This paper proposes a compiler-assisted approach to reduce power consumption and energy consumption in intermittent computing environments by selectively approximating floating-point operations. Instead of uniformly converting all floating-point instructions, this method introduces custom directives that allow programmers to mark candidate variables and code regions for approximation. During preprocessing, these directives are used as a hint to determine whether floating-point operations can be safely replaced with integer operation. The transformed code is fully integrated with the Immortal Threads runtime system to ensure resilience against power interruptions by preserving computation states in non-volatile memory. Experimental evaluation on the MSP430FR5994 platform demonstrates that the proposed directive-based transformation can reduce both energy consumption and average power usage, while preserving the programmability of conventional thread-based models.

This paper presents a machine learning-based automatic test point insertion method that leverages structural features such as extended fanout and fanin ranges, along with controllability and observability metrics of circuit nodes. The proposed approach iteratively performs training, prediction, and test point insertion while updating circuit data after each test point insertion, thereby enabling adaptive and effective selection of optimal test point locations. Experimental results on ISCAS'89 benchmark circuits shows that the proposed method improves fault coverage by 6.74% compared to circuits without test point insertion, by 5.98% compared to the conventional SCOAP-based test point selection method, and by 0.78% compared to a commercial EDA tool-based test point selection method.

Computer vision technology, which mimics human vision to extract and analyze information from images and videos, has undergone significant advancements alongside the rapid evolution of artificial intelligence. While image sensing has focused on enhancing technical resolution and sensitivity to produce high-quality images that closely resemble real objects, data-driven analysis has prioritized the rapid and accurate extraction of information from these objects for analytical purposes. The typical detectable range, often restricted to the visible spectrum of the human eye, can be extended to encompass a broader wavelength domain, enabling computer vision technology to address an expanding set of industrial and scientific needs. In this review, we examine recent developments in short-wavelength infrared image sensors. We also discuss broad research directions in this area and consider how advances in short-wavelength infrared imaging are contributing to the evolution of next-generation sensing technologies.

An Electrostatic Chuck is a device that secures substrates, such as wafers or Organic Light-Emitting Diode panels, by utilizing the electrostatic force generated between charges applied to two parallel plates. Unlike mechanical clamps like vises, the Electrostatic Chuck relies on electrostatic force, making it crucial to verify that it provides sufficient adhesion. In a practical Electrostatic Chuck, minute spaces exist between the substrate and the chuck due to surface roughness. This is represented as a gap. Previous analyses often did not define the distance of this gap. In this study, the analysis was conducted by defining the gap as a ratio of contact to non-contact between the chuck and the substrate. This approach reflects the fine clearances that occur in actual semiconductor processes and is set for various reasons, such as evaluating the uniformity of the chucking force. The dielectric constant of this gap can also be altered by introducing gases. The gap is an essential input for Electrostatic Chuck analysis, and a lower relative dielectric constant generally leads to a superior chucking force. In this research, the change in electrostatic force with respect to the gap distance was analyzed using a simplified 2D model in Ansys Maxwell, an electromagnetic field analysis software. Furthermore, the effect of the presence and modeling of the gap on the performance of the Electrostatic Chuck was investigated.

Recently, precision inspection and measurement equipment requires such high precision that its performance is affected by micro disturbances, such as floor vibration. For this reason, the need for technologies that effectively reduce micro-vibrations is increasing to implement high-precision systems. Passive vibration isolation using springs, pneumatic mounts, and dampers is one method for reducing vibration. In this study, vertical guides were applied to minimize horizontal disturbances and deformation caused by external forces. A single passive module was designed using coil spring and leaf spring guide mechanisms, and its performance was predicted through analysis. An actual single passive module was fabricated using the design values, and its performance was verified through experiments. Finally, a testbed was constructed using four passive modules for practical application, and the performance of the passive vibration isolation system was verified through modal analysis and experiments.

