Flexible design of double-layer grating polarization response integrated on an InSb detector using an artificial...

On-chip integrated polarimetric detectors based on subwavelength gratings (SWGs) are demanding in terms of grating structure design. However, traditional trial-and-error strategies and design solutions are computationally expensive and labor-intensive for finding an optimal structure. Here, we construct an artificial neural network (ANN) to accelerate the spectral prediction and design of high-performance polarization detectors. It is shown that the proposed ANN can accelerate the design process and exhibit high accuracy, with the prediction process taking only 0.15 s and the coefficients of determination of all three outputs greater than 0.94. On-demand integral matching enables the design of the optimal structure in 300 s, a 10⁴-fold time improvement compared to finite-difference time-domain (FDTD). The identified structure has more than 87% TM transmittance and a 65 dB extinction ratio across the entire mid-wave band. The mean square error (MSE) loss values for both the training and testing sets are low, reaching 2.93 × 10^-2 and 3.02 × 10^-2 after 100 epochs. Our results have great potential to accelerate the on-demand design of integrated photonic devices.