Hadron production is involved in all high-energy physics experiments and is a key element to understanding the formation of the visible matter in the universe. Fragmentation functions (FFs) are required to quantitatively describe hadron production; the key limitation is that they can not be calculated in theory and have to be purely determined from data. They are traditionally determined by making fits to experimental observables using pre-assumed functional forms. This study is the first to infer, directly from data, a functional form of fragmentation functions using a machine learning technique, symbolic regression. The learned function significantly describes data, is generalizable, and resembles the Lund string FF; thus, it could serve as a potential candidate for use in global fits of FFs.