Is it possible to justify why one CNN architecture outperforms other models?

I am using several pre-trained models to solve my problem. These models are VGG-16, Resnet-101, Inception-ResNet-v2, Densenet-201, and Xception. Xception has outperformed all of these models; is it possible to justify why this model has outperformed the rest? Or the reason is the hyper-parameter configuration of Xception is the best for my problem? Note I have used the same optimizer, batch size, etc. for all the models.