Artificial neural network for behavior learning from meso-scale simulations, application to multi-scale multimaterial flows

This thesis attempts to develop a framework to affect such a coupling of scales by "learning" from selected computational experiments at the meso-scale and transmitting the "learned" behavior to the macro-scale. The "learning" is performed by means of an artificial neural network that is trained using data extracted from the meso-scale direct numerical simulations. In particular, this thesis describes the use of an Artificial Neural Network (hereafter abbreviated to ANN), to learn and predict the transient forces on a particle in a compressible flow field to produce an accurate model for shocked particulate-laden flows. In the multi-scale sense, the ANN learns meso-scale information of particle-fluid interactions requiring expensive computations; once the behavior is learnt, the ANN can be interrogated to obtain information by a macro-scale model to accurately produce results without continuing to perform expensive computations in direct numerical simulations. Particle data is collected from a compressible Eulerian-Lagrangian solver and provided to the ANN for a range of control parameters, such as Mach number, particle radii, particle-fluid density ratio, position, and volume fraction. Beginning with a simple single stationary particle case and progressing to moving particle laden clouds, the ANN is able to evolve and reproduce correlations between the control parameters and particle dynamics. The trained ANN is then used in computing the macro-scale flow behavior in a model of shocked dusty gas advection. The model predicts particle motion and other macro-scale phenomena in agreement with experimental observations and with a very large reduction in time and computational expense.