Neural network model of two-dimensional silicon-based dielectric cylinder photonic crystal point defect microcavity

Aiming at the problems of high computational cost and long optimization cycle in numerical simulation for traditional photonic crystal microcavity design, this study proposes an intelligent inverse design method combining numerical calculation from the MIT Photonic-Bands (MPB) software package and Back Propagation Neural Network (BP Neural Network). Taking two-dimensional photonic crystals with silicon-based dielectric cylinders (permittivity =12) as the research object, 55 sets of band structure data are generated using MPB software by systematically adjusting the radius of dielectric cylinders (0.010.5 m with a step size of 0.005 m), constructing a "structural parameter-optical property" mapping dataset. A three-layer BP neural network model (with 9 neurons in the hidden layer) is designed, optimized and trained using the Levenberg-Marquardt Algorithm (LM algorithm), combined with the Hyperbolic Tangent Sigmoid Function (tansig function) to handle nonlinear features. Experimental results show that the coefficient of determination R of the model on the test set reaches 0.95309, with an average relative error of 0.08183% and a maximum relative error of 0.1419%. The design cycle is shortened from the traditional "day-level" to "second-level", with efficiency improved by more than 3 orders of magnitude.