Effective Data Assimilation with Machine Learning

Abstract

Data Assimilation (DA) is the approximation of the true state of some physical system at a given time by combining time-distributed observations with a dynamic model in an optimal way. DA incorporates observational data into a prediction model in order to improve numerical forecasted results. It allows for problems with uneven spatial and temporal data distribution and redundancy to be addressed such that models can ingest information. DA is a vital step in numerical modeling. In DA, one makes repeated corrections to data during a single run, to bring the code output into agreement with the latest observed data. In operational forecasting there is insufficient time to restart a run from the beginning with new data then, DA should enable real-time utilisation of data to improve predictions. This mandates the choice of efficient methods to opportunely develop and implement DA models. Due to the scale of the forecasting areas and the number of state variables used to model real world, DA is a large scale problem that should be solved in near real-time. This chapter shows how to accelerate DA simulations introducing Machine Learning models and Domain Decomposition methods in the DA process at the top layer of the math stack. Experimental results are provided for pollutant dispersion within an urban environment