Jaeyeop Choi

Digital twins act on exported machine data that may not certify the setpoint channel, the time base, or a hidden controller state. We propose an audit for integrators: a twin acts only when its decision is the same for every admissible value of the missing quantity, and otherwise abstains, naming the missing evidence. Its error is bounded by the miss rate of the admissible set. Building on set-based decision making, we measure this audit on published robot, drone, machine-tool and process data, one raster scanner and one industrial robot. Lead–lag role inference erred on at least 20 % of the windows where the check abstains in 6 of 7 motion datasets. A UR5 twin built on inferred roles missed 7.8 % to 31.4 % of heavier payloads (after a post-hoc unit correction; refuted as registered), and a delay commissioned at low speed misplaced 97.0 % to 99.5 % of fast scanner strokes at 1 mm (15.2 % under host load). Plan matching, whose reference favours it, erred on 4.5 % pooled and 10.3 % unweighted, about half the rate of random choice. Registered Granger tests, on levels and on differences, were right on some datasets and mostly wrong on others; pooled, both erred as often as random choice. Where an inferred filter removed the truth, the check proceeded wrongly on most such windows in four of eleven datasets and on none in six. For integrators, the audit states which twin decisions an interface supports, and at what error.

C-scan images in scanning acoustic microscopy (SAM) are crucial for ensuring the reliability and accuracy of semiconductor devices. These images provide detailed cross-sectional views of devices, allowing for precise identification and analysis of internal structures and defects without causing any damage. Recent studies have extensively employed C-scan imaging to enhance quality control processes, significantly improving the overall effectiveness of non-destructive testing (NDT). However, most samples typically possess flat surfaces and are examined under optimal scanning conditions. This article addresses the challenge of uneven or deformed sample surfaces that result in low-quality C-scan images. Specifically, we developed a surface segmentation model utilizing B-scan images from the SAM system and a fine-tuned version of the latest You Only Look Once (YOLO) architecture. Subsequently, the A-scan was recalculated to reconstruct the C-scan image. To evaluate the feasibility of this approach, we posed two research questions: (1) Can the improved YOLOv11-seg network effectively perform surface segmentation? and (2) How is the quality of the C-scan image improved after reconstruction? The results indicate that our enhanced model achieves promising performance in surface segmentation of B-scan images, with a mean Average Precision mAP_0.5−0.95 of approximately 0.938 for bounding boxes and 0.901 for masks. This study proposes an effective solution for the NDT of semiconductor samples with uneven surfaces that ensures high-quality C-scan images, supports quality control processes, and enhances the overall effectiveness of non-destructive evaluation methodologies.

Metallic blocks and components are indispensable across the aerospace, energy, and heavy-engineering industries, where undetected internal flaws such as cracks, voids, and inclusions may precipitate catastrophic structural failure. Reliable nondestructive evaluation (NDE) is essential to ensure their integrity and operational safety. Among the available NDE techniques, phased array ultrasonic testing (PAUT) has emerged as one of the most accessible and widely adopted, by virtue of its rapid scanning, electronic beam steering, and capacity to image subsurface defects without disassembly. However, the interpretation of PAUT B-scan images remains hindered by background reflections, material-dependent echo characteristics, and substantial variability in defect size. In this work, a fine-tuned encoder–decoder deep learning network is proposed for the automated segmentation of internal defects in PAUT B-scan images of metallic block specimens. The network couples a ResNet50 encoder with a shallow detail stem, a multi-scale feature fusion module, and a detail refinement block, designed to preserve small defect echoes and sharpen weak defect boundaries characteristic of internal flaws. The proposed approach was compared with five state-of-the-art segmentation architectures, namely FCN, PSPNet, DeepLabv3+, HRNet-OCR, and SegFormer, as well as a conventional Otsu-thresholding baseline representing standard PAUT screening practice. Experimental results demonstrate that the proposed network attained the highest Dice score of 0.7964, defect intersection-over-union of 0.6617, and precision of 0.7094 among all evaluated models, while the Otsu baseline yielded the lowest scores, confirming the benefit of learned segmentation over fixed amplitude thresholding. These findings indicate that the proposed network achieves a favorable trade-off between defect localization accuracy and false-positive suppression, underscoring its potential for reliably segmenting internal defects in PAUT B-scan imaging.

