New AI Method Improves Ocean Model Calibration via Localized Posterior Inference — Artificial Intelligence

Researchers have introduced a novel Bayesian calibration technique that addresses model discrepancies in complex physical systems by constructing localized posterior distributions. By weighting observations based on their proximity to specific contexts and utilizing conditional normalizing flows, the method significantly reduces predictive errors in ocean turbulence simulations compared to traditional global approaches.

A new study published on arXiv presents an advanced approach to model calibration designed to handle discrepancies that prevent a single parameter from accurately reproducing physical responses across varying conditions. The research, titled "Amortized Localized Posterior Inference for Model Calibration with Application to an Ocean Turbulence Closure," addresses the limitations of standard calibration methods where uncertainties in observations and contexts can skew results. Instead of relying on a global parameter, the authors propose constructing a family of localized posteriors that weight observation likelihoods according to their proximity to a target context.

The methodology incorporates Gaussian localization weights to account for context uncertainty through covariance inflation, while also managing domain boundaries via reflection techniques. To efficiently approximate this family of posteriors without repeating inference for every target context, the team employed conditional normalizing flows within an amortized variational inference framework. This allows for the rapid evaluation of posterior approximations at new contexts and enables the integration of localized posteriors under a prescribed context measure to derive a marginalized parameter distribution.

The effectiveness of this approach was validated through two distinct experiments. First, a twin experiment using the Branin-Hoo model compared global and localized inference under both uniform and skewed context distributions. The results indicated that localization reduced aggregate predictive errors relative to the global posterior. Second, the method was applied to calibrate a k-ε turbulence closure in a single-column ocean model, where atmospheric surface forcing served as the context. In this application, localized posterior calibration demonstrated lower temperature and zonal-velocity errors compared to both deterministic and global-posterior calibration methods.

The study highlights that while numerical models are essential for fields like climate modeling and ocean dynamics, they often involve parameters controlling unresolved mechanisms that vary with context. By assuming local compensability, where model discrepancy can be offset by calibration parameters within a specific neighborhood, the new technique offers a more robust way to quantify uncertainty. The authors note that their approach successfully reduces errors in complex scenarios while providing insights into parameter identifiability under varying conditions, such as weak-wind and heating environments.

Sources: arXiv cs.AI.