A04. Using neural network ensembles as model comparators

Title

Using neural network ensembles as model comparators

Authors

Christopher Holder (Unlicensed); Anand Gnanadesikan (Unlicensed)

Abstract

Biogeochemical simulations can differ across models either because the physical climate is different or because the intrinsic relationships between physical forcing and biogeochemical output are different. For example, models of ocean ecosystems may yield different distributions of chlorophyll because the location of the subpolar front is offset due to winds being offset or because the level of macronutrient required to have high biomass is different. We demonstrate that one can distinguish between these drivers of model difference using neural network ensembles (NNEs) to capture "apparent relationships" between physical forcing and biomass. Under changes in either subgridscale parameterization or greenhouse gas levels, NNEs can recover the fact that the different solutions are produced by the same underlying relationships between light, nutrient and biomass and thus that physical forcing is responsible for the difference. NNEs may also be used to distinguish biological models with different parameterizations of biology from each other. We propose that such tools may be useful in identifying the fundamental drivers explaining why different Earth System Models produce different distributions of biomass under both modern and climate change conditions.