Speaker
Description
We present a physics-informed geostatistical framework for modeling Sea-Surface Temperature (SST) variability, bridging the gap between deterministic numerical models and stochastic uncertainty quantification. While numerical models like ERSEM provide valuable point predictions, they often lack the probabilistic framework necessary for comprehensive risk assessment of rising water temperatures. We model the residuals of the point projections, leveraging an overlapping period of the model’s estimation and real satellite data observations.
We aim to characterize the covariance structure of the residuals, in order to embed future water temperature into a Gaussian Random Field (GRF), with mean given by the ERSEM projections and covariance operator estimated from the empirical residuals.
Traditional Euclidean modeling often fails in marine environments because it ignores complex domain geometries and the inherent anisotropy of water currents. We address this by discretizing the domain into a directed network that explicitly excludes landmasses. Within this network, we define a Markov chain where transition probabilities are derived from current velocity fields. We then construct a valid, positive definite covariance structure that respects this network topology and the advective transport induced by currents.
This approach allows for rigorous Monte Carlo simulations that remain physically grounded while ensuring mathematical consistency. The resulting framework enhances risk assessment capabilities—including early hot-spot detection and joint exceedance probability estimation—by providing full distributional forecasts. Incorporating physical flow constraints into the covariance operator yields a parsimonious yet powerful tool for characterizing uncertainty in complex marine environments.