Learning unknown physics of non-Newtonian fluids

Abstract

Non-Newtonian fluids have a shear-rate dependent viscosity that is difficult to measure in experiments. We present a physics-informed neural networks (PINN) approach for learning the viscosity using indirect measurements (such as velocity and pressure) subject to the momentum conservation and continuity equations constraints. We use the PINN approach to estimate viscosity of polymer melts and suspensions of particles using velocity measurements from two-dimensional shear flow simulations. The PINN-inferred viscosity models agree with empirical models for shear rates with large absolute values but deviate for shear rates near zero where the empirical models have an unphysical singularity.