The deflection distortion of an electron beam probe, caused by astigmatism and field curvature aberrations in electron microscopy using an electron column, is one of the key adjustment factors that determine raster image sharpness. This challenge must be addressed in an electrostatic ultra-miniaturized microcolumn specially designed and optimized for multi-array configurations to innovatively enhance throughput. Such correction is an essential requirement to satisfy the sharpness specifications of high-speed electron beam inspection equipment required in current sub-7 nm EUV semiconductor manufacturing processes. In this study, using an electrostatic ultra-miniaturized microcolumn, we investigated the feasibility of a dynamic focusing drive method capable of simultaneously correcting astigmatism and field curvature aberrations that occur at each point on the specimen surface when electron beam is deflected horizontally and vertically by the electric field of a deflector to form a raster image. The deflection distortion of the electron beam probe was evaluated by applying dynamic voltages synchronized with the deflection driving voltage to the eight electrodes of an octupole stigmator integrated into the electrostatic ultra-miniaturized microcolumn. As a result, both aberrations occurring at each scanning point of the raster image were effectively compensated, achieving a nearly circular electron beam shape and significantly improving field of view.

This paper aims to design crypthgraphic hardware that can operate efficiently in a variety of applications ranging from the Internet of Things to high-performance computing. To this end, we explore the design approaches of different hardware architectures around the KECCAK-512 hash function and suggest the optimal design direction to achieve a balance between performance and resource efficiency. The three proposed architectures take different structural perspectives to achieve computational efficiency, hardware resource utilization, and throughput maximization. They are divided into ways to secure computational flexibility through real-time round constant calculation, to enable fast access using pre-calculated data, and to improve processing speed through parallel processing. As a result of the analysis, it was confirmed that each architecture provides different advantages depending on the constraints of the target system-for example, throughput-centric design, area restriction design, or energy efficiency-centric design. Through this, this study proposes design guidelines for selecting and combining architectures to meet performance goals required by various application scenarios. These guidelines can provide practical reference points to designers' decision-making processes in future cryptographic hardware development.

As the operating frequency of semiconductor packages increases, predicting the high-frequency performance of pogo pin sockets has become critical. While simulations require modeling that considers actual operating conditions, manufacturers' datasheets typically only contain information on dielectric properties at low frequencies, which are inadequate for high-frequency analysis. In this paper, we measured the permittivity of a pogo pin socket housing and incorporated the measured data into the electromagnetic simulation model. Subsequently, the simulation results were compared with measurements and confirmed that the simulation based on measured permittivity achieved better agreement with the measurement results compared to the simulation based on the existing datasheet.

Deep learning-based computer vision, especially techniques for estimating human 3D postures from single-view images, plays a key role in quantitatively analyzing complex dynamic movements. These techniques have great potential to provide an objective indicator in the field of sports analysis, where subjective evaluation is prone to involvement. This study applies 3D human pose estimation techniques to the field of figure skating to develop an AI algorithm that automatically classifies six major jumps (axel, flip, loop, lutz, salchow, and toeloop) that require high skill and artistry but pose subjectivity issues of judgment. The research process is as follows. First, we applied RTMW3D, a 3D human pose estimation model, to a public figure skating video dataset to extract the skaters' 3D joint coordinates as time-series data. Next, based on the extracted 3D skeleton data, we quantified the key biomechanical features that distinguish each jump, such as edge just before leaping, aerial revolution, and leap method. Finally, these feature vectors were entered into Transformer, a deep learning model specialized for time series classification, to determine the type of jump. The model developed in this study showed 70.3% accuracy, suggesting that further improvements can provide objective feedback to athletes and coaches and improve training efficiency. Furthermore, this model is expected to contribute as an auxiliary tool to increase the accuracy and fairness of referees.

This study aims to enhance the accuracy of defective pixel detection and yield analysis in micro-LEDs by implementing a combined inspection technique utilizing Automated Optical Inspection (AOI) and Photoluminescence (PL), along with the application of a machine learning algorithm. During the image preprocessing stage, template matching and histogram normalization were performed, followed by the extraction of features based on the Gray-Level Co-occurrence Matrix (GLCM). A Random Forest model was employed to classify normal and defective pixels, with initial defect detection performed via outlier analysis using box plots. Based on these results, synthetic training data were generated by combining image segments, enabling the development of a model that achieved over 99% accuracy. To further categorize defect types, Principal Component Analysis (PCA) and K-means clustering were applied. In the PL inspection, defective pixels were identified through box plot analysis of the mean and standard deviation of brightness data. Finally, only devices classified as non-defective in both AOI and PL inspections were considered acceptable, thereby improving overall defect detection accuracy. Further validation with large-scale datasets and diverse defect types are required for industrial application.

