level: research
extreme ocean events like marine heatwaves are hard to predict and even harder to diagnose because forecasts often do not show the physical causes. a new model called oceancbm tackles this by using a concept bottleneck approach. it predicts mixed layer heat content, a key factor for marine heatwaves, while forcing the prediction to go through a set of human-understandable concepts. these concepts come from geophysical fluid dynamics, such as heat fluxes and ocean currents, plus one free concept that captures leftover physical processes.
the model uses mixed supervision, meaning some concepts are trained with direct labels and others are learned without explicit guidance. this soft physical structure avoids over-constraining the model while still making the reasoning transparent. by examining which concepts are activated, scientists can see what physical factors the model relies on for each forecast. the free concept helps regularize the predictions and can reveal new physical insights not pre-programmed into the model.
tests across multiple model initializations show that oceancbm maintains competitive predictive accuracy compared to black-box models. the interpretability comes without a major loss in skill, making it practical for real-world ocean monitoring. the approach could be extended to other earth system forecasting tasks where understanding the why behind a prediction is as important as the prediction itself.
why it matters: it gives ocean forecasters a way to trust and debug ai predictions by showing the physical reasoning, which is critical for early warnings of marine heatwaves.