DLDong Keun Lee

Dong Keun Lee

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Dong Keun Lee has a total of 46 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    A vision sensor for simultaneous imaging and distance sensing in real-time

    by Keewoong Haan, Kukjin Chun, Byung-Gook Park, Hyeon Cheol Kim, Bonghwan Kim

    A vision sensor for simultaneous imaging and distance sensing is proposed herein. This vision sensor uses two images with a depth of field by two different aperture sizes to extract distance data. The aperture was fabricated through the microelectromechanical systems process. The optical parameter related to making a blur was precisely selected to reduce the active voltage and response time. The aperture measurement result showed that the maximum displacement of 1170 µm was obtained when 12 V was applied which enlarge aperture size from 2.75 mm to 3.92 mm. The response time was 16.6 ms with a rising and falling time of 6.2 and 10.4 ms, respectively. The distance data was obtained using the depth from defocus method, which compares the blurriness of two images using the aperture size. Through deep learning, the image and distance information were simultaneously obtained in a single camera. The result of a three-dimensional depth map showed an average accuracy of 98.7% when sensing the maximum distance of 10 m. To examine the accuracy of the device, experiments were conducted for different colors, and the result showed that maximum and minimum error rates of 3.46% and 1.83% were achieved, respectively. In addition, the error rate according to brightness was investigated, and the average error rate was maintained at 2.64% between 10 000 and 200 lx. The proposed sensor can be installed in self-driving robots, drones, and various smart devices.

  • Analog synaptic devices applied to spiking neural networks for reinforcement learning applications

    by Jangsaeng Kim, Soochang Lee, Chul-Heung Kim, Byung-Gook Park, Jong-Ho Lee

    In this work, we implement hardware-based spiking neural network (SNN) using the thin-film transistor (TFT)-type flash synaptic devices. A hardware-based SNN architecture with synapse arrays and integrate-and-fire (I&F) neuron circuits is presented for executing reinforcement learning (RL). Two problems were used to evaluate the applicability of the proposed hardware-based SNNs to off-chip RL: the Cart Pole balancing problem and the Rush Hour problem. The neural network was trained using a deep Q-learning algorithm. The proposed hardware-based SNNs using the synapse model with measured characteristics successfully solve the two problems and show high performance, implying that the networks are suitable for executing RL. Furthermore, the effect of variations in non-ideal synaptic devices and neurons on the performance was investigated.


  • Lamination method for improved polarization-leakage current relation in HfO2-based metal/ferroelectric/insulator/semiconductor structure

    by Yeonwoo Kim, Kyung Kyu Min, Junsu Yu, Daewoong Kwon, Byung-Gook Park

    Ever since the ferroelectricity of complementary metal-oxide semiconductor (CMOS) compatible HfO2-based materials was discovered, numerous studies have been conducted on their ferroelectric (FE) properties and device applications. In particular, pure-HfO2 FE materials without external doping have attracted considerable attention owing to their excellent robustness against variation because variations that appear in conventional doped-HfO2 FEs are not observed in electrical characteristics induced by dopant fluctuations in pure-HfO2 FEs. Studies on metal/FE/insulator/semiconductor (MFIS) stack are required to apply the ferroelectricity of pure-HfO2 to memory devices that are completely compatible with Si-based CMOS processes. In pure-HfO2 based MFIS stacks, the polarization tends to reduce with increasing thickness of the HfO2, although the leakage current diminishes. To overcome the tradeoff between the polarization and leakage current with respect to the thickness of the HfO2, an Al2O3 layer was inserted between the HfO2 layers to form a laminated FE structure. By employing the laminated FE, leakage current was effectively suppressed by the Al2O3 and lower HfO2 layers, and polarization was enhanced by the FE sum of the upper and lower HfO2 layers. Therefore, an MFIS structure with maximized polarization and minimized leakage current was successfully demonstrated using laminated FE. In addition, the feasibility of the proposed MFIS with laminated FE for nonvolatile memory device applications was confirmed by verifying the multistate operations of a FE tunnel junction.


