Predicting extreme events in a data-driven model of turbulent shear flow using an atlas of charts
Andrew Fox, C. Ricardo Constante-Amores, Michael D. Graham
Abstract
Dynamical systems with extreme events are difficult to capture with data-driven modeling, due to the relative scarcity of data within extreme events. A recently developed technique called CANDyMan decomposes the system into separate charts containing extreme and non-extreme states, learning dynamical models in each chart via time-mapping neural networks, then stitching the charts into a global dynamical model. We apply CANDyMan to a low-dimensional model of turbulent shear flow which undergoes extreme intermittent quasi-laminarization events. We show that the multi-chart model better forecasts the dynamical system evolution than either a standard single-chart model or Koopman-based model.