What the study found
A probabilistic machine learning approach, Gaussian Process Regression (GPR), was used to reconstruct Atlantic Ocean δ13C of dissolved inorganic carbon (δ13CDIC), expanding sparse observations into a much denser dataset.
Why the authors say this matters
The authors say the reconstructed data improve spatial and temporal coverage, which they state offers advantages for decadal trend assessment and new opportunities to resolve regional carbon cycle dynamics, validate Earth system models, refine estimates of oceanic carbon uptake, and extend climate reanalysis records.
What the researchers tested
The researchers compiled δ13CDIC data from 51 historical cruises, including a high-resolution 2023 A16N transect, and used crossover analysis for secondary quality control, retaining 37 cruises for model training, validation, and testing. They then applied GPR using the GLODAPv2.2023 Atlantic dataset as predictor variables, and also used numerical model-based validation to support the reconstruction.
What worked and what didn't
The trained GPR model achieved an average bias of −0.007 ± 0.082 ‰ and an overall uncertainty of 0.11 ‰, with uncertainty attributed to measurement, mapping, and input-variable errors. The reconstruction increased acceptable δ13CDIC samples from 8,941 to 68,435, improved resolution in longitude, latitude, and depth, and reduced spatial discontinuities; the abstract also says it revealed finer vertical structures consistent with other high-resolution biogeochemical observations.
What to keep in mind
Validation was limited by sparse observations, which is why the authors supplemented the study with numerical model-based validation. The abstract does not describe additional limitations beyond the scope of the available Atlantic Ocean dataset and the reconstruction framework.
Key points
- The study reconstructed Atlantic Ocean δ13CDIC using Gaussian Process Regression, a probabilistic machine learning method.
- Quality-controlled data from 37 of 51 historical cruises were used for model training, validation, and testing.
- The model’s average bias was −0.007 ± 0.082 ‰, with overall uncertainty reported as 0.11 ‰.
- The reconstruction increased acceptable samples from 8,941 to 68,435 across the Atlantic basins.
- The abstract says the resulting dataset improves spatial and temporal coverage and supports decadal trend assessment.
Disclosure
- Research title:
- Atlantic δ13CDIC reconstructed with machine learning
- Authors:
- Hui Gao, Zelun Wu, Zhentao Sun, Diana Cai, Meibing Jin, Wei-Jun Cai
- Institutions:
- Guangdong Ocean University, University of Delaware, Flatiron Health (United States), University of Alaska Fairbanks, Nanjing University of Information Science and Technology
- Publication date:
- 2026-04-02
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with peer-reviewed research without reading academic papers — one filtered digest, every Friday.