Electronics Letters最新文献

Large language models (LLMs) have shown increasing potential in analogue circuit analysis automation, yet their ability to understand and identify functional blocks can vary significantly depending on the representation format of the circuit. This paper investigates how circuit representation, flat SPICE Netlists versus structured PySpice code, influences LLM performance in functional block recognition tasks. Using a benchmark of ten analogue comparator circuits derived from a standard educational collection, we evaluate five state-of-the-art LLMs across both representations. Each circuit is annotated with ground-truth sub-topologies, and models are prompted to extract these blocks in a standardized JSON format. Our results reveal that DeepSeek R1 achieves the highest average accuracy on Netlist inputs, while GPT 5 provides the most balanced performance across both formats. LLaMA 4 shows a slight advantage on PySpice compared to Netlist, indicating that semantic cues in structured code can benefit certain models. Overall, most models still perform better on Netlist than on PySpice, demonstrating that code-structured representations do not generically improve performance and highlighting the importance of representation format selection in LLM-driven electronic design automation.

This paper proposes a quantisation-driven data-compression framework for machine vision sensors by combining nonlinear quantisation and bit-shift masking. The proposed approach compresses pixel data to an effective 1–4 bits per pixel while preserving edge-salient information required for object detection and feature tracking. Experimental results show that the compressed/quantised images maintain detection and tracking performance without degradation, indicating the proposed framework can reduce sensor-to-processor bandwidth and optimise required bit resolution.

A D-band ultra-low-power and broadband two-stage, gain-boosted low-noise amplifier (LNA) in a 250-nm InP DHBT process is presented. The LNA adopts a simultaneous noise and input matching technique, minimizing loss at the input while boosting the gain over a broad frequency range. The gain boosting core uses a feedback network based on a long transmission line between the base and collector and feedback at the emitter to satisfy the gain, stability and bandwidth requirements at once. The LNA achieves 6.2 dB noise figure, 27 dB gain between 130 and 155 GHz while consuming only 2.5 mW DC power.

Due to the trade-off between sensing and communication performance in the integrated sensing and communication (ISAC) systems, the sensing ability will inevitably be reduced for guaranteed communication performance. This degradation can severely compromise target detection performance. Existing methods fail to effectively address this issue. To address this, we propose a two-step optimisation framework for robust beamforming design. First, we optimise the correlation matrix via manifold optimisation to maximise multi-target detection probability without communication constraint. Second, we jointly design the beamforming matrix under communication SINR constraints by projecting the solution to a convex set that satisfies the communication SINR constraint and power constraint. This approach strategically distributes transmit energy across targets while strictly guaranteeing communication performance. The numerical results demonstrate that the proposed scheme outperforms the baselines, significantly improving the probability of detection without sacrificing communication quality.

We successfully fabricated GaN-based red resonant-cavity light-emitting diodes by using two dielectric distributed Bragg reflectors as the top and bottom reflector mirrors. The device exhibited resonant modes up to 673 nm, being the longest emission wavelength ever reported for a GaN-based microcavity device. This achievement establishes spectral overlap between GaN-based light-emitting devices and GaAs-based devices. This work contributes significantly to the development of GaN-based microcavity devices in longer wavelengths.

A Drowsiness-Detection-Network (DDN) is proposed that combines a Convolutional Neural Network (CNN) for spatial feature extraction with a Long Short-Term Memory (LSTM) module for temporal dependencies. The 68-point model approach of Dlib is used to identify and extract facial features to detect drowsiness. Further, for enhancing the interpretability of predictions, Shapley Additive Explanations (SHAP) along with the gradient-weighted class activation mapping (Grad-CAM) technique are incorporated. Experimental evaluations using 5-fold cross validation on two benchmark datasets, Yawning Detection Dataset (YawDD) and University of Texas at Arlington Real-Life Drowsiness Dataset (UTA RLDD), demonstrate that the proposed DDN consistently outperforms Inception-V3, VGG-16, and VGG-19 in terms of accuracy, precision, recall, and F1-score.

Full-wave electromagnetic simulators, such as HFSS and CST, are essential in antenna design and analysis, but their high computational resources and time cost often limit the amount of training data available for building surrogate models. To enhance modelling accuracy, particularly in contexts where the constraint of limited data arises, this study proposes a data augmentation (DA) framework that integrates a conditional variational autoencoder (CVAE) with deep metric learning (DML), where a contrastive loss is employed to ensure the quality of the synthetic samples. This approach is experimentally validated on the proposed wideband circularly polarized S-shaped slot antenna (WB-SSA), where the surrogate model is built as an inverse design, mapping multi-objective performances, including S₁₁, axial ratio (AR), and gain, to the corresponding structural parameters. Specifically, the CVAE encoder learns to map the structural parameter sets to a latent distribution, conditioned on the multi-objective performances. A contrastive loss regularizes this latent space by separating latent vectors with divergent multi-objective performances. The decoder, in turn, generates synthetic samples from randomly generated multi-objective vectors, thereby producing more reliable synthetic samples for DA application. Experimental results demonstrate that the proposed method contributes to a clear improvement in the performance of the inverse model.

This work proposes an eccentric conductor plate design aimed at overcoming the pervasive issue of diminished shielding effectiveness (SE) at the boundaries of traditional plates, thereby enabling wider and more efficient low-frequency electromagnetic shielding. Compared to traditional flat and other shaped plates, the wave-type configuration demonstrates superior SE and effective shielding area, compared to traditional plat type, achieving a maximum increase of 12.36% in SE at the point of maximum magnetic field intensity and 94.42% in effective shielding coverage. Utilizing 1600 finite element analysis (FEA) simulations, a Deep Belief Networks—backpropagation model was developed to capture multi‑parameter relationships. This led to the establishment of a multi‑objective geometric optimization procedure based on central SE, SE at the point of maximum field strength, and effective shielding area, thereby laying a foundation for practical engineering applications.

Physical reservoir computing (RC) uses nonlinear device dynamics for energy-efficient temporal processing. Most MEMS RC was implemented with multi-component reservoirs or time-delayed feedback. In this study, we harness the intrinsic Duffing nonlinearity of a piezoelectric ceramic disc resonator to implement physical reservoir computing. The resonator's underdamped transients provide fading memory and self-masking, while Duffing nonlinearity maps the input data into a high-dimensional state space. The mechanism is quantified by establishing a nonlinear equivalent-circuit model, and the computational capability is validated via end-to-end simulations. The feasibility of a PZT disc resonator as a physical reservoir is experimentally verified. Driven near resonance, the device achieves 98.5% accuracy on a 3-bit parity-check task at 5000 b/s. The simple sensing-and-computing architecture provides high-throughput temporal processing and has potential for edge computing.

The system-level chopping (SLC) technique attenuates low-frequency noise, such as 1/f noise, particularly in an incremental analogue-to-digital converter (IADC). To achieve additional low-frequency noise attenuation, SLC can be utilised together with dynamic offset cancellation (DOC) techniques such as correlated double sampling (CDS) and auto-zeroing (AZ). In this letter, a novel weighted system-level chopping (WSLC) technique is proposed to achieve a further suppressed low-frequency noise characteristic. By applying the proposed WSLC with a dynamic circuit weight, a third-order high-pass noise transfer function (NTF) can be achieved efficiently. Consequently, with the noise weight of 0.9, it provides a 1/f noise suppression of 78.94% compared to the conventional second-order SLC technique while maintaining the same ADC latency.