Combined autoencoder and clustering-based approach to investigate extreme events in turbulent flows

Turbulent flows may exhibit extreme events, which are sudden large amplitude changes of the flow state. The physical understanding and prediction of extreme events are intricate due to chaos, high dimensionality, and the (relatively) infrequent occurrence of the events.

We propose a combined convolutional autoencoder and modularity-based clustering approach, named Quantised-CAE, to investigate the mechanism of extreme events in turbulent flows. First, the autoencoder is used to obtain a reduced latent representation of the flow dynamics. In the latent representation, a modularity-based clustering technique segregates between latent states representing normal, extreme and, importantly, precursor flow states. These precursors are the set of states linking the normal and extreme clusters, which represent the flow states that likely transition towards extreme states. By decoding these precursor states from the latent space back to the full states, physical insights into the flow structures that can foretell the occurrence of extreme events can be obtained. This approach is shown on the 2D Kolmogorov flow and the Minimal Flow Unit.