Uncertainty Quantification of an Earth System Model for risk assessment of Climate Tipping via Non-stationary Extreme Events

 

Predicting climate tipping points, such as the collapse of the Atlantic Meridional Overturning Circulation (AMOC), remains a formidable challenge due to the absence of direct observation records of such events and significant parametric and structural uncertainties in Earth System Models (ESMs). It is the grand challenge to explore whether climate tipping or its risk is foreseeable using ESMs and observation without any directs records of climate tipping in the past.

We performed Observation System Simulation Experiments (OSSE) using one of Earth system models of intermediate complexity, LOVECLIM. To assess the predictability of AMOC tipping under a freshwater hosing scenario, we employed a surrogate model-based uncertainty quantification approach to estimate five uncertain parameters related to atmospheric and oceanic physics. We introduced a new dimensionality reduction technique, Wasserstein GEV PCA, which maps the trends of mean climate and extreme events into a flattened statistical manifold using the Wasserstein metric. This allows for the  quantification of observation errors and likelihoods even under transient climate conditions.

Our results demonstrate that while parameters related to precipitation adjustment are critical for accurate AMOC projections, they are difficult to constrain. Comparative analysis reveals that among single-variable observations, Sea Surface Salinity is the most effective constraint for reducing the parametric uncertainty including the precipitation adjustment parameters and narrowing the projection spread of the AMOC. Furthermore, while standard mean-field PCA methods exhibit significant estimation errors when applied to non-stationary data the proposed GEV-based method maintains high robustness and estimation accuracy even with the nonstationary observation. This study highlights that tracking the geometry of extreme value distributions provides a superior pathway for non-stationary climate data, thereby enabling more reliable risk assessments of climate tipping.