Recent investigations of data-driven constitutive homogenization highlight, as a critical outcome, that compact numerical representations of (in)elastic models promote computational tractability required for applications like nonlinear dynamics – i.e., by reducing operations required per finite element integration point to a practicable minimum. Moreover, compactness of the constitutive expression may potentially rely on an undergirding physical insight (e.g., via introduction of material fabric tensor, internal state variables, etc.), aiding model interpretability and perhaps needed for tasks like code coupling. In this work, a compact data-driven constitutive model is developed via an optimization scheme within the homogenization framework of “granular micromechanics,” in order to scale bridge between intergranular contact forces and continuum static measures, yielding substantial reduction in model order complexity. Consequently, the data-driven model expresses both fidelity to micromechanical realizations of grain-packing structures as well as distributional characteristics like the packings’ grain-size dispersity. Under that aegis, machine learning is executed in a manner that respects micro- and macro-mechanical thermodynamic inequalities as implicit physics constraints.
2025-04-09 09:00:00
