Skip to main content
U.S. flag

An official website of the United States government

Decomposing Antarctic Sub-shelf Melt Variability using Generalized Clustering with Kernel Embeddings

Presentation Date
Friday, December 15, 2023 at 8:30am - Friday, December 15, 2023 at 12:50pm
MC - Poster Hall A-C - South



Ice loss from the Antarctic Ice Sheet is a dominant contributor to global sea level rise due to the recent warming of the ocean. To assess the uncertainty in the Antarctic contribution to sea level rise from internal variability, we need many realizations of spatiotemporal ocean melt variability, which are extremely limited and expensive to produce with climate models. To mitigate this bottleneck we can learn the variability of the stationary component of ice melt dynamics, and generate new samples using machine learning methods. However, existing work following this path has shown that due to the underlying ice melt dynamics’ complex and multimodal nature it is difficult for any single model to accurately represent the whole spectrum of behaviors, without being dominated by a few principal modes. Thus it is crucial to decompose the whole variability of the Antarctic ice melt behavior into homogeneous sub-components (clusters) that can be modeled independently. Once we have identified such sub-components, we can generate additional realizations from each, using temporally-endowed generative adversarial networks. This work studies the output from a state-of-the-art model of the Earth System (E3SM): a single 150-year simulation of pre-industrial global climate, with high-resolution melt rates calculated underneath floating Antarctic ice shelves. We propose a non-parametric kernel-based ML method for clustering time-series of melt flux that can identify similarities in the dynamics of sub-shelf melt even if they are separated spatially. By applying a similarity metric capable of handling local warping effects, it focuses more on the shape of the time series, and is less sensitive to distortions like scale and temporal shift. Also, the method uses a statistical stopping criterion which adaptively determines the number of clusters based on the distance of distributions as measured by the maximum mean discrepancy (MMD) criterion. We evaluate the recovered clusters using a non-parametric statistical test for comparing high-dimensional distributions, and via consistency with physical knowledge. This work opens the path towards generating additional realizations of sub-shelf melt that are physically consistent and retain the complex multi-modal dynamics, which is currently being explored.

Funding Program Area(s)
Additional Resources:
ALCC (ASCR Leadership Computing Challenge)