Feasibility of prediction in driven-dissipative system displaying effective ergodicity breaking

Using machine learning techniques we forecast the size of avalanches in the Olami-Feder-Christensen model using spatial information at different values of noise. We use the convolutional neural network and find that prediction is possible at low noise effective non-ergodic phase and not possible at the high noise effective ergodic phase. By looking at the higher level structures learned from the convolutional neural network, we can identify precursors of large events. Our goal is to understand the theoretical limitations of forecasting extreme events in complex systems.