  • A More Hardware-Oriented Spiking Neural Network Based on Leading Memory Technology and Its Application With Reinforcement Learning

    by Min-Hwi Kim, Sungmin Hwang, Suhyun Bang, Tae-Hyeon Kim, Dong Keun Lee, Hasan Raza Ansari, Seongjae Cho, Byung-Gook Park

    In recent days, more hardware-driven artificial intelligence system capable of brain-like low-energy consumption is gaining ever-increasing interest. The hardware-driven property lies in the low-power synaptic device and its array along with the area and energy-efficient neuron circuits. In this work, a spiking neural network (SNN) based on analog synaptic device of resistive-switching random access memory (RRAM) is constructed from the experimentally fabricated devices. Furthermore, the capability of the designed SNN hardware for sequential tasks through an optimal reinforcement learning (RL) algorithm is demonstrated. More specifically, the Rush Hour game is conducted as an example of applications for the sequential task for which an SNN architecture is plausibly suited. The rule of the game is simple but has not been demonstrated by a hardware-oriented artificial neural network (ANN) yet, and in this work, it is reported that the analog RRAM synaptic devices in the cross-point array architecture successfully solve the problem via the RL algorithm.


  • Optimization of post-deposition annealing temperature for improved signal-to-noise ratio in In 2 O 3 gas sensor

    by Wonjun Shin, Seongbin Hong, Yujeong Jeong, Gyuweon Jung, Jinwoo Park, Donghee Kim, Byung-Gook Park, Jong-Ho Lee

    This paper investigates the effects of post-deposition annealing (PDA) temperature on H 2 S gas sensing and low-frequency noise characteristics of In 2 O 3 gas sensors. In 2 O 3 thin-films are deposited using the radio frequency (RF) sputtering method at an RF power of 150 W and post-annealed at various temperatures (200, 300, and 400 °C). The response of the In 2 O 3 gas sensor to H 2 S decreases with increasing PDA temperature due to the increase of grain size. However, the In 2 O 3 post-annealed at 200 °C shows the largest 1/ f noise since the damaged sensing material-substrate interface is not fully recovered by the PDA. The sensors post-annealed at a higher temperature (300 °C and 400 °C) recover the damaged interface. Thanks to its moderate response and noise level, the sensor post-annealed at 300 °C shows the largest signal-to-noise ratio.


  • Distance sensor with a tunable aperture actuated by thermal expansion of silicon chevron beams

    by Chong Ho Hong, Hyeon Cheol Kim, Wisnu Jatmiko, Kurnianingsih, Kukjin Chun, Byung-Gook Park

    To measure the distance of an object, we propose a distance sensor that combines a tunable aperture operated by the thermal expansion of silicon chevron beams with a commercial camera similar to the automotive black box. The distance sensor is composed of a tunable aperture, a lens with an 8 mm focal length, and a 1/4-inch full high definition image sensor, where a tunable aperture is attached to the camera lens. To estimate the distance to the object, the blur difference is expressed in the form of a depth map using the depth from defocus algorithm. The mean error of distance estimation was 0.29 m. The area of the tunable aperture in the distance sensor is 1.18 mm 2 at OFF state and 2.79 mm 2 when the voltage is applied. The distance sensor can display an image and a distance simultaneously and can be applied to autonomous vehicles, drones, and robots.


  • 3-bit multilevel operation with accurate programming scheme in TiO x /Al 2 O 3 memristor crossbar array for quantized neuromorphic system

    by Tae-Hyeon Kim, Jaewoong Lee, Sungjoon Kim, Jinwoo Park, Byung-Gook Park, Hyungjin Kim

    As interest in artificial intelligence (AI) and relevant hardware technologies has been developed rapidly, algorithms and network structures have become significantly complicated, causing serious power consumption issues because an enormous amount of computation is required. Neuromorphic computing, a hardware AI technology with memory devices, has emerged to solve this problem. For this application, multilevel operations of synaptic devices are important to imitate floating point weight values in software AI technologies. Furthermore, weight transfer methods to desired weight targets must be arranged for off-chip training. From this point of view, we fabricate 32 × 32 memristor crossbar array and verify the 3-bit multilevel operations. The programming accuracy is verified for 3-bit quantized levels by applying a reset-voltage-control programming scheme to the fabricated TiO x /Al 2 O 3 -based memristor array. After that, a synapse composed of two differential memristors and a fully-connected neural network for modified national institute of standards and technology (MNIST) pattern recognition are constructed. The trained weights are post-training quantized in consideration of the 3-bit characteristics of the memristor. Finally, the effect of programming error on classification accuracy is verified based on the measured data, and we obtained 98.12% classification accuracy for MNIST data with the programming accuracy of 1.79% root-mean-square-error. These results imply that the proposed reset-voltage-control programming scheme can be utilized for a precise tuning, and expected to contribute for the development of a neuromorphic system capable of highly precise weight transfer.


