Complexity of extreme-event prediction in turbulent flows

The limitations of data-driven extreme-event forecasting are examined by finding the minimum computational cost of producing accurate forecasts. For this purpose, the information bottleneck method is applied to a very large dataset of direct numerical simulations of turbulent trajectories in two-dimensional Kolmogorov flow. This method is used to construct optimal models to predict extreme dissipation bursts, exploring the trade-off between model complexity and predictive skill. The results show that model complexity must increase exponentially with the forecast horizon to produce accurate predictions and that this is connected with uncertainty in the causal origin of extreme events.