Ocean dynamics are nonlinear, turbulent, and chaotic. Forecasting them over long horizons — years, not days — requires either expensive numerical simulations or surrogate models that inevitably accumulate error. The standard deep learning approach (Transformers, autoregressive networks) predicts the next timestep from the current one, then feeds the prediction back as input. Errors compound. Over 2,000 simulated days, Transformer-based models exhibit progressive energy drift and error amplification that make the forecasts physically meaningless.
Gao et al. (arXiv:2603.05560) take a different approach. Their Continuous-Time Koopman Autoencoder projects the nonlinear ocean dynamics into a latent space where the evolution is governed by a linear ordinary differential equation. The latent dynamics are solved exactly via matrix exponential, then decoded back to physical space. The result: bounded error growth and stable large-scale statistics over multi-year rollouts, at orders-of-magnitude faster inference than numerical solvers.
The cost is fine-scale turbulent structure. Small eddies are dissipated in the latent representation. But bulk energy spectra, enstrophy evolution, and autocorrelation structure remain physically consistent over the full forecast horizon.
This inverts the usual machine learning intuition that preserving all information yields better predictions. Here, discarding fine-grained detail produces a more reliable model, because the loss eliminates the high-frequency modes that seed error amplification in autoregressive schemes. The linearization isn't an approximation to be improved — it's a structural constraint that enforces stability by construction.
The forecast is lossy. The lossiness is the point. Keeping less of the present makes the future more accurate — not because the missing details don't matter, but because they're the source of the instability that destroys long-horizon predictions. Sometimes seeing less means seeing further.