Our project aims to develop advanced methods that will synergistical exploit multiple, complementing, heterogenous data sources to estimate true states of large dynamical systems. We are developing deep learning-based data fusion framework
that uses trained differentiable surrogates within a Bayesian inference optimization problem to quantify the posterior.
The primary benefit of the data-fusion paradigm is to rapidly adapt pre-trained data-driven models to real-time, potentially
sparse, observations providing increased robustness in nowcasting ability with quantified uncertainty.