The red emitting europium doped gallium oxide (Ga₂O₃:Eu3³⁺) film was synthesized by using the Ga₂O₃:Eu3³⁺ solgel. The device starts to emit at 6 V, which showed the red emission peaking at the 591, 614, 703 nm attributed to electronic transition of the Eu³⁺ ions. The electroluminescence intensity was increased as increasing the operating voltage and saturated at 800 Hz. The EL intensity and current density showed significant asymmetry according to applied polarities. The device had an FWHM of 15 nm at the main(614 nm) and sub(703 nm) peak in spite of voltage and frequency change. Thus, it showed strong potential as a uniformly surface emitting Red EL device.

Recently, various sensors are being used in autonomous manufacturing and smart systems. Therefore, signal acquisition and analysis technologies optimized for these sensors are being developed and demanded. To properly acquire and analyze specific sensor signals, an appropriate signal acquisition system must be selected for each sensor. In industrial sites, signal acquisition systems are required to acquire and analyze various analog signals in the 1-10 kHz range, such as vibration, sound, current, and voltage. Commercial signal acquisition systems are typically high-spec, bulky, and high power consumption compared to the required specifications, making them difficult to deploy in industrial settings. In this study, a low-cost and optimized data acquisition system is proposed to replace expensive commercial signal acquisition systems. A 16-bit ADC IC capable of processing analog signals was selected, and the design allowed the MCU to process digital signals. The performance of the developed data acquisition system was verified experimentally, confirming its ability to acquire data in the 10 kHz frequency band at low cost.

This study is about the internal flow distribution within a stabilization furnace used in carbon nanofiber production. The stabilization process, conducted under conditions that require applying temperatures ranging from 200℃ to 300℃, leads to the formation of ring structures in their polymer chains. For producing high-quality carbon nanofibers, ensuring uniform heat treatment within the stabilization furnace itself is important. This research focuses on analyzing the air flow since temperature distribution is highly dependent on the air flow. Computational fluid dynamics analyses results showed that airflow in the stabilization furnace was not uniform, with air tending to concentrate on side wall where it contacts the fibers. In order to solve this problem, we designed several perforated plates with many holes and had numerical analysis for various cases. As a result, a perforated plate with 5 mm and 10 mm diameter holes was installed in the furnace to achieve a more uniform internal air flow distribution. In addition, the stabilization furnace's bottom part was flattened to get more uniform air flow distribution.

Deep Neural Networks (DNNs) have achieved remarkable success across computer vision tasks but remain vulnerable to adversarial attacks that cause misclassification through imperceptible perturbations, posing critical security risks in safety-critical applications. This paper proposes a novel hierarchical defense architecture that synergistically combines multiple defense mechanisms to provide robust protection against diverse adversarial attacks. The proposed architecture consists of three complementary layers: (1) an ensemble detection module combining StarGAN discriminator with EfficientNet-B0, improving L2-norm attack detection from 40-71% (StarGAN alone) to 93-94%; (2) a StarGAN generator-based purification module that removes adversarial noise; and (3) a Hybrid Adversarial Training (HAT) enhanced classifier that maintains robustness without sacrificing clean accuracy. Comprehensive experiments on CIFAR-10 against six attack types (FGSM, PGD, BIM, MIM, CWL2, DeepFool) demonstrate our approach's effectiveness. The ensemble detection achieves 96.88% average detection rate with notable improvements on L2-based attacks (CWL2: 93.93%, DeepFool: 94.22%). The complete pipeline achieves 84.16% purification rate while maintaining 91.24% clean accuracy. Our hierarchical approach effectively addresses the accuracy-robustness trade-off in traditional adversarial training, though gradient-based attacks (FGSM, PGD, BIM, MIM) require further improvement.

This study analyzed the effects of surface treatment and surface damage on gas emission characteristics from aluminum surfaces within a vacuum chamber. Aluminum samples were prepared with and without surface treatment, and damage was induced on the treated samples to simulate surface damage. Surface morphology and roughness before and after heating in the vacuum chamber were measured using a digital microscope and a surface roughness tester. Vacuum performance and gas emission characteristics were evaluated using a residual gas analyzer (RGA) within the vacuum chamber. Experimental results showed differences in surface characteristics based on surface treatment, but no differences were observed due to surface damage. The vacuum performance evaluation confirmed that surface treatment effectively improves vacuum performance, while increasing surface damage negatively impacts vacuum performance. Furthermore, the residual gas analyzer confirmed that surface treatment resulted in lower emissions of moisture and hydrocarbon gases, indicating reduced gas release from the surface. It was also confirmed that the emission levels of these gases increased as surface damage worsened. These results demonstrate that surface treatment is effective in reducing gas emissions from the surface and improving vacuum performance, whereas surface damage diminishes these benefits. The findings of this study provide fundamental data for optimizing the use and maintenance of aluminum components within vacuum chambers.

