Data-Driven Particle Method for Multiscale Multiphase Flows

  • Win, Max (University of Pennsylvania)
  • Palathinkal, Cyril (University of Pennsylvania)
  • Hernandez, Quercus (EMMI AI)
  • Trask, Nathaniel (University of Pennsylvania)

Please login to view abstract download link

Bridging across multiple length scales and timescales introduces emergent behaviors such as nonlocality, dissipation, and stochasticity. Data-driven particle dynamics is a thermodynamically-consistent framework that uses metriplectic brackets to adhere to conservation laws and bridges across scales. [1–3] These metriplectic brackets are trained on particle trajectories in order to learn a mesh-free nonlocal model. We extend this framework to multiphase systems and for advection-diffusion problems by intro- ducing mass fraction as a new state variable, in addition to positions, velocities, and entropies. We use the model to learn and infer the dynamics of a smoothed dissipative particle dynamics for ideal mixtures, finite-volume Raleigh-Taylor dataset, and a LAMMPS-generated molecular dynamics trajectory of a binary fluid. [4–6] This work introduces a new data-driven model that can be trained on various kinds of coarse-grained data.