Yongkeun Lee
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yongkeun Lee has a total of 21 co-authors. Here are the top co-authors by number of shared publications:
- Jiayi Fan
Seoul National University of Science and Technology
11 shared publications - JangHyeon Lee
Department of Computer Science and Engineering, University of Minnesota, Minneapolis, MN, USA
4 shared publications - JongKwang Kim
Graduate School of Nano-IT Design Fusion, Seoul National University of Science and Technology, Seoul, South Korea
4 shared publications - ByoungGyu Kim
Department of Semiconductor Engineering, Seoul National University of Science and Technology, Seoul, South Korea
3 shared publications - Insu Jung
Medical Accelerator Research Team, Department of RI Application, Korea Institute of Radiological & Medical Sciences (KIRAMS), Seoul 01812, Korea
3 shared publications Jiayi Fan
2 shared publications
Latest Publications
- 22 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Microelectromechanical Systems
- Vol. 35 (3) , 550-556
Sub-30 nm Silicon Nanopatterning: Toward Reduced Reflectance and Enhanced Sensitivity in CMOS Image Sensors
This work investigated how subwavelength silicon patterning and dielectric passivation influence the optical and electrical performance of advanced CMOS image sensors (CIS). Hole arrays with sidewall angles of 73°, 79°, and 86° were fabricated using industrial plasma-etch processes, with and without aluminum-oxide/hafnium-oxide passivation and antireflective (AR) coatings. Reflectance measurements from 360–1000 nm showed that etch geometry and dielectric filling substantially reduced Fresnel reflection, with the 79° profile achieving the lowest broadband reflectance and film-filled structures providing additional suppression in the blue–green region. Device-level evaluations on 0.64 $\mu $ m-pixel CIS sensors demonstrated RGB sensitivity improvements of 0.7–3.3% under low illumination (23 lux, 30 ms), with no increase in crosstalk. These trends correlated closely with wafer-level reflectance, confirming reflectance as a reliable early predictor of device performance. Although certain etch conditions increased dark current due to plasma-induced damage, the aluminum-oxide/hafnium-oxide passivation stack effectively mitigated interface traps, enabling stable electrical behavior for optimized profiles. Overall, the results demonstrated that co-optimizing etch geometry, dielectric materials, and process conditions enhances both optical efficiency and electrical robustness, providing a manufacturable path toward higher-sensitivity CIS devices in submicron-pixel technology.[2025-0151]
- 22 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Microelectromechanical Systems
- Vol. 35 (3) , 550-556
Sub-30 nm Silicon Nanopatterning: Toward Reduced Reflectance and Enhanced Sensitivity in CMOS Image Sensors
This work investigated how subwavelength silicon patterning and dielectric passivation influence the optical and electrical performance of advanced CMOS image sensors (CIS). Hole arrays with sidewall angles of 73°, 79°, and 86° were fabricated using industrial plasma-etch processes, with and without aluminum-oxide/hafnium-oxide passivation and antireflective (AR) coatings. Reflectance measurements from 360–1000 nm showed that etch geometry and dielectric filling substantially reduced Fresnel reflection, with the 79° profile achieving the lowest broadband reflectance and film-filled structures providing additional suppression in the blue–green region. Device-level evaluations on 0.64 $\mu $ m-pixel CIS sensors demonstrated RGB sensitivity improvements of 0.7–3.3% under low illumination (23 lux, 30 ms), with no increase in crosstalk. These trends correlated closely with wafer-level reflectance, confirming reflectance as a reliable early predictor of device performance. Although certain etch conditions increased dark current due to plasma-induced damage, the aluminum-oxide/hafnium-oxide passivation stack effectively mitigated interface traps, enabling stable electrical behavior for optimized profiles. Overall, the results demonstrated that co-optimizing etch geometry, dielectric materials, and process conditions enhances both optical efficiency and electrical robustness, providing a manufacturable path toward higher-sensitivity CIS devices in submicron-pixel technology.[2025-0151]
- Open Access
- 10 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 56363-56373
Physics-Guided Machine Learning for Accurate Series and Shunt Resistance Estimation in Photovoltaic Modules
