Mathematics and Statistics Colloquium

Calvin University

Tristan Contant
Colorado State University

April 23, 2026
3:05pm   •   North Hall 276

Spatiotemporal Multi-Fidelity Emulation for Arctic Sea Ice Dynamics

Multi-fidelity (MF) emulation is a modeling approach that combines data from simulations of varying fidelity to construct an efficient and accurate surrogate model. Typically, these data sources are a small number of accurate but computationally expensive high-fidelity (HF) runs and a larger number of cheaper, less accurate low-fidelity (LF) runs. We develop an MF emulator using Tucker decomposition for dimensionality reduction, Gaussian process (GP) priors for flexible function approximation, and additive discrepancy modeling to systematically account for differences between LF and HF outputs. This approach enables scalable emulation not only across high-dimensional input parameter spaces but also for large spatiotemporal output fields, encompassing many locations and time points. It preserves strong predictive accuracy and uncertainty quantification while effectively capturing complex spatiotemporal trends. By leveraging the complementary strengths of LF and HF data, the emulator greatly reduces the computational cost associated with MPAS-Seaice, a complex and resource-intensive sea ice model, making it particularly suitable for prediction at untried inputs and for data assimilation tasks in climate science. This is especially important for sea ice research, where understanding ice dynamics under climate change requires both computational efficiency and robust uncertainty quantification across a wide range of scenarios.