects and then induces large property variation. Furthermore, macroscopic regulation introduces the strains to the lattices, including the deformation of flexible substrates (bend, tension and compression), external pressure (applied by the diamond anvil cell) and thermal expansion coefficient mismatch. Macroscopic regulation would be compatible with industrial manufacture to achieve strain regulation on an extremely large scale in the future. There are some other ways to achieve strain regulation such as the design of special stack structures and the induction of the external electrical field. After summarizing the methods of introducing strain to 2D TMDs, we presented the applications based on strain regulation, such as field effect transistors, flexible photodetectors and strain sensors. Finally, we pointed out the further development and challenges of strain regulation of 2D TMDs.

Non-Von Neumann computing application constituted by artificial synapses based on electrochemical random-access memory (ECRAM) has aroused tremendous attention owing to its capability to perform parallel operations, thus reducing the cost of time and energy spent [1-3]. Existing ECRAM synapses comprise two-terminal memristors and three-terminal synaptic transistors (SynT). While low cost, scalability, and high density are the highlights for memristors, their nonlinear, asymmetric state modulation, high ON current withdrawal, and sneak path in crossbar array integration prevent them from becoming the ideal synaptic elements for artificial neural networks (ANN) [4]. SynT configuration, on the other hand, offers an additional electrolyte-gated control from which ion doping content can be monitored via redox reactions, thus decoupling write-read actions and improving the linearity of programming states [5-6]. Nevertheless, existing SynTs suffer from different integration issues stemming from liquid-based ionic conductors and manually exfoliated channels. Moreover, several kinds of SynTs possess highly conductive channels in the range of µS to mS, significantly scaling up the energy spent for analog states reading. Despite having numerous communications on the performance of different ECRAM, a comprehensive electrochemical view of ion intercalation into the active material, the main root of conductance modulation, is clearly missing.In this work, we present the elaboration procedure of an all-solid-state synaptic transistor composed of nanoscale electrolyte and channel layers. The devices have been elaborated on 8’’ Silicon wafers using microfabrication processes compatible with conventional semiconductor technology and CMOS back end of line (BEoL) integration. (Figure 1a)We demonstrate the excellent synaptic plasticity properties of short-term potentiation (STP) and long-term potentiation (LTP) of our SynT. We performed tests to study the correlation between linearity, asymmetry, and the number of analog states. By averaging the amount of injected ions per write operation, we estimated the energy consumed for switching among adjacent states of this device is 22.5 pJ, yielding area-normalized energy of 4 fJ/µm2. In addition, operating in the range of nS, our SynTs meet the critical criteria of low energy consumption for both write and read operations. Endurance was highlighted by cycling in ambient conditions with 100 states of potentiation and depression for over 1000 cycles with only a slight variation of Gmax/Gmin ratio of 6.2 % (Figure 1b, c). Approximately 95 % accuracy in MNIST pattern recognition test on ANN in the crossbar array configuration has been obtained by simulation with SynTs as synaptic elements reassured SynT is a promising candidate for future neuromorphic computing hardware.To shed light on the properties of intercalation phenomena of Li ions into the TiO2 layer, a further electrochemical study on a cell comprising Ti/TiO2/LiPON/Li corresponding to the SynT gate stack was performed. This understanding will help to elucidate the correlation with conductance modulation characteristics for a synaptic transistor. Multiple tests were carried out, including cyclic voltammetry (CV) with different scan rates, rate capability with Galvanostatic cycling with potential limit (GCPL), and electrochemical impedance spectroscopy (EIS) on different states of charge. A circuit model was introduced to fit the frequency response of the cell, and it explained well the behavior of charging capability at different OCV (Figure 1d).