Accurate estimation of the series (Rs) and shunt (Rsh) resistances from photovoltaic (PV) I–V measurements is essential for device characterization and degradation assessment, yet remains highly sensitive to measurement noise, particularly for conventional slope-based extraction methods. This work proposes a machine-learning–assisted framework for robust estimation of Rs and Rsh from noisy I–V curves by combining physics-informed, normalized features with ensemble regression. A comprehensive set of irradiance-normalized environmental, electrical, slope-related, and current-distribution features is extracted from the full I–V curve and used to train an XGBoost regressor. To further enhance robustness, principal component analysis (PCA) is optionally applied to suppress correlated and noise-sensitive feature directions while preserving physically meaningful variance. Model performance is evaluated across multiple noise levels, with robustness and feature relevance examined using leave-one-feature-out ablation. The proposed framework achieves consistently high accuracy for Rs and significantly improves Rsh estimation under moderate-to-high noise conditions. Relative to the no-PCA baseline, PCA reduces relative error by up to 20–25% for Rs and 10–15% for Rsh, with optimal performance obtained by retaining approximately five principal components. Overall, the results demonstrate that resistance estimation is governed by a low-dimensional, noise-robust subspace of physics-informed features, providing a compact and practical alternative to conventional PV parameter extraction methods.
- 10 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 56363-56373
Open AccessPhysics-Guided Machine Learning for Accurate Series and Shunt Resistance Estimation in Photovoltaic Modules
Accurate estimation of the series (Rs) and shunt (Rsh) resistances from photovoltaic (PV) I–V measurements is essential for device characterization and degradation assessment, yet remains highly sensitive to measurement noise, particularly for conventional slope-based extraction methods. This work proposes a machine-learning–assisted framework for robust estimation of Rs and Rsh from noisy I–V curves by combining physics-informed, normalized features with ensemble regression. A comprehensive set of irradiance-normalized environmental, electrical, slope-related, and current-distribution features is extracted from the full I–V curve and used to train an XGBoost regressor. To further enhance robustness, principal component analysis (PCA) is optionally applied to suppress correlated and noise-sensitive feature directions while preserving physically meaningful variance. Model performance is evaluated across multiple noise levels, with robustness and feature relevance examined using leave-one-feature-out ablation. The proposed framework achieves consistently high accuracy for Rs and significantly improves Rsh estimation under moderate-to-high noise conditions. Relative to the no-PCA baseline, PCA reduces relative error by up to 20–25% for Rs and 10–15% for Rsh, with optimal performance obtained by retaining approximately five principal components. Overall, the results demonstrate that resistance estimation is governed by a low-dimensional, noise-robust subspace of physics-informed features, providing a compact and practical alternative to conventional PV parameter extraction methods.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 171667-171676
Classifier-Controlled Evaluation of Pointwise LoRA for Parameter-Efficient Fine-Tuning of Pretrained MobileNets
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 171667-171676
Open AccessClassifier-Controlled Evaluation of Pointwise LoRA for Parameter-Efficient Fine-Tuning of Pretrained MobileNets
- 31 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Electron Devices
- Vol. 73 (2) , 821-827
Interpretable XGBoost Estimation of Memristor Mobility and Thickness
Accurate estimation of device thickness and instantaneous mobility is essential for reliable characterization and optimization of memristors. Conventional curve-fitting methods, based on idealized physical assumptions, often produce substantial errors, with relative deviations of 10%–25% for thickness and 20%–40% for mobility in nanoscale TiO2 devices. This study presents a noise-resilient and physically interpretable XGBoost regression framework for direct estimation of thickness and instantaneous mobility from TiO2 memristor current–voltage ( $I$ – $V$ ) data. Evaluated under realistic noise conditions—including shot, thermal, flicker, and quantization noise modeled after a Keysight B1500A analyzer—the framework achieved ${R}^{{2}} = {0.9973}$ for thickness and ${R}^{{2}} = {0.9383}$ for instantaneous mobility, with relative errors (REs) of 1.90% and 18.9%, respectively. Feature-importance analysis identified ${R}_{\text {on}}$ , ${R}_{\text {off}}$ , temperature, and predicted thickness as the dominant predictors, aligning with their physical roles in doped-region dynamics and ionic drift. These results demonstrate that the proposed machine learning (ML) framework provides a robust, interpretable, and scalable solution for parameter extraction from noisy $I$ – $V$ measurements, substantially surpassing traditional fitting approaches and enabling reliable device modeling and optimization.
