Explaining and modeling the effect of particle distribution on mean forces and torques using hierarchical machine learning

Hydrodynamic force and torque experienced by each individual particle in a distribution, even at fixed mesoscale variables such as Reynolds number and particle volume fraction, is significantly influenced by the deterministic location of nearby particles (neighbors). Particle-resolved (PR) simulations provide the most accurate computational representation of these unique particle forces and torques. However, the associated computational cost typically enables consideration of only O(10⁴) static particle systems. The present work develops robust neural models with this limited available data using a physics-based hierarchical framework and symmetry-preserving networks. Among considered mesoscale conditions the models achieve a maximum accuracy of 85% and 96% in the prediction of neighbor-induced force and torque perturbations respectively. Furthermore, the generalizability of these models is thoroughly investigated using PR data of distinct particle distributions that are not involved in the training process. Upon establishing satisfactory generalizability, these relatively inexpensive models are deployed on several very large-scale, statistically-different particle distributions. A rigorous analysis and outcomes of this investigation will be discussed in explaining the effects of clustering and anisotropy on the observed mean and higher order statistics of force and torque.