Lost in Translation: Can mathematical relationships be extracted using neural networks?

Neural nets (NNs) are widely used in many fields of science, including particle physics. They are often trained to distinguish between signal and background physics processes, where the training samples come from Monte Carlo datasets where the physics principles are known. While NNs are very good at learning the differences between datasets, they are a black box and do not return any physics principles, just the probability of a given set of feature values belonging to dataset A or dataset B. We are interested in trying to understand if there is a way to extract the underlying mathematical relationships between the features (input values) and we start with a very naive approach of trying to visualize the NNs weights and biases to see if we can learn to "see" these relationships. One goal would be to extract the relativistic relationships between energy, mass, and momentum, from a dataset of 4-vectors. The current status of this work will be presented.