Graph workloads such as BFS, CC, DC, and SPath exhibit highly irregular memory accesses, making conventional stream/stride prefetchers ineffective. We propose Delta-based Region-Aware Feedback-Tuned Prefetcher (D-RAFT), a lightweight, zero-training prefetcher that adapts in real time using only runtime deltas and region-aware behavior. D-RAFT extracts frequent deltas, predicts using combined two-step and one-step patterns, and adjusts priority through feedback from hit/miss outcomes. Using a trace-driven simulator with graph processing workloads, D-RAFT consistently reduces miss count and improves prefetch coverage compared to Delta-Correlation, GHB, and Best-Offset. D-RAFT reduces misses by 21.3% on SPath relative to Delta-Correlation and achieves 1.29× coverage on BFS/DC when normalized to Best-Offset. These results demonstrate that D-RAFT provides efficient, low-overhead improvements for irregular memory environments in SoC and edge systems.

Accurate measurement of adhesive bead width is essential for ensuring reliability and yield in semiconductor packaging and other assembly processes. Conventional vision-based methods estimate bead width from binarized images using heuristic rules or distance transform (DT). These approaches are simple and interpretable, but they require empirical parameter tuning and provide only point estimates without any notion of reliability. This paper proposes the use of a Bayesian neural network (BNN) for adhesive bead width measurement with a monocular camera and compares it against conventional DT- and heuristic-based methods. A mask is used both to compute DT-based reference width and to determine the sampling locations of patches for training. The BNN takes the original gray-scale patches as input and is designed to output the mean width and the log-variance. By applying MC-Dropout at inference time to obtain multiple stochastic forward passes, the method separates aleatoric (data) uncertainty and epistemic (model) uncertainty and derives a 95% confidence interval (CI) for the width at each angular position. Comparative experiments show that the proposed BNN achieves bead-width estimation accuracy (MAE, RMSE, correlation) comparable to the DT-based reference, and at the same time it provides a probabilistic measure of confidence for each estimate.

TBDB processes rely on adhesive layers to support thin wafers, where thermo-mechanical warpage directly affects downstream alignment and process stability. While prior work has focused on glass-silicon thermal mismatch, the role of adhesive properties has not been quantitatively assessed. This study evaluates how adhesive Young's modulus, Poisson's ratio, and CTE influence warpage in a glass-adhesive-silicon stack using finite element analysis and a DOE approach. Results show that Young's modulus is the dominant factor, with Poisson's ratio contributing moderately and CTE having minimal effect within practical ranges. A limited interaction was observed between modulus and Poisson's ratio, while other combinations remained largely independent. Overall, reducing adhesive modulus was identified as the most effective strategy for minimizing warpage, offering practical guidelines for material selection and property tuning in TBDB applications.

With the growing demand for HPC, the significance of 2.5D and 3D packaging technologies is gaining global attention, leading to increasing recognition of thermo-mechanical stability as a key requirement in interposer-based structures. While silicon interposer is widely used, limitations in cost and scalability have led to growing interest in glass interposer due to their low dielectric constant, high electrical insulation, and compatibility with large-panel fabrication. However, the brittle nature of glass and CTE mismatch with surrounding materials can result in warpage and residual deformation during processing. In this study, the thermo-mechanical behavior of a glass interposer was evaluated through Finite Element Analysis. The analysis focused on examining warpage trends under thermal loading as a function of substrate thickness and overall interposer size. The results indicate that both variables significantly influence deformation, highlighting the need for design consideration to ensure mechanical stability in glass interposer.

