Iterate updating states and retraining parameters

[jzwart]

Data assimilation weighs information from observations and models to estimate the true state. Traditionally, ML models are trained on observations assuming they are the true state of the system, when in fact they are not due to error in the observations. An iterative state and parameter updating (similar to Brajard et al. 2020) would allow the model to get trained closer and closer to truth rather than training on observations.