Memristive Monte Carlo DropConnect crossbar array enabled by device and algorithm co-design.

Device and algorithm co-design aims to develop energy-efficient hardware that directly implements complex algorithms and optimizes algorithms to match the hardware's characteristics. Specifically, neuromorphic computing algorithms are constantly growing in complexity, necessitating an ongoing search for hardware implementations capable of handling these intricate algorithms. Here, we present a memristive Monte Carlo DropConnect (MC-DC) crossbar array developed through a hardware algorithm co-design approach. To implement the MC-DC neural network, stochastic switching and analog memory characteristics are required, and we achieved them using Ag-based diffusive selectors and Ru-based electrochemical metalization (ECM) memristors, respectively. The devices were integrated with a one-selector one-memristor (1S1M) structure, and their well-matched operating voltages and currents enabled stochastic readout and deterministic analog programming. With the integrated hardware, we successfully demonstrated the MC-DC operation. Additionally, the selector allowed for the control of switching polarity, and by understanding this hardware characteristic, we were able to modify the algorithm to fit it and further improve the network performance.

Mott neurons with dual thermal dynamics for spatiotemporal computing.

Heat dissipation is a natural consequence of operating any electronic system. In nearly all computing systems, such heat is usually minimized by design and cooling. Here, we show that the temporal dynamics of internally produced heat in electronic devices can be engineered to both encode information within a single device and process information across multiple devices. In our demonstration, electronic NbOx Mott neurons, integrated on a flexible organic substrate, exhibit 18 biomimetic neuronal behaviours and frequency-based nociception within a single component by exploiting both the thermal dynamics of the Mott transition and the dynamical thermal interactions with the organic substrate. Further, multiple interconnected Mott neurons spatiotemporally communicate purely via heat, which we use for graph optimization by consuming over 106 times less energy when compared with the best digital processors. Thus, exploiting natural thermal processes in computing can lead to functionally dense, energy-efficient and radically novel mixed-physics computing primitives.

Memristive Explainable Artificial Intelligence Hardware.

Artificial intelligence (AI) is often considered a black box because it provides optimal answers without clear insight into its decision-making process. To address this black box problem, explainable artificial intelligence (XAI) has emerged, which provides an explanation and interpretation of its decisions, thereby promoting the trustworthiness of AI systems. Here, a memristive XAI hardware framework is presented. This framework incorporates three distinct types of memristors (Mott memristor, valence change memristor, and charge trap memristor), each responsible for performing three essential functions (perturbation, analog multiplication, and integration) required for the XAI hardware implementation. Three memristor arrays with high robustness are fabricated and the image recognition of 3 × 3 testing patterns and their explanation map generation are experimentally demonstrated. Then, a software-based extended system based on the characteristics of this hardware is built, simulating a large-scale image recognition task. The proposed system can perform the XAI operations with only 4.32% of the energy compared to conventional digital systems, enlightening its strong potential for the XAI accelerator.

Fully Memristive Elementary Motion Detectors for a Maneuver Prediction.

Insects can efficiently perform object motion detection via a specialized neural circuit, called an elementary motion detector (EMD). In contrast, conventional machine vision systems require significant computational resources for dynamic motion processing. Here, a fully memristive EMD (M-EMD) is presented that implements the Hassenstein-Reichardt (HR) correlator, a biological model of the EMD. The M-EMD consists of a simple Wye (Y) configuration, including a static resistor, a dynamic memristor, and a Mott memristor. The resistor and dynamic memristor introduce different signal delays, enabling spatio-temporal signal integration in the subsequent Mott memristor, resulting in a direction-selective response. In addition, a neuromorphic system is developed employing the M-EMDs to predict a lane-changing maneuver by vehicles on the road. The system achieved a high accuracy (> 87%) in predicting future lane-changing maneuvers on the Next Generation Simulation (NGSIM) dataset while reducing the computational cost by 92.9% compared to the conventional neuromorphic system without the M-EMD, suggesting its strong potential for edge-level computing.

Probabilistic computing with NbOx metal-insulator transition-based self-oscillatory pbit.

