Non-intrusive surrogate modeling for parametrized time-dependent PDEs using convolutional autoencoders