Neuromorphic computing [1], inspired by the energy-efficient and parallelized architecture of the human brain, represents a paradigm shift in computing. This approach not only drives innovative hardware designs but also inspires novel network architectures like spiking neural networks (SNNs), which process information as discrete spike events. However, scalability in neuromorphic systems is constrained by the large hardware footprint. Realizing multimodal functionality using the same set of hardware primitives can offer a promising solution to these limitations.In this work, we present a dual encoder-neuron circuit based on a nanoporous graphene (NPG) memristor device [2]. The graphene leaky integrate-and-fire (GLIF) circuit has a versatile multimodal design that combines the roles of a LIF neuron and a spike encoder within the same architecture. This design mimics the dual functionality of biological neurons, which encode and process signals locally. The spike encoding capability of the circuit is further revealed using single and double layer SNN networks for pattern recognition tasks using the Modified National Institute of Standards and Technology (MNIST) dataset.Inset of Figure 1(a) shows the schematic of the fabricated NPG device with a channel length of 100 µm. The device has a lateral structure with NPG channel and gold (Au) electrodes. Figure 1(a) shows the current-voltage (I-V) characteristics of the fabricated NPG device which shows a threshold switching behavior with a wide hysteresis window and a threshold voltage (Vth) of 4.9 V. A behavioral SPICE model is developed based on a voltage-controlled switch model. It is seen from Figure 1(a) that the I-V curves simulated using the SPICE model matches the experimentally observed device switching characteristics. This reveals that the simplified SPICE model can be a good substitute for complex Verilog based models. Based on the SPICE model, we designed a simple neuron circuit with a resistor (Rs, 12 kΩ) in series with a parallel combination of a capacitor (Cm, 10 nF) and the NPG device (SPICE model). Figure 1(b) shows the variation of the membrane potential (Vm) and the output spike current in response to the input pulse voltage (Vin, 6 V). Vm shows a clear leaky integration behavior and full reset to zero voltage after the spike response. These are essential attributes of a bioplausible LIF neuron circuit. In order to develop the multimodal GLIF circuit, the output spike frequency of the LIF neuron circuit is plotted as a function of the Rs values. This mapping allows pixel values from MNIST input images to be directly translated into normalized conductance values (G/G0). Hence, each pixel is represented as a normalized conductance value, establishing a clear correlation between pixel intensity and spike frequency. This core idea is further used to develop the GLIF encoder based on the same neuron circuit as before. Based on the dual encoder-neuron circuit, we designed a single layer SNN for recognising the handwritten digits from the MNIST dataset. Figure 1(c) shows the spike-encoded images for the MNIST digits “2” and “5”. The probability of spiking in each timestep is proportional to the spiking frequency. The spike encoded images at each time step reveal the contours of the original input images using sparse representation. These sparse matrices highlight the energy efficiency and computational advantages of the GLIF system. The performance of the dual encoder-neuron system is evaluated by training the single layer SNN using both the GLIF encoder and neuron circuit as integral components. The single layer SNN recorded a high pattern recognition accuracy of 90.77% which is comparable to the purely software-based implementation (92.69%). To explore the scalability of the GLIF system, we implemented a double-layer SNN using the GLIF system, achieving a high recognition accuracy of 97.37%. The GLIF circuit exemplifies how reconfigurable primitive components can drive the development of scalable neuromorphic hardware.

Classical Convolutional Neural Networks, or ConvNets, have been the benchmarks for most object classification and face recognition tasks despite suffering from major limitations such as the inability to capture spatial co-locality between data points and favoring invariance over equivariance. Hence, a hierarchical routing layered architecture called Capsule Networks was proposed to overcome shortcomings of ConvNets. Capsules replace average or max pooling techniques of ConvNets with dynamic routing abilities between lower level and higher level neural units which better capture hierarchical relationships within the data and introduced reconstruction regularization mechanisms which deals with equivariance properties. By overcoming existing limitations, Capsules have proven themselves to be potential benchmarks in object segmentation, detection and reconstruction. Capsules have achieved state of the art results on the fundamental MNIST (Modified National Institute of Standards and Technology) handwritten digit dataset by reducing the ConvNets test error benchmark of 0.39% to 0.25%. In order to further augment this distinction, we experimented with five activation units such as sigmoid, e-Swish, Swish, variants of Rectified Linear Units (ReLU) like Parametric ReLU (PReLU), leaky ReLU (lReLU) and Scaled Exponential Linear Units (SELU), on two fundamental datasets - MNIST and cifar10. Based on these experimental results, we establish that e-Swish, and ReLU variants better optimize the Capsule architecture as compared to the currently used ReLU activation function.

