A new view of data assimilation is provided by the dynamical systems paradigm of synchronized chaos. Two chaotic systems fall into synchronized motion along their strange attractors if loosely coupled through only a few of many variables, despite sensitive dependence, under a wide variety of conditions. If one system is ``model" coupled to ``truth" in one direction, through a noisy channel, then synchronization realizes effective data assimilation by the model. It is shown that the optimal coupling scheme for synchronization is equivalent to 3DVar or Kalman filtering under an assumption of local linearity, but the synchronization view defines a natural treatment of nonlinearities, allowing the rough magnitudes of covariance inflation factors to be derived.
More interestingly, synchronization of states is readily extended to synchronization of parameters. If the parameters are connection coefficients on links between corresponding variables in different models, then the parameter adaptation scheme gives a recipe for model fusion. It is suggested that the resulting multi-model, trained on 20th century data, can give climate projections that are superior to those derived from averaging of model outputs, currently the state of the art in model fusion.