Predictive models in materials science aim to link processing conditions to microstructure evolution and emergent material properties. A central difficulty is the calibration of model parameters whose physical dependence on processing variables is often unknown. This issue is well documented in earlier work on constitutive and recrystallization modeling, where calibration parameters must be inferred from experimental data. Similar challenges have been addressed in flow-stress modeling using symbolic regression and genetic programming, for example in Kabliman et al. (2019) [2], Haghdadi et al. (2013) [1], and more recently in symbolic regression approaches for plastic deformation modeling [3, 4]. In the present work, we compare two complementary strategies to identify functional dependencies between calibration parameters and processing conditions in a physics-based recrystallization model. First, we establish a baseline by fitting the classical model to combined experimental and virtual test data. Second, we extend this framework by replacing the calibration parameters with symbolic expressions learned through genetic programming, following a conceptually similar strategy to previous work in flow-stress and constitutive model discovery [2, 3, 4]. This allows a flexible representation of functional parameter dependencies and improved numerical performance. Although the use of learned symbolic expressions increases computational cost, we observe an good predictive accuracy combined with additional insights into processes during recrystallzation compared to fixed-parameter calibration. These results highlight the strong synergy between mechanistic modeling and data-driven symbolic regression in uncovering process–structure relationships relevant for microstructure evolution.