UQ -- Fast Surrogates Key to New Methodologies in an Operational and Research Volcanic Hazard Forecasting System

Thursday, 18 December 2014
Christopher G Hughes1, Elena Ramona Stefanescu2, Abani K Patra2, Marcus I Bursik1, Reza Madankan2, Solene Pouget1, Matt Jones3, Puneet Singla2, Tarunraj Singh2, E Bruce Pitman4, Donald Morton5 and Peter Webley6, (1)SUNY Buffalo, Department of Geology, Buffalo, NY, United States, (2)SUNY Buffalo, Department of Mechanical & Aerospace Engineering, Buffalo, NY, United States, (3)University at Buffalo, Buffalo, NY, United States, (4)SUNY Buffalo, Department of Mathematics, Buffalo, NY, United States, (5)Boreal Scientific Computing, LLC, Fairbanks, AK, United States, (6)University of Alaska Fairbanks, Geophysical Institute, Fairbanks, AK, United States
As the decision to construct a hazard map is frequently precipitated by the sudden initiation of activity at a volcano that was previously considered dormant, timely completion of the map is imperative. This prohibits the calculation of probabilities through direct sampling of a numerical ash-transport and dispersion model. In developing a probabilistic forecast for ash cloud locations following an explosive volcanic eruption, we construct a number of possible meta-models (a model of the simulator) to act as fast surrogates for the time-expensive model. We will illustrate the new fast surrogates based on both polynomial chaos and multilevel sparse representations that have allowed us to conduct the Uncertainty Quantification (UQ) in a timely fashion. These surrogates allow orders of magnitude improvement in cost associated with UQ, and are likely to have a major impact in many related domains.

This work will be part of an operational and research volcanic forecasting system (see the Webley et al companion presentation) moving towards using ensembles of eruption source parameters and Numerical Weather Predictions (NWPs), rather than single deterministic forecasts, to drive the ash cloud forecasting systems. This involves using an Ensemble Prediction System (EPS) as input to an ash transport and dispersion model, such as PUFF, to produce ash cloud predictions, which will be supported by a Decision Support System. Simulation ensembles with different input volcanic source parameters are intelligently chosen to predict the average and higher-order moments of the output correctly.