Machine Learning Methods Suggest That Large Regional Changes in Phytoplankton Biomass Produced by Earth System Models Do Not Reflect Realistic Responses to Changing Climate

Surface phytoplankton biomass (measured in mol C m⁻³) represents a critical parameter within the Earth System that is measured from space and simulated in Earth System Models. Under climate change, the current generation of Earth System Models agrees that low-latitude biomass will decline and high-latitude biomass will increase. However, on a regional scale, the magnitude, phenology and spatial pattern of these changes are highly inconsistent across models. We use machine learning to investigate the sources of the divergence and evaluate the realism of the simulations. We train Random Forests driven by environmental drivers to simulate surface phytoplankton biomass under both pre-industrial control and SSP5-8.5 scenarios. Outside the Arctic, the bulk of the changes in biomass are driven by rearrangements in the spatiotemporal distribution of environmental predictors. Large regional changes in models, however, are associated either with unrealistically low pre-industrial levels of macronutrients or unrealistically strong responses to those macronutrients. Within the Arctic, relationships between environmental predictors and biomass change under global warming. While increased light drives increased biomass, the effect is largest in models with a high nutrient bias. Feeding inputs from an ensemble of models to an emulator trained on observations predicts observed biomass better than the ensemble of the models does, highlighting the fact that models do not produce the correct relationships between environmental predictors and biomass. However, this technique does not yield mechanistically consistent predictions of biomass under climate change. Skepticism of large regional changes in surface phytoplankton biomass produced by individual models is warranted.