The Smooth Failure
# The Smooth Failure
Machine learning atmospheric emulators achieve impressive standalone performance โ accurate forecasts, stable long integrations, correct climatology. When coupled to a full-depth dynamical ocean model for 70-year simulations, the ML atmosphere produces tropical Pacific oscillations of "very low amplitude." The ENSO-like variability that dominates tropical climate effectively disappears.
The failure mode is specific. The ML atmosphere is too smooth โ it does not generate the stochastic atmospheric forcing (westerly wind bursts, Madden-Julian Oscillation events) that triggers and sustains ENSO oscillations in the real system. In standalone mode, this smoothness produces accurate forecasts because the chaotic atmospheric variability is noise around the predictable signal. In coupled mode, this same smoothness kills the feedback loop: the ocean responds to atmospheric forcing, the atmosphere responds to ocean state, and ENSO emerges from the mutual amplification. Without the stochastic kicks, the amplification loop never engages.
The inversion is precise: the property that makes the ML emulator a good forecaster โ suppression of unpredictable variability โ makes it a bad climate simulator. Forecast skill and climate fidelity require opposite properties of the same atmospheric model. The forecast wants the predictable signal without the noise. The climate needs the noise because the noise drives the coupled oscillation.
The structural observation: a component that performs excellently in isolation becomes the weak link in a coupled system, specifically because of the property that makes it excellent. Optimization for standalone accuracy selects against the stochastic features that coupled dynamics require. The ML atmosphere is trained to predict, not to excite.
