In this work, we aim to optimize the all-optical phase-controlled phase modulators that incorporate lasers, optical directional couplers, optical amplifiers, and phase shifters. Two types of the systems are studied: linear type and sinusoidal type in which the phase of the output light beam of the systems follows a linear and a sinusoidal dependence on the phase difference between the input light and the control light. Using the inverse design approach based on the 2-layer multilayer perceptron through the pseudoinverse of matrix, the coupling ratio of directional couplers and the phase delay of the phase shifters of the systems can be obtained. The standard errors of phase and power between the output modulated light and the target output light are less than those of our previous reports. The non-linear phase-controlled phase modulator could be applied in optical neural networks as an activation function.
Multispectral optical coatings require simultaneous engineering of optical responses across widely separated spectral regions. Conventional inverse-design optimization is repeated when the target specification or design constraints change. Here, we present a generative inverse-design framework constructed on a multispectral encoder–decoder transformer (ME-DT) that supports reuse within the studied material and substrate domain by learning a reusable mapping between broadband optical spectra and multilayer structures. Rather than compressing the target spectrum into a single latent vector, ME-DT tokenizes the spectrum wavelength by wavelength, allowing a single model to span the entire ultraviolet to long-wave infrared (200 nm–13 μm) window. Once trained, the same model can be applied to diverse design problems by changing the spectral conditioning and inference-time objective. The ME-DT is trained using the training partition of an 11-million-record corpus of transfer matrix simulations based on six thin-film materials over the 200 nm–13 μm spectral range. Its capabilities are verified through representative inverse-design tasks, including structural coloration and dual-band visible–infrared camouflage. In both cases, multilayer structures with target-dependent fidelity are obtained using the same pretrained model, while practical constraints such as material availability, layer count, and total thickness are incorporated directly during inference. The proposed approach provides a scalable foundation for multifunctional optical coating design and suggests a promising direction toward more general-purpose artificial intelligence for computational photonics.
Multi-band frequency selective surfaces (FSSs) operating in the W-band (75–110 GHz) are essential for emerging millimeter-wave applications. However, their realization remains challenging due to the limited design flexibility of conventional predefined resonator geometries. In this work, we demonstrate that a single-layer digital metasurface optimized by an active-learning framework provides a flexible platform for realizing dual-, tri-, and quad-band FSSs in the W-band. Each unit cell is represented as a binary pixelated pattern, and the large design space is efficiently explored using factorization machine and simulated annealing based active learning. The optimized digital metasurfaces achieve target multi-band transmittance spectra with well-controlled peak positions and high valley-to-peak ratios, while offering greater design flexibility than metasurfaces based on predefined PSB-type elements. The proposed approach enables tuning of passband positions, reduced band ratios, and access to locally optimal designs with robust angle-insensitive behavior. Equivalent LC-circuit modeling, pixel-wise performance sensitivity, and electromagnetic-field analyses indicate that pixel clusters act as effective LC resonators responsible for the multi-band responses. Fabricated samples show good agreement between measured and simulated transmittance spectra, confirming the practical feasibility of the proposed single-layer digital metasurface platform. These results suggest that digital metasurfaces combined with active learning provide a compact and versatile framework for multi-band FSSs design at millimeter-wave frequencies.
Uni-traveling carrier photodiode (UTC-PD) shows broad application prospects in the domains of military and civilian, encompassing telecommunications networks, satellite communications, and data centers, because of its advantages of high speed and high saturation photocurrent. One of the main obstacles to high-speed UTC-PD is the low quantum efficiency due to the thin absorption layer required for fast response. In this work, a novel photo-trapping enhanced UTC-PD integrated with a metal nanograting was proposed, which enables ultrahigh efficiency without thickening the absorption layer. The effects of various metallic gratings with different structural designs on the light absorption performance of photodetectors were systematically studied in the near-infrared band. The optical absorption coefficients of uni-traveling carrier photodiodes (UTC-PDs) integrated with periodic Au nano-gratings at 1550 nm can be raised more than 5 times that of conventional vertically illuminated devices after proper optimization. The responsivity is increased from 0.24 A/W for the conventional planar structure to the highest of 0.88 A/W. This work provides a reliable method for enhancing the quantum efficiency of the vertically illuminated mesa-type photodetectors, which enables high-speed and high-responsivity performance.
