A Unified Current-Voltage Model for Metal Oxide-Based Resistive Random-Access Memory.

Resistive random-access memory (RRAM) is essential for developing neuromorphic devices, and it is still a competitive candidate for future memory devices. In this paper, a unified model is proposed to describe the entire electrical characteristics of RRAM devices, which exhibit two different resistive switching phenomena. To enhance the performance of the model by reflecting the physical properties such as the length index of the undoped area during the switching operation, the Voltage ThrEshold Adaptive Memristor (VTEAM) model and the tungsten-based model are combined to represent two different resistive switching phenomena. The accuracy of the I-V relationship curve tails of the device is improved significantly by adjusting the ranges of unified internal state variables. Furthermore, the unified model describes a variety of electrical characteristics and yields continuous results by using the device's current-voltage relationship without dividing its fitting conditions. The unified model describes the optimized electrical characteristics that reflect the electrical behavior of the device.

Circuit Optimization Method to Reduce Disturbances in Poly-Si 1T-DRAM.

A capacitorless one-transistor dynamic random-access memory device (1T-DRAM) is proposed to resolve the scaling problem in conventional one-transistor one-capacitor random-access memory (1T-1C-DRAM). Most studies on 1T-DRAM focus on device-level operation to replace 1T-1C-DRAM. To utilize 1T-DRAM as a memory device, we must understand its circuit-level operation, in addition to its device-level operation. Therefore, we studied the memory performance depending on device location in an array circuit and the circuit configuration by using the 1T-DRAM structure reported in the literature. The simulation results show various disturbances and their effects on memory performance. These disturbances occurred because the voltages applied to each device during circuit operation are different. We analyzed the voltage that should be applied to each voltage line in the circuit to minimize device disturbance and determine the optimized bias condition and circuit structure to achieve a large sensing margin and realize operation as a memory device. The results indicate that the memory performance improves when the circuit has a source line and the bias conditions of the devices differ depending on the write data at the selected device cell. Therefore, the sensing margin of the 1T-DRAM used herein can expectedly be improved by applying the proposed source line (SL) structure.

Analysis of Grain Boundary Dependent Memory Characteristics in Poly-Si One-Transistor Dynamic Random-Access Memory.

A capacitorless one-transistor dynamic random-access memory cell with a polysilicon body (poly-Si 1T-DRAM) has a cost-effective fabrication process and allows a three-dimensional stacked architecture that increases the integration density of memory cells. Also, since this device uses grain boundaries (GBs) as a storage region, it can be operated as a memory cell even in a thin body device. GBs are important to the memory characteristics of poly-Si 1T-DRAM because the amount of trapped charge in the GBs determines the memory's data state. In this paper, we report on a statistical analysis of the memory characteristics of poly-Si 1T-DRAM cells according to the number and location of GBs using TCAD simulation. As the number of GBs increases, the sensing margin and retention time of memory cells deteriorate due to increasing trapped electron charge. Also, "0" state current increases and memory performance degrades in cells where all GBs are adjacent to the source or drain junction side in a strong electric field. These results mean that in poly-Si 1T-DRAM design, the number and location of GBs in a channel should be considered for optimal memory performance.

Analysis of a Lateral Grain Boundary for Reducing Performance Variations in Poly-Si 1T-DRAM.

A capacitorless one-transistor dynamic random-access memory device that uses a poly-silicon body (poly-Si 1T-DRAM) has been suggested to overcome the scaling limit of conventional one-transistor one-capacitor dynamic random-access memory (1T-1C DRAM). A poly-Si 1T-DRAM cell operates as a memory by utilizing the charge trapped at the grain boundaries (GBs) of its poly-Si body; vertical GBs are formed randomly during fabrication. This paper describes technology computer aided design (TCAD) device simulations performed to investigate the sensing margin and retention time of poly-Si 1T-DRAM as a function of its lateral GB location. The results show that the memory's operating mechanism changes with the GB's lateral location because of a corresponding change in the number of trapped electrons or holes. We determined the optimum lateral GB location for the best memory performance by considering both the sensing margin and retention time. We also performed simulations to analyze the effect of a lateral GB on the operation of a poly-Si 1T-DRAM that has a vertical GB. The memory performance of devices without a lateral GB significantly deteriorates when a vertical GB is located near the source or drain junction, while devices with a lateral GB have little change in memory characteristics with different vertical GB locations. This means that poly-Si 1T-DRAM devices with a lateral GB can operate reliably without any memory performance degradation from randomly determined vertical GB locations.

Effect of Initial Synaptic State on Pattern Classification Accuracy of 3D Vertical Resistive Random Access Memory (VRRAM) Synapses.

