Mixed X-anion perovskites, such as CsPbX3 (X = Cl, Br, or I), play an important role in photovoltaic applications. The massive disordered structures associated with mixed anions produce the need for property calculations. However, traditional density functional theory (DFT) computational tools are limited by their computational efficiency to generate the properties of a large number of structures quickly. Researchers have proposed supervised deep learning to forecast crystal properties. For such a supervised convolutional neural network (CNN), we introduce an adversarial loss function that allows for consistent or lower errors with a fewer samples. Meanwhile, we have trained parameterized quantum circuits (PQCs) of CNNs and auto-encoder networks for extracting structural representations. PQCs of deep learning, also named quantum deep learning or quantum machine learning, have been first applied in the research of perovskites and obtained an RMSE (root mean squared error) of less than 1 meV. Our work demonstrates that adversarial learning training mechanisms and PQC-based quantum deep learning will emerge for extensive and deep exploration of data-driven material formation prediction tasks.