Polynomial Chaos-based Bayesian Inference of K-Profile Parametrization in a General Circulation Model of the Tropical Pacific

TL;DR

This paper introduces a Polynomial Chaos-based Bayesian inference framework to efficiently quantify uncertainties in the K-Profile Parametrization within a tropical Pacific circulation model, using surrogate modeling and advanced statistical techniques.

Contribution

It develops a novel PC surrogate model combined with BPDN for Bayesian inference of KPP parameters, reducing computational costs in complex climate modeling.

Findings

Good agreement of posteriors with default parameter values

Most parameters showed barely informative posteriors

Efficient uncertainty quantification method for climate models

Abstract

The authors present a Polynomial Chaos (PC)-based Bayesian inference method for quantifying the uncertainties of the K-Profile Parametrization (KPP) within the MIT General Circulation Model (MITgcm) of the tropical pacific. The inference of the uncertain parameters is based on a Markov Chain Monte Carlo (MCMC) scheme that utilizes a newly formulated test statistic taking into account the different components representing the structures of turbulent mixing on both daily and seasonal timescales in addition to the data quality, and filters for the effects of parameter perturbations over those due to changes in the wind. To avoid the prohibitive computational cost of integrating the MITgcm model at each MCMC iteration, we build a surrogate model for the test statistic using the PC method. To filter out the noise in the model predictions and avoid related convergence issues, we resort to a Basis-Pursuit-DeNoising (BPDN) compressed sensing approach to determine the PC coefficients of a representative surrogate model. The PC surrogate is then used to evaluate the test statistic in the MCMC step for sampling the posterior of the uncertain parameters. Results of the posteriors indicate good agreement with the default values for two parameters of the KPP model namely the critical bulk and gradient Richardson numbers; while the posteriors of the remaining parameters were barely informative.