Generative Adversarial Network for Turbulent Multiphase Flows

This project aims to employ Generative Adversarial Networks (GANs) in combination with other machine learning methods to predict and carry out data assimilation for turbulent multiphase flows, particularly slug flow. Predicting the occurrence and characteristics of specific multiphase flow regimes in pipes is crucial within the energy industry. However, numerical simulations of these problems carry substantial computational costs, making the development of data-driven methods essential for many practical applications.

GANs are an unsupervised machine learning framework where two neural networks compete against each other. A generator attempts to produce data following a certain distribution, and a discriminator attempts to distinguish between real and fake data produced by the generator. Current work is focussed on applying GANs to fluid dynamics as this adversarial training process can lead to improved results over conventional loss functions typically used with generative networks. When given a set of training data from a specific flow regime, a GAN can consistently generate realistic flow fields in that same regime.

This project uses IC-FERST(Imperial College Finite Element Reservoir Simulator) to carry out numerical simulations of single-phase and multiphase flows with adaptive meshing. Simulation data is compressed from a high dimensional space to a low dimensional space using convolutional autoencoders, and GANs are trained on this low dimensional representation of the flow. Predictions and data assimilation are then carried out using input optimisation techniques. Once trained on a flow regime, GANs can generate realistic flow fields from a limited amount of input information, and faster than with Computational Fluid Dynamics (CFD).