Title: Learning coupled rheology and segregation of granular flows from data

Author (Invited): Konstantinos Karapiperis, EPFL

Abstract:

We discuss a physics- and thermodynamics-informed machine-learning framework to uncover transient segregating granular flows. Using discrete element simulations of anisotropic bidisperse inclined-chute flows, we generate high-fidelity datasets and obtain continuum fields through coarse-graining. Carefully designed neural networks constrained by physical principles, are used to learn closure relations for the rheology and segregation dynamics, while downstream symbolic distillation extracts interpretable equations that reveal the dominant mechanisms driving the dynamics. The resulting framework enables accurate prediction of granular flows, rediscovers known theories and extends them to better accommodate the transient regime. The proposed framework paves the way for eventually uncovering and understanding a wider class of complex rheologies.