Machine learning and direct numerical simulation for loads prediction in particle suspensions

This internship addresses multiscale modelling of dispersed phase flows. In particular, knowledge of the stresses to which solid particles immersed in a liquid flow are subjected is of paramount importance in order to better model the dynamics of two-phase flows. This dynamic system is found in nature like the water treatment and in various industrial domains, such as fluidizer beds, bubble columns and flotation processes. Very often the continuous liquid phase and the dispersed solid phase are treated together through an Eulerian formalism. It first requires the mediation of transport equations and closure laws, including knowledge of the stresses on inclusions, which are of main importance for the precise prediction of pressure drops within certain industrial processes, as well as the prediction of their trajectories. Several models exist in the literature that can predict average stresses on solid particles, but very few models exist for local determination at the particle scale. It is for this purpose that we aim in this work to build an Artificial Neural Network model capable of giving us a prediction of the local efforts exerted on each inclusion. We first start with a simple benchmark model, and we then gradually improve upon this work using more DNS (Direct Numerical Simulation) training data, optimizing the model hyperparameter and transforming raw inputs into a symmetry preserving embedding.