Evaluation of the current state of mechanistic aquatic biogeochemical modeling

The need for predictive process-oriented planktonic ecosystem models is widely recognized by the aquatic science community. We conducted a meta-analysis of recent mechanistic aquatic biogeochemical models (153 studies published from 1990 to 2002), to assess their ability to predict spatial and temporal patterns in the physical, chemical and biological dynamics of planktonic systems. The selected modeling studies covered a wide range of model complexity, ecosystem-types, spatio-temporal scales and purposes for model development. Despite the heterogeneous nature of this data set, we were able to identify model behavior trends and illuminate aspects of current modeling practice that need to be reevaluated. Temperature and dissolved oxygen had the highest coefficients of determination (respective median r² values were 0.93 and 0.70) and the lowest relative error (median RE < 10%), nutrients and phytoplankton had intermediate predictability (median r² values ranging from 0.40 to 0.60 and median RE ~ 40%), whereas bacteria (median r² = 0.06) and zooplankton (median RE = 70%) dynamics were poorly predicted. Longer simulation periods (i.e. months to decades) reduced model predictability, and increased model complexity did not improve fit. Aquatic biogeochemical modelers need to be more consistent in how they apply conventional methodological steps during model development (i.e. sensitivity analysis, validation), and the aquatic modeling community should adopt generally accepted standards of model performance. Recent advancements in data assimilation techniques, the combination of the present family of models with goal functions (derived from non-equilibrium thermodynamics) and the development of models with a stronger physiological basis are promising frameworks for obtaining more accurate simulations of planktonic processes.