Models of systems in physics usually start with elementary processes. New work with a neural network shows how models can also be built by observing the system as a whole and deducing the underlying interactions.