Predicting Marine Biogeochemistry from Structured Representational Priors

The biogeochemical ocean is one of the least well represented parts of the Earth system, and policy-relevant seasonal forecasts remain scarce. We introduce a modular, data-driven system built around a learned representation of the water column. The model is trained on global ocean reanalysis, and it produces six-month, multivariate forecasts of nutrients, oxygen, biology, and carbon pools at a monthly, $1/4^\circ$ resolution. Each additional source of context contributes measurable skill, and the final model outperforms climatology and persistence for most variables at most lead times. We use our model's predictions to derive surface partial pressure of CO$_2$ (which determines air-sea exchange), demonstrating that the learned representations can support a downstream climate application without direct supervision.