Forward performance prediction for all-dielectric metasurface sensors via deep learning

An innovative deep neural network (DNN) model is proposed, in which structural parameters are directly mapped to sensing parameters, replacing the traditional process of extracting sensing metrics through the computation of the complete spectrum. A spectral compression strategy is employed, utilizing only three key wavelength points to accurately capture spectral features and morphological information. Redundant data are reduced while the error remains below 0.3 nm. This method achieves 99% prediction accuracy on the test dataset and accelerates the process by four orders of magnitude compared to traditional finite difference time domain (FDTD) simulations. The model supports forward prediction of sensing parameters, and experimental validation shows that the error in the fabricated devices is less than 5%. This method provides a fast and efficient solution for the construction of biosensors, demonstrating its potential application in lab-on-a-chip systems.