Pearse Buchanan: Optimization of Australia’s ocean biogeochemical model with a machine learning surrogate.
Optimisation of model parameters is crucial to ensure that model performance is based on process representation (i.e., functionality), rather than poor choices of input parameter values. However, for ocean biogeochemical models, standard optimisation techniques are not viable due to computational cost. Typically, (tens of) thousands of simulations are required to accurately estimate optimal parameter values. To overcome this persistent challenge, we apply surrogate machine learning methods to optimise a new version of the World Ocean Model of Biogeochemistry and Trophic dynamics (WOMBAT). WOMBAT underwent rigorous updates in functionality and required reoptimisation. A computationally inexpensive surrogate machine learning model based on Gaussian Process Regression was trained on a set of 512 simulations with WOMBAT that varied 26 uncertain parameters and assessed performance against 8 target datasets. Tens of thousands of synthetic simulations using the surrogate facilitated a global sensitivity analysis to identify the most important parameters and facilitated Bayesian parameter optimisation using Markov Chain Monte Carlo sampling. This process improved key model performance metrics, including chlorophyll-a concentrations, air-sea carbon dioxide fluxes and patterns of phytoplankton nutrient limitation in a complex, non-linear model. Overall, we show that surrogate-based calibration can deliver optimal parameter values for complex modules of Earth system models and can improve the simulation of key processes in the global carbon cycle.