Earth system models depend on physical parameters that are difficult to
constrain directly from observations, especially under a non-stationary,
changing climate. This work explores a manifold learning-aided approach to
offline parameter estimation, using the low-dimensional structure of model
output to make uncertainty quantification more tractable when only indirect,
time-varying observations are available. Here, I tried two approaches, one is
UMAP and the other one is mapping into the statistical manifold of generalized extreme value distribution.
Both approaches are not so much successful in the end, and they do not reach the quality for publishing.
Presented at JpGU Meeting 2025 (Makuhari, Oral and Poster) and AOGS 2025
(Singapore, Oral).