Optimization of the World Ocean Model of Biogeochemistry and Trophic dynamics (WOMBAT) using surrogate machine learning methods

Abstract

The introduction of new processes in biogeochemical models brings new model parameters that must be set. Optimization of the model parameters is crucial to ensure that model performance is based on process representation (i.e., functional forms) rather than poor choices of input parameter values. However, for most biogeochemical models, standard optimization techniques are not viable due to computational cost. Typically, (tens of) thousands of simulations are required to accurately estimate optimal parameter values of complex non-linear models. To overcome this persistent challenge, we apply surrogate machine learning methods to optimize the model parameters of a new version of the World Ocean Model of Biogeochemistry and Trophic dynamics (WOMBAT), which we call WOMBAT-lite. WOMBAT-lite has undergone numerous updates described herein with many new model parameters to prescribe. A computationally inexpensive surrogate machine learning model based on Gaussian process regression was trained on a set of 512 simulations with WOMBAT-lite and was used to produce synthetic results emulating tens of thousands of simulations. These simulations explored model fidelity to 8 observation-based target datasets by varying 26 uncertain parameters across their a priori ranges. The surrogate model, trained on these 512 simulations, facilitated a global sensitivity analysis to identify the most important parameters and facilitated Bayesian parameter optimization. Our approach returned constrained posterior distributions of 13 important parameters that, when sampled and input to WOMBAT-lite, ensured excellent fidelity to the target datasets. This process improved the representation of chlorophyll a concentrations, air-sea carbon dioxide fluxes and patterns of phytoplankton nutrient limitation. We present an optimal parameter set for use by the modeling community. Overall, we show that surrogate-based calibration can deliver optimal parameter values for the biogeochemical components of Earth system models and can improve the simulation of key processes in the global carbon cycle.