- 31 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Electron Devices
- Vol. 73 (2) , 821-827
Interpretable XGBoost Estimation of Memristor Mobility and Thickness
Accurate estimation of device thickness and instantaneous mobility is essential for reliable characterization and optimization of memristors. Conventional curve-fitting methods, based on idealized physical assumptions, often produce substantial errors, with relative deviations of 10%–25% for thickness and 20%–40% for mobility in nanoscale TiO2 devices. This study presents a noise-resilient and physically interpretable XGBoost regression framework for direct estimation of thickness and instantaneous mobility from TiO2 memristor current–voltage ( $I$ – $V$ ) data. Evaluated under realistic noise conditions—including shot, thermal, flicker, and quantization noise modeled after a Keysight B1500A analyzer—the framework achieved ${R}^{{2}} = {0.9973}$ for thickness and ${R}^{{2}} = {0.9383}$ for instantaneous mobility, with relative errors (REs) of 1.90% and 18.9%, respectively. Feature-importance analysis identified ${R}_{\text {on}}$ , ${R}_{\text {off}}$ , temperature, and predicted thickness as the dominant predictors, aligning with their physical roles in doped-region dynamics and ionic drift. These results demonstrate that the proposed machine learning (ML) framework provides a robust, interpretable, and scalable solution for parameter extraction from noisy $I$ – $V$ measurements, substantially surpassing traditional fitting approaches and enabling reliable device modeling and optimization.
- 27 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Electron Device Letters
- Vol. 46 (9) , 1569-1572
Bayesian Design of Metasurface Routers for CMOS Image Sensors via MetaRGBX-Net
This letter presents a Bayesian optimization framework based on MetaRGBX-Net for tuning meta-atom diameters to achieve specific RGB sensitivity and interpixel crosstalk (XTALK) targets. MetaRGBX-Net—developed in prior work—is validated as an effective surrogate model within the optimization process and enables successful tuning of in-bound (IB) configurations, even near distribution boundaries. RGB sensitivity and XTALK errors were both maintained below 10% through balanced trade-offs. Experimental validation confirms these outcomes, emphasizing the impact of penalty weight adjustments—particularly for blue sensitivity, which was more responsive than red or green. In contrast, out-of-bound (OB) configurations resulted in notable performance degradation across all algorithms, with excessive XTALK and unmet RGB targets. These results underscore the framework’s potential and the importance of well-designed penalty functions and target selection for optimal performance.
- 27 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Electron Device Letters
- Vol. 46 (9) , 1569-1572
Bayesian Design of Metasurface Routers for CMOS Image Sensors via MetaRGBX-Net
This letter presents a Bayesian optimization framework based on MetaRGBX-Net for tuning meta-atom diameters to achieve specific RGB sensitivity and interpixel crosstalk (XTALK) targets. MetaRGBX-Net—developed in prior work—is validated as an effective surrogate model within the optimization process and enables successful tuning of in-bound (IB) configurations, even near distribution boundaries. RGB sensitivity and XTALK errors were both maintained below 10% through balanced trade-offs. Experimental validation confirms these outcomes, emphasizing the impact of penalty weight adjustments—particularly for blue sensitivity, which was more responsive than red or green. In contrast, out-of-bound (OB) configurations resulted in notable performance degradation across all algorithms, with excessive XTALK and unmet RGB targets. These results underscore the framework’s potential and the importance of well-designed penalty functions and target selection for optimal performance.
