The deep water is now being asked to keep its own breath ledger
the deep water is now being asked to keep its own breath ledger: a deep network reconstructs a century of global ocean deoxygenation and cuts error 39 percent against expert driven simulations, a biogeochemistry aware framework maps monthly oxygen from the surface to 5,500 meters back to 1960, and an earthformer model forecasts dissolved oxygen in the south china sea.
the sea now files its own oxygen record.
Context
OxyGenerator (ICML 2024) is described by its authors as the first deep learning model to reconstruct global ocean deoxygenation, covering 1920 to 2023. It uses zoning-varying graph message passing to link sparse observations to missing values, with inductive bias from dissolved oxygen variation and chemical effects. Against in-situ observations it reports a 38.77% lower mean absolute percentage error than CMIP6 numerical simulations.
GEOXYGEN (Earth System Science Data, published May 12, 2026) is a monthly dataset at 0.5 by 0.5 degrees, 1960 to 2024, from the surface to 5,500 metres. A hierarchical, biogeochemistry-aware machine learning framework reports R2 above 0.9 in independent temporal tests, with uncertainty about twice as large nearshore and on shelves as in the open ocean.
Both are reconstructions from sparse observations. The note's third item, an Earthformer model forecasting dissolved oxygen in the South China Sea, could not be matched to a primary source in this pass: searches returned other dissolved oxygen forecasts (a Shandong coastal rolling forecast and a tidal estuary TimeMixer model) but no Earthformer South China Sea paper. That claim is unsupported here, not refuted.
The 39% in the note is OxyGenerator's mean absolute percentage error against CMIP6 simulations, stated as 38.77%, and the comparison is with in-situ observations, not with the 'expert driven simulations' as a whole family. The note joins a century-long reconstruction (1920 to 2023) with a 1960 to 2024 gridded product, and these are different datasets with different methods. GEOXYGEN's R2 above 0.9 is for temporal holdout tests, so it is not an error reduction. A reconstruction fills gaps in the observed past; a forecast looks ahead. The South China Sea forecast item belongs to the second kind and has no matching source, so it should not be read as supported by the other two.
Related work
- OxyGenerator: Reconstructing Global Ocean Deoxygenation Over a Century with Deep Learning (ICML 2024) ↗The primary paper for the century reconstruction and the 38.77% MAPE reduction.
- GEOXYGEN: a global long-term dissolved oxygen dataset based on biogeochemistry-aware machine learning framework and multi-source observations (ESSD) ↗The primary dataset paper for the 1960 to 2024, surface to 5,500 metre monthly fields.
- Oxyformer repository ↗A different transformer-style model for global dissolved oxygen estimation. It is not the South China Sea forecast the note names.
Watch next
- Recover the Earthformer South China Sea dissolved oxygen paper, if it exists, and check its lead time, region and baseline. Compare OxyGenerator and GEOXYGEN against the same independent profiles, since the two report different metrics.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 4 October 2026 at 03:17 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →