The hidden value of low-performers: ensemble design strategies for coupled ocean-circulation biogeochemical modelling

Abstract

Abstract Ensemble approaches in weather forecasting and climate science combine the output of several differing models to overcome the limitations of singular models with the idea to exploit the “power of many”. Ensemble members are selected to reflect the full range of the best available knowledge. This often motivates replacing low-performing ensemble members with superior alternatives. Here we propose an alternative approach that benefits from extracting the hidden value of low-performing members in the context of coupled ocean-circulation biogeochemical modelling. We demonstrate for a Baltic Sea coastal test site that, when leveraged with machine learning, a perturbed parameter ensemble of models selected for their capabilities to reproduce extreme dynamics can outperform conventional selection approaches - despite individual rather weak model performance. Implications for assessing global marine biogeochemical projections and ocean geo-engineering options are discussed.