OceanCBM routes marine heatwave forecasts through physics concepts for interpretability

OceanCBM is a concept bottleneck model that predicts mixed layer heat content—a key precursor of marine heatwaves—while routing predictions through an intermediate layer of prescribed geophysical fluid dynamics concepts plus a 'free' concept that captures residual processes. The architecture uses mixed supervision to impose soft physical structure without over-constraining the model, explicitly trading off interpretability and performance rather than treating them as incompatible goals.

The model addresses a persistent tension in machine learning for physical sciences: recent approaches achieve strong predictive skill on ocean dynamics but remain largely opaque, providing limited guarantees of fidelity to ground-truth physics. Accurate forecasts alone don't reveal which physical drivers are active during extreme events, a gap that matters for operational oceanography and climate adaptation. OceanCBM's concept bottleneck forces the model to explain its predictions through interpretable intermediate variables derived from geophysical fluid dynamics, making the reasoning process explicit rather than hidden in high-dimensional latent space. The free concept serves dual purposes: it regularizes the prescribed physics concepts and captures residual physical processes that the prescribed concepts don't cover. Across ensemble initializations, mixed supervision yields consistent mechanistic representations, whereas prediction-only and prescription-only baselines achieve similar forecast accuracy but learn highly variable latent structures—different random seeds produce different internal explanations for the same predictive performance. That inconsistency undermines trust in mechanistic interpretations. OceanCBM's mixed supervision stabilizes the learned concepts, making the model's reasoning reproducible across runs. The preprint appeared on arXiv on May 14, 2026.