CSE Community Seminar | September 6, 2024

Abstract

As a result of climate change, extreme weather events have increased in severity and frequency, leading to significant damage to critical infrastructure and numerous premature deaths. The ability to rapidly model various climate scenarios is essential for future resource management and planning. However, the broad range of dynamically relevant spatiotemporal scales in the atmosphere makes direct numerical simulations computationally expensive and simplified data-driven approaches inaccurate. Scientific machine learning methods are a promising substitute, but they are slow or intractable for large data sets, and model generalization is poor for systems with extreme events. To overcome these challenges, we introduce a model-agnostic active learning approach that sequentially selects an optimal subset of training points most relevant to the dynamics of extreme events. Points are chosen according to a likelihood-weighted uncertainty sampling acquisition function for which the uncertainty is estimated with probabilistic predictions from ensembles of neural networks. The function is weighted by the likelihood of the output to improve prediction in the tails of the distribution, i.e. extreme events. To reduce the black box nature of the method, we investigate the optimally selected points with clustering. When applied to a real-world climate data set, the likelihood-weighted active data selection algorithm is able to more accurately reproduce the statistics of extreme weather events with fewer points.