- Open Access
- 25 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Improved Position Sensorless Control of Switched Reluctance Motor by Phase Voltage Estimation Method
In this letter, an improved position sensorless control of switched reluctance motor (SRM) is proposed. In the conventional inductance model-based sensorless control scheme, the variation of the phase voltage is neglected, and thus, the accuracy of the position estimation suffers from inaccurate phase voltage, especially in low voltage drive applications. Therefore, this letter proposes using a simple phase voltage estimation method in the sensorless scheme, considering the voltage drop of switches and diodes, so that the phase voltage is accurate and the accuracy of the sensorless control can be improved. Simulations show that the accuracy of the position estimation is significantly improved compared with the conventional method.
- 25 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Open AccessImproved Position Sensorless Control of Switched Reluctance Motor by Phase Voltage Estimation Method
In this letter, an improved position sensorless control of switched reluctance motor (SRM) is proposed. In the conventional inductance model-based sensorless control scheme, the variation of the phase voltage is neglected, and thus, the accuracy of the position estimation suffers from inaccurate phase voltage, especially in low voltage drive applications. Therefore, this letter proposes using a simple phase voltage estimation method in the sensorless scheme, considering the voltage drop of switches and diodes, so that the phase voltage is accurate and the accuracy of the sensorless control can be improved. Simulations show that the accuracy of the position estimation is significantly improved compared with the conventional method.
- Open Access
- 25 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
A Pulse Injection based Position Detection Method of Switched Reluctance Motor Considering Effect of Back EMF
In this paper, an improved position detection method of switched reluctance motor (SRM) is proposed based on the pulse injection in the inactive phases. The conventional pulse injection method is normally used by neglecting the back EMF term due to it is small at standstill and very low speed. However, it is found out in this study that the exclusion of the back EMF term indeed affects the accuracy of the rotor position estimation. This paper offers a more accurate expression of the peak phase current rather than the oversimplified peak current equation in the conventional method, which is used to estimate the self-inductance, thus the rotor position can be obtained from a single inductance curve, and achieves better accuracy. Simulation results are given under different speed conditions, which show the improvement of accuracy of the position estimation is over wide speed range.
- 25 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in TechRxiv
Open AccessA Pulse Injection based Position Detection Method of Switched Reluctance Motor Considering Effect of Back EMF
In this paper, an improved position detection method of switched reluctance motor (SRM) is proposed based on the pulse injection in the inactive phases. The conventional pulse injection method is normally used by neglecting the back EMF term due to it is small at standstill and very low speed. However, it is found out in this study that the exclusion of the back EMF term indeed affects the accuracy of the rotor position estimation. This paper offers a more accurate expression of the peak phase current rather than the oversimplified peak current equation in the conventional method, which is used to estimate the self-inductance, thus the rotor position can be obtained from a single inductance curve, and achieves better accuracy. Simulation results are given under different speed conditions, which show the improvement of accuracy of the position estimation is over wide speed range.
- 18 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Electron Device Letters
- Vol. 46 (7) , 1235-1238
MetaRGBX-Net: RGB Sensitivity and Cross-Talk Prediction in CMOS Image Sensor
The CMOS image sensor (CIS) underpins optical applications, enabling high-resolution imaging across the visible and near-infrared spectra. Advances in nanofabrication have enhanced pixel density, improving resolution, but as pixel dimensions approach the diffraction limit, maintaining optical sensitivity without performance trade-offs remains challenging. Nanophotonic solutions, like nanophotonic-based color routers, address these limitations. This study builds on MetaRGB-Net, a machine learning framework achieving 98% prediction accuracy in optimizing RGB sensitivity. We introduce MetaRGBX-Net, which employs separate neural networks to optimize both RGB sensitivity and cross-talk, achieving 95% and 98% prediction accuracy, respectively. This enables precise optimization of critical parameters and serves as a foundation for Bayesian optimization to refine metasurface designs, ensuring efficient light routing through RGB channels while minimizing cross-talk. MetaRGBX-Net streamlines metasurface design and provides a scalable foundation for next-generation CIS applications in IoT, biomedical imaging, and environmental monitoring.
