An interpreter at a conference hears the noise of the hall but keeps translating. They don't repeat every word; instead, they find an intermediate language where complex phrases become simple, yet meaning is preserved. That's how the conditional Gaussian Koopman network works: it teaches a neural network to translate a chaotic system into a form where forecasting and data assimilation become fast and accurate.

Before this work, scientists could either predict chaos with deep neural networks or assimilate observations analytically, but not both at once. Deep models were black boxes: to weave observations into them, one had to run thousands of model copies and average the results, accumulating sampling errors. The authors propose a different path: they build a neural network that itself finds an intermediate language where hidden variables become linear if the observed variables are known. Then, closed-form equations for the mean and covariance of the hidden state can be written, turning data assimilation into a routine calculation rather than an expensive ensemble.

The model was tested on three systems where chaos is particularly strong. For the Burgers–Sivashinsky equation, which describes flame front instability, the forecast error was 0.284, compared to 0.246 for the true model. In the Lorenz-96 system with 40 variables, the forecast error dropped to 0.064, while the standard Koopman network gave 0.306. For El Niño, the model predicted sea surface temperature with a correlation above 0.5 at a 12-month horizon, outperforming the persistence forecast.

The interpreter who learned chaos's dialect not only hears the hall's noise but instantly weaves it into the translation. They don't translate everything word for word; they find an intermediate language where complex phrases become simple. In this language, hidden variables obey linear equations when the observed variables are known, making data assimilation analytical. This approach could accelerate climate models, geophysical forecasts, and control tasks that require fast estimates of hidden states from incomplete data.