The development of physiology and neurological diseases is normally accompanied by the production of acidic changes, making it crucial to develop biosensing system to detect pH variations to prevent early organ lesions. Due to the electrical double layer at the electrolyte/nanowire interface, artificial synaptic transistors can sensitively detect the pH variations. However, few studies focus on monitoring pH variations of the body based on synaptic transistors. In this study, electrolyte-gated synaptic transistors were fabricated using zinc tin oxide (ZnSnO) nanowires as channel materials. The fabricated devices can successfully mimic the functionality of acid-sensing ion channels (ASICs) in biological synapses, including excitatory postsynaptic currents (EPSCs), paired-pulse facilitations (PPFs), and short-term potentiations (STPs) under acidic conditions. Furthermore, <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${3}\times {3}$ </tex-math></inline-formula> ZnSnO synaptic transistor arrays were constructed to monitor acid stimulations of diverse intensities and different regions to mimic the tongue ability of human. This work can not only advance the understanding of ASIC functionality in artificial synaptic devices, but also provide a novel approach to develop bioinspired taste-sensing systems with integrated sensing and memory capabilities.

Nonlinear Nanophotonics With 2D Transition Metal Dichalcogenides

In the quest to create computing systems that mimic the brain’s remarkable ability to process data efficiently and in parallel, researchers have recognized the need for in-memory computing devices that can overcome the limitations of the traditional von Neumann architecture. Recently, neuromorphic devices like memristors and synaptic transistors have been developed to emulate the process of neurons. However, memristors face challenges in simultaneously handling signal transmission and learning tasks. As an alternative, synaptic transistors have gained attention owing to their adjustable weight and improved non-volatile synaptic functionality. Notably, synaptic transistors with electrolyte insulating layers can form electric double layers (EDLs) with high capacitance per unit area, because of mobile ions and dipoles. This enables efficient gate-controlled synaptic behavior. Additionally, amorphous oxynitride semiconductors, used as n-type channels, exhibit superior mobility compared to other materials like 2D transition metals, perovskites, organic materials, and oxides. In our investigation, therefore, we employed a cross-linked poly-4-vinyl phenol (PVP) gate insulator to enhance the synaptic properties of oxynitride thin film transistors.In our study, we employed zinc oxynitride (ZnON) as the channel material. ZnON offers greater stability compared to oxide materials due to its reduced oxygen vacancy concentration. We precisely controlled the solution composition and the curing process during the deposition of the electrolyte insulating layer. To optimize the formation of electric double layers (EDLs), we varied the weight ratio of poly(melamine-co-formaldehyde) methylated (PMF) and poly-4-vinyl phenol (PVP). Our investigation focused on synaptic properties, including excitatory postsynaptic current (EPSC), paired-pulse facilitation (PPF), long-term memory, and high-band filtering. All samples exhibited a counter-clockwise hysteresis curve and a memory window exceeding 2 V at 1 V of drain voltage (Vd), except for the PVP:PMF 2:1 sample. Notably, the device with a 10:1 weight ratio demonstrated the most effective EPSC response, showing a gradual increase in EPSC with repeated gate voltage (Vg) pulses. Additionally, we analyzed the EPSC characteristics under various Vg pulse conditions (width, frequency, period). To understand the behavior of internal ions and dipoles in the cross-linked PVP insulating layer, we measured the frequency-dependent capacitance. Interestingly, the capacitance decreased as the frequency increased, suggesting that internal ions could be easily influenced by electrical bias in the cross-linked PVP electrolyte, forming an EDL layer. These findings suggest that the cross-linked PVP/ZnON-based transistor could potentially serve as a significant synaptic transistor component in artificial intelligence technology.