Energy-based computing is a promising approach for addressing the rising demand for solving NP-hard problems across diverse domains, including logistics, artificial intelligence, cryptography, and optimization. Probabilistic computing utilizing pbits, which can be manufactured using the semiconductor process and seamlessly integrated with conventional processing units, stands out as an efficient candidate to meet these demands. Here, we propose a novel pbit unit using an NbOx volatile memristor-based oscillator capable of generating probabilistic bits in a self-clocking manner. The noise-induced metal-insulator transition causes the probabilistic behavior, which can be effectively modeled using a multi-noise-induced stochastic process around the metal-insulator transition temperature. We demonstrate a memristive Boltzmann machine based on our proposed pbit and validate its feasibility by solving NP-hard problems. Furthermore, we propose a streamlined operation methodology that considers the autocorrelation of individual bits, enabling energy-efficient and high-performance probabilistic computing.

Threshold Modulative Artificial GABAergic Nociceptor.

Gamma-aminobutyric acid (GABA) is a crucial inhibitory neurotransmitter of the central nervous system. It modifies the signal threshold of the nociceptor, allowing it to react to external stimuli in various circumstances. Thus, GABAergic behaviors are critical characteristics of adaptive behavior in life. Here, a threshold-modulative artificial GABAergic nociceptor is reported for the first time at a Pt/Ti/Nb2 O5- x /Al2 O3- y /Pt/Ti (top to bottom) of the double charge trapping structure. The Al2 O3- y layer contains deep defect states that function similarly to the GABA neurotransmitter in modulating the signal threshold. Meanwhile, the Nb2 O5- x layer traps volatile charges and produces nociceptive behaviors. The combined dynamics of the two layers readily offer threshold-modulative GABAergic nociceptive behaviors. Based on these GABAergic behaviors, a method of implementing hot- and cold-sensitive thermoreceptors is demonstrated and shows its potential applications in advanced sensory devices.

High Amplitude Spike Generator in Au Nanodot-Incorporated NbOx Mott Memristor.

NbOx-based Mott memristors exhibit fast threshold switching behaviors, making them suitable for spike generators in neuromorphic computing and stochastic clock generators in security devices. In these applications, a high output spike amplitude is necessary for threshold level control and accurate signal detection. Here, we propose a materialwise solution to obtain the high amplitude spikes by inserting Au nanodots into the NbOx device. The Au nanodots enable increasing the threshold voltage by modulating the oxygen contents at the electrode-oxide interface, providing a higher ON current compared to nanodot-free NbOx devices. Also, the reduction of the local switching region volume decreases the thermal capacitance of the system, allowing the maximum spike amplitude generation. Consequently, the Au nanodot incorporation increases the spike amplitude of the NbOx device by 6 times, without any additional external circuit elements. The results are systematically supported by both a numerical model and a finite-element-method-based multiphysics model.

Retention Secured Nonlinear and Self-Rectifying Analog Charge Trap Memristor for Energy-Efficient Neuromorphic Hardware.

A memristive crossbar array (MCA) is an ideal platform for emerging memory and neuromorphic hardware due to its high bitwise density capability. A charge trap memristor (CTM) is an attractive candidate for the memristor cell of the MCA, because the embodied rectifying characteristic frees it from the sneak current issue. Although the potential of the CTM devices has been suggested, their practical viability needs to be further proved. Here, a Pt/Ta2 O5 /Nb2 O5- x /Al2 O3- y /Ti CTM stack exhibiting high retention and array-level uniformity is proposed, allowing a highly reliable selector-less MCA. It shows high self-rectifying and nonlinear current-voltage characteristics below 1 µA of programming current with a continuous analog switching behavior. Also, its retention is longer than 105 s at 150 °C, suggesting the device is highly stable for non-volatile analog applications. A plausible band diagram model is proposed based on the electronic spectroscopy results and conduction mechanism analysis. The self-rectifying and nonlinear characteristics allow reducing the on-chip training energy consumption by 71% for the MNIST dataset training task with an optimized programming scheme.

