Decoupling Regime and Mechanistic Heterogeneity via Hierarchical Mixture-of-Experts for Global Carbon Flux Prediction

METHOD / FINDINGS

REM-MoE

A hierarchical mixture-of-experts model that separates environmental regimes from physical mechanisms for interpretable global ocean carbon-flux forecasting.

Level-1 routing represents geographical and seasonal regimes; Level-2 routing represents thermal, biological, and mixed processes. A global regime dictionary, affinity priors, and distinct latent channels help prevent expert collapse and entangled representations.

38.4% lower 12-month MSE than PINP. 12 historical months predicting the next 12 months; target is monthly delta pCO2. Relative reduction = (0.1112 - 0.0685) / 0.1112. (Table 1, p. 6.)