What information can we extract from neural networks? Do they tend to learn useful representations that can be interpreted by human experts? In this work, we investigate the learned representations of a model trained on nuclear physics data and show that it captures some of the underlying physical principles.