Comparison of different deep learning models for inverse design. In comparison, mean error characterizes the accuracy of a model, where the test error reflects the model’s generalization ability. The convergence time and output variety reflect the model’s optimization efficiency and the varied outputs for the need of multivalued design, respectively. From the table, we find that the ANN and TNN fail to optimize for multivalued function, although the TNN leads to very good accuracy. The GAN and VAE have output varieties but face the trade-off between the accuracy and time costs, whether Bayesian optimization is deployed or not. Only the PDN leads to a balanced performance in all three qualities.