Uncertainty Quantification in Data-Driven Ensemble Modeling of the Upper Atmosphere

Abstract:

The quest to construct a predictive first-principles model of the upper atmosphere has so far focused on reproducing observed driver-response relationships deterministically. Such modeling approaches, however, fall short of accounting for the role of uncertainties arising from dynamical and physical nonlinearity and the effects of initial conditions and driver uncertainty in determining predictability. Current approaches also do not routinely and systematically integrate observations into modeling to reduce uncertainties in initial conditions and drivers. These are major drawbacks of the current approach as the community grapples with finding a viable pathway to predictive modeling of the near-Earth space environment. Some uncertainties are more likely to lead to greater prediction error, through amplification of an originally rather small uncertainty due to nonlinear coupling processes. Data assimilation can help reduce these uncertainties more effectively if observability of internal states and drivers is better characterized. The impact of observations on observability and predictability can be evaluated heuristically using ensemble-based methods without the need of tangent-linear and adjoint models, for instance, through the reduction in ensemble spread, ensemble-based dimensionality analysis to assess reduction in dynamical imbalance and instability, and ensemble forecast sensitivity analysis. This paper illustrates these points by using examples from our development efforts towards constructing an ensemble-based probabilistic model of the impacts of uncertainties in forcing, initial conditions, model dynamics and physics on upper atmosphere predictability as it responds to nearly constant changes in forcing.