Improving terrestrial carbon flux simulations with machine learning and global Earth observations

Abstract

The land carbon cycle currently absorbs about one-third of anthropogenic CO2 emissions, but multi-model studies project a future weakening of this sink and a possible shift to a carbon source. Large inter-model differences limit confidence in these projections, and some of these discrepancies may arise from parameter uncertainty. Parameter optimization using global Earth observations could reduce this uncertainty, but it is computationally expensive and complicated by equifinality, where different parameter combinations yield similar performance through compensating effects. To address these challenges, this study uses a genetic algorithm to optimize 28 model parameters against 13 global observation datasets. A Gaussian process emulator is then used to approximate the relationship between model performance and parameter values, explore equifinality, and identify alternative parameter sets with comparable performance. These sets are used to generate an ensemble of simulations, providing an estimate of uncertainty associated with parameter optimization. Results show that optimization improves global model performance, especially for leaf area index and gross primary productivity (GPP). Optimized global GPP decreases by 5 %, resulting in a 61 % reduction in global net biome productivity (NBP) compared with the default simulation. While equifinality arises from compensating effects among many parameters, the reductions in GPP and NBP remain robust and are confirmed by the emulator-derived parameter sets. These findings highlight that parameter tuning can substantially alter carbon fluxes, and modelling groups should integrate advanced parameter optimization frameworks into their development cycle.