Journal Article

Learning‐Based Calibration of Ocean Carbon Models to Tackle Physical Forcing Uncertainties and Observation Sparsity

Littaye, J., Fablet, R., Memery, L.

2025

Overview

Biogeochemical (BGC) ocean models are simplified representations of complex coupled processes, usually resulting in a large number of parameters, that need to be calibrated. In general, these parameters are constrained relying on incomplete and very heterogeneous sets of data. In addition, as biogeochemical tracers strongly depend on ocean circulation, the spatio‐temporal uncertainties in the physical forcing can bias the circulation, which makes the calibration of ocean carbon models challenging. This study addresses the calibration of ocean biogeochemical models when dealing with imperfect physical forcings and sparse observations. We design a numerical testbed based on a simple BGC box model. It comprises different uncertainty scenarios for the physical forcing as well as different observation configurations of the considered nutrient, phytoplankton, zooplankton, detritus dynamics. We propose and benchmark a learning‐based scheme against a variational data assimilation (DA) approach. The former frames the calibration as learning a neural operator between observations and model parameters. The experiments revealed that the DA‐based calibration is highly sensitive to imperfect physical forcing and limited observations, often leading to significant estimation errors in BGC parameters. Conversely, the learning‐based approach demonstrated a greater robustness in parameter estimation and simulated BGC patterns. We discuss further how these results could transfer to more realistic BGC models and real observing systems. Plain Language Summary Ocean carbon models are essential tools for studying climate change, especially its role in the earth's carbon cycle. However, calibrating these models requires diverse data sources, including observational data sets that are often scarce and model‐based data sets that contain significant uncertainties. The quality of these data impacts the accuracy and validity of the models. This study explores the relationship between data quality and model validity using a simple ocean carbon model. Additionally, the study investigates emerging learning methods, as neural networks, that handle imperfect data to represent ocean processes. By comparing a traditional variational data assimilation method with a new learning‐based approach, the study evaluates their effectiveness in model calibration. The results show that the traditional method is highly sensitive to data quality, while the learning method is more robust. As these results are representative of an idealized framework with a simple carbon model, we conclude by discussing how this method could apply to more realistic models. Key Points Calibrating biogeochemical models through assimilation is sensitive to physical forcing uncertainties and to the observation configuration The calibration of the biogeochemical models can be stated as a learning problem The neural method shows a more robust calibration compared to data assimilation when dealing with imperfect forcings, sparse observations

• Publisher: John Wiley & Sons, Inc,American Geophysical Union (AGU)
• Subject: biogeochemical cycles, processes, and modeling / Biogeochemistry / Calibration / Carbon / Carbon cycle / Climate change / Data assimilation / Data collection / Datasets / Detritus / Estimation errors / Inverse problems / neural networks, fuzzy logic, machine learning / Nutrient dynamics / Ocean circulation / ocean data assimilation and reanalysis / Ocean models / Oceans / Parameter estimation / Parameters / Partial differential equations / physical and biogeochemical interactions / Phytoplankton / Tracers / uncertainty / Variables / Water circulation / Zooplankton