Sensitivity-driven simplification of complex ecosystem models: Integrating mechanistic insights for cost reduction and predictive accuracy

With the major challenges posed by climate change and significant shifts in Earth systems, the need for high-precision and diverse climate predictions has grown. These predictions aim to explore a variety of scenarios, such as the Shared Socioeconomic Pathways (SSPs). Advances in computational power have enabled the development of sophisticated coupled physical-biogeochemical-ecological models of marine systems. However, these models remain computationally intensive and energy-demanding, raising questions about the appropriate level of complexity relative to the availability of independent data for accurate calibration, and calls for simplification to reduce execution time. Here, we aim to simplify the Eco3M-MED model, which is a complex biogeochemical model representing the low trophic levels (up to mesozooplankton) in the ocean through 37 state variables, and which is intended to be run at the scale of the Mediterranean basin.

Common simplification methods include conservation analysis, lumping, time exploration, and sensitivity analysis. Since most of these simplification methods reduce or even penalize the ability to interpret model results, or require complex implementation, we have chosen a simple, classic method, based on the local sensitivity analysis (One-Factor-At-A-Time, OFAT) method that does not impair this ability. This work's originality lies in the approach adopted to obtain different declinations of the reduced model. This approach indeed benefits from an original strategy for parametrizing the Eco3M-MED model, initiated several years ago and recently implemented in practice. This strategy consists of the construction of a set of consistent parameters, resulting in the establishment of relations between the so-called core parameters and dependent parameters. Core parameters are perturbed based on the level of knowledge of each parameter. The main objective of this study is to apply this novel approach to identify the biogeochemical processes that can be removed with minimal impact on model performance, thereby enabling model simplification and reducing computational costs. We also apply the principle that a single simplified model is not necessarily the best solution, and aim instead to derive a family of simplified models associated with different usage objectives, ensuring that the simplified model reproduces certain quantities well in particular.  The criteria used to derive a simplified model from the sensitivity analysis are also subject to analysis to identify their influence on the degree of simplification. Finally, the computational efficiency and accuracy of simplified models were compared with the full model to determine optimal simplification for specific applications. Future research will focus on performing global sensitivity analysis on high-impact core parameters to assess uncertainties in both the full and simplified models.