This study investigates the influence of wafer trim width on the thermo-mechanical stress response during TBDB processing. Three trim width conditions (0, 250, and 500 ㎛) were evaluated through finite element analysis to examine how structural constraints at the wafer edge modify stress distributions across sequential processing stages. The results show that trim width has minimal effect when the stress field is dominated by global deformation, as observed in the initial and final stages. In contrast, a pronounced sensitivity appears in the intermediate stage, where edge-region stress redistribution is strongly affected by trim geometry. The 250 ㎛ trim width yielded the lowest stress, indicating a more balanced load transfer, while excessive removal at 500 ㎛ led to increased stress concentration. These findings demonstrate that trim width is a stage-dependent design parameter and that optimizing edge geometry can contribute to improving process stability in TBDB-based thin-wafer handling.

Multi-fisheye camera-based stitching is essential in ADAS and autonomous vehicles, yet severe distortion, non-uniform overlap, and poor seam selection still degrade image quality. Conventional photometric seam methods rely mainly on color or brightness differences, causing incorrect boundaries across objects or moving targets. This paper introduces an optimal seam extraction method using ORB-RANSAC geometric inlier consistency for a front-side fisheye camera system. After calibration and panoramic projection, ORB features are detected and RANSAC is applied to retain only geometrically valid correspondences. These inliers are projected into a unified panoramic space, and their spatial distribution is used to generate an initial seam via regression. A final seam is optimized using a minimum-cost function that combines photometric differences with geometric proximity, solved through dynamic programming assisted by graph-cut refinement. The experimental results demonstrate that the proposed method reduces seam-line visibility by approximately 21.7% compared to the conventional ORB + GraphCut approach and achieves an improvement of about 34.5% over the Fixed Seam method. In addition, the method supports real-time processing with a throughput of 32-45 FPS.

As cloud computing establishes itself as the core infrastructure of the AI era, the importance of security continues to grow. To overcome the limitations of the traditional perimeter-based security model, the Zero Trust (ZT) principle, guided by the philosophy of "never trust, always verify," has emerged as a new security paradigm. This study analyzes the limitations of the Cloud Security Assurance Program (CSAP) and proposes strategies to strengthen it based on the Zero Trust Architecture (ZTA). This study analyzes the legal and institutional foundations of CSAP, control items, and the core principles and technical elements of Zero Trust, focusing on the NIST SP 800-207 (ZTA) document, and proposes an integrated CSAP framework based on zero-trust principles. Furthermore, through a comparative analysis with advanced international practices, such as the US FedRAMP, specific policy and technical improvement methods for CSAP are derived. The proposed enhancement methods are expected to significantly contribute to dramatically improving the security of the public cloud environment, while simultaneously promoting technological innovation among Cloud Service Providers (CSPs) and strengthening the competitiveness of the domestic cloud industry.

In this paper, we propose a LiDAR point cloud-based pedestrian detection method and labeling pipeline as a 3D perception and semi-automatic labeling method for physical AI systems such as autonomous vehicles and mobile robots. The proposed method takes Velodyne data from the KITTI dataset as input, selects points within the front camera FOV, performs ground separation and plane estimation, and then applies DBSCAN clustering with distance-range-dependent optimal parameters. Using ground snapping and height-based candidate filtering, it first forms pedestrian candidate clusters, which are subsequently refined via single template matching with a seed pedestrian template. By performing GT-cluster matching based on center distance, the method quantitatively evaluates recall and precision for each distance range. STM is used as a module to further filter clusters with a high likelihood of being pedestrians among the DBSCAN results, and this work integrates a distance-range-optimized DBSCAN design framework into an STM-based pedestrian recognition system. Experimental results show that the proposed method provides quantitative DBSCAN evaluation metrics aimed at mitigating performance variation in pedestrian detection across distance ranges and, through semi-automatic 3D bounding box generation, can serve as an effective auxiliary tool for 3D object detection in the domain of physical AI.

High-reliability serial communication is increasingly required in industrial control and IoT nodes, yet conventional MCU UARTs often provide parity-level checking and shallow buffering, making them vulnerable to multi-bit/burst errors and data loss under burst traffic. Many systems implement stronger error checking and buffering in firmware, making robustness timing- and workload-dependent. This paper presents an FPGA-based UART IP core written in VHDL that integrates a hardware CRC-16/CCITT engine and RX/TX FIFOs (64-byte RX, 256-byte TX) and exposes a memory-mapped interface to an STM32 MCU. The proposed UART IP autonomously performs CRC accumulation and error detection inside the IP while providing frame-level integrity checking and burst-reception protection, so the MCU mainly reads CRC registers and monitors FIFO status rather than computing CRC in software. Quartus II VWF simulations and board tests show that a register-only baseline UART loses bytes during continuous reception (overwritten as 0x00), whereas the proposed UART preserves the received sequence ("123", 0x31-0x33) in the RXFIFO and reports a matching CRC value (e.g., 0x5BCE for "123"). Corrupted frames are clearly detected by CRC mismatch. Overall, integrating CRC and FIFO at the hardware IP level enables robust UART links with reduced dependence on MCU firmware timing.