- 18 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Electron Device Letters
- Vol. 46 (7) , 1235-1238
MetaRGBX-Net: RGB Sensitivity and Cross-Talk Prediction in CMOS Image Sensor
The CMOS image sensor (CIS) underpins optical applications, enabling high-resolution imaging across the visible and near-infrared spectra. Advances in nanofabrication have enhanced pixel density, improving resolution, but as pixel dimensions approach the diffraction limit, maintaining optical sensitivity without performance trade-offs remains challenging. Nanophotonic solutions, like nanophotonic-based color routers, address these limitations. This study builds on MetaRGB-Net, a machine learning framework achieving 98% prediction accuracy in optimizing RGB sensitivity. We introduce MetaRGBX-Net, which employs separate neural networks to optimize both RGB sensitivity and cross-talk, achieving 95% and 98% prediction accuracy, respectively. This enables precise optimization of critical parameters and serves as a foundation for Bayesian optimization to refine metasurface designs, ensuring efficient light routing through RGB channels while minimizing cross-talk. MetaRGBX-Net streamlines metasurface design and provides a scalable foundation for next-generation CIS applications in IoT, biomedical imaging, and environmental monitoring.
- Open Access
- 2 March 2025
- Published by National Technical University Kharkiv Polytechnic Institute in Electrical Engineering & Electromechanics
- No. 2,p. 3-6
Online detection of phase resistance of switched reluctance motor by sinusoidal signal injection
Introduction. Switched reluctance motors (SRMs) are widely used in various applications due to their simplicity, robustness, and cost-effectiveness. However, the performance of SRMs can be significantly influenced by variations in their phase resistance, especially under high current and saturated conditions. Accurate knowledge of this parameter is crucial for optimal control and efficient operation. Problem. During operation, SRM parameters, particularly phase resistance, can vary considerably. These variations pose challenges to control strategies that rely on precise parameter values, leading to potential inefficiencies and degraded performance. There is a need for an effective method to monitor and identify these changes in real-time. Goal. This paper aims to develop and validate a method for the online detection and identification of phase resistance in SRMs. The method should work under varying operational conditions without requiring additional hardware, thereby maintaining the system's simplicity and cost-effectiveness. Methodology. The proposed method injects a sinusoidal signal into the inactive phase of the SRM using Sinusoidal Pulse Width Modulation (SPWM) via the main converter. The phasor method is then applied to determine the impedance of the phase circuit, from which the phase resistance can be identified. This approach eliminates the need for extra circuits, making it an efficient solution. Results. Simulations were conducted to evaluate the proposed method. The results demonstrate that the method can accurately track the variation in phase resistance under different operational conditions, validating its effectiveness. Originality. The originality of this work lies in its innovative use of the phasor method combined with SPWM for online phase resistance detection in SRMs, without the need for additional hardware components. Practical value. This method provides a practical solution for real-time phase resistance identification in SRMs, enhancing the reliability and performance of control strategies in various industrial applications. References 17, table 1, figures 6.
