Micrometer-sized colloids coated with DNA can be programmed to self-assemble: particles bind when their DNA strands are complementary, so choosing the sequences sets the interactions. Yet some target clusters are far easier to design than others. The designability of a geometry, the number of interaction rules that assemble it, varies by orders of magnitude, with a few “magic” geometries reached by very many designs, much as a few protein structures are encoded by far more sequences than others. These results come from exhaustive enumeration, which quickly becomes infeasible as clusters grow. In this project you will use machine learning on an existing designability dataset to identify the structural features that make a geometry designable. You will test whether they hold for systems too large to enumerate, and compare them across assembly methods, to help guide particle and sequence design in experiments.