IEEE PhotonicsGlobal[at]Singapore

Optical biosensors based on surface plasmon resonance (SPR) provide accurate, economical, and real-time detection of biochemical analytes. The use of machine learning techniques in combination with SPR biosensors is growing in popularity because it enables faster, more precise, and highly sensitive predictions. This study introduces a cohesive physics-regularized neural network (PRNN) framework for the performance analysis and comparison of photonic crystal fiber (PCF)-based multi-analyte biosensors, with an emphasis on loss profile prediction. Initially, a comprehensive evaluation of current deep learning (DL) and machine learning (ML) methods is conducted to identify key shortcomings, such as insufficient cross-sensor validation and a lack of physics-regularized models. Four different PCF-SPR biosensor datasets with varying design-dependent and independent parameters are used to develop a common benchmarking technique to overcome these issues. The proposed PRNN employs residual connections, sinusoidal representation learning, and soft wavelength-gradient regularization to promote smooth and stable wavelength-dependent responses while preserving the non-monotonic SPR resonance profiles learned from the FEM-derived data. In addition, five diverse design- and application-based datasets were used to evaluate the performance of the PRNN model. The experimental findings demonstrate that the proposed model outperforms the others across several datasets. It achieved competitive results, with the highest $R^{2}$ value of 99.16%. Statistical analysis and effect-size evaluation provide additional evidence of consistent improvements in error measurements relative to baseline models. In addition, the proposed model assesses generalizability, stability, and various biosensing applications. These findings demonstrate that physics-regularized learning is a useful strategy for developing reliable and broadly applicable biosensor models. It also offers a scalable basis for the development of intelligent multi-analyte biosensors.

Ghost imaging reconstructs object images based on the correlation between random speckle patterns and bucket signals. Although conventional correlation-based algorithms and algorithms that model reconstruction as an inverse problem have been proposed, they often require either a large number of measurements or computationally expensive optimization for high quality images. In this study, we propose matrix-sketching-based pseudo-inverse ghost imaging (MSPGI), which reduces the computational complexity of pseudo-inverse ghost imaging through randomized dimensionality reduction. In contrast to conventional pseudo-inverse ghost imaging, MSPGI avoids computing the pseudo-inverse of the full-measurement matrix. The proposed method was evaluated on binary and grayscale objects using peak signal-to-noise ratio, structural similarity index measure, and computation time. The results were compared with those of conventional correlation-based algorithms, compressive sensing ghost imaging, and full-measurement pseudo-inverse ghost imaging. MSPGI consistently achieved higher reconstruction quality than the conventional correlation-based algorithms across all sampling ratios evaluated in this study. Compared with compressive sensing ghost imaging, MSPGI achieved comparable image quality with substantially lower computation time. Although its reconstruction quality was lower than that of full-measurement pseudo-inverse ghost imaging, MSPGI substantially reduced the computation time. These findings suggest that MSPGI provides an efficient reconstruction framework that balances reconstruction quality and computational efficiency for ghost imaging applications.

Free Space Optical (FSO) communication provides high data rate and license free wireless transmission but is highly susceptible to atmospheric impairments, particularly fog, which causes high attenuation and link degradation. This paper presents a machine learning based intelligent fog detection and adaptive power control framework to enhance FSO link reliability under varying fog conditions. A vision based fog classification system is developed using a Bidirectional Encoder Representations from Image Transformers (BEiT) model trained on labeled images representing low, moderate, and dense fog scenarios. The proposed model achieves a classification accuracy of 97.26%, outperforming conventional convolutional neural networks and other vision transformer architectures. To improve robustness, the vision based output is integrated with meteorological data using a logistic regression based fog probability model. The combined decision mechanism dynamically controls the laser transmission power of the FSO link via a Siemens S7-1200 programmable logic controller (PLC). Experimental results from multiple locations demonstrate enhanced link stability and improved energy efficiency, validating the effectiveness of the proposed adaptive FSO system in fog prone environments. Clear weather conditions achieve the lowest BER at relatively low transmit power, while dense fog requires a substantial increase in transmitted power to mitigate severe attenuation and maintain acceptable BER performance.

We present a freeform lens design methodology that integrates numerical feedback correction within the polar-grids-based source–target framework, specifically applied along the diagonal of refractive surfaces to achieve uniform target-plane irradiance. While conventional polar-grids-based source–target mapping provides an efficient forward design framework, it inherently lacks adaptability for iterative correction, often resulting in non-uniform illumination. To address this limitation, we integrate a feedback-driven optimization process that iteratively modifies the lens surface according to simulated irradiance distributions. This process is fully automated via the Zemax OpticStudio API, enabling seamless refinement of freeform geometries. Numerical results demonstrate that the proposed approach achieves a relative standard deviation (RSD) of 2.11% under point-source illumination and improves further to 1.4% with extended sources. The introduced framework not only surpasses the intrinsic limitations of polar-based mapping but also establishes a robust and scalable pathway for the automated design of freeform optical elements with superior irradiance uniformity.

In this paper, we propose and numerically validate a compact polarization-insensitive optical switch based on polarization rotators, mode-switching units, and a micro-ring resonator. Building on the wavelength selection capability of the micro-ring resonator, the design integrates the polarization rotation functionality of polarization rotators with the mode conversion ability of mode-switching units. This integration enables multiple functionalities, including polarization-insensitive wavelength selection, polarization-conversion wavelength selection, and broadband polarization conversion. Simulation results show that, when configured as a wavelength-selective switch, the proposed device exhibits an extinction ratio exceeding 10.46 dB, a polarization crosstalk below −18.1 dB at 1550 nm, and a 3-dB bandwidth of 0.79 nm. When functioning as a broadband polarization converter, the polarization extinction ratio is larger than 24.64 dB, and the polarization-dependent loss remains under 0.33 dB.