  • Effect of Ag source layer thickness on the switching mechanism of TiN/Ag/SiN x /TiN conductive bridging random access memory observed at sub- µ A current

    by Yeon-Joon Choi, Suhyun Bang, Tae-Hyeon Kim, Dong Keun Lee, Kyungho Hong, Sungjun Kim, Byung-Gook Park

    Experiments are conducted to compare the resistive switching characteristics for several samples with different amounts of Ag deposition in TiN/Ag/SiN x /TiN conductive bridging random access memory (CBRAM). The compliance current in TiN/Ag/SiN x /TiN CBRAM determines the volatile/non-volatile memory operation as the current level controls the strength of the filament made of Ag. The transient measurement showed that the effective thickness of Ag source layer in the TiN/Ag/SiN x /TiN controls the supply of the Ag atoms into the insulating layer, affecting the strength of the conductive bridge. The mechanism for the switching characteristics and the volatility trend with the amount of Ag deposition is closely investigated using transmission electron microscopy and scanning electron microscopy images. The device shows the conductance potentiation by a voltage pulse train under 1 µ A current level, and the higher potentiation rate is observed in the CBRAM with thick Ag source layer.


  • Open Access

    Insertion of Ag Layer in TiN/SiNx/TiN RRAM and Its Effect on Filament Formation Modeled by Monte Carlo Simulation

    by Yeon-Joon Choi, Min-Hwi Kim, Suhyun Bang, Tae-Hyeon Kim, Dong Keun Lee, Kyungho Hong, Chae Soo Kim, Sungjun Kim, Seongjae Cho, Byung-Gook Park

    In this study, the electrical characteristics of TiN/SiN x /TiN and TiN/Ag/SiN x /TiN RRAMs were thoroughly investigated through I-V measurements. Our novel Ag-inserted silicon nitride-based resistive switching memory (RRAM) achieved switching operation at lower voltages and lower current levels compared to the conventional silicon nitride RRAM that does not have an Ag insertion layer. These enhanced characteristics enable low power operation, one of the main goals of RRAM research. Gradual conductance modulation is also possible in this device with constant voltage pulses being 50% lower than that seen in conventional devices, this property fulfills an important requirement that makes our device suitable for use as a neuromorphic synapse device. The conduction mechanism of our Ag-inserted device was further analyzed through I-V measurements and compared with a control group. Based on these data, a 2D Monte Carlo simulation was implemented to investigate the mechanism behind the new behavior displayed by our Ag-inserted silicon nitride RRAM. A resistive network model was used to calculate the voltage and electric field distribution, and then the motion of the particles was simulated by taking into account its relation to the heat generated inside the device.


  • Improvement of Resistive Switching Characteristics of Titanium Oxide Based Nanowedge RRAM Through Nickel Silicidation

    by Dong Keun Lee, Min-Hwi Kim, Suhyun Bang, Tae-Hyeon Kim, Yeon-Joon Choi, Kyungho Hong, Sungjun Kim, Seongjae Cho, Jong-Ho Lee, Byung-Gook Park

    Low operation power and high endurance of resistive random access memory (RRAM) as a synaptic device are critical parameters for in-memory computing applications. Yet, high power consumption and reliability issue of silicon bottom electrode (BE) RRAM hinder its commercialization as a synaptic device. In this experiment, we report on the improvement of switching characteristics of silicon BE nanowedge RRAM via the Nickel (Ni) silicidation process. Existing highly doped Si-BE forms a SiO₂ interfacial layer (IL) during a switching layer deposition and increases an effective thickness, leading to increased voltage drop within the RRAM device and large cycle-to-cycle variations. By siliciding the Si-BE with Ni, the issue of IL formation is removed and the resistance of metallic NiSi BE is further reduced compared to Arsenic (As⁺) doped Si BE. Both dc and ac analyses of the fabricated NiSi-BE nanowedge RRAM have shown the reduction of overshoot and switching current down to 55% of the original value. Transmission electron microscopy (TEM) and energy-dispersive spectroscopy (EDS) analysis convinced the formation of NiSi BE. In addition, gradual switching characteristics, uniform low resistance state (LRS), and better endurance of NiSi-BE nanowedge RRAM enable the Si compatible approach to fabricate a large-size RRAM cross-point array for utilization in hardware-implemented neuromorphic computing applications.