Sea Ice Modeling and Data Assimilation (SIMDA)

Integrated Foundations of Sensing, Modeling, and Data Assimilation for Sea Ice Prediction
Department of Defense Multidisciplinary University Research Initiatives program through the Office of Naval Research

Physics driven latent space learning

17 Oct 2022 - Jihun Han


J. Han and his collaborators are working on Physics-driven Machine Learning algorithm to learn ice dynamics from sequential sea ice imagery. Instead of learning the nonlinear dynamics in the physical space, the proposed method learns the dynamics in the latent space driven by a linear PDE model. The pull back and push forward between the physical and latent spaces induces a simple dynamics in the latent space, which enables accurate and improved training and learning.

The two videos show the predicted dynamics using two images (left and right or top and bottom, which are 12 days apart).