This paper tackles a foundational problem in the intersection of logic and geometry: given an existential formula $F$ in linear integer-real arithmetic, compute the convex hull of the set of models of $F$. We introduce an algorithm that learns the convex hull of a formula from models sampled by a satisfiability-modulo-theories solver. The key idea is the notion of a \textit{local abstraction} operator, which generalizes a model relative to the input formula, enabling the algorithm to learn the convex hull from finitely many samples. We design local abstractions for linear real arithmetic (LRA), linear integer arithmetic (LIA), and linear integer-real arithmetic (LIRA), and thus obtain convex hull algorithms for each of these theories. Our approach is the first that is capable of computing convex hulls for LIRA formulas, and experimental results show that our algorithms for LRA and LIA are superior to baseline algorithms.