Mentors: Mohammad Farazmand & Bryan Chu
Team: Anna Asch, Ethan Brady, Hugo Gallardo, & John Hood
Rogue waves, earthquakes, epileptic seizures, and stock market crashes, although occur sporadically, have devastating humanitarian, environmental, and financial impacts. Real-time prediction of these extreme events would help largely mitigate their undesirable effects. Complete measurements of the relevant system information are often unavailable in practice, which hinders our ability to predict extreme events. Deep learning shows promise in predicting chaotic systems, but prior work studying extreme events rarely considers the constraint of incomplete information. Our goal is to use neural networks to predict extreme events from sparse observations. We implement feedforward, long-short term memory (LSTM), and reservoir computing networks to predict future time series. All networks display comparable predictive power when the input data is noise-free. However, when considering noisy observational data, LSTM networks perform best. We demonstrate our results on two chaotic dynamical systems that exhibit extreme events: synchronized firing of neurons and turbulent fluid flow.