Ultrasonic anemometers are widely used for wind measurement owing to their fast response, high temporal resolution, and long operational lifetime with minimal recalibration requirements. However, most configurations suffer from transducer shadowing, where cylindrical probes disrupt local airflow and introduce systematic errors in time-of-flight (TOF) measurements. While prior computational fluid dynamics (CFD)-based investigations have characterized this effect, their analyses remain confined to low wind speeds, and no existing study has explicitly proposed a method to mitigate shadowing-induced bias. This paper presents a coupled CFD and acoustic propagation framework to analyze wake-induced velocity deficits and their effects on TOF measurements for a three-transducer ultrasonic anemometer, with simulations spanning 5 to 75 m/s over the full 360° range. The results show that shadowing distortions are strongly direction-dependent, peaked within approximately ±5° angular sectors, with a near-constant velocity deficit of approximately 40% along affected paths. A shadowing-aware acoustic path selection method is then proposed that selectively excludes corrupted acoustic paths, reducing the average velocity root-mean-square error (RMSE) to 0.349 m/s and the directional RMSE to 1.14°, representing improvements of more than an order of magnitude over shadow-unaware methods. These findings provide a physically grounded, simulation-based framework for shadowing-aware wind measurement using ultrasonic anemometers.

Three-dimensional (3D) bioprinting has rapidly evolved into a controlling platform for the fabrication of patient-specific biomedical implants, with growing importance in advanced drug delivery systems. Beyond structural tissue engineering, bioprinted constructs now function as programmable therapeutic depots capable of localized, sustained, and stimuli-responsive drug release. This review focuses on recent biomaterial design strategies that enable precise control over drug encapsulation, retention, and release kinetics within 3D bioprinted architectures. The physicochemical and mechanical properties of bioinks, including crosslinking density, porosity, degradation behavior, viscoelasticity, and swelling characteristics, directly influence drug loading efficiency and release dynamics under physiological conditions. The rational tuning of these parameters allows the development of constructs that provide spatially controlled and temporally regulated therapeutic delivery. Recent advances in predictive modeling, such as finite element modeling (FEM), data-driven machine learning approaches, and ML, have significantly improved the ability to correlate material composition, printing parameters, and structural geometry with drug diffusion and degradation-mediated release mechanisms. These tools facilitate the optimization of printing variables including extrusion pressure, nozzle diameter, and layer resolution to ensure structural fidelity while maintaining therapeutic functionality. Emerging strategies incorporating multi-material printing, gradient architectures, and stimuli-responsive biomaterials have expanded the potential of 3D bioprinting for combination therapies and personalized medicine. This review discusses key challenges in translating bioprinted drug delivery systems into clinical applications, including the standardization of drug release characterization methods, and long-term stability assessment.

Heart rate monitoring (HRM) serves as a fundamental component of healthcare systems, providing essential insight into cardiovascular function and overall physiological status. Conventional HRM systems commonly rely on contact-based sensing modalities, such as electrocardiography (ECG) and photoplethysmography (PPG) wearables, which require direct skin contact and complex wiring. This work presents a noncontact ballistocardiogram (NCBCG) bed sensor system integrated with tiny machine learning (TinyML) to provide on-chip processing for noncontact HRM. The proposed system incorporates a polyvinylidene fluoride (PVDF) film sensor placed beneath the mattress, enabling comfortable, nonintrusive, and continuous monitoring. We design a TinyML model that offers on-chip processing on a microcontroller unit (MCU), reducing dependency on cloud-based or external. Experimental findings demonstrate that the system achieves accurate heart rate (HR) estimation, which is compliant with the AAMI standard. These results highlight the feasibility and reliability of accurate HR estimation using embedded artificial intelligence (AI). The work demonstrates potential for real-world applications in long-term remote healthcare and home-based monitoring.

Background/Objectives: Improper landing mechanics in Taekwondo can lead to non-contact injuries such as ankle sprains and knee ligament tears, highlighting the necessity for objective methods to evaluate landing stability and injury risk. Electromyography (EMG) enables the examination of muscle activation patterns; however, conventional analyses based on simple averages have limited predictive value. Methods: This study analyzed EMG signals recorded during single-leg landings (45 cm height) in 30 elite male Taekwondo athletes. Participants were divided into regular exercise groups (REG, n = 15) and non-exercise groups (NEG, n = 15). Signals were segmented into two phases. Eight features were extracted per muscle per phase. Classification models (Random Forest, XGBoost, Logistic Regression, Voting Classifier) were used to classify between groups, while regression models (Ridge, Random Forest, XGBoost) predicted continuous muscle activation changes as injury risk indicators. Results: The Random Forest Classifier achieved an accuracy of 0.8365 and an F1-score of 0.8547. For regression, Ridge Regression indicated high performance (R² = 0.9974, MAE = 0.2620, RMSE = 0.4284, 5-fold CV MAE: 0.2459 ± 0.0270), demonstrating strong linear correlations between EMG features and outcomes. Conclusions: The AI-enabled EMG analysis can be used as an objective measure of the study of the individual landing stability and risk of injury in Taekwondo athletes, but its clinical application has to be validated in the future by biomechanical injury indicators and prospective cohort studies.