Novel Deep Learning applications on Multi-Phase Flow regimes

Daniel Johan Aarstein: 

Within fluid mechanics, most interesting phenomena occur on the boundary between fluids of different densities, i.e. water+air, water+oil. Adding the constraints that the two fluids are insoluble, in addition to having the system take place in a pipe, we might experience what is known as a "slug".

Experimental and numerical study of slug behavior is a field within itself, this thesis aims to be a proof-of-concept that a novel, non-intrusive Deep Learning model can be used for real-time analysis. The model itself utilizes a Convolutional Neural Network in order to classify, and predict properties for a given slug in a pipe, based solely on acoustic emission from said pipe.

Current findings indicate that the classification on unseen data has an accuracy of ~93 %. The regression for velocity and length is, however, less precise with R2 scores of ~0.5 and ~0, respectively.