Amidst the considerable attention artificial intelligence (AI) has attracted in recent years, a neuromorphic chip that mimics the biological neuron has emerged as a promising technology. Memristor or Resistive random-access memory (RRAM) is widely used to implement a synaptic device. Recently, 3D vertical RRAM (VRRAM) has become a promising candidate to reducing resistive memory bit cost. This study investigates the operation principle of synapse in 3D VRRAM architecture. In these devices, the classification response current through a vertical pillar is set by applying a training algorithm to the memristors. The accuracy of neural networks with 3D VRRAM synapses was verified by using the HSPICE simulator to classify the alphabet in 7×7 character images. This simulation demonstrated that 3D VRRAMs are usable as synapses in a neural network system and that a 3D VRRAM synapse should be designed to consider the initial value of the memristor to prepare the training conditions for high classification accuracy. These results mean that a synaptic circuit using 3D VRRAM will become a key technology for implementing neural computing hardware.

Analysis of the Sensing Margin of Silicon and Poly-Si 1T-DRAM.

Recently, one-transistor dynamic random-access memory (1T-DRAM) cells having a polysilicon body (poly-Si 1T-DRAM) have attracted attention as candidates to replace conventional one-transistor one-capacitor dynamic random-access memory (1T-1C DRAM). Poly-Si 1T-DRAM enables the cost-effective implementation of a silicon-on-insulator (SOI) structure and a three-dimensional (3D) stacked architecture for increasing integration density. However, studies on the transient characteristics of poly-Si 1T-DRAM are still lacking. In this paper, with TCAD simulation, we examine the differences between the memory mechanisms in poly-Si and silicon body 1T-DRAM. A silicon 1T-DRAM cell's data state is determined by the number of holes stored in a floating body (FB), while a poly-Si 1T-DRAM cell's state depends on the number of electrons trapped in its grain boundary (GB). This means that a poly-Si 1T-DRAM can perform memory operations by using GB as a storage region in thin body devices with a small FB area.

Three-Dimensional (3D) Vertical Resistive Random-Access Memory (VRRAM) Synapses for Neural Network Systems.

Memristor devices are generally suitable for incorporation in neuromorphic systems as synapses because they can be integrated into crossbar array circuits with high area efficiency. In the case of a two-dimensional (2D) crossbar array, however, the size of the array is proportional to the neural network's depth and the number of its input and output nodes. This means that a 2D crossbar array is not suitable for a deep neural network. On the other hand, synapses that use a memristor with a 3D structure are suitable for implementing a neuromorphic chip for a multi-layered neural network. In this study, we propose a new optimization method for machine learning weight changes that considers the structural characteristics of a 3D vertical resistive random-access memory (VRRAM) structure for the first time. The newly proposed synapse operating principle of the 3D VRRAM structure can simplify the complexity of a neuron circuit. This study investigates the operating principle of 3D VRRAM synapses with comb-shaped word lines and demonstrates that the proposed 3D VRRAM structure will be a promising solution for a high-density neural network hardware system.

Memristor Neural Network Training with Clock Synchronous Neuromorphic System.

Memristor devices are considered to have the potential to implement unsupervised learning, especially spike timing-dependent plasticity (STDP), in the field of neuromorphic hardware research. In this study, a neuromorphic hardware system for multilayer unsupervised learning was designed, and unsupervised learning was performed with a memristor neural network. We showed that the nonlinear characteristic memristor neural network can be trained by unsupervised learning only with the correlation between inputs and outputs. Moreover, a method to train nonlinear memristor devices in a supervised manner, named guide training, was devised. Memristor devices have a nonlinear characteristic, which makes implementing machine learning algorithms, such as backpropagation, difficult. The guide-training algorithm devised in this paper updates the synaptic weights by only using the correlations between inputs and outputs, and therefore, neither complex mathematical formulas nor computations are required during the training. Thus, it is considered appropriate to train a nonlinear memristor neural network. All training and inference simulations were performed using the designed neuromorphic hardware system. With the system and memristor neural network, the image classification was successfully done using both the Hebbian unsupervised training and guide supervised training methods.

Analysis of the Memristor-Based Crossbar Synapse for Neuromorphic Systems.

In this study, we analyzed the memristor device typically used as a synapse in neuromorphic architecture and confirmed that the synaptic memristor device can be adopted to perform the machine learning algorithm. The nonlinear characteristics of the memristor complicates its use as the neuromorphic hardware in an artificial neural network (ANN) with a back-propagation algorithm. Using a memristor device with a nonlinear characteristic, we demonstrated that pattern classification can be implemented in ANNs using the Guide training algorithm without back-propagation. Furthermore, the memristor characteristics required to achieve accurate learning results are analyzed.