Abstract

The world is full of extreme events that vary in space and time. For example, a central question in wildfire modeling might concern the explosive (extreme) growth of a fire plume under meteorological conditions and fuel types. One way such processes are studied is through complex geophysical fluid dynamics simulators. Such simulators are very computationally expensive and do not provide measures of uncertainty. It has long been known that surrogate statistical/machine learning models can be used to emulate certain outputs from such models for the purposes of calibration and uncertainty quantification. Yet, existing emulators do not account explicitly for spatially-dependent extreme events – which is exactly what we are interested in. Here, we demonstrate that a novel, flexible, spatio-temporal-extremes surrogate model can be implemented efficiently through variational autoencoders, providing very efficient simulations of the simulator output. Such simulations then can allow for calibration and/or uncertainty quantification quite efficiently. We demonstrate the approach on simulations of complex geophysical phenomena.