Model parameter estimation is critical in computational science and engineering, as it connects numerical models to real-world observations. Granular materials can be simulated at the microscale with the discrete element method (DEM) or at the macroscale with the material point method (MPM). However, parameter inference faces two major hurdles: (1) the high computational expense of these models and (2) the complexity of parameter probability distributions. To overcome these challenges, we employ model-order reduction and clustering algorithms to emulate physics-based solvers while preserving uncertainty in a reduced parameter space. In this presentation, we introduce probabilistic and data-driven approaches for efficient parameter inference in granular column collapse. We show how machine learning and physics-based modeling can be integrated within a Bayesian framework to estimate material parameters and link them to macroscopic observables, such as the evolution of runout distance over time.