Predictability of Extreme Dynamic Systems using Data Mining

Anthropogenic global warming has made potentially dangerous shifts in climate and ocean systems, leading to severe and frequent climate extremes. To reduce the damages from these extremes in a future warming climate, we must improve our prediction capabilities with the development of new data mining perspectives. In this talk, I introduce new methods of supervised dimensionality reduction in the context of statistical downscaling. Using these novel methods, we can represent high-dimensional atmospheric predictors in lower-dimensional space without losing much information of data. I discuss the benefits of these methods in combination with machine learning models to improve the predictability of any hydro-climatic variables under a changing climate in the future. I also discuss Deep Learning methods, by which we can model temporal dependences in the ocean dynamic systems. By combining Convolutional Neural Network (CNN) and LSTM models, I show how we can learn from the memory in ocean features, and predict the destructiveness of hurricanes for a future warming climate. In the third section, I will introduce new multidimensional dynamic risk frameworks by developing Bayesian, dynamic copulas to model the dependence structure of climatic compound extremes. I will present some applications of these frameworks for quantifying the impact of anthropogenic global warming on various cases.  These cases include the increased risk of compound warm and dry conditions, weather and climate compound extremes which lead to more intense wildfires in California, and multiple hazards from hurricanes.