Modified Dynamic Physical Model of Valence Change Mechanism Memristors.

Valence change-type resistance switching behaviors in oxides can be understood by well-established physical models describing the field-driven oxygen vacancy distribution change. In those models, electroformed residual oxygen vacancy filaments are crucial as they work as an electric field concentrator and limit the oxygen vacancy movement along the vertical direction. Therefore, their movement outward by diffusion is negligible. However, this situation may not be applicable in the electroforming-free system, where the field-driven movement is less prominent, and the isotropic oxygen vacancy diffusion by concentration gradient is more significant, which has not been given much consideration in the conventional model. Here, we propose a modified physical model that considers the change in the oxygen vacancies' charged state depending on their concentrations and the resulting change in diffusivity during switching to interpret the electroforming-free device behaviors. The model suggests formation of an hourglass-shaped filament constituting a lower concentration of oxygen vacancies due to the fluid oxygen diffusion in the thin oxide. Consequently, the proposed model can explain the electroforming-free device behaviors, including the retention failure mechanism, and suggest an optimized filament configuration for improved retention characteristics. The proposed model can plausibly explain both the electroformed and the electroforming-free devices. Therefore, it can be a standard model for valence change memristors.

Ternary Logic with Stateful Neural Networks Using a Bilayered TaOX -Based Memristor Exhibiting Ternary States.

A memristive stateful neural network allowing complete Boolean in-memory computing attracts high interest in future electronics. Various Boolean logic gates and functions demonstrated so far confirm their practical potential as an emerging computing device. However, spatio-temporal efficiency of the stateful logic is still too limited to replace conventional computing technologies. This study proposes a ternary-state memristor device (simply a ternary memristor) for application to ternary stateful logic. The ternary-state implementable memristor device is developed with bilayered tantalum oxide by precisely controlling the oxygen content in each oxide layer. The device can operate 157 ternary logic gates in one operational clock, which allows an experimental demonstration of a functionally complete three-valued Lukasiewicz logic system. An optimized logic cascading strategy with possible ternary gates is ≈20% more efficient than conventional binary stateful logic, suggesting it can be beneficial for higher performance in-memory computing.

Self-clocking fast and variation tolerant true random number generator based on a stochastic mott memristor.

The intrinsic stochasticity of the memristor can be used to generate true random numbers, essential for non-decryptable hardware-based security devices. Here, we propose a novel and advanced method to generate true random numbers utilizing the stochastic oscillation behavior of a NbOx mott memristor, exhibiting self-clocking, fast and variation tolerant characteristics. The random number generation rate of the device can be at least 40 kb s-1, which is the fastest record compared with previous volatile memristor-based TRNG devices. Also, its dimensionless operating principle provides high tolerance against both ambient temperature variation and device-to-device variation, enabling robust security hardware applicable in harsh environments.

Defect-Engineered Electroforming-Free Analog HfOx Memristor and Its Application to the Neural Network.

The thin-film growth conditions in a plasma-enhanced atomic layer deposition for the (3.0-4.5) nm thick HfO2 film were optimized to use the film as the resistive switching element in a neuromorphic circuit. The film was intended to be used as a feasible synapse with analog-type conductance-tuning capability. The 4.5 nm thick HfO2 films on both conventional TiN and a new RuO2 bottom electrode required the electroforming process for them to operate as a feasible resistive switching memory, which was the primary source of the undesirable characteristics as the synapse. Therefore, electroforming-free performance was necessary, which could be accomplished by thinning the HfO2 film down to 3.0 nm. However, the device with only the RuO2 bottom electrode offered the desired functionality without involving too high leakage or shorting problems, which are due to the recovery of the stoichiometric composition of the HfO2 near the RuO2 layer. In conjunction with the Ta top electrode, which provided the necessary oxygen vacancies to the HfO2 layer, and the high functionality of the RuO2 as the scavenger of excessive incorporated oxygen vacancies, which appeared to be inevitable during the repeated switching operation, the Ta/3.0 nm HfO2/RuO2 provided a highly useful synaptic device component in the neuromorphic hardware system.