Inverse Design of Photonic Crystal Band Gap Based on BP Neural Network

To address the problems of heavy computational cost and time-consuming parameter traversal in traditional numerical methods for photonic crystal design, this paper proposes a BP neural network inverse design method that uses the relative band gap width as the sole input feature. Taking a two-dimensional square lattice photonic crystal as the research object, MPB (MIT Photonic Bands) software is used to generate 272 sets of energy band samples within the range of radius-to-period ratio r/a = 0.01~0.5 with a step size of 0.002, and the relative band gap width of each sample is extracted to form a single-parameter training set. A 1-9-1 three-layer BP neural network is constructed, with relative band gap width as input and dielectric column radius as output, using Levenberg-Marquardt algorithm for supervised training. The results show that the coefficient of determination R² of the model test set reaches 0.745, and the training takes only 12 seconds, achieving fast inversion from macroscopic band gap features to structural parameters, providing a new approach for lightweight inverse design of photonic crystals.