Multi-Resolution Augmented Artificial Neural Networks for Modeling of Particle-Laden Compressible Flows

The utilization of an artificial neural network to learn and predict transient drag coefficients of a cylinder at varying Mach numbers as well as multi -resolution analysis is a promising concept for greatly reducing large and fine scale fluid modeling. Training data gathered from a compressible Eulerian-Lagrangian algorithm for interface tracking fluid modeling code was decomposed by multi -resolution analysis and fed into an artificial neural network where it was normalized and learned. T he network then created prediction curves for the particles at varying Mach numbers or volume fraction not supplied during the learning process. Plots of these predictions by the multi -resolution augmented neural network were created and compared to numerical data. Prediction curves were obtained for simpler experiments of a post -shocked, a stationary, and a moving particle with small error. More complicated arrays with multiple particles also displayed minimal error, and the main characteristics of the drag curves were accurate. The multi-resolution augmented artificial neural network was able to accomplish this with a very large reduction in the computational power necessary.