Abstract: A prediction method based on a deep neural network (DNN) is studied in this paper, which constructs a complex fitting function by learning data-to-data mapping relationships to reveal the underlying physical principles between structural parameters and electromagnetic responses. The trained neural network can rapidly predict the maximum absorption rate of metamaterial absorber (MMA) unit structures at operating frequencies using input parameters. The influence of key parameters in the DNN model is thoroughly analyzed to minimize prediction errors. Upon completion of training, the network achieves millisecond-level prediction speeds with a model size of only 3 megabytes, significantly reducing storage requirements while maintaining an average prediction error below 5%. Comparative experiments on four sets of MMA unit parameters demonstrate that the method quickly and accurately generates predictions consistent with full-wave simulation results, verifying its validity and feasibility. Compared with existing prediction schemes in the literature, the proposed approach achieves low-error predictions while exhibiting reduced training costs.