Research map / Graph learning / Knowledge and temporal graphs
OxyGenerator: Reconstructing Global Ocean Deoxygenation Over a Century with Deep Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OxyGenerator, a graph hypernetwork with zoning-varying message passing and chemistry-informed gradient variance regularization, to reconstruct global ocean dissolved oxygen from 1920 to 2023 from sparse observations, beating CMIP6 simulations on MAPE.
Problem
Reconstructing century-scale global ocean dissolved oxygen (deoxygenation) from extremely sparse in-situ observations (about 3.7% observed), where CMIP6 numerical simulations do not correct for observations.
Approach
A spatio-temporal graph hypernetwork generates node-wise rotation and bias parameters from geographic and physical context, giving adaptive zoning-varying message passing over a 4D graph with proximity and information-hub neighbors, plus a regularizer that encourages constant DO-nitrate/phosphate gradients following Redfield stoichiometry.
Key results
- OxyGenerator achieves average MAPE 14.28 versus 23.32 for the best CMIP6 simulation (CESM2 omip1), a 38.77% reduction.
- Average R2 of 0.9026 versus 0.8228 for the best simulation (GFDL-ESM4 historical).
- Ablation: removing zoning-varying message passing raises MAPE from 14.28 to 23.84; removing chemistry regularization raises it to 14.85.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use OxyGenerator instead of the strongest CMIP6 baseline | Global ocean DO reconstruction · Global ocean DO reconstruction (4-fold cross-test) | MAPE: 23.32 -> 14.28; 38.77% reduction | Helped |
| use OxyGenerator instead of GFDL-ESM4 historical | Global ocean DO reconstruction · Global ocean DO reconstruction (4-fold cross-test) | R2: 0.8228 -> 0.9026 | Helped |
| use OxyGenerator instead of GFDL-ESM4 historical | Global ocean DO reconstruction · Global ocean DO reconstruction (4-fold cross-test) | RMSE: 35.45 -> 26.31 | Helped |
| use OxyGenerator instead of GFDL-ESM4 historical | Global ocean DO reconstruction · Global ocean DO reconstruction (4-fold cross-test) | MAE: 23.69 -> 17.57 | Helped |
| use zoning-varying graph message-passing (ablation) | OxyGenerator · Global ocean DO reconstruction (4-fold cross-test) | MAPE: 23.84 -> 14.28 | Helped |
| use chemistry-informed regularization (ablation) | OxyGenerator · Global ocean DO reconstruction (4-fold cross-test) | MAPE: 14.85 -> 14.28 | Helped |
| use environmental factors (ablation) | OxyGenerator · Global ocean DO reconstruction (4-fold cross-test) | MAPE: 15.70 -> 14.28 | Helped |
| use temporal DO observations (ablation) | OxyGenerator · Global ocean DO reconstruction (4-fold cross-test) | MAPE: 17.01 -> 14.28 | Helped |
| use the full OxyGenerator instead of vanilla MLP (ablation) | Global ocean DO reconstruction · Global ocean DO reconstruction (4-fold cross-test) | MAPE: 23.03 -> 14.28 | Helped |
Limitations
- Scarce observations make evaluation uncertain; for example, the South Pacific accounts for only 0.027% of complete observations in 1920-1950.
- Reconstruction results require validation against oceanographic laws and have not been fully verified by oceanographers.
Experimental setup
Datasets: World Ocean Database 2018 CCHDO Argo GLODAPv2 2022 GEOTRACES IDP CMIP6 CESM2 omip1/omip2 and GFDL-ESM4 historical simulations
Benchmarks: in-situ DO observations (4-fold cross testing, 25% held out)
Metrics: MAPE RMSE MAE R2
Baselines: CESM2 omip1 CESM2 omip2 GFDL-ESM4 historical MLP GNN Transformer
Builds on: Redfield ratio stoichiometry hypernetworks graph message passing
Closest papers by meaning
- FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution (NeurIPS 2025 · similarity 0.85)
- REACT: Physically and Chemically Consistent Reconstruction of Marine Active Tracers (NeurIPS 2026 · similarity 0.85)
- MFogHub: Bridging Multi-Regional and Multi-Satellite Data for Global Marine Fog Detection and Forecasting (CVPR 2025 · similarity 0.83)
- CoastalBench: A Decade-Long High-Resolution Dataset to Emulate Complex Coastal Processes (ICML 2025 · similarity 0.82)
- OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting (NeurIPS 2026 · similarity 0.82)
- CarbonSense: A Multimodal Dataset and Baseline for Carbon Flux Modelling (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.