Predicting Forecast Sensitivity: Observation Impact with Machine Learning

The Artificial Intelligence (AI) community has recently made tremendous progress in developing efficient techniques for parameter estimation and the JCSDA is interested in applying Machine Learning as a component of the Joint Effort for Data assimilation Integration framework. One key problem that lends itself to Machine Learning is the assessment of the impact of an observing system. The observations impact on NWP forecasts is only indirect, since observations are combined with a model background inside the Data Assimilation algorithm to provide a best-state estimate, which is used in turn to initialize the model forecast. Computing the exact (non-linear) impact of each observation and its associated processing components can be prohibitively expensive. However, we can rely on a proxy metric called Forecast Sensitivity - Observation Impact (FSOI). FSOI is a linearized estimate of instantaneous impact of observations, based either on adjoint code or ensemble approach. We used the NASA FSOI diagnostic capability as metric to compute the loss function for the training and evaluation of the machine learning procedure.

The main scientific questions that we are trying to address are:

1. What fraction of diagnosed FSOI is systematic (as opposed to random) and predictable via statistical inference? What skill can machine learning achieve to predict FSOI for each observation?
2. Given a prediction of FSOI for each observation, what are the most effective strategies to optimize the impact of observations?
3. How does FSOI optimization translate into real observation impact with the non-linear model?

Preliminary results with AMSUA data from NOAA 18 satellite seem to indicate that the correlation between the forecast sensitivity and the model state is limited. However, non-linear AI methods, such as Gradient Boosting and Neural Network, sensibly improve the prediction, when compared to simpler approaches, like multi-dimensional linear regression. This points out the potential of Machine Learning for the prediction of complex non-linear phenomena present in data assimilation problems.