A Global Machine Learning Modelfor BGC-Argo Profile Prediction

The GLOBIO (Bridging Global and Local Scales for Biogeochemical Profile Prediction) project has designed and trained a 1D convolutional neural network (CNN) for global-scale reconstruction of biogeochemical (BGC) Argo profiles of nitrate, chlorophyll-a, and bbp700.

The data for training and evaluation used the Global Data Assembly Center (GDAC) Coriolis dataset. The BGC-Argo profiles obtained from the Coriolis Data Center underwent an additional in-house quality-control procedure to ensure the consistency of the working dataset, with a resulting quality-controlled dataset comprising approximately 229,500 oxygen profiles, 101,000 chlorophyll-a profiles, and 63,000 nitrate profiles, all spanning the period 2010–2024.

To design the CNN, rather than manually designing multiple architectures through trial-and-error, the adopted strategy relied on the automatic discovery and optimization of models using evolutionary algorithms. In particular the DENSER framework, which works by exploring architectural variations across generations (iterations), jointly optimized Mean Absolute Error and model complexity. The evolution produced variable-specific CNNs that were  compared to a manually designed architecture previously employed only on the Mediterranean Sea (PPCon).  The results showed that this baseline was consistently outperformed, with the following evolved architectural choices:

• Moderate depths (12–19 layers) were most effective for nitrate and chlorophyll, whereas BBP700 reconstruction benefited from deeper networks.
• Larger convolutional kernels were consistently favored, underscoring the importance of capturing broader vertical features of the profiles.
• Evolved architectures tended to be deeper overall but employed fewer parameters per layer, resulting in more efficient networks compared to the PPCon design.

The resulting architecture was firstly trained and further tuned using only 1/50 of the global dataset, before proceeding with training on the full dataset. Spatial-temporal error mapping reveals heterogeneous performance, with larger deviations in ocean regions of sparse sampling or extreme seasonal conditions.  A comparison with other machine learning-based methods shows good results of the 1D CNN approach. The next step will provide uncertainty estimation and model localization to generate local models starting from the global ones, moving towards “model-as-a-service”.