- 2 March 2025
- Published by National Technical University Kharkiv Polytechnic Institute in Electrical Engineering & Electromechanics
- No. 2,p. 3-6
Open AccessOnline detection of phase resistance of switched reluctance motor by sinusoidal signal injection
Introduction. Switched reluctance motors (SRMs) are widely used in various applications due to their simplicity, robustness, and cost-effectiveness. However, the performance of SRMs can be significantly influenced by variations in their phase resistance, especially under high current and saturated conditions. Accurate knowledge of this parameter is crucial for optimal control and efficient operation. Problem. During operation, SRM parameters, particularly phase resistance, can vary considerably. These variations pose challenges to control strategies that rely on precise parameter values, leading to potential inefficiencies and degraded performance. There is a need for an effective method to monitor and identify these changes in real-time. Goal. This paper aims to develop and validate a method for the online detection and identification of phase resistance in SRMs. The method should work under varying operational conditions without requiring additional hardware, thereby maintaining the system's simplicity and cost-effectiveness. Methodology. The proposed method injects a sinusoidal signal into the inactive phase of the SRM using Sinusoidal Pulse Width Modulation (SPWM) via the main converter. The phasor method is then applied to determine the impedance of the phase circuit, from which the phase resistance can be identified. This approach eliminates the need for extra circuits, making it an efficient solution. Results. Simulations were conducted to evaluate the proposed method. The results demonstrate that the method can accurately track the variation in phase resistance under different operational conditions, validating its effectiveness. Originality. The originality of this work lies in its innovative use of the phasor method combined with SPWM for online phase resistance detection in SRMs, without the need for additional hardware components. Practical value. This method provides a practical solution for real-time phase resistance identification in SRMs, enhancing the reliability and performance of control strategies in various industrial applications. References 17, table 1, figures 6.
- Open Access
- 20 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 34613-34626
Toward Efficient Cancer Detection on Mobile Devices
Recent advancements in deep learning for cancer detection have achieved impressive accuracy, yet high computational costs and latency remain significant barriers for practical deployment on resource-constrained devices, such as smartphones and IoT platforms. This study focuses on optimizing MobileNetV1 and MobileNetV2 models to achieve more efficient, real-time cancer type identification. Through optimization strategies including selective layer unfreezing, pruning, and quantization, we demonstrate significant improvements in model size, inference time, and efficiency. For MobileNetV1, model size was reduced from 13.1 MB to 3.23 MB, and inference time was cut from 23 ms to 14 ms, with an F1 score above 0.99. For MobileNetV2, the model size was reduced from 9.41 MB to 2.82 MB, with inference times reduced from 24 ms to 13 ms, while maintaining a high F1 score of 0.98. The efficiency scores for MobileNetV1 and MobileNetV2 were 0.984 and 0.994, respectively, significantly outperforming other state-of-the-art neural networks such as VGG16 (efficiency score: 0.458), ResNet50 (0.418), and DenseNet121 (0.731). These findings demonstrate that with appropriate optimizations, MobileNet models can be deployed on edge devices, achieving high accuracy (above 95%), fast inference times (under one second), and superior efficiency, making them ideal candidates for real-time cancer detection in resource-constrained environments like mobile and IoT devices.
- 20 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 34613-34626
Open AccessToward Efficient Cancer Detection on Mobile Devices
Recent advancements in deep learning for cancer detection have achieved impressive accuracy, yet high computational costs and latency remain significant barriers for practical deployment on resource-constrained devices, such as smartphones and IoT platforms. This study focuses on optimizing MobileNetV1 and MobileNetV2 models to achieve more efficient, real-time cancer type identification. Through optimization strategies including selective layer unfreezing, pruning, and quantization, we demonstrate significant improvements in model size, inference time, and efficiency. For MobileNetV1, model size was reduced from 13.1 MB to 3.23 MB, and inference time was cut from 23 ms to 14 ms, with an F1 score above 0.99. For MobileNetV2, the model size was reduced from 9.41 MB to 2.82 MB, with inference times reduced from 24 ms to 13 ms, while maintaining a high F1 score of 0.98. The efficiency scores for MobileNetV1 and MobileNetV2 were 0.984 and 0.994, respectively, significantly outperforming other state-of-the-art neural networks such as VGG16 (efficiency score: 0.458), ResNet50 (0.418), and DenseNet121 (0.731). These findings demonstrate that with appropriate optimizations, MobileNet models can be deployed on edge devices, achieving high accuracy (above 95%), fast inference times (under one second), and superior efficiency, making them ideal candidates for real-time cancer detection in resource-constrained environments like mobile and IoT devices.