To reveal complex light?matter interactions, it is necessary to simplify the on?demand design of metamaterials for both forward and inverse applications. Deep learning, a popular data?driven approach, has recently alleviated to a large extent the time?consuming and empirical nature of widely used numerical simulations.A fully?connected deep neural network?based framework for inverse design and spectral prediction of broadband absorbers was proposed.The results demonstrate and validate the high accuracy of the proposed DNN model at 87.47%.The model not only outperform traditional numerical algorithms while ensuring accuracy, but also provides an important reference for on?demand design performance of metamaterials.