This article explores how machine learning, through a paradigm of mathematical data science, can enable the discovery of new mathematical structures by analyzing datasets of mathematical objects, illustrated by case studies on murmurations in number theory and partition loadings related to Kronecker coefficients.