A neural network based framework to model particle rebound and fracture

Abstract

The current state of the art approach in the simulation of particle-laden flow in turbomachinery is to handle particle–wall interactions via rebound and erosion models. Rebound models often require a priori parameter tuning to match experimental measurements. Moreover, the actual stochastic nature of the rebound is neglected, and the particle is assumed not to fracture upon impact. However, this affects the resulting particle trajectories and is particularly critical at high (normal) impact velocities, where particles in typical aero engine flow exhibit a high probability of fracture, as illustrated in our previous work. In this work, we propose a method to develop a generalized rebound model which is parameter-free for the user and considers the stochasticity of the rebound. To this end, state of the art methods from function approximation, more precisely, deep dense neural networks are employed. The networks are trained through a supervised learning approach, where the neural network maps the impacting particles’ characteristics to its new particle trajectory after rebound. For this, we present an efficient method to predict probability distributions in a supervised learning context without a priori parameter tuning of known PDFs. In a second step, we extend the network to account for particle fracture, where the particle breakage is based on a fracture probability distribution to determine whether a particle breaks. The performance of the proposed framework is illustrated by the use of experimental measurements of the statistical rebound of sand particles in an erosion test rig specifically designed to match flow and impact conditions (including particle fracture) as well as particle sizes in aero engine compressors.