Learning Calibrated and Samplable Predictive Distributions of Surface-Ocean CO Using Tabular Diffusion Models
Organizations: Department of Earth and Environmental Engineering, Columbia University, New York, NY, USA · CUNY New York City College of Technology, Brooklyn, NY, USA · Lamont-Doherty Earth Observatory, Columbia University, Palisades, NY, USA
Abstract
The ocean absorbs almost 30% of anthropogenic CO emissions, but quantifying this sink requires estimating surface-ocean CO from observations that are sparse in space and time. Most existing machine-learning estimates of this field are deterministic, giving a single value per grid point, with uncertainty approximated post hoc via residuals, ensembles, or prescribed distributions. We instead learn a conditional predictive distribution for each observation: using Treeffuser, a conditional diffusion model with gradient-boosted-tree score functions, we estimate the full predictive distribution of CO (the sea-air CO difference) from 13 environmental, spatial, and temporal predictors. On a held-out SOCAT test set, Treeffuser attains the lowest continuous ranked probability score among the uncertainty-quantification methods evaluated (8.229 atm) and the coverage closest to the nominal 68% / 95% levels (0.686 / 0.930), while retaining point-prediction skill comparable to a deterministic XGBoost baseline (RMSE 18.4 vs. 17.7 atm; 0.77 vs. 0.79). These results show that conditional diffusion can provide well-calibrated, observation-level uncertainties while largely preserving point-prediction skill. Because Treeffuser is generative, its samples can support probabilistic analyses of CO; extending this to spatially coherent field reconstruction will require modeling dependence across grid cells and evaluation on spatially held-out data.
Figures & tables
| Model | CRPS | Cov68 | Wid68 | IntScore68 | Cov95 | Wid95 | IntScore95 | RMSE | |
| Treeffuser | 8.229 | 0.686 | 27.452 | 45.8 | 0.930 | 55.04 | 81.7 | 18.41 | 0.77 |
| NGBoost | 9.717 | 0.659 | 31.015 | 53.0 | 0.927 | 61.13 | 91.2 | 20.34 | 0.72 |
| XGB-Quantile | 8.339 | 0.560 | 20.771 | 45.6 | 0.880 | 50.38 | 93.0 | 17.74 | 0.79 |
| XGB-Residual | 8.708 | 0.729 | 26.186 | 50.5 | 0.900 | 51.61 | 121.7 | 17.73 | 0.79 |
| XGB-Ensemble | 9.190 | 0.230 | 7.575 | 55.1 | 0.425 | 14.85 | 225.1 | 17.37 | 0.80 |
| XGB-Point | – | – | – | – | – | – | – | 17.73 | 0.79 |
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
| Most confident | Least confident | MAD vs. | ||||||
| Basin | Month | Month | XGB-Point | Obs. | ||||
| Atlantic | 10,497 | 16.89 | Jan | 11.98 | May | 22.68 | 0.87 | 1.31 |
| Pacific | 13,588 | 15.13 | Feb | 10.91 | Jul | 18.65 | 0.68 | 1.78 |
| Southern | 3,561 | 16.80 | Sep | 10.05 | Jan | 19.15 | 1.27 | 3.52 |
| Component | Hyperparameter | Value |
| LightGBM score model | n_repeats | 30 |
| n_estimators | 1000 | |
| early_stopping_rounds | 10 | |
| eval_percent | 0.1 | |
| num_leaves | 127 | |
| max_depth | 9 |
| Method | Hyperparameter | Value |
| XGB backbone (Point/Residual/Ensemble/Quantile) | n_estimators | 1000 |
| max_depth | 9 | |
| learning_rate ( ) | 0.05 | |
| seed | 380 | |
| LightGBM-Point | n_estimators | 1000 |
| max_depth | 9 |
| Draws | Steps | CRPS | Cov68 | Wid68 | Cov95 | Wid95 | Time (s) |
| 50 | 50 | 8.245 | 0.674 | 27.045 | 0.914 | 52.964 | 447 |
| 100 | 50 | 8.229 | 0.686 | 27.452 | 0.930 | 55.042 | 897 |
| 200 | 50 | 8.236 | 0.692 | 27.634 | 0.936 | 55.945 | 1804 |
| 300 | 50 | 8.234 | 0.693 | 27.717 | 0.939 | 56.367 | 2689 |
| 50 | 100 | 8.221 | 0.670 | 26.796 | 0.911 | 52.405 | 1005 |
| 100 | 100 | 8.214 | 0.681 | 27.256 | 0.927 | 54.649 | 1990 |
| Method | Predictive object | Interval | CRPS | |
| XGB-Point | mean only | mean | – | – |
| LightGBM-Point | mean only | mean | – | – |
| ANN-Point | mean only | mean | – | – |
| XGB-Quantile | quantile grid | median | grid quantiles | grid |
| XGB-Residual | Gaussian | Gaussian | ||
| XGB-Ensemble | members | member mean | member quantiles | samples |
Uncertainty Calibration
Uncertainty calibration aligns model confidence and predictive intervals with observed correctness and outcome frequencies.
See all 295 papers