Most modern engineering materials exhibit a complex microstructure that underpins the properties of the material in beneficial – or sometimes detrimental – ways. The on-going, rapid growth of available data from imaging experiments that resolve the microstructure, and the rise of machine-learning and artificial intelligence offer novel ways for doing scientific research with such data, but also challenge the traditional model-based understanding. Using examples from scanning transmission electron microscopy (STEM) and atom probe tomography (APT), I will show how combining data-centric methods and domain knowledge yields tools that reliably detect recurrent patterns in noisy raw data without a priori training, and generate coarse-grained descriptors for them.