Automatic image colorization has achieved remarkable progress with deep learning approaches. However, existing methods still suffer from three critical limitations: (1) lack of long-range color consistency within the same object, (2) frequent confusion between semantically similar colors (e.g., red and blue), and (3) tendency to produce desaturated sepia tones in complex indoor scenes. In this paper, we propose SAC-Net (Scene-Aware Consistent Network), a novel framework that addresses these limitations through three key innovations: Global-Local Attention Module (GLAM) for capturing long-range dependencies, Color Contrast Learning (CCL) mechanism to distinguish similar colors, and Scene Context Module (SCM) for adaptive colorization based on scene understanding. Extensive experiments on ImageNet and Places datasets demonstrate that our method achieves state-of-the-art performance, fooling human observers 42.3% of the time in perceptual studies (compared to 32% of previous methods), while significantly improving color consistency and reducing color confusion errors.

LiDAR-based 3D object detection suffers from severe sparsity for small moving objects such as pedestrians and cyclists. Although multi-sweep fusion is widely used, the number of fused sweeps is typically treated as a fixed global hyper-parameter, and its object-wise impact on IoU remains unclear. We propose a ground-truth-gated multi-sweep analysis framework on the KITTI Raw dataset: past LiDAR sweeps are aligned to an anchor frame, only points inside the ground-truth 3D box are accumulated, and DBSCAN clustering in an inflated region of interest yields pseudo detections whose 2D/3D IoU and geometric errors are evaluated as functions of the sweep count. Through experiments, it was verified that an adaptively fused multi-sweep strategy based on object distance or point density is more appropriate than using a globally fixed sweep length.

This study aims to develop a comprehensive evaluation index for cloud-based information systems focusing on the public sector as the 'digital platform government' initiative to transform public systems into cloud-native environments by 2030. Based on the traditional information system evaluation model, this study proposes six evaluation areas: technical quality and performance reflecting the core characteristics of cloud computing, user satisfaction and utilization, cost-effectiveness and financial management, security and regulatory compliance, strategic contribution and innovation, and sustainability. The proposed evaluation areas and evaluation items were verified for reliability and validity through Cronbach alpha analysis through a survey of cloud experts and practitioners in the public sector. Finally, 6 evaluation areas and 19 evaluation items for cloud-based information systems were developed. This study is expected to provide practical guidelines for organizations to systematically measure and manage the value of cloud introduction by extending the traditional information system evaluation theory to the cloud era.

Garnet-type Li₇La₃Zr2O₁₂(LLZO) solid electrolytes are considered promising candidates for all-solid-state batteries; however, their practical application is often limited by high interfacial resistance arising from poor solid-solid contact at electrolyte-electrode interfaces. In this study, we propose a carbon quantum dot (CQD)-reinforced interfacial engineering strategy to enhance the electrochemical performance of Ga-doped LLZO (Ga-LLZO)-based all-solid-state batteries. Ga-LLZO solid electrolytes were synthesized via a solution-based process followed by high-temperature sintering, resulting in a stable garnet structure. CQDs were introduced to both electrolyte and electrode microparticles to regulate interparticle and interfacial contact at the nanoscale. Microstructural analyses reveal that CQDs are uniformly distributed on particle surfaces, forming nanostructured interfacial layers that improve physical contact without altering the bulk crystal structure of Ga-LLZO. Electrochemical impedance spectroscopy using symmetric cells demonstrates that the CQD-modified Ga-LLZO exhibits reduced interfacial resistance and enhanced ionic conductivity under applied external pressure, indicating improved Li⁺ transport across solid-solid interfaces. These results indicate that CQDs function as interfacial regulators rather than bulk conductive additives, effectively mitigating interfacial degradation and mechanical mismatch during repeated cycling.