In this talk I will describe recent work which is applying unsupervised machine learning to build models in particle physics and string theory. After a brief review of reinforcement learning, I will discuss applications to two areas of model building. First, I explain how a neural network can be trained to build phenomenologically viable Froggatt-Nielson models of quark masses. Secondly, I will show how heterotic string theory models based on monad bundles can be explored via reinforcement learning.