Fig. 3

(A) Observed and predicted surface seawater pCO2 (μatm) across the training data sequence. (B) Relationship between observed and predicted surface seawater pCO2 (μatm) for the best-performing configuration. (C) Validation and training loss (mean squared error) for the best fold in the Bi-LSTM model, with the shaded area indicating the standard deviation among the five folds.