Two-dimensional (2D) transition metal dichalcogenides (TMDs) are attractive semiconductors for use in electronic, optoelectronic, and spintronic devices. This study examines how the electronic properties of 2D TMDs can be tuned for specific applications through a combination of alloying and applying strain. Group VIB TMDs (MoS2, MoSe2, WS2, and WSe2) are alloyed by mixing in the metal or chalcogen sublattices. Density functional theory is used to model the structures of the alloys at varying compositions and examine the electronic structure of the alloys under biaxial tensile and compressive strain. Alloying results in the continuous monotonic tuning of the direct bandgap between the limits of the pure components, with low bowing coefficients for all alloys. Applying strain results in a transition of the bandgap from direct to indirect at low values of tensile strain and higher values of compressive strain. Strain can also be used to increase or decrease the bandgap with low compressive strain or tensile strain, respectively. The shift rate, or the rate at which the bandgap changes with applied strain, changes monotonically with alloy composition. MoS2 is identified as the 2D TMD with the highest shift rate.

This study presents the Flexible Representation for Quantum Images (FRQI) Pairs method, a novel approach that leverages Quantum Recurrent Neural Networks (QRNN) for image classification. The proposed method achieves an accuracy of 74.60% on the full Modified National Institute of Standards and Technology (MNIST) handwritten digit data set, demonstrating its effectiveness in handling quantum encoded data for classification tasks.By reducing the size of the QRNN by the exponential factor, the FRQI Pairs method highlights the potential of integrating quantum computing principles with neural network architectures, offering a promising direction for advancing quantum machine learning.The research evaluates the FRQI Pairs method against existing quantum and classical models, demonstrating its competitive performance against other state-of-the-art approaches and showing potential for future advancements in the field. This research opens avenues for further exploration of quantum preprocessing and hybrid model architectures, marking a step forward in the application of quantum machine learning.

Synaptic devices based neuromorphic computing applications in artificial intelligence

SummaryAtomically thin two-dimensional (2D) transition metal dichalcogenides (TMDCs) have attracted significant attention owing to their prosperity in material research. The inimitable features of TMDCs triggered the emerging applications in diverse areas. In this review, we focus on the tailored and engineering of the crystal lattice of TMDCs that finally enhance the efficiency of the material properties. We highlight several preparation techniques and recent advancements in compositional engineering of TMDCs structure. We summarize different approaches for TMDCs such as doping and alloying with different materials, alloying with other 2D metals, and scrutinize the technological potential of these methods. Beyond that, we also highlight the recent significant advancement in preparing 2D quasicrystals and alloying the 2D TMDCs with MAX phases. Finally, we highlight the future perspectives for crystal engineering in TMDC materials for structure stability, machine learning concept marge with materials, and their emerging applications.

Since excitonic quasiparticles, including excitons, trions and charges, have a great influence on the photoelectric characteristics of two-dimensional (2D) transition metal dichalcogenides (TMDs), systematic explorations of the trion dynamics and charge photogeneration in 2D TMDs are important for their future optoelectronic applications. Here, broadband femtosecond transient absorption spectroscopic experiments are performed first to investigate the peak shifting and broadening kinetics in MoS2 nanosheets in solution prepared by liquid phase exfoliation (LPE-MoS2, ∼9 layers, 9L), which reveal that the binding energies for the A-, B-, and C-exciton states are ∼77 meV, ∼76 meV, and -70 meV (the energy difference between free charges and excitons; the negative sign for C-excitons means a spontaneous dissociation nature in band-nesting regions), respectively. Then, the trion dynamics and charge photogeneration in LPE-MoS2 nanosheets have been studied in detail, demonstrating that they are comparable to those in chemical vapor deposition grown MoS2 films (1L-, 3L- and 7L-MoS2). These experimental results suggest that LPE-TMD nanosheets also have the potential for use in charge-related optoelectronic devices based on 2D TMDs.