Recent advancements in physics- and thermodynamics-informed machine learning have reshaped the constitutive modeling of materials. Yet, granular materials continue to pose considerable challenges due to their complex history-dependent behavior, particularly in conjunction with identifying meaningful thermodynamic internal state variables. In this presentation, we discuss the integration of thermodynamics-informed machine learning with concepts from micromechanics with a view to upscaling the behavior of granular materials in an interpretable manner. We demonstrate this approach across various regimes of granular material behavior and assess its generalization capabilities. This work has the potential to advance the modeling of geomaterials more broadly while establishing direct connections between internal state variables and underlying dissipative processes.